Re: [BUG] cgroup/cpuset: Concurrent WARN_ON_ONCE triggers during remote partition stress-test

[Zhang Qiao <zhangqiao22@huawei.com>]

在 2026/8/31 22:18, Waiman Long 写道:
> On 8/31/26 5:05 AM, Zhang Qiao wrote:
>> Hi,
>>
>> While stress-testing cpuset partitions under KASAN (kernel 7.2-rc1,
>> dc59e4fea9d83 "Linux 7.2-rc1"), four distinct WARN_ON_ONCE() in
>> kernel/cgroup/cpuset.c fire within a ~1s window. They all concern the
>> remote partition / effective-cpumask invariants and are triggered by
>> concurrent CPU hotplug, cpuset.cpus/cpuset.cpus.exclusive/
>> cpuset.cpus.partition writes, task migration and (un)partitioning of
>> nested subgroups.
>>
>> Environment
>> -----------
>> - Kernel : Linux 7.2-rc1, generic KASAN enabled, x86_64, PREEMPT
>> - server   : Intel Xeon Platinum 8380 @ 2.30GHz, 2 NUMA nodes, 160 CPUs
>> - Cmdline: ... cgroup_disable=files apparmor=0    
>> systemd.unified_cgroup_hierarchy=1
>>
>> Reproduction
>> ------------
>> A pure-shell self-contained script (attached below) runs several concurrent
>> "disturbance" loops. The issue is extremely easy to reproduce; the script
>> consistently triggers the WARNs within seconds of execution.
>>
>>    - CPU hotplug toggle of a helper pool (CPU online/offline)
>>    - random writes to cpuset.cpus / cpuset.cpus.exclusive
>>      (single CPU, empty, or range)
>>    - random root <-> member switching of cpuset.cpus.partition
>>    - migrating burner tasks between subgroups via cgroup.procs
>>    - creating and removing nested sub-partitions under S0-S3
> 
> Thank for reporting the bug. It is recently known that there are bugs in the
> handling of nested sub-partitions. We are in the process of getting it fixed.
> 
> Cheers,
> Longman

Hi Waiman,

Thanks for the update!

I'm looking forward to the fix. Please CC me when the patch is ready, and I'll
be happy to test it on my end.

Thanks,
Zhang Qiao

> 
> 
> .