Re: [PATCH RFC] sched/fair: Avoid unnecessary IPIs for ILB

[Joel Fernandes]

On Mon, Oct 9, 2023 at 7:25 AM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
>
> * Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
>
> > > > Fixes: 7fd7a9e0caba ("sched/fair: Trigger nohz.next_balance updates when a CPU goes NOHZ-idle")
> > >
> > > Hurm.. does this really warrant a Fixes tag? Afaict nothing is currently
> > > broken -- this is a pure optimization question, no?
> >
> > IMHO it is a breakage as it breaks NOHZ -- a lot of times the ILB kicks
> > back the CPU stopping the tick out of idle (effectively breaking NOHZ).
> > The large number of IPIs also wrecks power and it happens only on 6.1 and
> > after. Having the fixes tag means it will also goto all stable kernels >=
> > 6.1. Hope that sounds reasonable and thank you for taking a look!
>
> So it's basically a fix of a NOHZ performance regression, introduced by
> 7fd7a9e0caba or so, correct?

Yes.

> As long as the fixes have a good hope of being backported with a low amount
> of overhead, a Fixes: tag for a ~2 years old performance regression is
> unusual but not unprecedented.
>
> We just need to make sure we don't put too much of a burden on the
> shoulders of -stable maintainers ...

Sure, that sounds good. After we upstream, I am happy to assist in any
stable backporting work related to this as well. In any case, I have
to backport it to ChromeOS kernels which are based on stable. There's
only the 6.1 stable kernel at the moment that is affected.

thanks,

 - Joel