Re: [PATCH] sched/ext: Suppress warning in __this_cpu_write() by disabling preemption

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jul 16, 2025 at 04:26:23PM +0200, Andrea Righi wrote:
> On Wed, Jul 16, 2025 at 03:36:31PM +0200, Peter Zijlstra wrote:
> > On Wed, Jul 16, 2025 at 03:15:12PM +0200, Andrea Righi wrote:
> > 
> > > The idea is to track the scx callbacks that are invoked with a rq lock held
> > > and, in those cases, store the locked rq. However, some callbacks may also
> > > be invoked from an unlocked context, where no rq is locked and in this case
> > > rq should be NULL.
> > > 
> > > In the latter case, it's acceptable for preemption to remain enabled, but
> > > we still want to explicitly set locked_rq = NULL. If during the execution
> > > of the callback we jump on another CPU, it'd still be in an unlocked state,
> > > so it's locked_rq is still NULL.
> > 
> > Right; but doing superfluous NULL stores seems pointless. So better to
> > avoid the store entirely, rather than making it more expensive and no
> > less pointless, right?
> 
> Right, we can definitely avoid rewriting NULL.
> The following should do the trick.
> 
> Breno, can you give it a try?
> 
> Thanks,
> -Andrea
> 
>  kernel/sched/ext.c | 14 ++++++++++----
>  1 file changed, 10 insertions(+), 4 deletions(-)
> 
> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> index e231450768897..c76d6c9e497b4 100644
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -1290,12 +1290,15 @@ static inline void update_locked_rq(struct rq *rq)
>  	 */
>  	if (rq)
>  		lockdep_assert_rq_held(rq);
> +
> +	WARN_ON_ONCE(preemptible());
>  	__this_cpu_write(scx_locked_rq_state, rq);
>  }

__this_cpu_*() implies that check, which is how we got here ;-)