From 8b1f3c54f930c3aeda0b5bad97bc317fc80267fd Mon Sep 17 00:00:00 2001 From: Chen Ridong Date: Thu, 29 Jan 2026 06:45:16 +0000 Subject: [PATCH] cpuset: fix overlap of partition effective CPUs 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: 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 Reviewed-by: Waiman Long Signed-off-by: Tejun Heo --- 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 003232dc9d2e..92a51316225d 100644 --- a/kernel/cgroup/cpuset.c +++ b/kernel/cgroup/cpuset.c @@ -2222,27 +2222,20 @@ 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; -- 2.47.3