test: let kernel OOM-kill a child process instead of the main one

This test occasionally fails due to a race where systemd processes
kernel's SIGKILL before the OOM notification, so the test service dies
with Result=signal instead of the expected Result=oom-kill:

[   51.008765] TEST-55-OOMD.sh[906]: + systemd-run --wait --unit oom-kill -p OOMPolicy=kill -p Delegate=yes -p DelegateSubgroup=init.scope /tmp/script.sh
[   51.048747] TEST-55-OOMD.sh[907]: Running as unit: oom-kill.service; invocation ID: 456645347d554ea2878463404b181bd8
[   51.066296] sysrq: Manual OOM execution
[   51.066596] kworker/1:0 invoked oom-killer: gfp_mask=0xcc0(GFP_KERNEL), order=-1, oom_score_adj=0
[   51.066915] CPU: 1 UID: 0 PID: 27 Comm: kworker/1:0 Not tainted 6.17.1-arch1-1 #1 PREEMPT(full)  d2b229857b2eb4001337041f41d3c4f131433540
[   51.066919] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS Arch Linux 1.17.0-2-2 04/01/2014
[   51.066921] Workqueue: events moom_callback
[   51.066928] Call Trace:
[   51.066931]  <TASK>
[   51.066936]  dump_stack_lvl+0x5d/0x80
[   51.066942]  dump_header+0x43/0x1aa
<...snip...>
[   51.087814] 47583 pages reserved
[   51.087969] 0 pages cma reserved
[   51.088208] 0 pages hwpoisoned
[   51.088519] Out of memory: Killed process 908 (sleep) total-vm:3264kB, anon-rss:256kB, file-rss:1916kB, shmem-rss:0kB, UID:0 pgtables:44kB oom_score_adj:1000
[   51.090263] TEST-55-OOMD.sh[907]:           Finished with result: signal
[   51.094416] TEST-55-OOMD.sh[907]: Main processes terminated with: code=killed, status=9/KILL
[   51.094898] TEST-55-OOMD.sh[907]:                Service runtime: 58ms
[   51.095436] TEST-55-OOMD.sh[907]:              CPU time consumed: 22ms
[   51.095854] TEST-55-OOMD.sh[907]:                    Memory peak: 1.6M (swap: 0B)
[   51.096722] TEST-55-OOMD.sh[912]: ++ systemctl show oom-kill -P Result
[   51.106549] TEST-55-OOMD.sh[879]: + assert_eq signal oom-kill
[   51.107394] TEST-55-OOMD.sh[913]: + set +ex
[   51.108256] TEST-55-OOMD.sh[913]: FAIL: expected: 'oom-kill' actual: 'signal'
[FAILED] Failed to start TEST-55-OOMD.service.

To mitigate this, let's spawn a child process and move it to the
subcgroup to get killed instead of the main process, so systemd has more
time to react to the OOM notification and terminate the service with the
expected oom-kill result.
This commit is contained in:
Frantisek Sumsal
2025-10-16 13:06:51 +02:00
committed by Daan De Meyer
parent b8ee7c2206
commit 1ec5d43a0e

View File

@@ -301,6 +301,7 @@ testcase_reload() {
testcase_kernel_oom() {
cat >/tmp/script.sh <<"EOF"
#!/usr/bin/env bash
set -x
choom --adjust '+1000' -- bash -c 'echo f >/proc/sysrq-trigger && exec sleep infinity'
choom --adjust '+1000' -p $$
echo f >/proc/sysrq-trigger
@@ -324,12 +325,22 @@ EOF
cat >/tmp/script.sh <<"EOF"
#!/usr/bin/env bash
set -x
echo '+memory' >/sys/fs/cgroup/system.slice/oom-kill.service/cgroup.subtree_control
mkdir /sys/fs/cgroup/system.slice/oom-kill.service/sub
echo 1 >/sys/fs/cgroup/system.slice/oom-kill.service/sub/memory.oom.group
echo $$ >/sys/fs/cgroup/system.slice/oom-kill.service/sub/cgroup.procs
choom --adjust '+1000' -p $$
echo f >/proc/sysrq-trigger
# Start a child process in the subcgroup that will trigger OOM and be killed but keep the main process
# outside the subcgroup to avoid a race condition where the kernel SIGKILLs the main process before systemd
# can process the OOM notification. With the main process still alive, systemd should have time to receive
# the OOM event and enter the 'oom-kill' state before the service exits.
(
echo $BASHPID >/sys/fs/cgroup/system.slice/oom-kill.service/sub/cgroup.procs
choom --adjust '+1000' -p $BASHPID
echo f >/proc/sysrq-trigger
exec sleep infinity
) &
wait $! || :
exec sleep infinity
EOF
chmod +x /tmp/script.sh