Re: [PATCH v2 0/4] PM: domains: Avoid boilerplate code for DVFS in subsystem/drivers

[Stephan Gerhold <stephan@gerhold.net>]

On Thu, Jun 03, 2021 at 05:27:30PM +0200, Ulf Hansson wrote:
> On Thu, 3 Jun 2021 at 13:13, Stephan Gerhold <stephan@gerhold.net> wrote:
> > I think this might also go into the direction of my problem with the OPP
> > core for CPU DVFS [1] since the OPP core currently does not "power-on"
> > the power domains, it just sets a performance state. I got kind of stuck
> > with all the complexity of power domains in Linux so I think we never
> > solved that.
> 
> Hmm, that issue is in a way related.
> 
> Although, if I understand correctly, that was rather about at what
> layer it makes best sense to activate the device (from runtime PM
> point of view). And this was needed due to the fact that the
> corresponding genpd provider, requires the PM domain to be power on to
> allow changing a performance state for it. Did I get that correct?
> 

Yes, mostly. But I guess I keep coming back to the same question:

When/why does it make sense to vote for a "performance state" of
a power domain that is or might be powered off?

"Powered off" sounds like the absolutely lowest possible performance
state to me, it's just not on at all. And if suddenly a device comes and
says "I want performance state X", nothing can change until the power
domain is also "powered on".

I think my "CPU DVFS" problem only exists because in many other
situations it's possible to rely on one of the following side effects:

  1. The genpd provider does not care if it's powered on or not.
     (i.e. it's always-on or implicitly powers on if state > 0).
  2. There is some other device that votes to keep the power domain on.

And that's how the problem relates to my comment for this patch series ...

>
> >
> > Do I understand your patch set correctly that you basically make the
> > performance state votes conditional to the "power-on" vote of the device
> > (which is automatically toggled during runtime/system PM)?
> 
> The series can be considered as a step in that direction, but no, this
> series doesn't change that behaviour.
> 
> Users of dev_pm_genpd_set_performance_state() are still free to set a
> performance state, orthogonally to whether the PM domain is powered on
> or off.
> 
> >
> > If yes, I think that's a good thing. It was always really confusing to me
> > that a device can make performance state votes if it doesn't actually
> > want the power domain to be powered on.
> 
> I share your view, it's a bit confusing.
> 
> Just adding the condition internally to genpd to prevent the caller of
> dev_pm_genpd_set_performance() from succeeding to set a new state,
> unless the genpd is powered on, should be a rather simple thing to
> add.
> 
> However, to change this, we first need to double check that all the
> callers are making sure they have turned on the PM domain (typically
> via runtime PM).
> 

... because if performance state votes would be conditional to the
"power-on" vote of the device, it would no longer be possible
to rely on the side effects mentioned above. So this would most
certainly break some code that (incorrectly?) relies on these side
effects, but would also prevent such code.

My (personal) feeling so far is that just dropping performance votes
during runtime/system suspend just makes the entire situation even more
confusing.

> >
> > What happens if a driver calls dev_pm_genpd_set_performance_state(...)
> > while the device is suspended? Will that mess up the performance state
> > when the device resumes?
> 
> Good question. The idea is:
> 
> If genpd in genpd_runtime_suspend() are able to drop an existing vote
> for a performance state, it should restore the vote in
> genpd_runtime_resume(). This also means, if there is no vote to drop
> in genpd_runtime_suspend(), genpd should just leave the vote as is in
> genpd_runtime_resume().
> 

But the next time the device enters runtime suspend that vote would be
dropped, wouldn't it? That feels kind of strange to me.

Stephan