Re: [PATCH] sched/fair: Limit sched_cfs_period_timer loop to avoid hard lockup

From: Phil Auld
Date: Fri Mar 15 2019 - 12:17:12 EST


On Fri, Mar 15, 2019 at 05:03:47PM +0100 Peter Zijlstra wrote:
> On Fri, Mar 15, 2019 at 11:30:42AM -0400, Phil Auld wrote:
>
> > In my defense here, all the fair.c imbalance pct code also uses 100 :)
>
> Yes, I know, I hate on that too ;-) Just never got around to fixing
> that.
>
>
> > with the below:
> >
> > [ 117.235804] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 2492, cfs_quota_us = 143554)
> > [ 117.346807] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 2862, cfs_quota_us = 164863)
> > [ 117.470569] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 3286, cfs_quota_us = 189335)
> > [ 117.574883] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 3774, cfs_quota_us = 217439)
> > [ 117.652907] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 4335, cfs_quota_us = 249716)
> > [ 118.090535] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 4978, cfs_quota_us = 286783)
> > [ 122.098009] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 5717, cfs_quota_us = 329352)
> > [ 126.255209] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 6566, cfs_quota_us = 378240)
> > [ 126.358060] cfs_period_timer[cpu2]: period too short, scaling up (new cfs_period_us 7540, cfs_quota_us = 434385)
> > [ 126.538358] cfs_period_timer[cpu9]: period too short, scaling up (new cfs_period_us 8660, cfs_quota_us = 498865)
> > [ 126.614304] cfs_period_timer[cpu9]: period too short, scaling up (new cfs_period_us 9945, cfs_quota_us = 572915)
> > [ 126.817085] cfs_period_timer[cpu9]: period too short, scaling up (new cfs_period_us 11422, cfs_quota_us = 657957)
> > [ 127.352038] cfs_period_timer[cpu9]: period too short, scaling up (new cfs_period_us 13117, cfs_quota_us = 755623)
> > [ 127.598043] cfs_period_timer[cpu9]: period too short, scaling up (new cfs_period_us 15064, cfs_quota_us = 867785)
> >
> >
> > Plus on repeats I see an occasional
> >
> > [ 152.803384] sched_cfs_period_timer: 9 callbacks suppressed
>
> That should be fine, right? It's a fallback for an edge case and
> shouldn't trigger too often anyway.

It doesn't hit the NMI, just takes a bit longer to get out. It is a little messier
output, but as you say, it's a fallback. If you're okay with it do you want to
just use your patch?

Otherwise, I'm happy to do a fixup v2.

>
> >> I'll rework the maths in the averaged version and post v2 if that makes sense.
> >
> > It may have the extra timer fetch, although maybe I could rework it so that it used the
> > nsstart time the first time and did not need to do it twice in a row. I had originally
> > reverted the hrtimer_forward_now() to hrtimer_forward() but put that back.
>
> Sure; but remember, simpler is often better, esp. for code that
> typically 'never' runs.
>
> > Also, fwiw, this was reported earlier by Anton Blanchard in https://lkml.org/lkml/2018/12/3/1047
>
> Bah, yes, I sometimes loose track of things :/


No worries. I just meant that to show I was not the only one with these low settings,
and to give credit, or whatever :)


Cheers,
Phil

--