In general, all fields of public NMPlatform* structs must be
plain/simple. Meaning: copying the struct must be possible without
caring about cloning/duplicating memory.
In other words, if there are fields which lifetime is limited,
then these fields cannot be inside the public part NMPlatform*.
That is why
- "NMPlatformLink.kind", "NMPlatformQdisc.kind", "NMPlatformTfilter.kind"
are set by platform code to an interned string (g_intern_string())
that has a static lifetime.
- the "ingress_qos_map" field is inside the ref-counted struct NMPObjectLnkVlan
and not NMPlatformLnkVlan. This field requires managing the lifetime
of the array and NMPlatformLnkVlan cannot provide that.
See also for example NMPClass.cmd_obj_copy() which can deep-copy an object.
But this is only suitable for fields in NMPObject*. The purpose of this
rule is that you always can safely copy a NMPlatform* struct without
worrying about the ownership and lifetime of the fields (the field's
lifetime is unlimited).
This rule and managing of resource lifetime is the main reason for the
NMPlatform*/NMPObject* split. NMPlatform* structs simply have no mechanism
for copying/releasing fields, that is why the NMPObject* counterpart exists
(which is ref-counted and has a copy and destructor function).
This is violated in tc_commit() for the "kind" strings. The lifetime
of these strings is tied to the setting instance.
We cannot intern the strings (because these are arbitrary strings
and interned strings are leaked indefinitely). We also cannot g_strdup()
the strings, because NMPlatform* is not supposed to own strings.
So, just add comments that warn about this ugliness.
The more correct solution would be to move the "kind" fields inside
NMPObjectQdisc and NMPObjectTfilter, but that is a lot of extra effort.
There is only one caller, hence it's simpler to see it all in one place.
I prefer this, because then I can read the code top to bottom and
see what's happening, without following helper functions.
Also, this way we can "reuse" the nla_put_failure label and assertion. Previously,
if the assertion was hit we would not rewind the buffer but continue
constructing the message (which is already borked). Not that it matters
too much, because this was on an "failed-assertion" code path.
Arguably, the structure is used inside a union with another (larger)
struct, hence no memory is saved.
In fact, it may well be slower performance wise to access a boolean bitfield
than a gboolean (int).
Still, boolean fields in structures should be bool:1 bitfields for
consistency.
Kernel calls the netlink attribute TCA_FQ_CODEL_MEMORY_LIMIT. Likewise,
iproute2 calls this "memory_limit".
Rename because TC parameters are inherrently tied to the kernel
implementation and we should use the familiar name.
iproute2 uses the special value ~0u to indicate not to set
TCA_FQ_CODEL_CE_THRESHOLD in RTM_NEWQDISC. When not explicitly
setting the value, kernel treats the threshold as disabled.
However note that 0xFFFFFFFFu is not an invalid threshold (as far as
kernel is concerned). Thus, we should not use that as value to indicate
that the value is unset. Note that iproute2 uses the special value ~0u
only internally thereby making it impossible to set the threshold to
0xFFFFFFFFu). But kernel does not have this limitation.
Maybe the cleanest way would be to add another field to NMPlatformQDisc:
guint32 ce_threshold;
bool ce_threshold_set:1;
that indicates whether the threshold is enable or not.
But note that kernel does:
static void codel_params_init(struct codel_params *params)
{
...
params->ce_threshold = CODEL_DISABLED_THRESHOLD;
static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
struct netlink_ext_ack *extack)
{
...
if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
}
static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
{
...
if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
codel_time_to_us(q->cparams.ce_threshold)))
goto nla_put_failure;
This means, kernel internally uses the special value 0x83126E97u to indicate
that the threshold is disabled (WTF). That is because
(((guint64) 0x83126E97u) * NSEC_PER_USEC) >> CODEL_SHIFT == CODEL_DISABLED_THRESHOLD
So in kernel API this value is reserved (and has a special meaning
to indicate that the threshold is disabled). So, instead of adding a
ce_threshold_set flag, use the same value that kernel anyway uses.
The memory-limit is an unsigned integer. It is ugly (if not wrong) to compare unsigned
values with "-1". When comparing with the default value we must also use an u32 type.
Instead add a define NM_PLATFORM_FQ_CODEL_MEMORY_LIMIT_UNSET.
Note that like iproute2 we treat NM_PLATFORM_FQ_CODEL_MEMORY_LIMIT_UNSET
to indicate to not set TCA_FQ_CODEL_MEMORY_LIMIT in RTM_NEWQDISC. This
special value is entirely internal to NetworkManager (or iproute2) and
kernel will then choose a default memory limit (of 32MB). So setting
NM_PLATFORM_FQ_CODEL_MEMORY_LIMIT_UNSET means to leave it to kernel to
choose a value (which then chooses 32MB).
See kernel's net/sched/sch_fq_codel.c:
static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
struct netlink_ext_ack *extack)
{
...
q->memory_limit = 32 << 20; /* 32 MBytes */
static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
struct netlink_ext_ack *extack)
...
if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
Note that not having zero as default value is problematic. In fields like
"NMPlatformIP4Route.table_coerced" and "NMPlatformRoutingRule.suppress_prefixlen_inverse"
we avoid this problem by storing a coerced value in the structure so that zero is still
the default. We don't do that here for memory-limit, so the caller must always explicitly
set the value.
When using nm_utils_strbuf_*() API, the buffer gets always moved to the current
end. We must thus remember and return the original start of the buffer.
In practice, there is no difference when representing 0 or 1 as signed/unsigned 32
bit integer. But still use the correct type that also kernel uses.
Also, the implicit conversation from uint32 to bool was correct already.
Still, explicitly convert the uint32 value to boolean in _new_from_nl_qdisc().
It's no change in behavior.
"NM_CMP_FIELD (a, b, fq_codel.ecn == TRUE)" is quite a hack as it relies on
the implementation of the macro in a particular way. The problem is, that
NM_CMP_FIELD() uses typeof() which cannot be used with bitfields. So, the
nicer solution is to use NM_CMP_FIELD_UNSAFE() which exists exactly for bitfields
(it's "unsafe", because it evaluates arguments more than once as it avoids
the temporary variable with typeof()).
Same with nm_hash_update_vals() which uses typeof() to avoid evaluating
arguments more than once. But that again does not work with bitfields.
The "proper" way is to use NM_HASH_COMBINE_BOOLS().
ethtool/mii API is based on the ifname. As an interface can be renamed,
this API is inherently racy. We would prefer to use the ifindex instead.
The ifindex of a device cannot change (altough it can repeat, which opens a
different race *sigh*).
Anyway, we were already trying to minimize the race be resolving the
name from ifindex immediately before the call to ethtool/mii.
Do better than that. Now resolve the name before and after the call. If
the name changed in the meantime, we have an indication that a race
might have happend (but we cannot be sure).
Note that this can not catch every possible kind of rename race. If you are very
unlucky a swapping of names cannot be detected.
For getters this is relatively straight forward. Just retry when we
have an indication to fall victim to a race (up to a few times). Yes, we
still cannot be 100% sure, but this should be very reliable in practice.
For setters (that modify the device) we also retry. We do so under the
assumption that setting the same options multiple times has no bad effect.
Note that for setters the race of swapping interface names is particularly
bad. If we hit a very unlucky race condition, we might set the setting on
the wrong interface and there is nothing we can do about it. The retry only
ensures that eventually we will set it on the right interface.
Note that this involves one more if_indextoname() call for each operation (in
the common case when there is no renaming race). In cases where we make
multiple ioctl calls, we cache and reuse the information though. So, for such
calls the overhead is even smaller.
On Ubuntu 14.04 kernel (4.4.0-146-generic, x86_64) this easily causes
test failures:
make -j 8 src/platform/tests/test-route-linux \
&& while true; do \
NMTST_SEED_RANDOM= ./tools/run-nm-test.sh src/platform/tests/test-route-linux -p /route/rule \
|| break; \
done
outputs:
...
/route/rule/1:
nmtst: initialize nmtst_get_rand() with NMTST_SEED_RAND=22892021
OK
/route/rule/2: >>> failing...
>>> no fuzzy match between: [routing-rule,0x205ab30,1,+alive,+visible; [6] 0: from all suppress_prefixlen 8 none]
>>> and: [routing-rule,0x205c0c0,1,+alive,+visible; [6] 0: from all suppress_prefixlen -1579099242 none]
**
test:ERROR:src/platform/tests/test-route.c:1695:test_rule: code should not be reached
Why didn't we get a compiler warning about this bug?
At least clang (3.8.0-2ubuntu4, Ubuntu 16.04) warns:
CC src/platform/src_libNetworkManagerBase_la-nm-platform.lo
../src/platform/nm-platform.c:5389:14: error: data argument not used by format string [-Werror,-Wformat-extra-args]
lnk->remote ? nm_sprintf_buf (str_remote, " remote %s", nm_utils_inet4_ntop (lnk->remote, str_remote1)) : "",
^
Fixes: 4c2862b958 ('platform: add gretap tunnels support')
(cherry picked from commit dfb899f465)
"libnm-core" implements common functionality for "NetworkManager" and
"libnm".
Note that clients like "nmcli" cannot access the internal API provided
by "libnm-core". So, if nmcli wants to do something that is also done by
"libnm-core", , "libnm", or "NetworkManager", the code would have to be
duplicated.
Instead, such code can be in "libnm-libnm-core-{intern|aux}.la".
Note that:
0) "libnm-libnm-core-intern.la" is used by libnm-core itsself.
On the other hand, "libnm-libnm-core-aux.la" is not used by
libnm-core, but provides utilities on top of it.
1) they both extend "libnm-core" with utlities that are not public
API of libnm itself. Maybe part of the code should one day become
public API of libnm. On the other hand, this is code for which
we may not want to commit to a stable interface or which we
don't want to provide as part of the API.
2) "libnm-libnm-core-intern.la" is statically linked by "libnm-core"
and thus directly available to "libnm" and "NetworkManager".
On the other hand, "libnm-libnm-core-aux.la" may be used by "libnm"
and "NetworkManager".
Both libraries may be statically linked by libnm clients (like
nmcli).
3) it must only use glib, libnm-glib-aux.la, and the public API
of libnm-core.
This is important: it must not use "libnm-core/nm-core-internal.h"
nor "libnm-core/nm-utils-private.h" so the static library is usable
by nmcli which couldn't access these.
Note that "shared/nm-meta-setting.c" is an entirely different case,
because it behaves differently depending on whether linking against
"libnm-core" or the client programs. As such, this file must be compiled
twice.
(cherry picked from commit af07ed01c0)
From the files under "shared/nm-utils" we build an internal library
that provides glib-based helper utilities.
Move the files of that basic library to a new subdirectory
"shared/nm-glib-aux" and rename the helper library "libnm-core-base.la"
to "libnm-glib-aux.la".
Reasons:
- the name "utils" is overused in our code-base. Everything's an
"utils". Give this thing a more distinct name.
- there were additional files under "shared/nm-utils", which are not
part of this internal library "libnm-utils-base.la". All the files
that are part of this library should be together in the same
directory, but files that are not, should not be there.
- the new name should better convey what this library is and what is isn't:
it's a set of utilities and helper functions that extend glib with
funcitonality that we commonly need.
There are still some files left under "shared/nm-utils". They have less
a unifying propose to be in their own directory, so I leave them there
for now. But at least they are separate from "shared/nm-glib-aux",
which has a very clear purpose.
(cherry picked from commit 80db06f768)
We built (among others) two libraries from the sources in "shared/nm-utils":
"libnm-utils-base.la" and "libnm-utils-udev.la".
It's confusing. Instead use directories so there is a direct
correspondence between these internal libraries and the source files.
(cherry picked from commit 2973d68253)
"shared/nm-utils" contains general purpose utility functions that only
depend on glib (and extend glib with some helper functions).
We will also add code that does not use glib, hence it would be good
if the part of "shared/nm-utils" that does not depend on glib, could be
used by these future projects.
Also, we use the term "utils" everywhere. While that covers the purpose
and content well, having everything called "nm-something-utils" is not
great. Instead, call this "nm-std-aux", inspired by "c-util/c-stdaux".
(cherry picked from commit b434b9ec07)
Also, in nm_platform_routing_rule_cmp() always compare the routing
table field, also if l3mdev is set. For kernel, we cannot set table and
l3mdev together, hence such rules don't really exist (or if we try to
configure it, it will be rejected by kernel). But as far as
nm_platform_routing_rule_cmp() is concerned, if the table is set,
always compare it.
(cherry picked from commit b6ff02e76f)
For routes and routing rules, kernel uses a certain (not stictly defined) set
of attributes to decide whether to routes/rules are identical.
That is a problem, as different kernel versions disagree on whether
two routes/rules are the same (EEXIST) or not.
Note that when NetworkManager tries to add a rule with protocol set to
anything but RTPROT_UNSPEC, then kernel will ignore the attribute if it
doesn't have support for it. Meaning: the added rule will have a
different protocol setting then intended.
Note that NMPRulesManager will add a rule if it doesn't find it in the
platform cache so far. That means, when looking into the platform cache
we must ignore or honor the protocol like kernel does.
This does not only affect FRA_PROTOCOL, but all attributes where kernel
and NetworkManager disagrees. But the protocol is the most prominent
one, because the rules tracked by nmp_rules_manager_track_default()
specify the protocol.
(cherry picked from commit ef4f8ccf6d)
Next we will need to detect more kernel features. First refactor the
handling of these to require less code changes and be more efficient.
A plain nm_platform_kernel_support_get() only reqiures to access an
array in the common case.
The other important change is that the function no longer requires a
NMPlatform instance. This allows us to check kernel support from
anywhere. The only thing is that we require kernel support to be
initialized before calling this function. That means, an NMPlatform
instance must have detected support before.
(cherry picked from commit ee269b318e)
In some cases it is convenient to specify ranges of bridge vlans, as
already supported by iproute2 and natively by kernel. With this commit
it becomes possible to add a range in this way:
nmcli connection modify eth0-slave +bridge-port.vlans "100-200 untagged"
vlan ranges can't be PVIDs because only one PVID vlan can exist.
https://bugzilla.redhat.com/show_bug.cgi?id=1652910
(cherry picked from commit 7093515777)
Policy routing rules are global, and unlike routes not tied to an interface by ifindex.
That means, while we take full control over all routes of an interface during a sync,
we need to consider that multiple parties can contribute to the global set of rules.
That might be muliple connection profiles providing the same rule, or rules that are added
externally by the user. NMPRulesManager mediates for that.
This is done by NMPRulesManager "tracking" rules.
Rules that are not tracked by NMPRulesManager are completely ignored (and
considered externally added).
When tracking a rule, the caller provides a track-priority. If multiple
parties track a rule, then the highest (absolute value of the) priority
wins.
If the highest track-priority is positive, NMPRulesManager will add the rule if
it's not present.
When the highest track-priority is negative, then NMPRulesManager will remove the
rule if it's present (enforce its absence).
The complicated part is, when a rule that was previously tracked becomes no
longer tracked. In that case, we need to restore the previous state.
If NetworkManager added the rule earlier, then untracking the rule
NMPRulesManager will remove the rule again (restore its previous absent
state).
By default, if NetworkManager had a negative tracking-priority and removed the
rule earlier (enforced it to be absent), then when the rule becomes no
longer tracked, NetworkManager will not restore the rule.
Consider: the user adds a rule externally, and then activates a profile that
enforces the absence of the rule (causing NetworkManager to remove it).
When deactivating the profile, by default NetworkManager will not
restore such a rule! It's unclear whether that is a good idea, but it's
also unclear why the rule is there and whether NetworkManager should
really restore it.
Add weakly tracked rules to account for that. A tracking-priority of
zero indicates such weakly tracked rules. The only difference between an untracked
rule and a weakly tracked rule is, that when NetworkManager earlier removed the
rule (due to a negative tracking-priority), it *will* restore weakly
tracked rules when the rules becomes no longer (negatively) tracked.
And it attmpts to do that only once.
Likewise, if the rule is weakly tracked and already exists when
NMPRulesManager starts posively tracking the rule, then it would not
remove again, when no longer positively tracking it.
- fix the argument type to be "gint32" and not "int".
- assert in nmp_rules_manager_track_default() for the input
arguments.
- use boolean bitfield in private data.
The name "priority" is overused. Also rules have a "priority", but that'
something else.
Rename the priority of how rules are tracked by NMPRuleManager to
"track_priority".
All that setting track-default does, is calling nmp_rules_manager_track_default()
when the rules are first accessed.
That is not right API. Since nmp_rules_manager_track_default() is already public
API (good), every caller that wishes this behavior should track these routes explicitly.
Seems on a busy system, we can hit this timeout. Increase it.
ERROR:../src/platform/tests/test-common.c:939:_ip_address_add: code should not be reached
'sriov_drivers_autoprobe' was added in kernel 4.12. With previous
kernel versions NM is currently unable to set any SR-IOV parameter
because it tries to read 'sriov_drivers_autoprobe' which doesn't
exist, assumes that current value is -1 and tries to change it,
failing.
When the file doesn't exist, drivers are automatically probed so we
can assume the value is 1. In this way NM is able to activate a
connection with sriov.autoprobe-drivers=1 (the default) even on older
kernel versions.
Fixes: 1e41495d9a ('platform: sriov: write new values when we can't read old ones')
https://bugzilla.redhat.com/show_bug.cgi?id=1695093
# NetworkManager-MESSAGE: <warn> [1553100541.6609] platform-linux: do-add-rule: failure 17 (File exists)
>>> failing... errno=-17, rule=[routing-rule,0xe9c540,1,+alive,+visible; [6] 4294967295: from all suppress_prefixlen 3 none goto-target 2955537847]
0: from all to 73.165.79.8/2 iif nm-test-device 178
0: from all 109
0: from all tos 0x13 lookup 10004 suppress_prefixlength 0 none
0: from all none
4294967295: not from all none
test:ERROR:../src/platform/tests/test-route.c:1607:test_rule: assertion failed (r == 0): (-17 == 0)
Possibly fixed by https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=7c8f4e6dc30996bff806285730a0bb4e714d3d52
The routing-rule tests generate a number of routing rules and tries to
add and delete them.
For that, _rule_create_random() sets random fields of the rule.
Note that especially interesting are rules that leave most fields
unset (at zero), because they trigger kernel issues rh#1686075 and
rh#1685816.
But a rule has many fields, so in order to generate rules that have most
fields unset, we need to use low probabilities when rolling the dice for
setting a field. Otherwise, most rules end up with several fields set
and don't reproduce the kernel issue (especially the test failed to hit
rh#1686075).
Routing rules are unlike addresses or routes not tied to an interface.
NetworkManager thinks in terms of connection profiles. That works well
for addresses and routes, as one profile configures addresses and routes
for one device. For example, when activating a profile on a device, the
configuration does not interfere with the addresses/routes of other
devices. That is not the case for routing rules, which are global, netns-wide
entities.
When one connection profile specifies rules, then this per-device configuration
must be merged with the global configuration. And when a device disconnects later,
the rules must be removed.
Add a new NMPRulesManager API to track/untrack routing rules. Devices can
register/add there the routing rules they require. And the sync method will
apply the configuration. This is be implemented on top of NMPlatform's
caching API.
