Re: [PATCH v2 sched_ext/for-7.1-fixes] sched_ext: Fix deadlock between scx_root_disable() and concurrent forks

[Andrea Righi <arighi@nvidia.com>]

Hi Tejun,

On Sun, May 17, 2026 at 07:43:16AM -1000, Tejun Heo wrote:
> scx_root_disable() enters SCX_DISABLING before it grabs scx_enable_mutex to
> clear __scx_switched_all and scx_switching_all. task_should_scx() short-circuits on DISABLING,
> so forks in that window land on fair while next_active_class() still skips
> fair - the new tasks stall.
> 
> This can deadlock the disable path itself: scx_alloc_and_add_sched() runs
> under scx_enable_mutex and creates a helper kthread; if that new kthread is
> one of the stalled fair tasks, the mutex holder waits forever and
> scx_root_disable() can never make progress. Only sub-sched support exposes
> this, since sub-sched enables are the only path where
> scx_alloc_and_add_sched() can race the root's disable.
> 
> Move the DISABLING check after @scx_switching_all. @scx_switching_all
> serves as a proxy for __scx_switched_all, so while it's set, forks keep
> going to scx. Once cleared, DISABLING applies normally.
> 
> v2: Reword in-source comment and description. (Andrea)
> 
> Fixes: 337ec00b1d9c ("sched_ext: Implement cgroup sub-sched enabling and disabling")
> Signed-off-by: Tejun Heo <tj@kernel.org>
> Reviewed-by: Andrea Righi <arighi@nvidia.com>
> ---
>  kernel/sched/ext.c |   22 +++++++++++++++++++++-
>  1 file changed, 21 insertions(+), 1 deletion(-)
> 
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -5092,10 +5092,30 @@ static const struct kset_uevent_ops scx_
>   */
>  bool task_should_scx(int policy)
>  {
> -	if (!scx_enabled() || unlikely(scx_enable_state() == SCX_DISABLING))
> +	/* if disabled, nothing should be on it */
> +	if (!scx_enabled())
>  		return false;
> +
> +	/* scx is taking over all SCHED_OTHER and SCHED_EXT tasks */
>  	if (READ_ONCE(scx_switching_all))
>  		return true;
> +
> +	/*
> +	 * scx is tearing down - keep new SCHED_EXT tasks out.
> +	 *
> +	 * Must come after scx_switching_all test, which serves as a proxy
> +	 * for __scx_switched_all. While __scx_switched_all is set, we must
> +	 * return true via the branch above: a fork routed to fair would
> +	 * stall because next_active_class() skips fair.
> +	 *
> +	 * This can develop into a deadlock - scx holds scx_enable_mutex across
> +	 * kthread_create() in scx_alloc_and_add_sched(); if the new kthread is
> +	 * the stalled task, the disable path can never grab the mutex to clear
> +	 * scx_switching_all.
> +	 */

Yeah, this is much better than my comment (that was quite confusing).

To make sure I understand: what fixes the deadlock is checking scx_switching_all
before DISABLING in task_should_scx(), because in this way the sched_ext_helper
kthread goes to scx (not fair), runs, the enable path completes, releases the
mutex and the disable path moves forward.

When I wrote my comment I was looking at the ordering of [__]scx_switched_all in
scx_root_disable():

  static_branch_disable(&__scx_switched_all);
  WRITE_ONCE(scx_switching_all, false);

And I was wondering, if we invert those we'd have a similar issue: a small
window where __scx_switched_all == ON and scx_switching_all == false. But the
current order is already the safe one, so no change needed.

Thanks,
-Andrea

> +	if (unlikely(scx_enable_state() == SCX_DISABLING))
> +		return false;
> +
>  	return policy == SCHED_EXT;
>  }
>