Re: [PATCH 1/2] cgroup/cpuset: Fix cpuset.cpus update in a remote partition

[Guopeng Zhang]

在 2026/10/22 11:23, Waiman Long 写道:
> On 10/10/26 3:59 AM, Guopeng Zhang wrote:
>> From: Guopeng Zhang <zhangguopeng@xxxxxxxxxx>
>>
>> When the CPU list of a remote partition is reduced, a removed CPU may
>> remain unavailable to the top cpuset.
>>
>> This can be reproduced on a cgroup v2 system with CPUs 0-7 online:
>>
>>      cd /sys/fs/cgroup
>>      echo +cpuset > cgroup.subtree_control
>>      mkdir A
>>      echo 1-7 > A/cpuset.cpus
>>      echo 1-7 > A/cpuset.cpus.exclusive
>>      echo +cpuset > A/cgroup.subtree_control
>>      mkdir A/B
>>      echo 1-4 > A/B/cpuset.cpus
>>      echo root > A/B/cpuset.cpus.partition
>>      echo 1-3 > A/B/cpuset.cpus
>>      cat cpuset.cpus.effective
>>
>> The last command should show 0,4-7, but CPU 4 is missing.
>>
>> update_cpumask() calls partition_cpus_change() while cs->cpus_allowed
>> still contains the old mask. remote_cpus_update() first releases CPU 4,
>> but the hierarchy update revisits B and compute_excpus() still sees the
>> old mask, so CPU 4 is removed from the top cpuset again.
>>
>> Move the cpus_allowed update before partition_cpus_change() so that
>> the hierarchy update sees the new mask when it revisits the cpuset.
>>
>> Fixes: f62a5d39368e ("cgroup/cpuset: Remove remote_partition_check() & make update_cpumasks_hier() handle remote partition")
>> Signed-off-by: Guopeng Zhang <zhangguopeng@xxxxxxxxxx>
>> ---
>>   kernel/cgroup/cpuset.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>> index 0ebba2d646c3..de7553459adc 100644
>> --- a/kernel/cgroup/cpuset.c
>> +++ b/kernel/cgroup/cpuset.c
>> @@ -2478,10 +2478,13 @@ static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
>>        */
>>       force = !cpumask_equal(cs->effective_xcpus, trialcs->effective_xcpus);
>>   +    spin_lock_irq(&callback_lock);
>> +    cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
>> +    spin_unlock_irq(&callback_lock);
>> +
>>       partition_cpus_change(cs, trialcs, &tmp);
>>         spin_lock_irq(&callback_lock);
>> -    cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
>>       cpumask_copy(cs->effective_xcpus, trialcs->effective_xcpus);
>>       if ((old_prs > 0) && !is_partition_valid(cs))
>>           reset_partition_data(cs);
> 
> Sorry for the late reply.
> 
> Thanks for reporting this problem. I suspect similar problem may exist with change to cpuset.cpus.exclusive as well. The root cause of this problem is the recursive call to remote_cpus_update() with the second call missing information leading to the revert of the first call. I have posted another patch to address this recursive call problem. Please take a look to see if you agree with that or not.
> 
Thanks for the explanation and for looking further into the root cause.

I'll take a look at your patch and run some local tests once I've taken care of some urgent work.

Thanks,
Guopeng