Re: [PATCH] cgroup/cpuset: Call rebuild_sched_domains() directly in hotplug

[Chen Ridong <chenridong@huaweicloud.com>]

On 2026/3/5 2:41, Waiman Long wrote:
> Besides deferring the call to housekeeping_update(), commit 6df415aa46ec
> ("cgroup/cpuset: Defer housekeeping_update() calls from CPU hotplug
> to workqueue") also defers the rebuild_sched_domains() call to
> the workqueue. So a new offline CPU may still be in a sched domain
> or new online CPU not showing up in the sched domains for a short
> transition period. That could be a problem in some corner cases and
> can be the cause of a reported test failure[1]. Fix it by calling

Miss Link [1]?

> rebuild_sched_domains_cpuslocked() directly in hotplug as before. If
> isolated partition invalidation or recreation is being done, the
> housekeeping_update() call to update the housekeeping cpumasks will
> still be deferred to a workqueue.
> 
> In commit 3bfe47967191 ("cgroup/cpuset: Move
> housekeeping_update()/rebuild_sched_domains() together"),
> housekeeping_update() is called before rebuild_sched_domains() because
> it needs to access the HK_TYPE_DOMAIN housekeeping cpumask. That is now
> changed to use the static HK_TYPE_DOMAIN_BOOT cpumask as HK_TYPE_DOMAIN
> cpumask is now changeable at run time.  As a result, we can move the
> rebuild_sched_domains() call before housekeeping_update() with
> the slight advantage that it will be done in the same cpus_read_lock
> critical section without the possibility of interference by a concurrent
> cpu hot add/remove operation.
> 

Nice.

> As it doesn't make sense to acquire cpuset_mutex/cpuset_top_mutex after
> calling housekeeping_update() and immediately release them again, move
> the cpuset_full_unlock() operation inside update_hk_sched_domains()
> and rename it to cpuset_update_sd_hk_unlock() to signify that it will
> release the full set of locks.
> 
> Fixes: 6df415aa46ec ("cgroup/cpuset: Defer housekeeping_update() calls from CPU hotplug to workqueue")
> Tested-by: Jon Hunter <jonathanh@nvidia.com>
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  kernel/cgroup/cpuset.c | 59 ++++++++++++++++++++++--------------------
>  1 file changed, 31 insertions(+), 28 deletions(-)
> 
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index 271bb99b1b9d..f7657b325490 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -881,7 +881,7 @@ static int generate_sched_domains(cpumask_var_t **domains,
>  	/*
>  	 * Cgroup v2 doesn't support domain attributes, just set all of them
>  	 * to SD_ATTR_INIT. Also non-isolating partition root CPUs are a
> -	 * subset of HK_TYPE_DOMAIN housekeeping CPUs.
> +	 * subset of HK_TYPE_DOMAIN_BOOT housekeeping CPUs.
>  	 */
>  	for (i = 0; i < ndoms; i++) {
>  		/*
> @@ -890,7 +890,7 @@ static int generate_sched_domains(cpumask_var_t **domains,
>  		 */
>  		if (!csa || csa[i] == &top_cpuset)
>  			cpumask_and(doms[i], top_cpuset.effective_cpus,
> -				    housekeeping_cpumask(HK_TYPE_DOMAIN));
> +				    housekeeping_cpumask(HK_TYPE_DOMAIN_BOOT));
>  		else
>  			cpumask_copy(doms[i], csa[i]->effective_cpus);
>  		if (dattr)
> @@ -1331,17 +1331,22 @@ static bool prstate_housekeeping_conflict(int prstate, struct cpumask *new_cpus)
>  }
>  
>  /*
> - * update_hk_sched_domains - Update HK cpumasks & rebuild sched domains
> + * cpuset_update_sd_hk_unlock - Rebuild sched domains, update HK & unlock
>   *
> - * Update housekeeping cpumasks and rebuild sched domains if necessary.
> - * This should be called at the end of cpuset or hotplug actions.
> + * Update housekeeping cpumasks and rebuild sched domains if necessary and
> + * then do a cpuset_full_unlock().
> + * This should be called at the end of cpuset operation.
>   */
> -static void update_hk_sched_domains(void)
> +static void cpuset_update_sd_hk_unlock(void)
> +	__releases(&cpuset_mutex)
> +	__releases(&cpuset_top_mutex)
>  {
> +	/* force_sd_rebuild will be cleared in rebuild_sched_domains_locked() */
> +	if (force_sd_rebuild)
> +		rebuild_sched_domains_locked();
> +
>  	if (update_housekeeping) {
> -		/* Updating HK cpumasks implies rebuild sched domains */
>  		update_housekeeping = false;
> -		force_sd_rebuild = true;
>  		cpumask_copy(isolated_hk_cpus, isolated_cpus);
>  
>  		/*
> @@ -1352,22 +1357,19 @@ static void update_hk_sched_domains(void)
>  		mutex_unlock(&cpuset_mutex);
>  		cpus_read_unlock();
>  		WARN_ON_ONCE(housekeeping_update(isolated_hk_cpus));
> -		cpus_read_lock();
> -		mutex_lock(&cpuset_mutex);
> +		mutex_unlock(&cpuset_top_mutex);
> +	} else {
> +		cpuset_full_unlock();
>  	}
> -	/* force_sd_rebuild will be cleared in rebuild_sched_domains_locked() */
> -	if (force_sd_rebuild)
> -		rebuild_sched_domains_locked();
>  }
>  
>  /*
> - * Work function to invoke update_hk_sched_domains()
> + * Work function to invoke cpuset_update_sd_hk_unlock()
>   */
>  static void hk_sd_workfn(struct work_struct *work)
>  {
>  	cpuset_full_lock();
> -	update_hk_sched_domains();
> -	cpuset_full_unlock();
> +	cpuset_update_sd_hk_unlock();
>  }
>  
>  /**
> @@ -3232,8 +3234,7 @@ ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
>  
>  	free_cpuset(trialcs);
>  out_unlock:
> -	update_hk_sched_domains();
> -	cpuset_full_unlock();
> +	cpuset_update_sd_hk_unlock();
>  	if (of_cft(of)->private == FILE_MEMLIST)
>  		schedule_flush_migrate_mm();
>  	return retval ?: nbytes;
> @@ -3340,8 +3341,7 @@ static ssize_t cpuset_partition_write(struct kernfs_open_file *of, char *buf,
>  	cpuset_full_lock();
>  	if (is_cpuset_online(cs))
>  		retval = update_prstate(cs, val);
> -	update_hk_sched_domains();
> -	cpuset_full_unlock();
> +	cpuset_update_sd_hk_unlock();
>  	return retval ?: nbytes;
>  }
>  
> @@ -3515,8 +3515,7 @@ static void cpuset_css_killed(struct cgroup_subsys_state *css)
>  	/* Reset valid partition back to member */
>  	if (is_partition_valid(cs))
>  		update_prstate(cs, PRS_MEMBER);
> -	update_hk_sched_domains();
> -	cpuset_full_unlock();
> +	cpuset_update_sd_hk_unlock();
>  }
>  
>  static void cpuset_css_free(struct cgroup_subsys_state *css)
> @@ -3925,11 +3924,13 @@ static void cpuset_handle_hotplug(void)
>  		rcu_read_unlock();
>  	}
>  
> -
>  	/*
> -	 * 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.
> +	 * rebuild_sched_domains() will always be called directly if needed
> +	 * to make sure that newly added or removed CPU will be reflected in
> +	 * the sched domains. However, if isolated partition invalidation
> +	 * or recreation is being done (update_housekeeping set), a work item
> +	 * will be queued to call housekeeping_update() to update the
> +	 * corresponding housekeeping cpumasks after some slight delay.
>  	 *
>  	 * 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
> @@ -3938,8 +3939,10 @@ static void cpuset_handle_hotplug(void)
>  	 * 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);
> +	if (force_sd_rebuild)
> +		rebuild_sched_domains_cpuslocked();
> +	if (update_housekeeping)
> +		queue_work(system_dfl_wq, &hk_sd_work);
>  
>  	free_tmpmasks(ptmp);
>  }

This means that rebuild schedule domains are decoupled from HK updates, right?

I'm wondering again whether we can do the same for changes to
cpus/partition/cpus.exclusive. If we can defer the HK update, then the
cpuset_top_mutex might no longer be necessary.

This patch looks good to me.

Reviewed-by: Chen Ridong <chenridong@huaweicloud.com>

-- 
Best regards,
Ridong