Re: [PATCH v6 7/8] cgroup/cpuset: Defer housekeeping_update() calls from CPU hotplug to workqueue

[Waiman Long <longman@redhat.com>]

On 2/26/26 11:06 AM, Frederic Weisbecker wrote:
> Le Sat, Feb 21, 2026 at 01:54:17PM -0500, Waiman Long a écrit :
>> The cpuset_handle_hotplug() may need to invoke housekeeping_update(),
>> for instance, when an isolated partition is invalidated because its
>> last active CPU has been put offline.
>>
>> As we are going to enable dynamic update to the nozh_full housekeeping
>> cpumask (HK_TYPE_KERNEL_NOISE) soon with the help of CPU hotplug,
>> allowing the CPU hotplug path to call into housekeeping_update() directly
>> from update_isolation_cpumasks() will likely cause deadlock.
> I am a bit confused here. Why would CPU hotplug path need to call
> update_isolation_cpumasks() -> housekeeping_update() for
> HK_TYPE_KERNEL_NOISE?

Oh, this is not the current behavior. However, to make nohz_full fully 
dynamically changeable in the near future, we will have to do that 
eventually.

Cheers,
Longman


>> So we
>> have to defer any call to housekeeping_update() after the CPU hotplug
>> operation has finished. This is now done via the workqueue where
>> the update_hk_sched_domains() function will be invoked via the
>> hk_sd_workfn().
>>
>> An concurrent cpuset control file write may have executed the required
>> update_hk_sched_domains() function before the work function is called. So
>> the work function call may become a no-op when it is invoked.
>>
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>   kernel/cgroup/cpuset.c                        | 31 ++++++++++++++++---
>>   .../selftests/cgroup/test_cpuset_prs.sh       | 11 ++++++-
>>   2 files changed, 36 insertions(+), 6 deletions(-)
>>
>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>> index 3d0d18bf182f..2c80bfc30bbc 100644
>> --- a/kernel/cgroup/cpuset.c
>> +++ b/kernel/cgroup/cpuset.c
>> @@ -1323,6 +1323,16 @@ static void update_hk_sched_domains(void)
>>   		rebuild_sched_domains_locked();
>>   }
>>   
>> +/*
>> + * Work function to invoke update_hk_sched_domains()
>> + */
>> +static void hk_sd_workfn(struct work_struct *work)
>> +{
>> +	cpuset_full_lock();
>> +	update_hk_sched_domains();
>> +	cpuset_full_unlock();
>> +}
>> +
>>   /**
>>    * rm_siblings_excl_cpus - Remove exclusive CPUs that are used by sibling cpusets
>>    * @parent: Parent cpuset containing all siblings
>> @@ -3795,6 +3805,7 @@ static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
>>    */
>>   static void cpuset_handle_hotplug(void)
>>   {
>> +	static DECLARE_WORK(hk_sd_work, hk_sd_workfn);
>>   	static cpumask_t new_cpus;
>>   	static nodemask_t new_mems;
>>   	bool cpus_updated, mems_updated;
>> @@ -3877,11 +3888,21 @@ static void cpuset_handle_hotplug(void)
>>   	}
>>   
>>   
>> -	if (update_housekeeping || force_sd_rebuild) {
>> -		mutex_lock(&cpuset_mutex);
>> -		update_hk_sched_domains();
>> -		mutex_unlock(&cpuset_mutex);
>> -	}
>> +	/*
>> +	 * Queue a work to call housekeeping_update() & rebuild_sched_domains()
>> +	 * There will be a slight delay before the HK_TYPE_DOMAIN housekeeping
>> +	 * cpumask can correctly reflect what is in isolated_cpus.
>> +	 *
>> +	 * We rely on WORK_STRUCT_PENDING_BIT to not requeue a work item that
>> +	 * is still pending. Before the pending bit is cleared, the work data
>> +	 * is copied out and work item dequeued. So it is possible to queue
>> +	 * the work again before the hk_sd_workfn() is invoked to process the
>> +	 * previously queued work. Since hk_sd_workfn() doesn't use the work
>> +	 * item at all, this is not a problem.
>> +	 */
>> +	if (update_housekeeping || force_sd_rebuild)
>> +		queue_work(system_unbound_wq, &hk_sd_work);
> Nit about recent wq renames:
>
> s/system_unbound_wq/system_dfl_wq
Good point. Will send additional patch to do the rename.
>
> But what makes sure this work is executed by the end of the hotplug operations?
> Is there a risk for a stale hierarchy to be observed when it shouldn't? Or a
> stale housekeeping cpumask?

If you look at the work function, it will make a copy of HK_TYPE_DOMAIN 
cpumask while holding rcu_read_lock(). So the current hotplug operation 
must have finished at that point. Of course, if there is another 
hot-add/remove operation right after the rcu_read_lock is released, the 
cpumask passed down to housekeeping_update() may not be the latest one. 
In this case, another work will be scheduled to call 
housekeeping_update() with the new cpumask again.

Cheers,
Longman