Re: [RFC PATCH RESEND] cpuidle: undelaying cpuidle in dpm_{suspend|resume}()

[Rafael J. Wysocki]

On Tue, Nov 12, 2019 at 6:10 AM Ikjoon Jang <ikjn@xxxxxxxxxxxx> wrote:
>
> On Fri, Nov 8, 2019 at 7:22 PM Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> >
> > On Wednesday, October 30, 2019 3:21:05 AM CET Ikjoon Jang wrote:
> > > cpuidle is paused only during dpm_suspend_noirq() ~ dpm_resume_noirq().
> > > But some device drivers need random sized IOs in dpm_{suspend|resume}()
> > > stage (e.g. re-downloading firmware in resume).
> > > And with such a device, cpuidle's latencies could be critical to
> > > response time of system suspend/resume.
> > >
> > > To minimize those latencies, we could apply pm_qos to such device drivers,
> > > but simply undelaying cpuidle from dpm_suspend_noirq() to dpm suspend()
> > > seems no harm.
> >
> > While the patch is generally acceptable, the changelog is not.
> >
> > First, what does "undelying" mean?
>
> You're right, that should be fixed,
> actually I used 'undelaying' from commit: 8651f97bd951d
> (PM / cpuidle: System resume hang fix with cpuidle),
> when the first time cpuidle_{pause|resume} is introduced:
>
> "Since we are dealing with drivers it seems best to call this function
> during dpm_suspend(). Delaying the call till dpm_suspend_noirq() does
> no harm, as long as it is before cpu_hotplug_begin() to avoid race
> conditions with cpu hotpulg operations."
>
> Delaying does no harm, but I think that there had been no specific
> reason of this
> delay from the beginning. Undelaying does no harm too.

I see.

It would be good to mention commit 8651f97bd951d in the changelog.
And while "delaying" is a proper word in English, "undelaying" isn't
AFAICS, so maybe say "avoid delaying" or something to that effect
instead.

> >
> > Second, you seem to be talking about the cases in which exit latencies of
> > idle states are not small relative to the system suspend/resume time, so
> > without any specific examples this is not really convincing.
> >
> > Also, potentially, there is another reason to make this change, which is
> > that on some systems i2c (or similar) controllers may be requisite for
> > idle state entry and exit, so it may make sense in general to prevent
> > cpuidle from being used over the entire suspend and resume of the
> > system.  However, without any example of a system in which that matters
> > it still is not convincing enough IMO.
> >
>
> Currently I've got only one specific device for examples.
> Maybe this patch needs more generalized examples for applying to all
> other machines.

One example would be enough, but please provide it in the changelog of
the patch.