Re: [PATCH v2 03/11] sched_ext: Add an event, SELECT_CPU_FALLBACK

[Changwoo Min]

Hello,

On 25. 1. 28. 05:01, Tejun Heo wrote:
> Hello,
>
> On Sun, Jan 26, 2025 at 07:16:06PM +0900, Changwoo Min wrote:
> ...
> >   struct scx_event_stats {
> > +	/*
> > +	 * If ops.select_cpu() returns a CPU which can't be used by the task,
> > +	 * the core scheduler code silently picks a fallback CPU.
> > +	 */
> > +	u64		SELECT_CPU_FALLBACK;
>
> As C has one global namespace, we're gonna have to prefix these. Hopefully,
> something not too long. SCX_EV_?

Sure, I will add SCX_EV_ as a prefix for all events.

> > @@ -3663,6 +3668,10 @@ static int select_task_rq_scx(struct task_struct *p, int prev_cpu, int wake_flag
> >   		cpu = SCX_CALL_OP_TASK_RET(SCX_KF_ENQUEUE | SCX_KF_SELECT_CPU,
> >   					   select_cpu, p, prev_cpu, wake_flags);
> >   		*ddsp_taskp = NULL;
> > +
> > +		if (unlikely(!is_cpu_allowed(p, cpu)))
> > +			__scx_add_event(SELECT_CPU_FALLBACK, 1);
>
> This is trying to guess what select_task_rq() is going to do and then count
> that as an event, which doesn't seem great. Can't we just record the picked
> CPU here into a field in p->scx and then compare that against cpu_of(rq) in
> enqueue_task_scx()? That way, we can reliably count events where CPU
> selection was modified by the fallback logic regardless of why or how that
> happens.

That sounds good. I will add a field (say, selected_cpu?) to struct 
sched_ext_entity.

Regards,
Changwoo Min