Re: [PATCH 1/2] cpuidle: Avoid calls to cpuidle_resume|pause() for s2idle

[Ulf Hansson]

On Wed, 6 Oct 2021 at 12:22, Maulik Shah <mkshah@xxxxxxxxxxxxxx> wrote:
>
> Hi,
>
> On 9/29/2021 8:14 PM, Ulf Hansson wrote:
> > In s2idle_enter(), cpuidle_resume|pause() are invoked to re-allow calls to
> > the cpuidle callbacks during s2idle operations. This is needed because
> > cpuidle is paused in-between in dpm_suspend_noirq() and dpm_resume_noirq().
> >
> > However, calling cpuidle_resume|pause() from s2idle_enter() looks a bit
> > superfluous, as it also causes all CPUs to be waken up when the first CPU
> > wakes up from s2idle.
>
> Thanks for the patch. This can be good optimization to avoid waking up
> all CPUs always.
>
> >
> > Therefore, let's drop the calls to cpuidle_resume|pause() from
> > s2idle_enter(). To make this work, let's also adopt the path in the
> > cpuidle_idle_call() to allow cpuidle callbacks to be invoked for s2idle,
> > even if cpuidle has been paused.
> >
> > Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
> > ---
> >   drivers/cpuidle/cpuidle.c |  7 ++++++-
> >   include/linux/cpuidle.h   |  2 ++
> >   kernel/power/suspend.c    |  2 --
> >   kernel/sched/idle.c       | 40 ++++++++++++++++++++++-----------------
> >   4 files changed, 31 insertions(+), 20 deletions(-)
> >
> > diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c
> > index ef2ea1b12cd8..c76747e497e7 100644
> > --- a/drivers/cpuidle/cpuidle.c
> > +++ b/drivers/cpuidle/cpuidle.c
> > @@ -49,7 +49,12 @@ void disable_cpuidle(void)
> >   bool cpuidle_not_available(struct cpuidle_driver *drv,
> >                          struct cpuidle_device *dev)
> >   {
> > -     return off || !initialized || !drv || !dev || !dev->enabled;
> > +     return off || !drv || !dev || !dev->enabled;
> > +}
> > +
> > +bool cpuidle_paused(void)
> > +{
> > +     return !initialized;
> >   }
> >
> >   /**
> > diff --git a/include/linux/cpuidle.h b/include/linux/cpuidle.h
> > index fce476275e16..51698b385ab5 100644
> > --- a/include/linux/cpuidle.h
> > +++ b/include/linux/cpuidle.h
> > @@ -165,6 +165,7 @@ extern void cpuidle_pause_and_lock(void);
> >   extern void cpuidle_resume_and_unlock(void);
> >   extern void cpuidle_pause(void);
> >   extern void cpuidle_resume(void);
> > +extern bool cpuidle_paused(void);
> >   extern int cpuidle_enable_device(struct cpuidle_device *dev);
> >   extern void cpuidle_disable_device(struct cpuidle_device *dev);
> >   extern int cpuidle_play_dead(void);
> > @@ -204,6 +205,7 @@ static inline void cpuidle_pause_and_lock(void) { }
> >   static inline void cpuidle_resume_and_unlock(void) { }
> >   static inline void cpuidle_pause(void) { }
> >   static inline void cpuidle_resume(void) { }
> > +static inline bool cpuidle_paused(void) {return true; }
> >   static inline int cpuidle_enable_device(struct cpuidle_device *dev)
> >   {return -ENODEV; }
> >   static inline void cpuidle_disable_device(struct cpuidle_device *dev) { }
> > diff --git a/kernel/power/suspend.c b/kernel/power/suspend.c
> > index eb75f394a059..388a5de4836e 100644
> > --- a/kernel/power/suspend.c
> > +++ b/kernel/power/suspend.c
> > @@ -97,7 +97,6 @@ static void s2idle_enter(void)
> >       raw_spin_unlock_irq(&s2idle_lock);
> >
> >       cpus_read_lock();
> > -     cpuidle_resume();
> >
> >       /* Push all the CPUs into the idle loop. */
> >       wake_up_all_idle_cpus();
>
> wake_up_all_idle_cpus() will still cause all CPUs to be woken up when
> first cpu wakes up.
>
> say for example,
> 1. device goes to s2idle suspend.
> 2. one CPU wakes up to handle irq (irq is not a wake irq but left
> enabled at GIC because of IRQF_NOSUSPEND flag) so such irq will not
> break suspend.
> 3. The cpu handles the irq.
> 4. same cpu don't break s2idle_loop() and goes to s2idle_enter() where
> it wakes up all existing idle cpus due to wake_up_all_idle_cpus()
> 5. all of CPUs again enter s2idle.
>
> to avoid waking up all CPUs in above case, something like below snip may
> help (i have not tested yet),
>
> when CPUs are in s2idle_loop(),
>
> 1. set the s2idle state to enter.
> 2. wake up all cpus from shallow state, so that they can re-enter
> deepest state.
> 3. Forever loop until a break with some wake irq.
> 4. clear the s2idle state.
> 5. wake up all cpus from deepest state so that they can now stay in
> shallow state/running state.
>
> void s2idle_loop(void)
> {
>
> +       s2idle_state = S2IDLE_STATE_ENTER;
> +       /* Push all the CPUs to enter deepest available state */
> +       wake_up_all_idle_cpus();
>          for (;;) {
>                  if (s2idle_ops && s2idle_ops->wake) {
>                          if (s2idle_ops->wake())
>                                 ..
>                  s2idle_enter();
>          }
> +       s2idle_state = S2IDLE_STATE_NONE;
> +       /* Push all the CPUs to enter default_idle() from this point */
> +       wake_up_all_idle_cpus();
> }

Overall, I follow your reasoning above and I think it makes sense to
me, but maybe Rafael has some concerns about it.

Even if the above code needs some polishing, the logic seems
reasonable to me. I suggest you post a patch, based on top of my small
series, so we can discuss your suggested improvements separately. Or
just tell me, if you would like me to do it.

>
> Thanks,
> Maulik

Thanks for reviewing!

[...]

Kind regards
Uffe