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

[Daniel Lezcano]

On 08/11/2019 11:56, Rafael J. Wysocki wrote:
> 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").

Why not simply ?

int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
				struct cpuidle_device *dev)
{
	int latency = cpuidle_governor_latency_req(dev->cpu);

        return find_deepest_state(drv, dev, latency_req, 0, false);
}

>> 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
> 


-- 
 <http://www.linaro.org/> Linaro.org â Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog