Re: [patch 2/2] sched/idle: Make default_idle_call() NOHZ aware

[Qais Yousef]

On 03/02/26 22:25, Rafael J. Wysocki wrote:
> On Mon, Mar 2, 2026 at 12:04 PM Christian Loehle
> <christian.loehle@xxxxxxx> wrote:
> >
> > On 3/1/26 19:30, Thomas Gleixner wrote:
> > > Guests fall back to default_idle_call() as there is no cpuidle driver
> > > available to them by default. That causes a problem in fully loaded
> > > scenarios where CPUs go briefly idle for a couple of microseconds:
> > >
> > > tick_nohz_idle_stop_tick() is invoked unconditionally which means unless
> > > there is timer pending in the next tick, the tick is stopped and a couple
> > > of microseconds later when the idle condition goes away restarted. That
> > > requires to program the clockevent device twice which implies a VM exit for
> > > each reprogramming.
> > >
> > > It was suggested to remove the tick_nohz_idle_stop_tick() invocation from
> > > the default idle code, but would be counterproductive. It would not allow
> > > the host to go into deeper idle states when the guest CPU is fully idle as
> > > it has to maintain the periodic tick.
> > >
> > > Cure this by implementing a trivial moving average filter which keeps track
> > > of the recent idle recidency time and only stop the tick when the average
> > > is larger than a tick.
> > >
> > > Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxx>
> > > ---
> > >  kernel/sched/idle.c |   65 +++++++++++++++++++++++++++++++++++++++++++++-------
> > >  1 file changed, 57 insertions(+), 8 deletions(-)
> > >
> > > --- a/kernel/sched/idle.c
> > > +++ b/kernel/sched/idle.c
> > > @@ -105,12 +105,7 @@ static inline void cond_tick_broadcast_e
> > >  static inline void cond_tick_broadcast_exit(void) { }
> > >  #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
> > >
> > > -/**
> > > - * default_idle_call - Default CPU idle routine.
> > > - *
> > > - * To use when the cpuidle framework cannot be used.
> > > - */
> > > -static void __cpuidle default_idle_call(void)
> > > +static void __cpuidle __default_idle_call(void)
> > >  {
> > >       instrumentation_begin();
> > >       if (!current_clr_polling_and_test()) {
> > > @@ -130,6 +125,61 @@ static void __cpuidle default_idle_call(
> > >       instrumentation_end();
> > >  }
> > >
> > > +#ifdef CONFIG_NO_HZ_COMMON
> > > +
> > > +/* Limit to 4 entries so it fits in a cache line */
> > > +#define IDLE_DUR_ENTRIES     4
> > > +#define IDLE_DUR_MASK                (IDLE_DUR_ENTRIES - 1)
> > > +
> > > +struct idle_nohz_data {
> > > +     u64             duration[IDLE_DUR_ENTRIES];
> > > +     u64             entry_time;
> > > +     u64             sum;
> > > +     unsigned int    idx;
> > > +};
> > > +
> > > +static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
> > > +
> > > +/**
> > > + * default_idle_call - Default CPU idle routine.
> > > + *
> > > + * To use when the cpuidle framework cannot be used.
> > > + */
> > > +static void default_idle_call(void)
> > > +{
> > > +     struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
> > > +     unsigned int idx = nd->idx;
> > > +     s64 delta;
> > > +
> > > +     /*
> > > +      * If the CPU spends more than a tick on average in idle, try to stop
> > > +      * the tick.
> > > +      */
> > > +     if (nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES)
> > > +             tick_nohz_idle_stop_tick();
> > > +
> > > +     __default_idle_call();
> > > +
> > > +     /*
> > > +      * Build a moving average of the time spent in idle to prevent stopping
> > > +      * the tick on a loaded system which only goes idle briefly.
> > > +      */
> > > +     delta = max(sched_clock() - nd->entry_time, 0);
> > > +     nd->sum += delta - nd->duration[idx];
> > > +     nd->duration[idx] = delta;
> > > +     nd->idx = (idx + 1) & IDLE_DUR_MASK;
> > > +}
> > > +
> > > +static void default_idle_enter(void)
> > > +{
> > > +     this_cpu_write(nohz_data.entry_time, sched_clock());
> > > +}
> > > +
> > > +#else  /* CONFIG_NO_HZ_COMMON */
> > > +static inline void default_idle_call(void { __default_idle_call(); }
> > > +static inline void default_idle_enter(void) { }
> > > +#endif /* !CONFIG_NO_HZ_COMMON */
> > > +
> > >  static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
> > >                              struct cpuidle_device *dev,
> > >                              u64 max_latency_ns)
> > > @@ -186,8 +236,6 @@ static void cpuidle_idle_call(void)
> > >       }
> > >
> > >       if (cpuidle_not_available(drv, dev)) {
> > > -             tick_nohz_idle_stop_tick();
> > > -
> > >               default_idle_call();
> > >               goto exit_idle;
> > >       }
> > > @@ -276,6 +324,7 @@ static void do_idle(void)
> > >
> > >       __current_set_polling();
> > >       tick_nohz_idle_enter();
> > > +     default_idle_enter();
> > >
> > >       while (!need_resched()) {
> > >
> > >
> >
> > How does this work? We don't stop the tick until the average idle time is larger,
> > but if we don't stop the tick how is that possible?
> >
> > Why don't we just require one or two consecutive tick wakeups before stopping?
> 
> Exactly my thought and I think one should be sufficient.

I concur. From our experience with TEO util threshold these averages can
backfire. I think one tick is sufficient delay to not be obviously broken. But
IMO the setup is broken too. No cpuidle driver and nohz is enabled but
performance is important is not a good combination. Since this has proven to
have both power and performance impact, ensuring there's a sensible cpuidle
driver is the right thing to do IMHO.