Don't build the same sources multiple times. The test code should
statically link against the tested code, just like the device plugin
that uses the code in production.
Oddly enough, valgrind was not complaining about this leak...
Fixes: 87b2d783b6 ('core: accept 'ssids':aay option in RequestScan() dictionary parameter')
Previously NetworkManager would wrongly add a broadcast address for the
network prefix that would collide with the IP address of the host on
the other end of the point-to-point link thus exhausting the IP address
space of the /31 network and preventing communication between the two
nodes.
Configuring a /31 address before this commit:
IP addr -> 10.0.0.0/31, broadcast addr -> 10.0.0.1
If 10.0.0.1 is configured as a broadcast address the communication
with host 10.0.0.1 will not be able to take place.
Configuring a /31 address after this commit:
IP addr -> 10.0.0.0/31, no broadcast address
Thus 10.0.0.0/31 and 10.0.0.1/31 are able to correctly communicate.
See RFC-3021. https://tools.ietf.org/html/rfc3021https://gitlab.freedesktop.org/NetworkManager/NetworkManager/issues/295https://bugzilla.redhat.com/show_bug.cgi?id=1764986
Keyfile support was initially added under GPL-2.0+ license as part of
core. It was moved to "libnm-core" in commit 59eb5312a5 ('keyfile: merge
branch 'th/libnm-keyfile-bgo744699'').
"libnm-core" is statically linked with by core and "libnm". In
the former case under terms of GPL-2.0+ (good) and in the latter case
under terms of LGPL-2.1+ (bad).
In fact, to this day, "libnm" doesn't actually use the code. The linker
will probably remove all the GPL-2.0+ symbols when compiled with
gc-sections or LTO. Still, linking them together in the first place
makes "libnm" only available under GPL code (despite the code
not actually being used).
Instead, move the GPL code to a separate static library
"shared/nm-keyfile/libnm-keyfile.la" and only link it to the part
that actually uses the code (and which is GPL licensed too).
This fixes the license violation.
Eventually, it would be very useful to be able to expose keyfile
handling via "libnm". However that is not straight forward due to the
licensing conflict.
https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/381
We don't need a separate "GSList *chains" to track the NMAuthChain
requests for the agents. Every agent should only have one auth-chain in
fly at any time. We can attach that NMAuthChain to the secret-agent.
Also, fix a race where:
1) A secret agent registers. We would start an auth-chain check, but not
yet track the secret agent.
2) Then the secret agent unregisters. The unregistration request will fail,
because the secret agent is not yet in the list of fully registered agents.
The same happens if the secret agent disconnects at this point.
agent_disconnect_cb() would not find the secret agent to remove.
3) afterwards, authentication completes and we register the
secret-agent, although we should not.
There is also another race: if we get authority_changed_cb() we would
not restart the authentication for the secret-agent that is still
registering. Hence, we don't know whether the result once it completes
would already contain the latest state.
Don't access the singleton getter here. Pass the agent-manager argument
instead to maybe_remove_agent_on_error().
Also, don't lookup the agent by name. We already know, whether the agent
is still tracked or not. Look at agent->agent_lst.
nm_agent_manager_get_agent_by_user() would only return the first
matching secret agent for the user. This way, we might miss an agent
that has permissions.
Instead, add nm_agent_manager_has_agent_with_permission() and search
all agents.
There was literally only one place where we would make use of
O(1) lookup of secret-agents: during removal.
In all other cases (which are the common cases) we had to iterate the
known agents. CList is more efficient and more convenient to use when
the main mode of operation is iterating.
Also note that handling secret agents inevitably scales linear with
the number of agents. That is, because for every check we will have
to sort the list of agents and send requests to them. It would be
very complicated (and probably less efficient for reasonable numbers
of secret agents) to avoid O(n).
NMAgentManager and NMSecretAgent work closely together. In particular,
the NMAgentManager creates and tracks the NMSecretAgents and controls
it.
Move NMSecretAgent struct to the header, so that some fields may become
accessible to NMAgentManager. In particular, we will track secret agents
with a CList, and this CList element can be embedded in the
NMSecretAgent structure.
Move it to shared as it's useful for clients as well.
Move and rename nm_dbus_manager_new_auth_subject_from_context() and
nm_dbus_manager_new_auth_subject_from_message() in nm-dbus-manager.c
as they're needed there.
If we know the address used previously, also tell the client to start
from the init-reboot phase, so that it will start with a DHCP request
instead of a discover.
We call svFindFirstNumberedKey() to check whether we have any NETMASK
set. Since commit 9085c5c3a9 ('ifcfg-rh: rename
svFindFirstKeyWithPrefix() to svFindFirstNumberedKey() for finding
NETMASK') that function would no longer find the "NETMASK" without
number.
Fix that, by letting nms_ifcfg_rh_utils_is_numbered_tag() return TRUE
for the tag itself. This also makes more sense, because it matches our
common understanding what numbered tags are.
Adjust the other callers that don't want this behavior to explicitly
check.
Fixes: 9085c5c3a9 ('ifcfg-rh: rename svFindFirstKeyWithPrefix() to svFindFirstNumberedKey() for finding NETMASK')
Previously, setting or getting a variable required to scan all lines.
Note that frequently we would look up variables that didn't actually
exist, which we could only determine after searching the entire list.
Also, since we needed to handle having the same variable specified
multiple times (where the last occurrence wins), we always had to search
all keys and couldn't stop when finding the first key. Well, technically
we could have searched in reverse order for the getter, but that wasn't
done. For the setter we wanted to delete all but the last occurrences,
so to find them, we really had to search them all.
We want to support profiles with hundreds or thousands of addresses and routes.
This does not scale well.
Add an hash table to find the variables in constant time.
Test this commit and the parent commit:
$ git clean -fdx &&
CFLAGS=-O2 ./autogen.sh --with-more-asserts=0 &&
./tools/run-nm-test.sh -m src/settings/plugins/ifcfg-rh/tests/test-ifcfg-rh &&
perf stat -r 50 -B src/settings/plugins/ifcfg-rh/tests/test-ifcfg-rh 1>/dev/null
Before:
Performance counter stats for 'src/settings/plugins/ifcfg-rh/tests/test-ifcfg-rh' (50 runs):
330.94 msec task-clock:u # 0.961 CPUs utilized ( +- 0.33% )
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
1,081 page-faults:u # 0.003 M/sec ( +- 0.07% )
1,035,923,116 cycles:u # 3.130 GHz ( +- 0.29% )
1,800,084,022 instructions:u # 1.74 insn per cycle ( +- 0.01% )
362,313,301 branches:u # 1094.784 M/sec ( +- 0.02% )
6,259,421 branch-misses:u # 1.73% of all branches ( +- 0.13% )
0.34454 +- 0.00116 seconds time elapsed ( +- 0.34% )
Now:
Performance counter stats for 'src/settings/plugins/ifcfg-rh/tests/test-ifcfg-rh' (50 runs):
329.78 msec task-clock:u # 0.962 CPUs utilized ( +- 0.39% )
0 context-switches:u # 0.000 K/sec
0 cpu-migrations:u # 0.000 K/sec
1,084 page-faults:u # 0.003 M/sec ( +- 0.05% )
1,036,130,698 cycles:u # 3.142 GHz ( +- 0.13% )
1,799,851,979 instructions:u # 1.74 insn per cycle ( +- 0.01% )
360,374,338 branches:u # 1092.756 M/sec ( +- 0.01% )
6,160,796 branch-misses:u # 1.71% of all branches ( +- 0.08% )
0.34287 +- 0.00133 seconds time elapsed ( +- 0.39% )
So, not much difference. But this is not surprising, because test-ifcfg-rh loads and
writes predominantly ifcfg files with few variables. The difference should be visible
when having large files.
svFindFirstKeyWithPrefix() only had one caller: to find whether there are
any NETMASK variables set. NETMASK is a numbered variable, so we should only
find variables that indeed follow the pattern. Since there was only
one caller, rename and repurpose the function.
When we write a connection profile to ifcfg-rh file, we first load the
possibly existing file and modify it. The purpose is to preserve
variables that we don't know about, keep comments and preserve the order
of the variables.
Note that the writer sets a bunch of variables according to the
profile's setting. At various places the writer would explicitly
clear variables with svUnsetValue(). However, that was problematic:
- we would not unset all variables that we care about. We really should
not leave previous variables if they make no sense anymore for the
profile. The only thing we want to preserve are entirely unknown keys
and comments. Note that when the writer omits to clear an unset variable,
it usually does so assuming that the reader would anyway ignore the
key, become some other key renders it irrelevant. Given the complexity
of the reader and writer, that is often not the case and hard to ensure.
We might have simply forgotten a svUnsetValue(), which was an easy
to make mistake and hard to find (because you'd have to test with
a pre-existing profile that happens to contain that key, which leaves
countless combinations for testing.
That means, a profile written by the writter might be interpreted
differently by the reader depending on which pre-existing keys were set.
- it was cumbersome to explicitly call svUnsetValue().
Note that for numbered tags in particular we would iterate the keys
trying to unset them. For example for addresses (like "IPADDR5") we
would iterate over the first 256 IPADDR keys, trying to unset them.
That is horrible. For one, it doesn't cover the case where there might
be more than 256 addresses. Also, it adds a significant overhead every
time.
While writing a ifcfg file currently is O(n^2) because setting one key
is O(l), with l being the number of keys/lines. So, if you set n keys
in a file with l lines, you get O(n*l). Which is basically O(n^2),
because the number of lines and the number of keys to set usually
corresponds.
So when setting 256 times IPADDR, the overall complexity was still
O(n^2 + 256 * n) and didn't change. However, the 256 factor here can
be very significant.
We should not explicitly unset variables, we should always unset all
known variables that we don't explicitly set.
The svUnsetValue() calls are still there. They will be dropped next.
Helper function to remove all variables that are still dirty (not
visited) and well-known.
Also add svWriteFileWithoutDirtyWellknown() to clear the lines
before persisting to disk.
This adds a lot of meta-data about how we handle ifcfg-rh.
We will use this to prune/delete all variables that are not explicitly
set (dirty) but also well-known.
We could now easily emit a warning when an ifcfg-rh file contains
unused key.
We also could add more meta-data for each key. For example, we write
different files (ifcfg- and keys- files). We could add flags to indicate
that variables are valid in certain files. Currently that's not done.
Also, for simple properties we could associate the key with the
NMSetting property, and treat does generically, like keyfile does.
Anyway, there are potentials. For now, we will use this to clear dirty
variables.
Previously, IS_NUMBERED_TAG() could only be called with a C literal.
Add is_numbered_tag() which can be called with any C string.
Also, IS_NUMBERED_TAG_PARSE() and IS_NUMBERED_TAG() didn't do exactly
the same. I think they should. The only difference was if the number
was larger than 2^63-1. Now IS_NUMBERED_TAG() starts ignoring such
keys, which is fine.
By default, all lines are now marked as dirty. Whenever we modify/set
a line, it becomes non-dirty. That will be used later to prune lines
that are dirty, that is, not yet visited.
Found by covscan:
NetworkManager-1.22.0/src/dhcp/nm-dhcp-nettools.c:945: check_return:
Calling "g_file_set_contents" without checking return value (as is
done elsewhere 16 out of 20 times).
Fixes: 9f89516928 ('dhcp: nettools: read/write lease files')
The autoconnection for virtual devices currently works in two
phases. First we detect that there is suitable profile that can
autoconnect and we realize the device. Then, when the device becomes
'disconnected', autoconnect kicks in and starts the activation.
However, if autoconnect is blocked for a device, currently we do step
1 without step 2, leaving a stale interface around. Fix this by also
checking that autoconnect is not blocked during step 1.
https://bugzilla.redhat.com/show_bug.cgi?id=1765047https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/360
The abbreviations "ns" and "ms" seem not very clear to me. Spell them
out to nsec/msec. Also, in parts we already used the longer abbreviations,
so it wasn't consistent.
If a device is being autoactivated and requires a parent that is
blocked due to user request, the autoactivation attempt should fail
because NM shouldn't overrule the user decision.
https://bugzilla.redhat.com/show_bug.cgi?id=1765566
Many device types take the MTU value from the wired setting; usually
they don't implement the can_reapply_change() method and so the MTU
can't be changed with the Reapply() API.
Instead of implementing the method for all such devices to support the
same property (adding a lot of duplicated code), add a check in
NMDevice to allow the reapply of MTU when we recognize that the device
uses the MTU from the wired setting.
Device types can still decide to implement can_reapply_change() and
support whatever properties they want, even from the wired setting.
