Re: [PATCH v3 2/3] cgroup/cpuset: Allow top-level partition CPU replacement

[Guopeng Zhang <guopeng.zhang@linux.dev>]

在 2026/11/10 20:11, Ridong Chen 写道:
> 
> 
> On 11/10/2026 1:03 PM, Guopeng Zhang wrote:
>> From: Guopeng Zhang <zhangguopeng@kylinos.cn>
>>
>> When replacing a partition's CPUs, tasks_nocpu_error() checks whether
>> the parent's current effective_cpus is a subset of the new exclusive
>> CPU mask. The parent mask does not yet include CPUs being returned by
>> the update, so this can invalidate an otherwise valid partition.
>>
>> For example, on an 8-CPU system without boot isolation:
>>
>>    echo 2-7 > A/cpuset.cpus
>>    echo isolated > A/cpuset.cpus.partition
>>    echo 0-4 > A/cpuset.cpus
>>    cat A/cpuset.cpus.partition
>>    isolated invalid (Parent unable to distribute cpu downstream)
>>
>> The update should return CPUs 5-7 to the parent and leave A as a
>> valid isolated partition.
>>
>> Skip this subset check for top_cpuset. top_cpuset_sd_nocpu() already
>> accounts for returned CPUs and ensures that an active boot domain
>> housekeeping CPU remains.
>>
>> Add a regression test for this case.
>>
>> Fixes: 0c7f293efc87 ("cgroup/cpuset: Add cpuset.cpus.exclusive.effective for v2")
>> Signed-off-by: Guopeng Zhang <zhangguopeng@kylinos.cn>
> 
> Hi Guopeng,
> 
> Patch 2 depends on patch 1. Skipping the cpumask_subset() check for
> top_cpuset is only safe because top_cpuset_sd_nocpu() added by patch 1
> now covers it. Without patch 1, patch 2 alone would let a partition
> take away all the CPUs of top_cpuset. So patch 2 can't be backported
> on its own, and the two have different Fixes tags. I think it would be
> better to fold patch 2 into patch 1.
> 

Thanks for the review. That makes sense. I hadn't considered the backport implications.

> Also, the fix in patch 2 seems incomplete. The same false PERR_NOCPUS
> happens when the parent is a local partition, as cpumask_subset()
> there doesn't take the CPUs given back by the update into account
> either. On an 8-CPU system:
> 
>     cd /sys/fs/cgroup
>     echo +cpuset > cgroup.subtree_control
>     mkdir A1
>     echo 1-7 > A1/cpuset.cpus
>     echo root > A1/cpuset.cpus.partition
>     echo +cpuset > A1/cgroup.subtree_control
>     mkdir A1/A2
>     echo 3-7 > A1/A2/cpuset.cpus
>     echo isolated > A1/A2/cpuset.cpus.partition
>     sleep 1000 &
>     echo $! > A1/cgroup.procs
>     echo 1-4 > A1/A2/cpuset.cpus
>     cat A1/A2/cpuset.cpus.partition
>     isolated invalid (Parent unable to distribute cpu downstream)
> 

You're right. I'll fix this in the next version.

Thanks，
Guopeng
> A2 takes CPUs 1-2 from A1 and gives CPUs 5-7 back, so A1 still has
> CPUs for its task. Instead of skipping the check for top_cpuset only,
> how about making the check in tasks_nocpu_error() account for the CPUs
> given back by cs for all parents? The hotplug caller would then need
> to not count cs->effective_xcpus as given back, similar to what patch 3
> does for top-level partitions.
>