Re: [PATCH v2] cgroup/cpuset: Invalidate remote partition on housekeeping conflict

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

在 2026/9/29 02:44, Tejun Heo 写道:
Hi，Tejun

> Hello, Guopeng.
> 
> The following is a Claude-generated review.
> 
> On Sun, Sep 27, 2026 at 05:57:21PM +0800, Guopeng Zhang wrote:
>> +	bool updating = is_remote_partition(cs);
>> +
>> +	if (!capable(CAP_SYS_ADMIN))
>> +		return PERR_ACCESS;
>> +
>> +	if (!updating &&
>> +	    (!cpumask_intersects(excpus, cpu_active_mask) ||
>> +	     cpumask_subset(top_cpuset.effective_cpus, addcpus)))
>> +		return PERR_INVCPUS;
> 
> cs->remote_partition can be stale when remote_partition_enable() is
> called. If a root <-> isolated switch of a valid remote partition fails
> in update_prstate(), the partition is made invalid without
> remote_partition_disable(), so the flag stays set and its CPUs stay in
> subpartitions_cpus. A later enable then skips the PERR_INVCPUS checks.
> With isolcpus=domain,4:
> 
> 1. A is a member with cpuset.cpus 2-6 and cpuset.cpus.exclusive 2-4. B
>    has cpuset.cpus 2-4 and no cpuset.cpus.exclusive.
> 2. "isolated" to B's cpuset.cpus.partition. B becomes a valid remote
>    partition with effective_xcpus 2-4.
> 3. "root" to B's cpuset.cpus.partition fails with PERR_HKEEPING.
>    effective_xcpus is cleared but remote_partition stays set.
> 4. 5-6 to A's cpuset.cpus.exclusive. B's excpus becomes empty, which
>    matches the cleared effective_xcpus, so update_cpumasks_hier() skips
>    B.
> 5. "isolated" to B's cpuset.cpus.partition. This used to fail with
>    PERR_INVCPUS. Now B becomes a valid isolated partition with empty
>    effective_xcpus and the WARN_ON_ONCE() at the end of update_prstate()
>    triggers.
> 
> This is from reading the code, not reproduced. Maybe have the callers
> pass whether it's an enable or an update instead of deriving it from
> cs->remote_partition?

You're right. I'll fix it in v3.

Thanks,
Guopeng
> 
> The stale flag itself is a separate, pre-existing bug. Switching B back
> to member after step 3 trips the WARN_ON_ONCE(old_prs < 0) in
> partition_xcpus_del().
> 
> Thanks.
> 
> --
> tejun