RE: [PATCH] sched/fair: fix pick_eevdf to always find the correct se

[Biju Das]

Hi Chen Yu,

> Subject: Re: [PATCH] sched/fair: fix pick_eevdf to always find the correct
> se
> 
> On 2023-10-06 at 21:24:45 +0200, Peter Zijlstra wrote:
> > On Fri, Oct 06, 2023 at 05:55:01PM +0200, Peter Zijlstra wrote:
> >
> > > And yeah, min_deadline is hosed somehow.
> > >
> > > migration/28-185     [028] d..2.    70.264274: validate_cfs_rq: --- /
> > > migration/28-185     [028] d..2.    70.264277: __print_se:
> ffff88845cf48080 w: 1024 ve: -58857638 lag: 870381 vd: -55861854 vmd: -
> 66302085 E (11372/tr)
> > > migration/28-185     [028] d..2.    70.264280: __print_se:
> ffff88810d165800 w: 25 ve: -80323686 lag: 22336429 vd: -41496434 vmd: -
> 66302085 E (-1//autogroup-31)
> > > migration/28-185     [028] d..2.    70.264282: __print_se:
> ffff888108379000 w: 25 ve: 0 lag: -57987257 vd: 114632828 vmd: 114632828 N
> (-1//autogroup-33)
> > > migration/28-185     [028] d..2.    70.264283: validate_cfs_rq:
> min_deadline: -55861854 avg_vruntime: -62278313462 / 1074 = -57987256
> > >
> > > I need to go make dinner (kids hungry), but I'll see if I can figure
> > > out how this happens...
> >
> > *sigh*, does the below help?
> >
> > ---
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index
> > 04fbcbda97d5..6a670f119efa 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -3632,6 +3747,7 @@ static void reweight_entity(struct cfs_rq *cfs_rq,
> struct sched_entity *se,
> >  		 */
> >  		deadline = div_s64(deadline * old_weight, weight);
> >  		se->deadline = se->vruntime + deadline;
> > +		min_deadline_cb_propagate(&se->run_node, NULL);
> >  	}
> >
> >  #ifdef CONFIG_SMP
> 
> Is it because without this patch, the se->deadline is not always synced
> with se->min_deadline, then in pick_eevdf() the following condition could
> not be met, thus we get a NULL candidate and has to pick the leftmost one?
> 	if (se->deadline == se->min_deadline)
> 
> Regarding the circular locking warning triggered by printk, does it mean we
> should not get a NULL candidate from __pick_eevdf() in theory? And besides,
> we should not use printk with rq lock hold?

Is it not a useful error log? At least from the initial report Marek Szyprowski doesn't see "EEVDF scheduling fail, picking leftmost" but seen only warning triggered by this in the logs.

Cheers,
Biju