Re: [PATCH v8 01/34] opp: Add dev_pm_opp_sync() helper

[Ulf Hansson <ulf.hansson@xxxxxxxxxx>]

On Fri, 20 Aug 2021 at 07:18, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
>
> On 19-08-21, 16:55, Ulf Hansson wrote:
> > Right, that sounds reasonable.
> >
> > We already have pm_genpd_opp_to_performance_state() which translates
> > an OPP to a performance state. This function invokes the
> > ->opp_to_performance_state() for a genpd. Maybe we need to allow a
> > genpd to not have ->opp_to_performance_state() callback assigned
> > though, but continue up in the hierarchy to see if the parent has the
> > callback assigned, to make this work for Tegra?
> >
> > Perhaps we should add an API dev_pm_genpd_opp_to_performance_state(),
> > allowing us to pass the device instead of the genpd. But that's a
> > minor thing.
>
> I am not concerned a lot about how it gets implemented, and am not
> sure as well, as I haven't looked into these details since sometime.
> Any reasonable thing will be accepted, as simple as that.
>
> > Finally, the precondition to use the above, is to first get a handle
> > to an OPP table. This is where I am struggling to find a generic
> > solution, because I guess that would be platform or even consumer
> > driver specific for how to do this. And at what point should we do
> > this?
>
> Hmm, I am not very clear with the whole picture at this point of time.
>
> Dmitry, can you try to frame a sequence of events/calls/etc that will
> define what kind of devices we are looking at here, and how this can
> be made to work ?
>
> > > > Viresh, please take a look at what I did in [1]. Maybe it could be done
> > > > in another way.
> > >
> > > I looked into this and looked like too much trouble. The
> > > implementation needs to be simple. I am not sure I understand all the
> > > problems you faced while doing that, would be better to start with a
> > > simpler implementation of get_performance_state() kind of API for
> > > genpd, after the domain is attached and its OPP table is initialized.
> > >
> > > Note, that the OPP table isn't required to be fully initialized for
> > > the device at this point, we can parse the DT as well if needed be.
> >
> > Sure, but as I indicated above, you need some kind of input data to
> > figure out what OPP table to pick, before you can translate that into
> > a performance state. Is that always the clock rate, for example?
>
> Eventually it can be clock, bandwidth, or pstate of anther genpd, not
> sure what all we are looking for now. It should be just clock right
> now as far as I can imagine :)
>
> > Perhaps, we should start with adding a dev_pm_opp_get_from_rate() or
> > what do you think? Do you have other suggestions?
>
> We already have similar APIs, so that won't be a problem. We also have
> a mechanism inside the OPP core, frequency based, which is used to
> guess the current OPP. Maybe we can enhance and use that directly
> here.

After reading the last reply from Dmitry, I am starting to think that
the problem he is facing can be described and solved in a much easier
way.

If I am correct, it looks like we don't need to add APIs to get OPPs
for a clock rate or set initial performance state values according to
the HW in genpd.

See my other response to Dmitry, let's see where that leads us.

Kind regards
Uffe