Re: [PATCH v8 1/6] sched: Annotate rq->rd with __rcu and update lockless readers

[Aaron Tomlin]

On Thu, Aug 27, 2026 at 03:40:09PM -0400, Aaron Tomlin wrote:
> @@ -3355,10 +3363,7 @@ static void set_cpus_allowed_dl(struct task_struct *p,
>  bool dl_task_needs_bw_move(struct task_struct *p,
>  			   const struct cpumask *new_mask)
>  {
> -	if (!dl_task(p))
> -		return false;
> -
> -	return !cpumask_intersects(task_rq(p)->rd->span, new_mask);
> +	return !cpumask_intersects(rcu_dereference_root_domain(task_rq(p)->rd)->span, new_mask);
>  }

Hi Peter, Vincent,

Please ignore; I will address this in the next iteration.

Sashiko [1] correctly reported that dl_task_needs_bw_move(), is invoked
while holding cpuset_mutex in process context, with neither
sched_domains_mutex held nor an active RCU read-side critical section.

I subsequently confirmed this using virtme-ng:

    [  170.335047] WARNING: suspicious RCU usage
    [  170.335052] 7.2.0-rc7-virtme-00041-g6509e33120f4 #23 Not tainted
    [  170.335054] -----------------------------
    [  170.335056] kernel/sched/deadline.c:3366 suspicious rcu_dereference_check() usage!

[1]: https://sashiko.dev/#/patchset/20260827194014.977758-1-atomlin%40atomlin.com

Additionally, direct dereferences of the __rcu pointer still require
suitable conversion.

Kind regards,
-- 
Aaron Tomlin