This handles a corner case when joining a container having all
the processes running exclusively on isolated CPU cores to force
the kernel to schedule runc process on the first CPU core within the
cgroups cpuset.
The introduction of the kernel commit
46a87b3851f0d6eb05e6d83d5c5a30df0eca8f76 has affected this deterministic
scheduling behavior by distributing tasks across CPU cores within the
cgroups cpuset. Some intensive real-time application are relying on this
deterministic behavior and use the first CPU core to run a slow thread
while other CPU cores are fully used by real-time threads with SCHED_FIFO
policy. Such applications prevents runc process from joining a container
when the runc process is randomly scheduled on a CPU core owned by a
real-time thread.
Introduces isolated CPU affinity transition OCI runtime annotation
org.opencontainers.runc.exec.isolated-cpu-affinity-transition to restore
the behavior during runc exec.
Fix issue with kernel >= 6.2 not resetting CPU affinity for container processes.
Signed-off-by: Cédric Clerget <cedric.clerget@gmail.com>
In code we have frozen the cgroup to avoid the processes get
an occasional "permission denied" error, while the systemd's application of device
rules is done disruptively. When the processes in the container can not
be frozen over 2 seconds (which defined in fs/freezer.go),
we still update the cgroup which resulting the container get an occasional
"permission denied" error in some cases.
Return error directly without updating cgroup, when freeze fails.
Fixes: #3803
Signed-off-by: Zoe <hi@zoe.im>
Commit 343951a22b added a call to os.Stat for the device path
when generating systemd device properties, to avoid systemd warning for
non-existing devices. The idea was, since systemd uses stat(2) to look
up device properties for a given path, it will fail anyway. In addition,
this allowed to suppress a warning like this from systemd:
> Couldn't stat device /dev/char/10:200
NOTE that this was done because:
- systemd could not add the rule anyway;
- runs puts its own set of rules on top of what systemd does.
Apparently, the above change broke some setups, resulting in inability
to use e.g. /dev/null inside a container. My guess is this is because
in cgroup v2 we add a second eBPF program, which is not used if the
first one (added by systemd) returns "access denied".
Next, commit 3b9582895b fixed that by adding a call to os.Stat for
"/sys/"+path (meaning, if "/dev/char/10:200" does not exist, we retry
with "/sys/dev/char/10:200", and if it exists, proceed with adding a
device rule with the original (non-"/sys") path).
How that second fix ever worked was a mystery, because the path we gave
to systemd still doesn't exist.
Well, I think now I know.
Since systemd v240 (commit 74c48bf5a8005f20) device access rules
specified as /dev/{block|char}/MM:mm are no longer looked up on the
filesystem, instead, if possible, those are parsed from the string.
So, we need to do different things, depending on systemd version:
- for systemd >= v240, use the /dev/{char,block}/MM:mm as is, without
doing stat() -- since systemd doesn't do stat() either;
- for older version, check if the path exists, and skip passing it on
to systemd otherwise.
- the check for /sys/dev/{block,char}/MM:mm is not needed in either
case.
Pass the systemd version to the function that generates the rules, and
fix it accordingly.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Commit d223e2adae ("Ignore error when starting transient unit
that already exists" modified the code handling errors from startUnit
to ignore UnitExists error.
Apparently it was done so that kubelet can create the same pod slice
over and over without hitting an error (see [1]).
While it works for a pod slice to ensure it exists, it is a gross bug
to ignore UnitExists when creating a container. In this case, the
container init PID won't be added to the systemd unit (and to the
required cgroup), and as a result the container will successfully
run in a current user cgroup, without any cgroup limits applied.
So, fix the code to only ignore UnitExists if we're not adding a process
to the systemd unit. This way, kubelet will keep working as is, but
runc will refuse to create containers which are not placed into a
requested cgroup.
[1] https://github.com/opencontainers/runc/pull/1124
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Sometimes, the init process is not in the root cgroup.
This can be noted by GetInitPath, which already scrubs the path of `init.scope`.
This was encountered when trying to patch the Kubelet to handle systemd being in a separate cpuset
from root (to allow load balance disabling for containers). At present, there's no way to have libcontainer or runc
manage cgroups in a hierarchy outside of the one init is in (unless the path contains `init.scope`, which is limiting)
Signed-off-by: Peter Hunt <pehunt@redhat.com>
It's more idiomatic Go to define interfaces on the receiver, and constructors to
return concrete types.
This patch changes various constructors to return a concrete type, with the
exceptions of NewWithPaths, which needs the abstraction as it switches between
implementations.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
This moves the functionality related to devices, SkipDevices, and
SkipFreezeOnSet to a separate file, in preparation for the next commit.
No code changes.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Currently the parent process of the container is moved to the right
cgroup v2 tree when systemd is using a hybrid model (last line with 0::):
$ runc --systemd-cgroup run myid
/ # cat /proc/self/cgroup
12:cpuset:/system.slice/runc-myid.scope
11:blkio:/system.slice/runc-myid.scope
10:devices:/system.slice/runc-myid.scope
9:hugetlb:/system.slice/runc-myid.scope
8:memory:/system.slice/runc-myid.scope
7:rdma:/
6:perf_event:/system.slice/runc-myid.scope
5:net_cls,net_prio:/system.slice/runc-myid.scope
4:freezer:/system.slice/runc-myid.scope
3:pids:/system.slice/runc-myid.scope
2:cpu,cpuacct:/system.slice/runc-myid.scope
1:name=systemd:/system.slice/runc-myid.scope
0::/system.slice/runc-myid.scope
However, if a second process is executed in the same container, it is
not moved to the right cgroup v2 tree:
$ runc exec myid /bin/sh -c 'cat /proc/self/cgroup'
12:cpuset:/system.slice/runc-myid.scope
11:blkio:/system.slice/runc-myid.scope
10:devices:/system.slice/runc-myid.scope
9:hugetlb:/system.slice/runc-myid.scope
8:memory:/system.slice/runc-myid.scope
7:rdma:/
6:perf_event:/system.slice/runc-myid.scope
5:net_cls,net_prio:/system.slice/runc-myid.scope
4:freezer:/system.slice/runc-myid.scope
3:pids:/system.slice/runc-myid.scope
2:cpu,cpuacct:/system.slice/runc-myid.scope
1:name=systemd:/system.slice/runc-myid.scope
0::/user.slice/user-1000.slice/session-8.scope
This commit makes that processes executed with exec are placed into the
right cgroup v2 tree. The implementation checks if systemd is using a
hybrid mode (by checking if cgroups v2 is mounted in
/sys/fs/cgroup/unified), if yes, the path of the cgroup v2 slice for
this container is saved into the cgroup path list.
The fs group driver has a similar issue, in this case none of the runc
run or runc exec commands put the process in the right cgroups v2. This
commit also fixes that.
Having the processes of the container in its own cgroup v2 is useful
for any BPF programs that rely on bpf_get_current_cgroup_id(), like
https://github.com/kinvolk/inspektor-gadget/ for instance.
[@kolyshkin: rebased]
Signed-off-by: Mauricio Vásquez <mauricio@kinvolk.io>
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Cgroup controllers should never panic, and yet sometimes they do.
Add a unit test to check that controllers never panic when called with
nil arguments and/or resources, and fix a few found cases.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This way we
- won't re-initialize the paths if they were provided;
- will always have paths ready for every method.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
1. Make Rootless and Systemd flags part of config.Cgroups.
2. Make all cgroup managers (not just fs2) return error (so it can do
more initialization -- added by the following commits).
3. Replace complicated cgroup manager instantiation in factory_linux
by a single (and simple) libcontainer/cgroups/manager.New() function.
4. getUnifiedPath is simplified to check that only a single path is
supplied (rather than checking that other paths, if supplied,
are the same).
[v2: can't -> cannot]
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
As ExpandSlice("system.slice") returns "/system.slice", there is no need
to call it for such paths (and the slash will be added by path.Join
anyway).
The same optimization was already done for v2 as part of commit
bf15cc99b1.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
We were checking if a unit is a slice two times. Consolidate those
checks, and improve comments while we're at it.
The code is the same in v1 and v2 but it's too complicated to factor it
out, thus we just do the same changes in v1.go and v2.go.
No functional change.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Only some libcontainer packages can be built on non-linux platforms
(not that it make sense, but at least go build succeeds). Let's call
these "good" packages.
For all other packages (i.e. ones that fail to build with GOOS other
than linux), it does not make sense to have linux build tag (as they
are broken already, and thus are not and can not be used on anything
other than Linux).
Remove linux build tag for all non-"good" packages.
This was mostly done by the following script, with just a few manual
fixes on top.
function list_good_pkgs() {
for pkg in $(find . -type d -print); do
GOOS=freebsd go build $pkg 2>/dev/null \
&& GOOS=solaris go build $pkg 2>/dev/null \
&& echo $pkg
done | sed -e 's|^./||' | tr '\n' '|' | sed -e 's/|$//'
}
function remove_tag() {
sed -i -e '\|^// +build linux$|d' $1
go fmt $1
}
SKIP="^("$(list_good_pkgs)")"
for f in $(git ls-files . | grep .go$); do
if echo $f | grep -qE "$SKIP"; then
echo skip $f
continue
fi
echo proc $f
remove_tag $f
done
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This is helpful to kubernetes in cases it knows for sure that the freeze
is not required (since it created the systemd unit with no device
restrictions).
As the code is trivial, no tests are required.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This fixes the behavior intended to avoid freezing containers/control
groups without it being necessary. This is important for end users of
libcontainer who rely on the behavior of no freeze.
The previous implementation would always get error trying to get
DevicePolicy from the Unit via dbus, since the Unit interface doesn't
contain DevicePolicy.
Signed-off-by: Odin Ugedal <odin@uged.al>
This was initially added by commits 41d9d26513 and 4a8f0b4db4,
apparently to implement docker run --cgroup container:ID, which was
never merged. Therefore, this code is not and was never used.
It needs to be removed mainly because having it makes it much harder to
understand how cgroup manager works (because with this in place we have
not one or two but three sets of cgroup paths to think about).
Note if the paths are known and there is a need to add a PID to existing
cgroup, cgroup manager is not needed at all -- something like
cgroups.WriteCgroupProc or cgroups.EnterPid is sufficient (and the
latter is what runc exec uses in (*setnsProcess).start).
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Introduce freezeBeforeSet, which contains the logic of figuring out
whether we need to freeze/thaw around setting systemd unit properties.
In particular, if SkipDevices is set, and the current unit properties
allow all devices, there is no need to freeze and thaw, as systemd
won't write any device rules in this case.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
m.Freeze method changes m.cgroups.Resources.Freezer field, which should
not be done while we're temporarily freezing the cgroup in Set. If this
field is changed, and r == m.cgroups.Resources (as it often happens),
this results in inability to freeze the container using Set().
To fix, add and use a method which does not change r.Freezer field.
A test case for the bug will be added separately.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
1. The meaning of SkipDevices is what it is -- do not set any
device-related options.
2. Reverts the part of commit 108ee85b82 which skipped the freeze
when the SkipDevices is set. Apparently, the freeze is needed on
update even if no Device* properties are being set.
3. Add "runc update" to "runc run [device cgroup deny]" test.
Fixes: 752e7a8249
Fixes: 108ee85b82
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
gofumpt (mvdan.cc/gofumpt) is a fork of gofmt with stricter rules.
Brought to you by
git ls-files \*.go | grep -v ^vendor/ | xargs gofumpt -s -w
Looking at the diff, all these changes make sense.
Also, replace gofmt with gofumpt in golangci.yml.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Commit 108ee85b82 adds SkipDevices flag, which is used by kubernetes
to create cgroups for pods.
Unfortunately the above commit falls short, and systemd DevicePolicy and
DeviceAllow properties are still set, which requires kubernetes to set
"allow everything" rule.
This commit fixes this: if SkipDevices flag is set, we return
Device* properties to allow all devices.
Fixes: 108ee85b82
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Historically, we never returned an error from failed startUnit
or stopUnit. The startUnit case was fixed by commit 3844789.
It is time to fix stopUnit, too. The reasons are:
1. Ignoring an error from stopUnit means an unexpected trouble down the
road, for example a failure to create a container with the same name:
> time="2021-05-07T19:51:27Z" level=error msg="container_linux.go:380: starting container process caused: process_linux.go:385: applying cgroup configuration for process caused: Unit runc-test_busybox.scope already exists."
2. A somewhat short timeout of 1 second means the cgroup might
actually be removed a few seconds later but we might have a
race between removing the cgroup and creating another one
with the same name, resulting in the same error as amove.
So, return an error if removal failed, and increase the timeout.
Now, modify the systemd cgroup v1 manager to not mask the error from
stopUnit (stopErr) with the subsequent one from cgroups.RemovePath,
as stopErr is most probably the reason why RemovePath failed.
Note that for v1 we do want to remove the paths even in case of a
failure from stopUnit, as some were not created by systemd.
There's no need to do that for v2, thanks to unified hierarchy,
so no changes there.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
A cgroup manager's Set method sets cgroup resources, but historically
it was accepting configs.Cgroups.
Refactor it to accept resources only. This is an improvement from the
API point of view, as the method can not change cgroup configuration
(such as path to the cgroup etc), it can only set (modify) its
resources/limits.
This also lays the foundation for complicated resource updates, as now
Set has two sets of resources -- the one that was previously specified
during cgroup manager creation (or the previous Set), and the one passed
in the argument, so it could deduce the difference between these. This
is a long term goal though.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Instead of reconnecting to dbus after some failed operations, and
returning an error (so a caller has to retry), reconnect AND retry
in place for all such operations.
This should fix issues caused by a stale dbus connection after e.g.
a dbus daemon restart.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
[@kolyshkin: doc nits, use dbus.ErrClosed and isDbusError]
Signed-off-by: Shiming Zhang <wzshiming@foxmail.com>
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This is somewhat radical approach to deal with kernel memory.
Per-cgroup kernel memory limiting was always problematic. A few
examples:
- older kernels had bugs and were even oopsing sometimes (best example
is RHEL7 kernel);
- kernel is unable to reclaim the kernel memory so once the limit is
hit a cgroup is toasted;
- some kernel memory allocations don't allow failing.
In addition to that,
- users don't have a clue about how to set kernel memory limits
(as the concept is much more complicated than e.g. [user] memory);
- different kernels might have different kernel memory usage,
which is sort of unexpected;
- cgroup v2 do not have a [dedicated] kmem limit knob, and thus
runc silently ignores kernel memory limits for v2;
- kernel v5.4 made cgroup v1 kmem.limit obsoleted (see
https://github.com/torvalds/linux/commit/0158115f702b).
In view of all this, and as the runtime-spec lists memory.kernel
and memory.kernelTCP as OPTIONAL, let's ignore kernel memory
limits (for cgroup v1, same as we're already doing for v2).
This should result in less bugs and better user experience.
The only bad side effect from it might be that stat can show kernel
memory usage as 0 (since the accounting is not enabled).
[v2: add a warning in specconv that limits are ignored]
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
This unconditionally enables TasksAccounting for systemd unified (v2)
cgroup driver, making it work the same way as the legacy (v1) driver.
Practically, it is probably a no-op since DefaultTasksAccounting is
usually true.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
In some cases, container init fails to start because it is killed by
the kernel OOM killer. The errors returned by runc in such cases are
semi-random and rather cryptic. Below are a few examples.
On cgroup v1 + systemd cgroup driver:
> process_linux.go:348: copying bootstrap data to pipe caused: write init-p: broken pipe
> process_linux.go:352: getting the final child's pid from pipe caused: EOF
On cgroup v2:
> process_linux.go:495: container init caused: read init-p: connection reset by peer
> process_linux.go:484: writing syncT 'resume' caused: write init-p: broken pipe
This commits adds the OOM method to cgroup managers, which tells whether
the container was OOM-killed. In case that has happened, the original error
is discarded (unless --debug is set), and the new OOM error is reported
instead:
> ERRO[0000] container_linux.go:367: starting container process caused: container init was OOM-killed (memory limit too low?)
Also, fix the rootless test cases that are failing because they expect
an error in the first line, and we have an additional warning now:
> unable to get oom kill count" error="no directory specified for memory.oom_control
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
All cgroup managers has Apply() and Set() methods:
- Apply is used to create a cgroup (and, in case of systemd,
a systemd unit) and/or put a PID into the cgroup (and unit);
- Set is used to set various cgroup resources and limits.
The fs/fs2 cgroup manager implements the functionality as described above.
The systemd v1/v2 manager deviate -- it sets *most* of cgroup limits
(those that can be projected to systemd unit properties) in Apply(),
and then again *all* cgroup limits in Set (first indirectly via systemd
properties -- same as in Apply, then via cgroupfs).
This commit removes setting the cgroup limits from Apply,
so now the systemd manager behaves the same way as the fs manager.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Support for systemd properties AllowedCPUs and AllowedMemoryNodes
was added by commit 13afa58d0e, but only for unified resources
of systemd v2 driver.
This adds support for Cpu.Cpus and Cpu.Mems resources to
both systemd v1 and v2 cgroup drivers.
An integration test is added to check that the settings work.
[v2: check for systemd version]
[v3: same in the test]
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Commit 27d3dd3df3 ("don't fail when subsystem not mounted") added
ignoring "not found" error to enableKmem, and as a result the function
now tries to call Mkdir with an empty path, which results in a weird
error message. For example, this is a failure from a
libcontainer/integration test:
> === RUN TestRunWithKernelMemorySystemd
> exec_test.go:704: runContainer failed with kernel memory limit: container_linux.go:370: starting container process caused: process_linux.go:327: applying cgroup configuration for process caused: mkdir : no such file or directory
I am not entirely sure if it is a good idea to silently ignore set
limits, but at least let's fix the error handling.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Commit a1d5398afa ("Respect container's cgroup path") added a
cgroupPath argument to FindCgroupMountpoint to make runc/libcontainer
work in a custom multitenant environment with multiple cgroup mount
points.
It also added passing c.Path as an argument to FindCgroupMountpoint
for systemd (v1) controller. This is wrong, because
1. systemd controller do not use c.Path at all (and c.Path is never set
by specconv) -- instead, it uses Name and Parent.
2. c.Path, if set, is not absolute -- it is relative to /sys/fs/cgroup
-- but it is used as an absolute path here.
Since c.Path is never set, the change did not result in any breakage, so
this code sit quietly for some time and the issue might not have been
discovered -- until we started running libcontainer/integration tests
in a CentOS 7 VM, which resulted in a following weird error:
> FAIL: TestPidsSystemd: utils_test.go:55: exec_test.go:630: unexpected error: container_linux.go:353: starting container process caused: process_linux.go:326: applying cgroup configuration for process caused: mountpoint for devices not found
The error was "fixed" in commit f57bb2fe3d by changing the tests'
cgroups Path to be "/sys/fs/cgroup/". This actually resulted in
creation of cgroup directories like /sys/fs/cgroup/memory/sys/fs/cgroup,
/sys/fs/cgroup/devices/sys/fs/cgroup and so on.
The proper fix to the test case is implemented in the previous commit,
which sets c.Name and c.Parent.
This commit just removes the invalid use of c.Path, and tells the whole
story.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
In all these cases, getSubsystemPath() was already called, and its
result stored in m.paths map. It makes no sense to not reuse it.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
We call joinCgroups() from Apply, and in there we iterate through the
list of subsystems, calling getSubsystemPath() for each. This is
expensive, since every getSubsystemPath() involves parsing mountinfo.
At the end of Apply(), we iterate through the list of subsystems to fill
the m.paths, again calling getSubsystemPath() for every subsystem.
As a result, we parse mountinfo about 20 times here.
Let's find the paths first and reuse m.paths in joinCgroups().
While at it, since join() is just two calls now, inline it.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
When paths are set, we only need to place the PID into proper
cgroups, and we do know all the paths already.
Both fs/d.path() and systemd/v1/getSubsystemPath() parse
/proc/self/mountinfo, and the only reason they are used
here is to check whether the subsystem is available.
Use a much simpler/faster check instead.
Frankly, I am not sure why the check is needed at all. Maybe it should
be dropped.
Also, for fs driver, since d is no longer used in this code path,
move its initialization to after it.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Cédric Clerget