Re: [PATCH v3 1/2] sched/fair: Add NOHZ balancer flag for nohz.next_balance updates

[Peter Zijlstra]

On Mon, Aug 23, 2021 at 03:53:16PM +0200, Dietmar Eggemann wrote:
> >> I'm a bit puzzled by this; that function has:
> >>
> >>   SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
> >>
> >> Which:
> >>
> >>  - isn't updated
> >>  - implies STATS must be set when BALANCE
> > 
> > Yup
> > 
> >>
> >> the latter gives rise to my confusion; why add that gate on STATS? It
> >> just doesn't make sense to do a BALANCE and not update STATS.
> > 
> > AFAIA that warning was only there to catch BALANCE && !STATS, so I didn't
> > tweak it.
> > 
> > Now, you could still end up with
> > 
> >   flags == NOHZ_NEXT_KICK
> > 
> > (e.g. nohz.next_balance is in the future, but a new CPU entered NOHZ-idle
> > and needs its own rq.next_balance collated into the nohz struct)
> > 
> > in which case you don't do any blocked load update, hence the
> > gate. In v1 I had that piggyback on NOHZ_STATS_KICK, but Vincent noted
> > that might not be the best given blocked load updates can be time
> > consuming - hence the separate flag.
> 
> Maybe the confusion stems from the fact that the NOHZ_NEXT_KICK-set
> changes are only introduced in 2/2?
> 
> @@ -10417,6 +10418,9 @@ static void nohz_balancer_kick(struct rq *rq)
>  unlock:
>  	rcu_read_unlock();
>  out:
> +	if (READ_ONCE(nohz.needs_update))
> +		flags |= NOHZ_NEXT_KICK;
> +

The confusion was about how we'd ever get there and not have STATS set,
but i guess having it all nicely gated does make it saner.

Thanks!