Re: [PATCH 1/2] cgroup/cpuset: Make cpuset hotplug processing synchronous

[Waiman Long]

On 4/3/24 08:02, Michal Koutný wrote:
> On Tue, Apr 02, 2024 at 11:30:11AM -0400, Waiman Long <longman@xxxxxxxxxx> wrote:
> > Yes, there is a potential that a cpus_read_lock() may be called leading to
> > deadlock. So unless we reverse the current cgroup_mutex --> cpu_hotplug_lock
> > ordering, it is not safe to call cgroup_transfer_tasks() directly.
> I see that cgroup_transfer_tasks() has the only user -- cpuset. What
> about bending it for the specific use like:
>
> diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
> index 34aaf0e87def..64deb7212c5c 100644
> --- a/include/linux/cgroup.h
> +++ b/include/linux/cgroup.h
> @@ -109,7 +109,7 @@ struct cgroup *cgroup_get_from_fd(int fd);
>   struct cgroup *cgroup_v1v2_get_from_fd(int fd);
>   
>   int cgroup_attach_task_all(struct task_struct *from, struct task_struct *);
> -int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from);
> +int cgroup_transfer_tasks_locked(struct cgroup *to, struct cgroup *from);
>   
>   int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts);
>   int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts);
> diff --git a/kernel/cgroup/cgroup-v1.c b/kernel/cgroup/cgroup-v1.c
> index 520a11cb12f4..f97025858c7a 100644
> --- a/kernel/cgroup/cgroup-v1.c
> +++ b/kernel/cgroup/cgroup-v1.c
> @@ -91,7 +91,8 @@ EXPORT_SYMBOL_GPL(cgroup_attach_task_all);
>    *
>    * Return: %0 on success or a negative errno code on failure
>    */
> -int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
> +int cgroup_transfer_tasks_locked(struct cgroup *to, struct cgroup *from)
>   {
>   	DEFINE_CGROUP_MGCTX(mgctx);
>   	struct cgrp_cset_link *link;
> @@ -106,9 +106,11 @@ int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
>   	if (ret)
>   		return ret;
>   
> -	cgroup_lock();
> -
> -	cgroup_attach_lock(true);
> +	/* The locking rules serve specific purpose of v1 cpuset hotplug
> +	 * migration, see hotplug_update_tasks_legacy() and
> +	 * cgroup_attach_lock() */
> +	lockdep_assert_held(&cgroup_mutex);
> +	lockdep_assert_cpus_held();
> +	percpu_down_write(&cgroup_threadgroup_rwsem);
>   
>   	/* all tasks in @from are being moved, all csets are source */
>   	spin_lock_irq(&css_set_lock);
> @@ -144,8 +146,7 @@ int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
>   	} while (task && !ret);
>   out_err:
>   	cgroup_migrate_finish(&mgctx);
> -	cgroup_attach_unlock(true);
> -	cgroup_unlock();
> +	percpu_up_write(&cgroup_threadgroup_rwsem);
>   	return ret;
>   }
>   
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index 13d27b17c889..94fb8b26f038 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -4331,7 +4331,7 @@ static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
>   			nodes_empty(parent->mems_allowed))
>   		parent = parent_cs(parent);
>   
> -	if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
> +	if (cgroup_transfer_tasks_locked(parent->css.cgroup, cs->css.cgroup)) {
>   		pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
>   		pr_cont_cgroup_name(cs->css.cgroup);
>   		pr_cont("\n");
> @@ -4376,21 +4376,9 @@ hotplug_update_tasks_legacy(struct cpuset *cs,
>   
>   	/*
>   	 * Move tasks to the nearest ancestor with execution resources,
> -	 * This is full cgroup operation which will also call back into
> -	 * cpuset. Execute it asynchronously using workqueue.
>   	 */
> -	if (is_empty && css_tryget_online(&cs->css)) {
> -		struct cpuset_remove_tasks_struct *s;
> -
> -		s = kzalloc(sizeof(*s), GFP_KERNEL);
> -		if (WARN_ON_ONCE(!s)) {
> -			css_put(&cs->css);
> -			return;
> -		}
> -
> -		s->cs = cs;
> -		INIT_WORK(&s->work, cpuset_migrate_tasks_workfn);
> -		schedule_work(&s->work);
> +	if (is_empty)
> +		remove_tasks_in_empty_cpuset(cs);
>   	}
>   }
>   

It still won't work because of the possibility of mutiple tasks 
involving in a circular locking dependency. The hotplug thread always 
acquire the cpu_hotplug_lock first before acquiring cpuset_mutex or 
cgroup_mtuex in this case (cpu_hotplug_lock --> cgroup_mutex). Other 
tasks calling into cgroup code will acquire the pair in the order 
cgroup_mutex --> cpu_hotplug_lock. This may lead to a deadlock if these 
2 locking sequences happen at the same time. Lockdep will certainly 
spill out a splat because of this. So unless we change all the relevant 
cgroup code to the new cpu_hotplug_lock --> cgroup_mutex locking order, 
the hotplug code can't call cgroup_transfer_tasks() directly.

Cheers,
Longman