Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1409775

Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to driver frequency

Show all headers | View raw

2016-05-30 22:25 GMT+08:00 Rafael J. Wysocki <rafael@kernel.org>:
> On Mon, May 30, 2016 at 12:18 PM, Viresh Kumar <viresh.kumar@linaro.org> wrote:
>> On 29-05-16, 02:40, Rafael J. Wysocki wrote:
>>> I can't really parse the above question, so I'm not going to try to
>>> answer it. :-)
>>
>> Sorry about that :(
>>
>> IOW, I think that we should make this change into the sched-governor (I will
>> send a patch separately if you agree to this):
>
> I don't.
>
>> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
>> index 14c4aa25cc45..5934b14aa21c 100644
>> --- a/kernel/sched/cpufreq_schedutil.c
>> +++ b/kernel/sched/cpufreq_schedutil.c
>> @@ -66,11 +66,6 @@ static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
>>
>>         if (unlikely(sg_policy->need_freq_update)) {
>>                 sg_policy->need_freq_update = false;
>> -               /*
>> -                * This happens when limits change, so forget the previous
>> -                * next_freq value and force an update.
>> -                */
>> -               sg_policy->next_freq = UINT_MAX;
>>                 return true;
>>         }
>>
>> And here is my reasoning behind this.
>>
>> Can you show me any case, where the above code (as present in mainline
>> today) leads to a freq-change? I couldn't find any.
>>
>> Let me demonstrate.
>>
>> Following only talks about the fast-switch path, the other path is
>> same as well.
>>
>> Suppose this is the current range of frequencies supported by a
>> driver: 200, 400, 600, 800, 1000 (in MHz).
>>
>> And policy->cur = next_freq = 400 MHz.
>>
>> A.) Suppose that we change policy->min to 400 MHz from userspace.
>>     -> sugov_limits()
>>        This will find everything in order and simply set
>>        need_freq_update, without updating the frequency.
>>
>>     On next util-callback, we will forcefully return true from
>>     sugov_should_update_freq() and reach sugov_update_commit().
>>
>>     We calculate next_freq and that comes to 400 MHz again (that's the
>>     case we are trying to target with the above code).
>>
>>     With the current code, we will forcefully end up calling
>>     cpufreq_driver_fast_switch().
>>
>>     Because the new and current frequencies are same,
>>     cpufreq_driver->fast_switch() will simply return.
>>
>>     NOTE: I also think that cpufreq_driver_fast_switch() should have a
>>     check like (policy->cur == target_freq). I will add that too, in
>>     case you agree.
>>
>>     So, forcefully updating next_freq to UINT_MAX will end up wasting
>>     some cycles, but wouldn't do any useful stuff.
>
> It will, but there's no way to distinguish this case from B in the
> governor with the current min/max synchronization mechanism.  That is,
> it only knows that something has changed, but checking what exactly
> has changed would be racy.
>
>> B.) Suppose that we change policy->min to 600 MHz from userspace.
>>     -> sugov_limits()
>>        This will find that policy->cur is less than 600 and will set
>>        that to 600 MHz by calling __cpufreq_driver_target(). We will
>>        also set need_freq_update.
>>
>>        Note that next_freq and policy->cur are not in sync anymore and
>>        perhaps this is the most important case for the above code.
>
> It is.
>
> Moreover, please note that __cpufreq_driver_target() is only called in
> sugov_limits() when policy->fast_switch_enabled is unset.
>
>>     On next util-callback, we will forcefully return true from
>>     sugov_should_update_freq() and reach sugov_update_commit().
>>
>>     We calculate next_freq and lets say that comes to 400 MHz again
>>     (as that's the case we are trying to target with the above code).
>>
>>     With the current code, we will forcefully end up calling
>>     cpufreq_driver_fast_switch().
>>
>>     Because next_freq() is not part of the new range, we will clamp it
>>     and set it to 600 MHz eventually. Again, next and current
>>     frequencies are same, cpufreq_driver->fast_switch() will simply
>>     return.
>
> Not really (as per the above).
>
> And even in the !fast_switch_enabled case, if next_freq stays at 400
> MHz (which is different from policy->cur), it may lead to suboptimal
> decisions going forward (eg. if it goes to 600 MHz next time and the
> governor will think that the frequency has changed, although in fact
> it hasn't).

Does set next_freq = UNIT_MAX has same effect as next_freq stays at
400MHz since both means that frequency has changed?

Regards,
Wanpeng Li

Back to linux.kernel | Previous | Next — Previous in thread | Next in thread | Find similar | Unroll thread

Thread

[PATCH v2 0/3] cpufreq: avoid redundant driver calls in schedutil Steve Muckle <steve.muckle@linaro.org> - 2016-05-26 05:00 +0200
  [PATCH v2 1/3] cpufreq: add resolve_freq driver callback Steve Muckle <steve.muckle@linaro.org> - 2016-05-26 05:00 +0200
    Re: [PATCH v2 1/3] cpufreq: add resolve_freq driver callback Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-26 08:30 +0200
      Re: [PATCH v2 1/3] cpufreq: add resolve_freq driver callback Steve Muckle <steve.muckle@linaro.org> - 2016-05-30 17:40 +0200
        Re: [PATCH v2 1/3] cpufreq: add resolve_freq driver callback Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-31 07:40 +0200
          Re: [PATCH v2 1/3] cpufreq: add resolve_freq driver callback Steve Muckle <steve.muckle@linaro.org> - 2016-05-31 20:50 +0200
    Re: [PATCH v2 1/3] cpufreq: add resolve_freq driver callback Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-31 13:20 +0200
      Re: [PATCH v2 1/3] cpufreq: add resolve_freq driver callback Steve Muckle <steve.muckle@linaro.org> - 2016-05-31 20:20 +0200
  [PATCH v2 2/3] cpufreq: acpi-cpufreq: add resolve_freq callback Steve Muckle <steve.muckle@linaro.org> - 2016-05-26 05:00 +0200
    Re: [PATCH v2 2/3] cpufreq: acpi-cpufreq: add resolve_freq callback Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-26 08:50 +0200
      Re: [PATCH v2 2/3] cpufreq: acpi-cpufreq: add resolve_freq callback Steve Muckle <steve.muckle@linaro.org> - 2016-05-30 18:30 +0200
        Re: [PATCH v2 2/3] cpufreq: acpi-cpufreq: add resolve_freq callback Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-31 13:40 +0200
  [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to driver frequency Steve Muckle <steve.muckle@linaro.org> - 2016-05-26 05:00 +0200
    Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-26 09:20 +0200
      Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency "Rafael J. Wysocki" <rafael@kernel.org> - 2016-05-29 02:50 +0200
        Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-30 12:20 +0200
          Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency "Rafael J. Wysocki" <rafael@kernel.org> - 2016-05-30 16:30 +0200
            Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Viresh Kumar <viresh.kumar@linaro.org> - 2016-05-30 17:40 +0200
              Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency "Rafael J. Wysocki" <rafael@kernel.org> - 2016-05-30 21:10 +0200
              Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Peter Zijlstra <peterz@infradead.org> - 2016-05-31 11:10 +0200
            Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Wanpeng Li <kernellwp@gmail.com> - 2016-05-31 03:50 +0200
      Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Steve Muckle <steve.muckle@linaro.org> - 2016-05-30 18:40 +0200
        Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Viresh Kumar <viresh.kumar@linaro.org> - 2016-06-01 13:00 +0200
    Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Wanpeng Li <kernellwp@gmail.com> - 2016-05-27 07:50 +0200
      Re: [PATCH v2 3/3] cpufreq: schedutil: map raw required frequency to  driver frequency Steve Muckle <steve.muckle@linaro.org> - 2016-05-30 18:50 +0200

csiph-web