Re: [RFC PATCH 3/4] cpufreq: Add Active Stats calls tracking frequency changes

[Lukasz Luba <lukasz.luba@arm.com>]

On 6/22/21 3:59 PM, Rafael J. Wysocki wrote:
> On Tue, Jun 22, 2021 at 4:51 PM Rafael J. Wysocki <rafael@kernel.org> wrote:
>>
>> On Tue, Jun 22, 2021 at 4:09 PM Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>
>>>
>>>
>>> On 6/22/21 2:51 PM, Rafael J. Wysocki wrote:
>>>> On Tue, Jun 22, 2021 at 3:42 PM Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 6/22/21 1:28 PM, Rafael J. Wysocki wrote:
>>>>>> On Tue, Jun 22, 2021 at 9:59 AM Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>>
>>>>>>> The Active Stats framework tracks and accounts the activity of the CPU
>>>>>>> for each performance level. It accounts the real residency, when the CPU
>>>>>>> was not idle, at a given performance level. This patch adds needed calls
>>>>>>> which provide the CPU frequency transition events to the Active Stats
>>>>>>> framework.
>>>>>>>
>>>>>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
>>>>>>> ---
>>>>>>>     drivers/cpufreq/cpufreq.c | 5 +++++
>>>>>>>     1 file changed, 5 insertions(+)
>>>>>>>
>>>>>>> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
>>>>>>> index 802abc925b2a..d79cb9310572 100644
>>>>>>> --- a/drivers/cpufreq/cpufreq.c
>>>>>>> +++ b/drivers/cpufreq/cpufreq.c
>>>>>>> @@ -14,6 +14,7 @@
>>>>>>>
>>>>>>>     #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
>>>>>>>
>>>>>>> +#include <linux/active_stats.h>
>>>>>>>     #include <linux/cpu.h>
>>>>>>>     #include <linux/cpufreq.h>
>>>>>>>     #include <linux/cpu_cooling.h>
>>>>>>> @@ -387,6 +388,8 @@ static void cpufreq_notify_transition(struct cpufreq_policy *policy,
>>>>>>>
>>>>>>>                    cpufreq_stats_record_transition(policy, freqs->new);
>>>>>>>                    policy->cur = freqs->new;
>>>>>>> +
>>>>>>> +               active_stats_cpu_freq_change(policy->cpu, freqs->new);
>>>>>>>            }
>>>>>>>     }
>>>>>>>
>>>>>>> @@ -2085,6 +2088,8 @@ unsigned int cpufreq_driver_fast_switch(struct cpufreq_policy *policy,
>>>>>>>                                policy->cpuinfo.max_freq);
>>>>>>>            cpufreq_stats_record_transition(policy, freq);
>>>>>>>
>>>>>>> +       active_stats_cpu_freq_fast_change(policy->cpu, freq);
>>>>>>> +
>>>>>>
>>>>>> This is quite a bit of overhead and so why is it needed in addition to
>>>>>> the code below?
>>>>>
>>>>> The code below is tracing, which is good for post-processing. We use in
>>>>> our tool LISA, when we analyze the EAS decision, based on captured
>>>>> trace data.
>>>>>
>>>>> This new code is present at run time, so subsystems like our thermal
>>>>> governor IPA can use it and get better estimation about CPU used power
>>>>> for any arbitrary period, e.g. 50ms, 100ms, 300ms, ...
>>>>
>>>> So can it be made not run when the IPA is not using it?
>>>
>>> I can make a Kconfig for IPA to select this ACTIVE_STATS.
>>> Also, I can add description that this framework is mostly needed
>>> for IPA, so don't enable it if you don't use IPA (default is 'n'
>>> so it shouldn't harm others).
>>>
>>> This Active Stats shouldn't be stopped when thermal zone is switching
>>> between governors at run time, e.g. IPA -> step_wise -> IPA
>>> because when IPA is set next time, it might not have correct CPU
>>> stats (what is the current frequency and for how long it has been
>>> actively used).
>>
>> But after a while it will collect enough useful data I suppose?
>>
>>> Beside, switching governors at run time is not a good idea
>>> (apart from stress testing them ;) ).
>>>
>>>>
>>>>>>
>>>>>> And pretty much the same goes for the idle loop change.  There is
>>>>>> quite a bit of instrumentation in that code already and it avoids
>>>>>> adding new locking for a reason.  Why is it a good idea to add more
>>>>>> locking to that code?
>>>>>
>>>>> This active_stats_cpu_freq_fast_change() doesn't use the locking, it
>>>>> relies on schedutil lock in [1].
>>>>
>>>> Ah, OK.
>>>>
>>>> But it still adds overhead AFAICS.
>>>
>>> Agree, it's an extra code. For platforms which use IPA it's a
>>> justifiable cost, weighted by better estimation thanks to this calls.
>>> For other platforms, this framework will be set to default 'n' option.
>>
>> A general problem with build-time configuration is for distros that
>> want to ship one kernel binary to run on multiple hardware platforms.
>> They need to enable those options anyway and then get the full cost on
>> the platforms that don't need it, but want to use the common binary
>> kernel.
>>
>> Again, please consider making this new code run only when it is needed
>> even if configured in and if it runs, make it as low-overhead as
>> possible.
> 
> Also, why don't you add these hooks to the drivers that are generally
> worked with by the IPA?

In Arm world (especially 32-bit world) there is 'a lot' custom idle
and cpufreq drivers. It's probably even not feasible to do.
We also has this CPU_IDLE_MULTIPLE_DRIVERS mechanism. It's a pain,
especially when not having all possible platfroms.

> 
> That you won't need to worry about the possible impact on everybody else.
> 

I'll try to make it as low-overhead as possible and turn off if there is
no client subsystem (like IPA) currently using it. That might be
feasible.

Thank you Rafael for valuable comments.