Re: [PATCH V6 2/3] cpuidle: play_idle: Specify play_idle with an idle state

[Rafael J. Wysocki]

On Fri, Nov 8, 2019 at 11:47 AM Daniel Lezcano
<daniel.lezcano@xxxxxxxxxx> wrote:
>
> On 08/11/2019 02:20, Rafael J. Wysocki wrote:
> > On Wednesday, October 30, 2019 8:51:40 AM CET Daniel Lezcano wrote:
> >> Currently, the play_idle function does not allow to tell which idle
> >> state we want to go. Improve this by passing the idle state as
> >> parameter to the function.
> >>
> >> Export cpuidle_find_deepest_state() symbol as it is used from the
> >> intel_powerclamp driver as a module.
> >>
> >> There is no functional changes, the cpuidle state is the deepest one.
> >>
> >> Signed-off-by: Daniel Lezcano <daniel.lezcano@xxxxxxxxxx>
> >> Acked-by: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> >> Reviewed-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
> >> ---
> >>   V6:
> >>    - Change variable name 'state' -> 'index':
> >>      https://lkml.org/lkml/2019/10/28/874
> >>   V4:
> >>    - Add EXPORT_SYMBOL_GPL(cpuidle_find_deepest_state) for the
> >>      intel_powerclamp driver when this one is compiled as a module
> >>   V3:
> >>    - Add missing cpuidle.h header
> >> ---
> >>  drivers/cpuidle/cpuidle.c                | 1 +
> >>  drivers/powercap/idle_inject.c           | 4 +++-
> >>  drivers/thermal/intel/intel_powerclamp.c | 4 +++-
> >>  include/linux/cpu.h                      | 2 +-
> >>  kernel/sched/idle.c                      | 4 ++--
> >>  5 files changed, 10 insertions(+), 5 deletions(-)
> >>
> >> diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c
> >> index 18523ea6b11b..b871fc2e8e67 100644
> >> --- a/drivers/cpuidle/cpuidle.c
> >> +++ b/drivers/cpuidle/cpuidle.c
> >> @@ -126,6 +126,7 @@ int cpuidle_find_deepest_state(void)
> >>
> >>      return find_deepest_state(drv, dev, UINT_MAX, 0, false);
> >>  }
> >> +EXPORT_SYMBOL_GPL(cpuidle_find_deepest_state);
> >
> > That doesn't appear to be really necessary to me.
> >
> >>
> >>  #ifdef CONFIG_SUSPEND
> >>  static void enter_s2idle_proper(struct cpuidle_driver *drv,
> >> diff --git a/drivers/powercap/idle_inject.c b/drivers/powercap/idle_inject.c
> >> index cd1270614cc6..233c878cbf46 100644
> >> --- a/drivers/powercap/idle_inject.c
> >> +++ b/drivers/powercap/idle_inject.c
> >> @@ -38,6 +38,7 @@
> >>  #define pr_fmt(fmt) "ii_dev: " fmt
> >>
> >>  #include <linux/cpu.h>
> >> +#include <linux/cpuidle.h>
> >>  #include <linux/hrtimer.h>
> >>  #include <linux/kthread.h>
> >>  #include <linux/sched.h>
> >> @@ -138,7 +139,8 @@ static void idle_inject_fn(unsigned int cpu)
> >>       */
> >>      iit->should_run = 0;
> >>
> >> -    play_idle(READ_ONCE(ii_dev->idle_duration_us));
> >> +    play_idle(READ_ONCE(ii_dev->idle_duration_us),
> >> +              cpuidle_find_deepest_state());
> >
> > The next patch changes this again and I'm not sure why this intermediate
> > change is useful.
> >
> >>  }
> >>
> >>  /**
> >> diff --git a/drivers/thermal/intel/intel_powerclamp.c b/drivers/thermal/intel/intel_powerclamp.c
> >> index 53216dcbe173..b55786c169ae 100644
> >> --- a/drivers/thermal/intel/intel_powerclamp.c
> >> +++ b/drivers/thermal/intel/intel_powerclamp.c
> >> @@ -29,6 +29,7 @@
> >>  #include <linux/delay.h>
> >>  #include <linux/kthread.h>
> >>  #include <linux/cpu.h>
> >> +#include <linux/cpuidle.h>
> >>  #include <linux/thermal.h>
> >>  #include <linux/slab.h>
> >>  #include <linux/tick.h>
> >> @@ -430,7 +431,8 @@ static void clamp_idle_injection_func(struct kthread_work *work)
> >>      if (should_skip)
> >>              goto balance;
> >>
> >> -    play_idle(jiffies_to_usecs(w_data->duration_jiffies));
> >> +    play_idle(jiffies_to_usecs(w_data->duration_jiffies),
> >> +              cpuidle_find_deepest_state());
> >
> > I don't see a reason for changing the code here like this.
> >
> > What you really need is to have a way to set a limit on the idle
> > state exit latency for idle injection on ARM.
>
> Mmh, yes you are right. The idle state number is part of the internals
> of the cpuidle framework while the exit latency is an input (from user
> or kernel).
>
> > For that you can pass the exit latency limit to play_idle(), but then
> > you need to change powerclamp to pass UNIT_MAX or similar which is
> > ugly, or you can redefine cpuidle_use_deepest_state() to take the
> > exit latency limit as the arg (with 0 meaning use_deepest_state == false).
>
> Should it make sense to just get the resume latency in
> cpuidle_use_deepest_state() and pass the value to find_deepest_state()?

Yes, I would change cpuidle_use_deepest_state() to take the max exit
latency as the arg (maybe with 0 meaning "don't use the deepest state
only any more").

> It is the only code path where the constraint is not taken into account
> AFAICT.
>
> With this simple change, we can manage everything from the pm_qos API
> then and this series is no longer needed.

OK