Re: [PATCH sched_ext/for-6.12] sched_ext: Handle cases where pick_task_scx() is called without preceding balance_scx()

[Peter Zijlstra]

On Thu, Sep 05, 2024 at 03:17:13PM -1000, Tejun Heo wrote:
> On Thu, Sep 05, 2024 at 06:41:42AM -1000, Tejun Heo wrote:
> > > @@ -12716,6 +12716,12 @@ static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf)
> > >  	if (this_rq->cfs.h_nr_running && !pulled_task)
> > >  		pulled_task = 1;
> > >  
> > > +	/*
> > > +	 * We pulled a task, but it got stolen before we re-acquired rq->lock.
> > > +	 */
> > > +	if (!this_rq->cfs.h_nr_running && pulled_task)
> > > +		pulled_task = 0;
> > > +
> > 
> > Lemme test that.
> 
> Did a bit of testing and it seems like it's mostly coming from delayed
> dequeue handling. pick_next_entity() does this:
> 
> 	struct sched_entity *se = pick_eevdf(cfs_rq);
> 	if (se->sched_delayed) {
> 		dequeue_entities(rq, se, DEQUEUE_SLEEP | DEQUEUE_DELAYED);
> 		SCHED_WARN_ON(se->sched_delayed);
> 		SCHED_WARN_ON(se->on_rq);
> 		return NULL;
> 	}
> 
> rq->cfs.nr_running includes the number of delay dequeued tasks which aren't
> really runnable, so it seems like balance_fair() saying yes and
> pick_next_entity() then hitting a delayed task.

Duh, yes.

> Maybe the solution is
> tracking the number of delayed ones and subtracting that from nr_running?

That came up yesterday for something else as well. Let me see if I can
make that happen.


Anyway, I suppose you keep your patch for now until I've managed to sort
this out.