Re: [PATCH] sched/fair: Fix nohz.next_balance update

[Vincent Guittot]

On Tue, 5 May 2020 at 17:16, Peng Liu <iwtbavbm@xxxxxxxxx> wrote:
>
> On Tue, May 05, 2020 at 04:27:11PM +0200, Vincent Guittot wrote:
> > Le mardi 05 mai 2020 Ã 21:40:56 (+0800), Peng Liu a Ãcrit :
> > > On Mon, May 04, 2020 at 05:17:11PM +0200, Vincent Guittot wrote:
> > > > On Sun, 3 May 2020 at 10:34, Peng Liu <iwtbavbm@xxxxxxxxx> wrote:
> > > > >
>
> [...]
>
> > > Yes, you're right. When need_resched() returns true, things become out
> > > of expectation. We haven't really got the earliest next_balance, abort
> > > the update immediately and let the successor to help. Doubtless this
> > > will incur some overhead due to the repeating work.
> >
> > There should not be some repeating works because CPUs and sched_domain, which
> > have already been balanced, will not be rebalanced until the next load balance
> > interval.
> >
> > Futhermore, there is in fact still work to do bcause not all the idle CPUs got
> > a chance to pull work
> >
> > >
> > >
> > > About the "tick is not stopped when entering idle" case, defer the
> > > update to nohz_balance_enter_idle() would be a choice too.
> > >
> > >
> > > Of course, only update nohz.next_balance in rebalance_domains() is the
> > > simpliest way, but as @Valentin put, too many write to it may incur
> > > unnecessary overhead. If we can gather the earliest next_balance in
> >
> > This is not really possible because we have to move it to the next interval.
> >
> > > advance, then a single write is considered to be better.
> > >
> > > By the way, remove the redundant check in nohz_idle_balance().
> > >
> > > FWIW, how about the below?
> >
> > Your proposal below looks quite complex. IMO, one solution would be to move the
> > update of nohz.next_balance before calling rebalance_domains(this_rq, CPU_IDLE)
> > so you are back to the previous behavior.
> >
> > The only difference is that in case of an break because of need_resched, it
> > doesn't update nohz.next_balance. But on the other hand, we haven't yet
> > finished run rebalance_domains for all CPUs and some load_balance are still
> > pending. In fact, this will be done during next tick by an idle CPU.
> >
> > So I would be in favor of something as simple as :
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 04098d678f3b..e028bc1c4744 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -10457,6 +10457,14 @@ static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
> >                 }
> >         }
> >
> > +       /*
> > +        * next_balance will be updated only when there is a need.
> > +        * When the CPU is attached to null domain for ex, it will not be
> > +        * updated.
> > +        */
> > +       if (likely(update_next_balance))
> > +               nohz.next_balance = next_balance;
> > +
> >         /* Newly idle CPU doesn't need an update */
> >         if (idle != CPU_NEWLY_IDLE) {
> >                 update_blocked_averages(this_cpu);
> > @@ -10477,14 +10485,6 @@ static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
> >         if (has_blocked_load)
> >                 WRITE_ONCE(nohz.has_blocked, 1);
> >
> > -       /*
> > -        * next_balance will be updated only when there is a need.
> > -        * When the CPU is attached to null domain for ex, it will not be
> > -        * updated.
> > -        */
> > -       if (likely(update_next_balance))
> > -               nohz.next_balance = next_balance;
> > -
> >         return ret;
> >  }
> >
>
> Indeed, simple and straightforward, it's better.
>
> > > ***********************************************
> > > * Below code is !!!ENTIRELY UNTESTED!!!, just *
>
> [...]
>
> > > @@ -10354,9 +10350,7 @@ static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
> > >  {
> > >     int this_cpu = this_rq->cpu;
> > >     unsigned int flags;
> > > -
> > > -   if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
> > > -           return false;
> >
> > why did you remove this ?
> >
>
> It seems that below 'if' do the same thing, isn't?

The test above is an optimization for the most common case

>
> /* could be _relaxed() */
> flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
> if (!(flags & NOHZ_KICK_MASK))
>         return false;
>
> > > +   bool done;
> > >
> > >     if (idle != CPU_IDLE) {
> > >             atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
> > > @@ -10368,9 +10362,16 @@ static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
> > >     if (!(flags & NOHZ_KICK_MASK))
> > >             return false;
> > >
>
> [...]
>
> > >  static void nohz_newidle_balance(struct rq *this_rq)