Re: [PATCH RFC v4] cpufreq: schedutil: Make iowait boost more energy efficient

From: Viresh Kumar
Date: Wed Jul 12 2017 - 05:46:41 EST

On 12-07-17, 11:36, Peter Zijlstra wrote:
> On Wed, Jul 12, 2017 at 10:30:35AM +0530, Viresh Kumar wrote:
> > On 11-07-17, 07:14, Joel Fernandes wrote:
> > > Another approach than setting min in sugov_set_iowait_boost, is, since
> > > we have already retrieved the current util, we can check if flags ==
> > > SCHED_CPUFREQ_IOWAIT, then set initial the iowait_boost such that
> > > (iowait_boost / iowait_boost_max) is aleast equal to (util / max) or
> > > iowait_boost_min, which ever is lower.
> >
> > So my concerns weren't only about the initial min value, but also that you
> > reduce the freq from sugov_set_iowait_boost(). We can discuss what the ideal
> > value to start with can be.
> I'm not sure I see that. He only mucks with iowait_boost, not the actual
> frequency afaict.
> And sugov_iowait_boost() picks the highest of util vs iowait_boost,
> which wasn't changed.
> Or am I completely missing something? (that code is a bit hard to
> follow)

No, I wasn't clear enough. Sorry about that. Lemme try again:

Suppose min freq is 500 MHz and Max is 2 GHz. The iowait-boost is
set to 1 GHz right now (because of previous events with IOWAIT flag
set), and sugov_set_iowait_boost() gets called again with IOWAIT flag,
we boost the iowait-boost value to 2 GHz. We are in the rate_limit_us
window right now, we return without changing the frequency.

If the next call into the schedutil governor happens due to normal
util-update, flags will be passed as 0. With the current patch, we
will bring iowait-boost back to 1 GHz (before updating the real
frequency to 2 GHz) as the prev-iowait-boost boolean would be set.

And even if the task is periodically getting queued after IOWAIT,
actual boosting may not happen at all in some cases.