Re: [PATCH v2 1/3] sched/fair: Fix warning if NEXT_BUDDY enabled

[K Prateek Nayak]

Hello Adam,

On 11/29/2024 8:51 AM, Adam Li wrote:
> On 11/28/2024 3:29 PM, K Prateek Nayak wrote:
> > Hello Adam,
> Hi Prateek,
> Thanks for comments.
>
> > On 11/27/2024 11:26 AM, Adam Li wrote:
> > > Enabling NEXT_BUDDY triggers warning, and rcu stall:
> > >
> > > [  124.977300] cfs_rq->next->sched_delayed
> >
> > I could reproduce this with a run of "perf bench sched messaging" but
> > given that we hit this warning, it also means that either
> > set_next_buddy() has incorrectly set a delayed entity as next buddy, or
> > clear_next_buddy() did not clear a delayed entity.
> Yes. The logic of this patch is a delayed entity should not be set as next buddy.
>
> > I also see PSI splats like:
> >
> >      psi: inconsistent task state! task=2524:kworker/u1028:2 cpu=154 psi_flags=10 clear=14 set=0
> >
> > but the PSI flags it has set "(TSK_MEMSTALL_RUNNING | TSK_MEMSTALL)" and
> > the flags it is trying to clear
> > "(TSK_MEMSTALL_RUNNING | TSK_MEMSTALL | TSK_RUNNING)" seem to be only
> > possible if you have picked a dequeued entity for running before its
> > wakeup, which is also perhaps why the "nr_running" computation goes awry
> > and pick_eevdf() returns NULL (which it should never since
> > pick_next_entity() is only called when rq->cfs.nr_running is > 0)
> IIUC, one path for pick_eevdf() to return NULL is:
> pick_eevdf():
> <snip>
> 	if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
> 		curr = NULL; <--- curr is set to NULL

"on_rq" is only cleared when the entity is dequeued so "curr" is in fact
going to sleep (proper sleep) and we've reached at pick_eevdf(),
otherwise, if "curr" is not eligible, there is at least one more tasks
on the cfs_rq which implies best has be found and will be non-null.

> <snip>
> found:
> 	if (!best || (curr && entity_before(curr, best)))
> 		best = curr; <--- curr and best are both NULL

Say "curr" is going to sleep, and there is no "best", in which case
"curr" is already blocked and "cfs_rq->nr_running" should be 0 and it
should have not reached pick_eevdf() in the first place since
pick_next_entity() is only called by pick_task_fair() if
"cfs_rq->nr_running" is non-zero.

So as long as "cfs_rq->nr_running" is non-zero, pick_eevdf() should
return a valid runnable entity. Failure to do so perhaps points to
"entity_eligible()" check going sideways somewhere or a bug in
"nr_running" accounting.

Chenyu had proposed a similar fix long back in
https://lore.kernel.org/lkml/20240226082349.302363-1-yu.c.chen@xxxxxxxxx/
but the consensus was it was covering up a larger problem which
then boiled down to avg_vruntime being computed incorrectly
https://lore.kernel.org/lkml/ZiAWTU5xb%2FJMn%2FHs@chenyu5-mobl2/

> 	return best;  <--- return NULL
>
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index fbdca89c677f..cd1188b7f3df 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -8748,6 +8748,8 @@ static void set_next_buddy(struct sched_entity *se)
> > >                return;
> > >            if (se_is_idle(se))
> > >                return;
> > > +        if (se->sched_delayed)
> > > +            return;
> >
> > I tried to put a SCHED_WARN_ON() here to track where this comes from and
> > seems like it is usually from attach_task() in the load balancing path
> > pulling a delayed task which is set as the next buddy in
> > check_preempt_wakeup_fair()
> >
> > Can you please try the following diff instead of the first two patches
> > and see if you still hit these warnings, stalls, and pick_eevdf()
> > returning NULL?
> Tested. Run specjbb with NEXT_BUDDY enabled, warnings, stalls and panic disappear.

Thank you for testing. I'll let Peter come back on which approach he
prefers :)

> Regards,
> -adam

--
Thanks and Regards,
Prateek