Date:   Thu, 29 Oct 2020 09:43:20 +0800
From:   "zhuguangqing83" <zhuguangqing83@...il.com>
To:     "'Viresh Kumar'" <viresh.kumar@...aro.org>
Cc:     <rjw@...ysocki.net>, <mingo@...hat.com>, <peterz@...radead.org>,
        <juri.lelli@...hat.com>, <vincent.guittot@...aro.org>,
        <dietmar.eggemann@....com>, <rostedt@...dmis.org>,
        <bsegall@...gle.com>, <mgorman@...e.de>, <bristot@...hat.com>,
        <linux-pm@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
        "'zhuguangqing'" <zhuguangqing@...omi.com>
Subject: Re: [PATCH] cpufreq: schedutil: set sg_policy->next_freq to the final cpufreq


> On 28-10-20, 19:03, zhuguangqing83 wrote:
> > Thanks for your comments. Maybe my description was not clear before.
> >
> > If I understand correctly, when policy->min/max get changed in the time
> > Window between get_next_freq() and sugov_fast_switch(), to be more
> > precise between cpufreq_driver_resolve_freq() and
> > cpufreq_driver_fast_switch(), the issue may happen.
> >
> > For example, the first time schedutil callback gets called from the
> > scheduler, we reached get_next_freq() and calculate the next_freq,
> > suppose next_freq is 1.0 GHz, then sg_policy->next_freq is updated
> > to 1.0 GHz in sugov_update_next_freq(). If policy->min/max get
> > change right now, suppose policy->min is changed to 1.2 GHz,
> > then the final next_freq is 1.2 GHz for there is another clamp
> > between policy->min and policy->max in cpufreq_driver_fast_switch().
> > Then sg_policy->next_freq(1.0 GHz) is not the final next_freq(1.2 GHz).
> >
> > The second time schedutil callback gets called from the scheduler, there
> > are two issues:
> > (1) Suppose policy->min is still 1.2 GHz, we reached get_next_freq() and
> > calculate the next_freq, because sg_policy->limits_changed gets set to
> > true by sugov_limits() and there is a clamp between policy->min and
> > policy->max, so this time next_freq will be greater than or equal to 1.2
> > GHz, suppose next_freq is also 1.2 GHz. Now next_freq is 1.2 GHz and
> > sg_policy->next_freq is 1.0 GHz,  then we find
> > "if (sg_policy->next_freq == next_freq)" is not satisfied and we call
> > cpufreq driver to change the cpufreq to 1.2 GHz. Actually it's already
> > 1.2 GHz, it's not necessary to change this time.
> 
> This isn't that bad, but ...
> 
> > (2) Suppose policy->min was changed again to 1.0 GHz before, we reached
> > get_next_freq() and calculate the next_freq, suppose next_freq is also
> > 1.0 GHz. Now next_freq is 1.0 GHz and sg_policy->next_freq is also 1.0 GHz,
> > then we find "if (sg_policy->next_freq == next_freq)" is satisfied and we
> > don't change the cpufreq. Actually we should change the cpufreq to 1.0 GHz
> > this time.
> 
> This is a real problem we can get into. What about this diff instead ?
> 
> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> index 0c5c61a095f6..bf7800e853d3 100644
> --- a/kernel/sched/cpufreq_schedutil.c
> +++ b/kernel/sched/cpufreq_schedutil.c
> @@ -105,7 +105,6 @@ static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
>         if (sg_policy->next_freq == next_freq)
>                 return false;
> 
> -       sg_policy->next_freq = next_freq;
>         sg_policy->last_freq_update_time = time;
> 
>         return true;

It's a little strange that sg_policy->next_freq and 
sg_policy->last_freq_update_time are not updated at the same time.

> @@ -115,7 +114,7 @@ static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
>                               unsigned int next_freq)
>  {
>         if (sugov_update_next_freq(sg_policy, time, next_freq))
> -               cpufreq_driver_fast_switch(sg_policy->policy, next_freq);
> +               sg_policy->next_freq = cpufreq_driver_fast_switch(sg_policy->policy, next_freq);
>  }
> 

Great, it also takes into account the issue that 0 is returned by the
driver's ->fast_switch() callback to indicate an error condition.

For policy->min/max may be not the real CPU frequency in OPPs, so
next_freq got from get_next_freq() which is after clamping between
policy->min and policy->max may be not the real CPU frequency in OPPs.
In that case, if we use a real CPU frequency in OPPs returned from
cpufreq_driver_fast_switch() to compare with next_freq,
"if (sg_policy->next_freq == next_freq)" will never be satisfied, so we
change the CPU frequency every time schedutil callback gets called from
the scheduler. I see the current code in get_next_freq() as following,
the issue mentioned above should not happen. Maybe it's just one of my
unnecessary worries.

static unsigned int get_next_freq(struct sugov_policy *sg_policy,
				  unsigned long util, unsigned long max)
{
......
	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
		return sg_policy->next_freq;
......
}

>  static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
> @@ -124,6 +123,7 @@ static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
>         if (!sugov_update_next_freq(sg_policy, time, next_freq))
>                 return;
> 
> +       sg_policy->next_freq = next_freq;
>         if (!sg_policy->work_in_progress) {
>                 sg_policy->work_in_progress = true;
>                 irq_work_queue(&sg_policy->irq_work);
> 
> --
> viresh

Powered by blists - more mailing lists