Re: [PATCH] sched/isolation: Publish new HK_TYPE_DOMAIN cpumask under cpuset_mutex

From: Waiman Long

Date: Wed Nov 18 2026 - 12:16:19 EST


On 11/18/26 9:20 AM, Frederic Weisbecker wrote:
Le Tue, Nov 17, 2026 at 04:39:15PM +0800, Ridong a écrit :
From: Ridong Chen <chenridong@xxxxxxxxxx>

Commit a84097e625f2 ("cgroup/cpuset: Call housekeeping_update() without
holding cpus_read_lock") calls housekeeping_update() after dropping
cpus_read_lock and cpuset_mutex, so the HK_TYPE_DOMAIN cpumask is now
switched and freed outside of those locks. But readers holding either
lock may still access it without rcu_read_lock() as allowed by
housekeeping_dereference_check(), and synchronize_rcu() doesn't wait
for them:

CPU A CPU B
rebuild_sched_domains() cpuset_update_sd_hk_unlock()
mutex_unlock(&cpuset_mutex)
cpus_read_unlock()
mutex_lock(&cpuset_mutex)
p = housekeeping_cpumask(DOMAIN)
housekeeping_update()
rcu_assign_pointer()
synchronize_rcu()
kfree(old)
access p <- UAF

Fix it by moving the pointer switch into housekeeping_publish(), which
is called before dropping the locks. Then CPU A either finishes before
CPU B switches the pointer, or only sees the new cpumask:

CPU A CPU B
cpuset_update_sd_hk_unlock()
housekeeping_publish()
rcu_assign_pointer()
mutex_lock(&cpuset_mutex) <- wait
mutex_unlock(&cpuset_mutex)
cpus_read_unlock()
p = housekeeping_cpumask(DOMAIN) housekeeping_update()
<- new cpumask synchronize_rcu()
kfree(old)

The rest of housekeeping_update(), including the workqueue flushes,
still runs without those locks.

Fixes: a84097e625f2 ("cgroup/cpuset: Call housekeeping_update() without holding cpus_read_lock")
Signed-off-by: Ridong Chen <chenridong@xxxxxxxxxx>
Ok that should work but, IIRC this whole locking puzzle is there because
cpuset_handle_hotplug() can change isolated partitions. Why is it the case
since cpuset.cpus shouldn't care about offline CPUs?

I remember Waiman said it could be fixed in the long run. I don't remember
the details though.

cpuset_handle_hotplug() has to remove the newly offlined CPU or to add back the newly online CPUs to the effective_cpus of all the cpusets that include the affected CPUs whether it is in an isolated or root partition or not. In the worst case that the last online CPU in an isolated partition is offline, we have to disable it. That is the main reason why hotplug can affect isolated partitions. I don't remember saying it is something that can be fixed long term. Perhaps, it is not something related but not exactly this.

Cheers,
Longman