Re: [PATCH -next] cpuset: fix overlap of partition effective CPUs

[Waiman Long <llong@redhat.com>]

On 1/29/26 1:45 AM, Chen Ridong wrote:
> From: Chen Ridong <chenridong@huawei.com>
>
> A warning was detect:
>
>   WARNING: kernel/cgroup/cpuset.c:825 at rebuild_sched_domains_locked
>   Modules linked in:
>   CPU: 12 UID: 0 PID: 681 Comm: rmdir  6.19.0-rc6-next-20260121+
>   RIP: 0010:rebuild_sched_domains_locked+0x309/0x4b0
>   RSP: 0018:ffffc900019bbd28 EFLAGS: 00000202
>   RAX: ffff888104413508 RBX: 0000000000000008 RCX: ffff888104413510
>   RDX: ffff888109b5f400 RSI: 000000000000ffcf RDI: 0000000000000001
>   RBP: 0000000000000002 R08: ffff888104413508 R09: 0000000000000002
>   R10: ffff888104413508 R11: 0000000000000001 R12: ffff888104413500
>   R13: 0000000000000002 R14: ffffc900019bbd78 R15: 0000000000000000
>   FS:  00007fe274b8d740(0000) GS:ffff8881b6b3c000(0000) knlGS:
>   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>   CR2: 00007fe274c98b50 CR3: 00000001047a9000 CR4: 00000000000006f0
>   Call Trace:
>    <TASK>
>    update_prstate+0x1c7/0x580
>    cpuset_css_killed+0x2f/0x50
>    kill_css+0x32/0x180
>    cgroup_destroy_locked+0xa7/0x200
>    cgroup_rmdir+0x28/0x100
>    kernfs_iop_rmdir+0x4c/0x80
>    vfs_rmdir+0x12c/0x280
>    filename_rmdir+0x19e/0x200
>    __x64_sys_rmdir+0x23/0x40
>    do_syscall_64+0x6b/0x390
>
> It can be reproduced by steps:
>
>    # cd /sys/fs/cgroup/
>    # mkdir A1
>    # mkdir B1
>    # mkdir C1
>    # echo 1-3 > A1/cpuset.cpus
>    # echo root > A1/cpuset.cpus.partition
>    # echo 3-5 > B1/cpuset.cpus
>    # echo root > B1/cpuset.cpus.partition
>    # echo 6 > C1/cpuset.cpus
>    # echo root > C1/cpuset.cpus.partition
>    # rmdir A1/
>    # rmdir C1/
>
> Both A1 and B1 were initially configured with CPU 3, which was exclusively
> assigned to A1's partition. When A1 was removed, CPU 3 was returned to the
> root pool. However, B1 incorrectly regained access to CPU 3 when
> update_cpumasks_hier was triggered during C1's removal, which also updated
> sibling configurations.
>
> The update_sibling_cpumasks function was called to synchronize siblings'
> effective CPUs due to changes in their parent's effective CPUs. However,
> parent effective CPU changes should not affect partition-effective CPUs.
>
> To fix this issue, update_cpumasks_hier should only be invoked when the
> sibling is not a valid partition in the update_sibling_cpumasks.
>
> Fixes: 2a3602030d80 ("cgroup/cpuset: Don't invalidate sibling partitions on cpuset.cpus conflict")
> Signed-off-by: Chen Ridong <chenridong@huawei.com>
> ---
>   kernel/cgroup/cpuset.c | 19 ++++++-------------
>   1 file changed, 6 insertions(+), 13 deletions(-)
>
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index cf67d3524c75..31ba74044155 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -2227,27 +2227,20 @@ static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
>   	 * It is possible a change in parent's effective_cpus
>   	 * due to a change in a child partition's effective_xcpus will impact
>   	 * its siblings even if they do not inherit parent's effective_cpus
> -	 * directly.
> +	 * directly. It should not impact valid partition.
>   	 *
>   	 * The update_cpumasks_hier() function may sleep. So we have to
>   	 * release the RCU read lock before calling it.
>   	 */
>   	rcu_read_lock();
>   	cpuset_for_each_child(sibling, pos_css, parent) {
> -		if (sibling == cs)
> +		if (sibling == cs || is_partition_valid(sibling))
>   			continue;
> -		if (!is_partition_valid(sibling)) {
> -			compute_effective_cpumask(tmp->new_cpus, sibling,
> -						  parent);
> -			if (cpumask_equal(tmp->new_cpus, sibling->effective_cpus))
> -				continue;
> -		} else if (is_remote_partition(sibling)) {
> -			/*
> -			 * Change in a sibling cpuset won't affect a remote
> -			 * partition root.
> -			 */
> +
> +		compute_effective_cpumask(tmp->new_cpus, sibling,
> +					  parent);
> +		if (cpumask_equal(tmp->new_cpus, sibling->effective_cpus))
>   			continue;
> -		}
>   
>   		if (!css_tryget_online(&sibling->css))
>   			continue;

Thanks for fixing this.

Reviewed-by: Waiman Long <longman@redhat.com>