by Ingo Molnar

[permalink] [raw]

Subject: Re: [PATCH 5/5] sched/fair: Rename set_next_buddy() to set_next_pick()

* Peter Zijlstra <[email protected]> wrote:

> On Sun, Apr 07, 2024 at 10:43:19AM +0200, Ingo Molnar wrote:
> > This is a mechanism to set the next task_pick target,
> > 'buddy' is too ambiguous and refers to a historic feature we
> > don't have anymore.
> >
> > Signed-off-by: Ingo Molnar <[email protected]>
> > ---
> > kernel/sched/fair.c | 28 +++++++++++++---------------
> > 1 file changed, 13 insertions(+), 15 deletions(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 93ea653065f5..fe730f232ffd 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -3200,7 +3200,16 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
> > hrtick_update(rq);
> > }
> >
> > -static void set_next_buddy(struct sched_entity *se);
> > +static void set_next_pick(struct sched_entity *se)
> > +{
> > + for_each_sched_entity(se) {
> > + if (SCHED_WARN_ON(!se->on_rq))
> > + return;
> > + if (se_is_idle(se))
> > + return;
> > + cfs_rq_of(se)->next = se;
> > + }
> > +}
> >
> > /*
> > * The dequeue_task method is called before nr_running is
> > @@ -3240,7 +3249,7 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
> > * p is sleeping when it is within its sched_slice.
> > */
> > if (task_sleep && se && !throttled_hierarchy(cfs_rq))
> > - set_next_buddy(se);
> > + set_next_pick(se);
> > break;
> > }
> > flags |= DEQUEUE_SLEEP;
> > @@ -4631,17 +4640,6 @@ balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
> > static inline void set_task_max_allowed_capacity(struct task_struct *p) {}
> > #endif /* CONFIG_SMP */
> >
> > -static void set_next_buddy(struct sched_entity *se)
> > -{
> > - for_each_sched_entity(se) {
> > - if (SCHED_WARN_ON(!se->on_rq))
> > - return;
> > - if (se_is_idle(se))
> > - return;
> > - cfs_rq_of(se)->next = se;
> > - }
> > -}
> > -
>
> Hurmmm.. afaict the only actual user of cfs_rq->next left is task_hot(),
> no? Is that thing worth it?

Yeah, so:

1)

While you are correct in the context of my patch, I think that might be a
bug - the yield_to() methods are intending to use ->next:

/* Tell the scheduler that we'd really like se to run next. */
set_next_buddy(se);

yield_task_fair(rq);

.. and yield_to() would rather fundamentally rely on ->next overriding the
next-task-pick selection, but it won't due to NEXT_BUDDY being false:

static struct sched_entity *
pick_next_entity(struct cfs_rq *cfs_rq)
{
/*
* Enabling NEXT_BUDDY will affect latency but not fairness.
*/
if (sched_feat(NEXT_BUDDY) &&
cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
return cfs_rq->next;

> That is, should we not totally nuke the thing?

I don't think we want to nuke it - there's 3 users:

- yield()
- CFS bandwidth
- wakeup

I think the yield() and CFS bandwidth ones are genuine, but non-working due
to NEXT_BUDDY at 0. Wakeup was the original intended NEXT_BUDDY logic, but
it got turned off due to some performance or latency considerations that
might or might not be valid & relevant today.

2)

Even the task_hot() use of ->next isn't spurious: if a task has been marked
as run-next, then presumably the current task is descheduling and we should
probably not tear its ->next away in load-balancing.

3)

Side note: a set rq->next should probably reduce a candidate runqueue's
weight both in periodic load-balancing and in idle-balancing, by rq->curr's
weight or so?

So what I think we should do is to keep ->next and fix all its intended
uses, and make it all unconditional by removing both NEXT_BUDDY and
CACHE_HOT_BUDDY. I can cook up a series if you agree in principle.

Thanks,

Ingo

2024-04-09 09:28:19

by Peter Zijlstra

[permalink] [raw]

Subject: Re: [PATCH 5/5] sched/fair: Rename set_next_buddy() to set_next_pick()

On Tue, Apr 09, 2024 at 10:32:59AM +0200, Ingo Molnar wrote:

> I don't think we want to nuke it - there's 3 users:
>
> - yield()
> - CFS bandwidth
> - wakeup
>
> I think the yield() and CFS bandwidth ones are genuine, but non-working due
> to NEXT_BUDDY at 0. Wakeup was the original intended NEXT_BUDDY logic, but
> it got turned off due to some performance or latency considerations that
> might or might not be valid & relevant today.
>
> 2)
>
> Even the task_hot() use of ->next isn't spurious: if a task has been marked
> as run-next, then presumably the current task is descheduling and we should
> probably not tear its ->next away in load-balancing.
>
> 3)
>
> Side note: a set rq->next should probably reduce a candidate runqueue's
> weight both in periodic load-balancing and in idle-balancing, by rq->curr's
> weight or so?
>
> So what I think we should do is to keep ->next and fix all its intended
> uses, and make it all unconditional by removing both NEXT_BUDDY and
> CACHE_HOT_BUDDY. I can cook up a series if you agree in principle.

So yes on fixing those yield_to() and cfs_bandwidth thingies, but put
then under a new knob -- if the fix regresses we can simply flip it.

Ack on removing the current knobs, for them not having been changed in
forever.