Re: [PATCH/for-next 1/2] cgroup/cpuset: Defer housekeeping_update() call from CPU hotplug to task_work

[Waiman Long <llong@redhat.com>]

On 1/29/26 2:15 AM, Chen Ridong wrote:
>
> On 2026/1/29 12:03, Chen Ridong wrote:
>>
>> On 2026/1/28 12:42, Waiman Long wrote:
>>> The update_isolation_cpumasks() function can be called either directly
>>> from regular cpuset control file write with cpuset_full_lock() called
>>> or via the CPU hotplug path with cpus_write_lock and cpuset_mutex held.
>>>
>>> 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 cause deadlock. So we
>>> have to defer any call to housekeeping_update() after the CPU hotplug
>>> operation has finished. This can be done via the task_work_add(...,
>>> TWA_RESUME) API where the actual housekeeping_update() call, if needed,
>>> will happen right before existing back to userspace.
>>>
>>> Since the HK_TYPE_DOMAIN housekeeping cpumask should now track the
>>> changes in "cpuset.cpus.isolated", add a check in test_cpuset_prs.sh to
>>> confirm that the CPU hotplug deferral, if needed, is working as expected.
>>>
>>> Signed-off-by: Waiman Long <longman@redhat.com>
>>> ---
>>>   kernel/cgroup/cpuset.c                        | 49 ++++++++++++++++++-
>>>   .../selftests/cgroup/test_cpuset_prs.sh       |  9 ++++
>>>   2 files changed, 56 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>>> index 7b7d12ab1006..98c7cb732206 100644
>>> --- a/kernel/cgroup/cpuset.c
>>> +++ b/kernel/cgroup/cpuset.c
>>> @@ -84,6 +84,10 @@ static cpumask_var_t	isolated_cpus;
>>>    */
>>>   static bool isolated_cpus_updating;
>>>   
>>> +/* Both cpuset_mutex and cpus_read_locked acquired */
>>> +static bool cpuset_full_locked;
>>> +static bool isolation_task_work_queued;
>>> +
>>>   /*
>>>    * A flag to force sched domain rebuild at the end of an operation.
>>>    * It can be set in
>>> @@ -285,10 +289,12 @@ void cpuset_full_lock(void)
>>>   {
>>>   	cpus_read_lock();
>>>   	mutex_lock(&cpuset_mutex);
>>> +	cpuset_full_locked = true;
>>>   }
>>>   
>>>   void cpuset_full_unlock(void)
>>>   {
>>> +	cpuset_full_locked = false;
>>>   	mutex_unlock(&cpuset_mutex);
>>>   	cpus_read_unlock();
>>>   }
>>> @@ -1285,25 +1291,64 @@ static bool prstate_housekeeping_conflict(int prstate, struct cpumask *new_cpus)
>>>   	return false;
>>>   }
>>>   
>>> +static void __update_isolation_cpumasks(bool twork);
>>> +static void isolation_task_work_fn(struct callback_head *cb)
>>> +{
>>> +	cpuset_full_lock();
>>> +	__update_isolation_cpumasks(true);
>>> +	cpuset_full_lock();
>>> +}
>>> +
>>>   /*
>>> - * update_isolation_cpumasks - Update external isolation related CPU masks
>>> + * __update_isolation_cpumasks - Update external isolation related CPU masks
>>> + * @twork - set if call from isolation_task_work_fn()
>>>    *
>>>    * The following external CPU masks will be updated if necessary:
>>>    * - workqueue unbound cpumask
>>>    */
>>> -static void update_isolation_cpumasks(void)
>>> +static void __update_isolation_cpumasks(bool twork)
>>>   {
>>>   	int ret;
>>>   
>>> +	if (twork)
>>> +		isolation_task_work_queued = false;
>>> +
>>>   	if (!isolated_cpus_updating)
>>>   		return;
>>>   
>>> +	/*
>>> +	 * This function can be reached either directly from regular cpuset
>>> +	 * control file write (cpuset_full_locked) or via hotplug
>>> +	 * (cpus_write_lock && cpuset_mutex held). In the later case, we
>>> +	 * defer the housekeeping_update() call to a task_work to avoid
>>> +	 * the possibility of deadlock. The task_work will be run right
>>> +	 * before exiting back to userspace.
>>> +	 */
>>> +	if (!cpuset_full_locked) {
>>> +		static struct callback_head twork_cb;
>>> +
>>> +		if (!isolation_task_work_queued) {
>>> +			init_task_work(&twork_cb, isolation_task_work_fn);
>>> +			if (!task_work_add(current, &twork_cb, TWA_RESUME))
>>> +				isolation_task_work_queued = true;
>>> +			else
>>> +				/* Current task shouldn't be exiting */
>>> +				WARN_ON_ONCE(1);
>>> +		}
>>> +		return;
>>> +	}
>>> +
>>>   	ret = housekeeping_update(isolated_cpus);
>>>   	WARN_ON_ONCE(ret < 0);
>>>   
>>>   	isolated_cpus_updating = false;
>>>   }
>>>   
>> The logic is not straightforward; perhaps we can simplify it as follows,
>> maybe I missed something, just correct me.
>>
>> static void isolation_task_work_fn(struct callback_head *cb)
>> {
>> 	guard(mutex)(&isolcpus_update_mutex);
>> 	WARN_ON_ONCE(housekeeping_update(isolated_cpus) < 0);
>> }
>>
>> /*
>>   * __update_isolation_cpumasks - Update external isolation related CPU masks
>>   * @twork - set if call from isolation_task_work_fn()
>>   *
>>   * The following external CPU masks will be updated if necessary:
>>   * - workqueue unbound cpumask
>>   */
>> static void __update_isolation_cpumasks(bool twork)
>> {
>> 	if (!isolated_cpus_updating)
>> 		return;
>>
>> 	/*
>> 	 * This function can be reached either directly from regular cpuset
>> 	 * control file write (cpuset_full_locked) or via hotplug
>> 	 * (cpus_write_lock && cpuset_mutex held). In the later case, we
>> 	 * defer the housekeeping_update() call to a task_work to avoid
>> 	 * the possibility of deadlock. The task_work will be run right
>> 	 * before exiting back to userspace.
>> 	 */
>> 	if (twork) {
>> 		static struct callback_head twork_cb;
>>
>> 		init_task_work(&twork_cb, isolation_task_work_fn);
>> 		if (task_work_add(current, &twork_cb, TWA_RESUME))
>> 			/* Current task shouldn't be exiting */
>> 			WARN_ON_ONCE(1);
>>
>> 		return;
>> 	}
>>
>> 	lockdep_assert_held(&isolcpus_update_mutex);
>> 	/*
>> 	 * Release cpus_read_lock & cpuset_mutex before calling
>> 	 * housekeeping_update() and re-acquiring them afterward if not
>> 	 * calling from task_work.
>> 	 */
>>
>> 	cpuset_full_unlock();
>> 	WARN_ON_ONCE(housekeeping_update(isolated_cpus) < 0);
>> 	cpuset_full_lock();
>>
>> 	isolated_cpus_updating = false;
>> }
>>
>> static inline void update_isolation_cpumasks(void)
>> {
>> 	__update_isolation_cpumasks(false);
>> }
>>
> It can be much clearer:
>
> static void isolation_task_work_fn(struct callback_head *cb)
> {
> 	guard(mutex)(&isolcpus_update_mutex);
> 	WARN_ON_ONCE(housekeeping_update(isolated_cpus) < 0);
> }
>
> /*
>   * __update_isolation_cpumasks - Update external isolation related CPU masks
>   * @defer
>   *
>   * The following external CPU masks will be updated if necessary:
>   * - workqueue unbound cpumask
>   */
> static void __update_isolation_cpumasks(bool defer)
> {
> 	if (!isolated_cpus_updating)
> 		return;
>
> 	/*
> 	 * This function can be reached either directly from regular cpuset
> 	 * control file write (cpuset_full_locked) or via hotplug
> 	 * (cpus_write_lock && cpuset_mutex held). In the later case, we
> 	 * defer the housekeeping_update() call to a task_work to avoid
> 	 * the possibility of deadlock. The task_work will be run right
> 	 * before exiting back to userspace.
> 	 */
> 	if (defer) {
> 		static struct callback_head twork_cb;
>
> 		init_task_work(&twork_cb, isolation_task_work_fn);
> 		if (task_work_add(current, &twork_cb, TWA_RESUME))
> 			/* Current task shouldn't be exiting */
> 			WARN_ON_ONCE(1);
>
> 		return;
> 	}
>
> 	lockdep_assert_held(&isolcpus_update_mutex);
> 	lockdep_assert_cpus_held();
> 	lockdep_assert_cpuset_lock_held();
>
> 	/*
> 	 * Release cpus_read_lock & cpuset_mutex before calling
> 	 * housekeeping_update() and re-acquiring them afterward if not
> 	 * calling from task_work.
> 	 */
>
> 	cpuset_full_unlock();
> 	WARN_ON_ONCE(housekeeping_update(isolated_cpus) < 0);
> 	cpuset_full_lock();
>
> 	isolated_cpus_updating = false;
> }
>
> static inline void update_isolation_cpumasks(void)
> {
> 	__update_isolation_cpumasks(false);
> }
>
> static inline void asyn_update_isolation_cpumasks(void)
> {
> 	__update_isolation_cpumasks(true);
> }
>
> The hotplug path just calls asyn_update_isolation_cpumasks(), cpuset_full_locked
> and isolation_task_work_queued can be removed.
>
Thanks for the suggestions. I will adopt some of them. BTW, I am 
switching to use workqueue instead of the task_work as the latter is 
suitable in this use case.

Cheers,
Longman