Re: [RFC PATCH 5/6] sched/fair: trigger the update of blocked load on newly idle cpu

[Vincent Guittot]

On Tue, 9 Feb 2021 at 14:09, Valentin Schneider
<valentin.schneider@xxxxxxx> wrote:
>
> On 05/02/21 12:48, Vincent Guittot wrote:
> > Instead of waking up a random and already idle CPU, we can take advantage
> > of this_cpu being about to enter idle to run the ILB and update the
> > blocked load.
> >
> > Signed-off-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> > ---
> >  include/linux/sched/nohz.h |  2 ++
> >  kernel/sched/fair.c        | 11 ++++++++---
> >  kernel/sched/idle.c        |  6 ++++++
> >  3 files changed, 16 insertions(+), 3 deletions(-)
> >
> > diff --git a/include/linux/sched/nohz.h b/include/linux/sched/nohz.h
> > index 6d67e9a5af6b..74cdc4e87310 100644
> > --- a/include/linux/sched/nohz.h
> > +++ b/include/linux/sched/nohz.h
> > @@ -9,8 +9,10 @@
> >  #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
> >  extern void nohz_balance_enter_idle(int cpu);
> >  extern int get_nohz_timer_target(void);
> > +extern void nohz_run_idle_balance(int cpu);
> >  #else
> >  static inline void nohz_balance_enter_idle(int cpu) { }
> > +static inline void nohz_run_idle_balance(int cpu) { }
> >  #endif
> >
> >  #ifdef CONFIG_NO_HZ_COMMON
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 935594cd5430..3d2ab28d5736 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -10461,6 +10461,11 @@ static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
> >       return true;
> >  }
> >
> > +void nohz_run_idle_balance(int cpu)
> > +{
> > +     nohz_idle_balance(cpu_rq(cpu), CPU_IDLE);
> > +}
> > +
> >  static void nohz_newidle_balance(struct rq *this_rq)
> >  {
> >       int this_cpu = this_rq->cpu;
> > @@ -10482,10 +10487,10 @@ static void nohz_newidle_balance(struct rq *this_rq)
> >               return;
> >
> >       /*
> > -      * Blocked load of idle CPUs need to be updated.
> > -      * Kick an ILB to update statistics.
> > +      * Set the need to trigger ILB in order to update blocked load
> > +      * before entering idle state.
> >        */
> > -     kick_ilb(NOHZ_STATS_KICK);
> > +     this_rq->nohz_idle_balance = NOHZ_STATS_KICK;
> >  }
> >
> >  #else /* !CONFIG_NO_HZ_COMMON */
> > diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
> > index 305727ea0677..52a4e9ce2f9b 100644
> > --- a/kernel/sched/idle.c
> > +++ b/kernel/sched/idle.c
> > @@ -261,6 +261,12 @@ static void cpuidle_idle_call(void)
> >  static void do_idle(void)
> >  {
> >       int cpu = smp_processor_id();
> > +
> > +     /*
> > +      * Check if we need to update some blocked load
> > +      */
> > +     nohz_run_idle_balance(cpu);
> > +
>
> What do we gain from doing this here vs having a stats update in
> newidle_balance()?

As mentioned by Joel, newidle_balance is called in the schedule
context with preempt and irq off  which prevent any local activity
like irq/timer. Whereas in this new place, we have the same condition
as during ILB with only preemptoff and _nohz_idle_balance() regularly
checks if it has to abort because something has been scheduled on the
cpu.


>
> The current approach is to have a combined load_balance() + blocked load
> update during newidle, and I get that this can take too long. But then,
> we could still have what you're adding to do_idle() in the tail of
> newidle_balance() itself, no? i.e.
>
>   newidle_balance()
>     ...
>     for_each_domain(this_cpu, sd) {
>        ...
>        pulled_task = load_balance(...);
>        ...
>     }
>     ...
>     if (!pulled_task && !this_rq->nr_running) {
>       this_rq->nohz_idle_balance = NOHZ_STATS_KICK;
>       _nohz_idle_balance();
>     }
>
> or somesuch.
>
> >       /*
> >        * If the arch has a polling bit, we maintain an invariant:
> >        *
> > --
> > 2.17.1