Re: [PATCH 2/5] cpufreq: schedutil: support scheduler cpufreq callbacks on remote CPUs

[Rafael J. Wysocki]

On Thu, May 19, 2016 at 8:40 PM, Steve Muckle <steve.muckle@xxxxxxxxxx> wrote:
> On Thu, May 19, 2016 at 01:24:41AM +0200, Rafael J. Wysocki wrote:
>> On Mon, May 9, 2016 at 11:20 PM, Steve Muckle <steve.muckle@xxxxxxxxxx> wrote:
>> > In preparation for the scheduler cpufreq callback happening on remote
>> > CPUs, add support for this in schedutil.
>> >
>> > Schedutil currently requires the callback occur on the CPU being
>> > updated in order to support fast frequency switches. Remove this
>> > limitation by checking for the current CPU being outside the target
>> > CPU's cpufreq policy and if this is the case, enqueuing an irq_work on
>> > the target CPU. The irq_work for schedutil is modified to carry out a
>> > fast frequency switch if that is enabled for the policy.
>> >
>> > If the callback occurs on a CPU within the target CPU's policy, the
>> > transition is carried out on the local CPU.
>> >
>> > Signed-off-by: Steve Muckle <smuckle@xxxxxxxxxx>
>> > ---
>> >  kernel/sched/cpufreq_schedutil.c | 86 ++++++++++++++++++++++++++++++----------
>> >  1 file changed, 65 insertions(+), 21 deletions(-)
>> >
>> > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
>> > index 154ae3a51e86..c81f9432f520 100644
>> > --- a/kernel/sched/cpufreq_schedutil.c
>> > +++ b/kernel/sched/cpufreq_schedutil.c
>> > @@ -76,27 +76,61 @@ static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
>> >         return delta_ns >= sg_policy->freq_update_delay_ns;
>> >  }
>> >
>> > -static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
>> > +static void sugov_fast_switch(struct sugov_policy *sg_policy, int cpu,
>> > +                             unsigned int next_freq)
>> > +{
>> > +       struct cpufreq_policy *policy = sg_policy->policy;
>> > +
>> > +       next_freq = cpufreq_driver_fast_switch(policy, next_freq);
>> > +       if (next_freq == CPUFREQ_ENTRY_INVALID)
>> > +               return;
>> > +
>> > +       policy->cur = next_freq;
>> > +       trace_cpu_frequency(next_freq, cpu);
>> > +}
>> > +
>> > +#ifdef CONFIG_SMP
>>
>> schedutil depends on CONFIG_SMP now, so that's not necessary at least
>> for the time being.
>
> Will remove.
>
>>
>> > +static inline bool sugov_queue_remote_callback(struct sugov_policy *sg_policy,
>> > +                                        int cpu)
>> > +{
>> > +       struct cpufreq_policy *policy = sg_policy->policy;
>> > +
>> > +       if (!cpumask_test_cpu(smp_processor_id(), policy->cpus)) {
>>
>> This check is overkill for policies that aren't shared (and we have a
>> special case for them already).
>
> I don't see why it is overkill -

Because it requires more computation, memory accesses etc than simply
comparing smp_processor_id() with cpu.

> regardless of whether the policy is
> shared, we need to determine whether or not we are running on one of the
> CPUs (or in the case of a non-shared policy, the single CPU) within that
> policy to know whether we can immediately change the frequency in this
> context or a remote call is required.
>
>> > +               sg_policy->work_in_progress = true;
>> > +               irq_work_queue_on(&sg_policy->irq_work, cpu);
>> > +               return true;
>> > +       }
>> > +
>> > +       return false;
>> > +}
>> > +#else
>> > +static inline bool sugov_queue_remote_callback(struct sugov_policy *sg_policy,
>> > +                                        int cpu)
>> > +{
>> > +       return false;
>> > +}
>> > +#endif
>> > +
>> > +static void sugov_update_commit(struct sugov_cpu *sg_cpu, int cpu, u64 time,
>>
>> It looks like you might pass hook here instead of the sg_cpu, cpu pair.
>
> I can do that but it means having to do the comtainer_of operation
> again. Strictly speaking this seems slightly less efficient than passing
> the above values which are already available in the callers.

Well, it seems a bit odd to pass two things referring to the same CPU,
but then I don't care that much.

>> >                                 unsigned int next_freq)
>> >  {
>> > +       struct sugov_policy *sg_policy = sg_cpu->sg_policy;
>> >         struct cpufreq_policy *policy = sg_policy->policy;
>> >
>> >         sg_policy->last_freq_update_time = time;
>> >
>> > +       if (sg_policy->next_freq == next_freq) {
>> > +               trace_cpu_frequency(policy->cur, cpu);
>> > +               return;
>> > +       }
>>
>> There was a reason why I put the above under the fast_switch_enabled
>> branch and it was because this check/trace is not necessary otherwise.
>
> I remember asking about this tracepoint earlier. You had said it was
> required because powertop would not work without it (reporting the CPU
> as idle in certain situations).
>
> I'm not sure why that is only true for the fast switch enabled case

Because in the other case cpufreq stats are used by powertop and then
this problem is not visible.

> but it seems like an odd inconsistency for the governor to trace unchanged
> frequencies when fast switches are enabled but not otherwise. It'd be
> useful I think for profiling and tuning if the tracing was consistent.

Well, fair enough.

> This behavioral change is admittedly not part of the purpose of the
> patch and could be split out if needbe.

No need to split IMO, but it might be prudent to mention that change
in behavior in the changelog.