Re: [PATCH v2] sched/fair: Make SCHED_IDLE entity be preempted in strict hierarchy

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Jul 08, 2024 at 02:47:34PM +0200, Vincent Guittot wrote:
> On Mon, 8 Jul 2024 at 14:02, Peter Zijlstra <peterz@infradead.org> wrote:

> > > @@ -8409,6 +8400,15 @@ static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int
> > >       if (cse_is_idle != pse_is_idle)
> > >               return;
> > >
> > > +     /*
> > > +      * Batch tasks do not preempt non-idle tasks (their preemption
> > > +      * is driven by the tick).
> > > +      * We've done the check about "only one of the entities is idle",
> > > +      * so cse must be non-idle if p is a batch task.
> > > +      */
> > > +     if (unlikely(entity_is_task(pse) && p->policy == SCHED_BATCH))
> > > +             return;
> >
> > I'm not convinced this condition is right. The current behaviour of
> > SCHED_BATCH doesn't care about pse, only p.
> >
> > That is, if p is SCHED_BATCH it will not preempt -- except an
> > SCHED_IDLE.
> >
> > So I'm tempted to delete this first part of your condition and have it
> > be:
> >
> >         if (p->policy == SCHED_BATCH)
> >                 return;
> >
> > That is, suppose you have:
> >
> >                         root
> >                          |
> >               ------------------------
> >               |                      |
> >         normal_cgroup          normal_cgroup
> >               |                      |
> >         SCHED_BATCH task_A     SCHED_BATCH task_B
> >
> > Then the preemption crud will end up comparing the groups to one another
> > and still allow A to preempt B -- except we explicitly do not want this.
> >
> > The 'problem' is that the whole BATCH thing isn't cgroup aware ofcourse,
> > but I'm not sure we want to go fix that -- esp. not in this patch.
> >
> > Hmm?
> 
> Good question, but do we want to make SCHED_BATCH tasks behave
> differently than SCHED_IDLE tasks in a group in this case ?

I suspect we'll have to. People added the idle-cgroup thing, but never
did the same for batch. With the result that they're now fundamentally
different.

> With this patch, we don't want task_B to preempt task_A for the case
> but task_A will preempt task_B whereas task A is SCHED_IDLE
> 
>                          root
>                           |
>                ------------------------
>                |                      |
>          normal_cgroup          idle_cgroup
>                |                      |
>          SCHED_IDLE task_A     SCHED_NORMAL task_B
> 
> As we only look at the common level of hierarchy between the tasks,
> the below will make A to preempt B whereas both are SCHED_IDLE
> 
>                          root
>                           |
>                ------------------------
>                |                      |
>          normal_cgroup          normal_cgroup
>                |                      |
>          SCHED_IDLE task_A     SCHED_IDLE task_B

So we can make the last test be:

	if (unlikely(p->policy != SCHED_NORMAL))
		return;

Much like the original condition removed here:

-       if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
+       if (!sched_feat(WAKEUP_PREEMPTION))

Except now after all that cgroup nonsense. Then the OP case will preempt
because normal_cgroup vs idle_cgroup, my BATCH example will not preempt,
because BATCH != NORMAL, your IDLE example will not preempt either,
because IDLE != NORMAL.