Re: [RFC PATCH] sched/cpufreq/schedutil: handling urgent frequency requests

[Joel Fernandes <joel@joelfernandes.org>]

On Wed, May 09, 2018 at 11:06:24AM +0200, Rafael J. Wysocki wrote:
> On Wed, May 9, 2018 at 10:51 AM, Joel Fernandes <joel@joelfernandes.org> wrote:
> > On Wed, May 09, 2018 at 10:30:37AM +0200, Rafael J. Wysocki wrote:
> >> On Wed, May 9, 2018 at 10:06 AM, Joel Fernandes <joel@joelfernandes.org> wrote:
> >> > On Wed, May 09, 2018 at 08:45:30AM +0200, Juri Lelli wrote:
> >> >> On 08/05/18 21:54, Joel Fernandes wrote:
> >> >>
> >> >> [...]
> >> >>
> >> >> > Just for discussion sake, is there any need for work_in_progress? If we can
> >> >> > queue multiple work say kthread_queue_work can handle it, then just queuing
> >> >> > works whenever they are available should be Ok and the kthread loop can
> >> >> > handle them. __cpufreq_driver_target is also protected by the work lock if
> >> >> > there is any concern that can have races... only thing is rate-limiting of
> >> >> > the requests, but we are doing a rate limiting, just not for the "DL
> >> >> > increased utilization" type requests (which I don't think we are doing at the
> >> >> > moment for urgent DL requests anyway).
> >> >> >
> >> >> > Following is an untested diff to show the idea. What do you think?
> >> >> >
> >> >> > thanks,
> >> >> >
> >> >> > - Joel
> >> >> >
> >> >> > ----8<---
> >> >> > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> >> >> > index d2c6083304b4..862634ff4bf3 100644
> >> >> > --- a/kernel/sched/cpufreq_schedutil.c
> >> >> > +++ b/kernel/sched/cpufreq_schedutil.c
> >> >> > @@ -38,7 +38,6 @@ struct sugov_policy {
> >> >> >     struct                  mutex work_lock;
> >> >> >     struct                  kthread_worker worker;
> >> >> >     struct task_struct      *thread;
> >> >> > -   bool                    work_in_progress;
> >> >> >
> >> >> >     bool                    need_freq_update;
> >> >> >  };
> >> >> > @@ -92,16 +91,8 @@ static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
> >> >> >         !cpufreq_can_do_remote_dvfs(sg_policy->policy))
> >> >> >             return false;
> >> >> >
> >> >> > -   if (sg_policy->work_in_progress)
> >> >> > -           return false;
> >> >> > -
> >> >> >     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;
> >> >> >     }
> >> >> >
> >> >> > @@ -129,7 +120,6 @@ static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
> >> >> >             policy->cur = next_freq;
> >> >> >             trace_cpu_frequency(next_freq, smp_processor_id());
> >> >> >     } else {
> >> >> > -           sg_policy->work_in_progress = true;
> >> >> >             irq_work_queue(&sg_policy->irq_work);
> >> >>
> >> >> Isn't this potentially introducing unneeded irq pressure (and doing the
> >> >> whole wakeup the kthread thing), while the already active kthread could
> >> >> simply handle multiple back-to-back requests before going to sleep?
> >> >
> >> > How about this? Will use the latest request, and also doesn't do unnecessary
> >> > irq_work_queue:
> >> >
> >> > (untested)
> >> > -----8<--------
> >> > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> >> > index d2c6083304b4..6a3e42b01f52 100644
> >> > --- a/kernel/sched/cpufreq_schedutil.c
> >> > +++ b/kernel/sched/cpufreq_schedutil.c
> >> > @@ -38,7 +38,7 @@ struct sugov_policy {
> >> >         struct                  mutex work_lock;
> >> >         struct                  kthread_worker worker;
> >> >         struct task_struct      *thread;
> >> > -       bool                    work_in_progress;
> >> > +       bool                    work_in_progress; /* Has kthread been kicked */
> >> >
> >> >         bool                    need_freq_update;
> >> >  };
> >> > @@ -92,9 +92,6 @@ static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
> >> >             !cpufreq_can_do_remote_dvfs(sg_policy->policy))
> >> >                 return false;
> >> >
> >> > -       if (sg_policy->work_in_progress)
> >> > -               return false;
> >> > -
> >>
> >> Why this change?
> >>
> >> Doing the below is rather pointless if work_in_progress is set, isn't it?
> >
> > The issue being discussed is that if a work was already in progress, then new
> > frequency updates will be dropped. So say even if DL increased in
> > utilization, nothing will happen because if work_in_progress = true and
> > need_freq_update = true, we would skip an update.  In this diff, I am
> > allowing the frequency request to be possible while work_in_progress is true.
> > In the end the latest update will be picked.
> 
> I'm not sure if taking new requests with the irq_work in flight is a good idea.

That's the point of the original $SUBJECT patch posted by Claudio :) In that
you can see if urgent_request, then work_in_progress isn't checked.

Also I don't see why we cannot do this with this small tweak as in my diff.
It solves a real problem seen with frequency updates done with the
slow-switch as we discussed at OSPM.

But let me know if I missed your point or something ;)

> 
> >>
> >> You'll drop the results of it on the floor going forward anyway then AFAICS.
> >
> > Why?
> 
> Because you cannot queue up a new irq_work before the previous one is complete?

We are not doing that. If you see in my diff, I am not queuing an irq_work if
one was already queued. What we're allowing is an update to next_freq. We
still use work_in_progress but don't use it to ban all incoming update
requests as done previously. Instead we use work_in_progress to make sure
that we dont unnecessarily increase the irq pressure and have excessive wake
ups (as Juri suggested).

I can clean it up and post it as a patch next week after some testing incase
that's less confusing.
This week I'm actually on vacation and the diff was pure vacation hacking ;-)

thanks,

- Joel