Re: [PATCH-next v10 10/11] cgroup/cpuset: Support multiple destination cpusets for cpuset_*attach()

[Waiman Long <longman@redhat.com>]

On 7/6/26 8:57 PM, Tejun Heo wrote:
> Hello, Waiman.
>
> On Thu, Jul 02, 2026 at 05:47:56PM -0400, Waiman Long wrote:
>> It is assumed that a given cpuset cannot be both a source and a
>> destination cpuset. [...] it will print a warning and fail the attach
>> operation in these unexpected cases [...]
> This assumption doesn't hold - the WARN_ON_ONCE() and -EINVAL fire on a
> legitimate migration. It's the case sashiko flagged, and it does reach
> can_attach().
>
> Threaded subtree with partial cpuset delegation:
>
>    P (+cpuset)
>    |- R (cpuset)        <- destination
>    |  `- C (no cpuset)  -> effective cpuset == R
>    `- W (cpuset)
>
> Group leader in R, thread_a in C, thread_b in W; migrate the whole
> process into R (echo $PID > R/cgroup.procs). thread_a moves C->R: its
> cgroup changes so compare_css_sets() keeps it in the taskset, but its
> cpuset css is unchanged (C inherits R's), so task_cs() == cs == R. cpuset
> is in ss_mask because thread_b (W->R) changed. can_attach() then tags R
> as a source (thread_a) and the destination (thread_b):
>
>    WARNING: kernel/cgroup/cpuset.c:3054 at cpuset_can_attach_check+0xcd/0x130
>     cpuset_can_attach+0x131/0x2f0
>     cgroup_migrate_execute+0x367/0x450
>    cgroup.procs write returns -EINVAL
>
> So a thread already in the destination's effective cpuset does show up in
> the iteration when it reaches that cpuset through a different, non-cpuset
> child - compare_css_sets() keys on the cgroup, not the cpuset css.
>
> 1-9 handle this correctly (the migration succeeds); the patch 10 guard
> regresses it. I've taken 1-9 into for-7.3 - please respin 10-11. One
> option: when oldcs == cs the pair is a cpuset no-op, so skip it and add
> it to neither list.

Sorry for the late reply. I was offline for a few days because my labtop 
SSD was having problem.

Yes, I realize that my assumption about oldcs != cs is incorrect. I will 
revise the remaining patch to address the issue.

Thanks,
Longman