Re: [PATCH v6 0/6] clk: qcom: use power-domain for sm8250's clock controllers

From: Dmitry Baryshkov
Date: Wed Sep 08 2021 - 04:50:50 EST


On Tue, 7 Sept 2021 at 17:34, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
>
> On Sun, 29 Aug 2021 at 17:54, Dmitry Baryshkov
> <dmitry.baryshkov@xxxxxxxxxx> wrote:
> >
> > On Sun, 29 Aug 2021 at 06:51, Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
> > >
> > > Quoting Dmitry Baryshkov (2021-08-26 14:56:23)
> > > > On 26/08/2021 21:31, Stephen Boyd wrote:
> > > > > Quoting Dmitry Baryshkov (2021-07-27 13:19:56)
> > > > >> On SM8250 both the display and video clock controllers are powered up by
> > > > >> the MMCX power domain. Handle this by linking clock controllers to the
> > > > >> proper power domain, and using runtime power management to enable and
> > > > >> disable the MMCX power domain.
> > > > >>
> > > > >> Dependencies:
> > > > >> - https://lore.kernel.org/linux-arm-msm/20210703005416.2668319-1-bjorn.andersson@xxxxxxxxxx/
> > > > >> (pending)
> > > > >
> > > > > Does this patch series need to go through the qcom tree? Presumably the
> > > > > dependency is going through qcom -> arm-soc
> > > >
> > > > It looks like Bjorn did not apply his patches in the for-5.15 series, so
> > > > we'd have to wait anyway. Probably I should rebase these patches instead
> > > > on Rajendra's required-opps patch (which is going in this window).
> > > >
> > >
> > > Ok. Thanks. I'll drop it from my queue for now.
> >
> > Just for the reference. I've sent v7 of this patchset. After thinking
> > more about power domains relationship, I think we have a hole in the
> > abstraction here. Currently subdomains cause power domains to be
> > powered up, but do not dictate the performance level the parent domain
> > should be working in.
>
> That's not entirely true. In genpd_add_subdomain() we verify that if
> the child is powered on, the parent must already be powered on,
> otherwise we treat this a bad setup and return an error code.
>
> What seems to be missing though, is that if there is a performance
> state applied for the child domain, that should be propagated to the
> parent domain too. Right?
>
> > While this does not look like an issue for the
> > gdsc (and thus it can be easily solved by the Bjorn's patches, which
> > enforce rpmhpd to be powered on to 'at least lowest possible'
> > performance state, this might be not the case for the future links. I
> > think at some point the pd_add_subdomain() interface should be
> > extended with the ability to specify minimum required performance
> > state when the link becomes on.
>
> I guess that minimum performance state could be considered as a
> "required-opp" in the DT node for the power-domain provider, no?

Yes, up to some point. But this enforces a particular driver code
(that I've had to change from v6 to v7).

In v6 the gdsc's power_on code would pm_runtime_get() the provider
device, power on the domain and the pm_runtime_put() the provider
device. Thus the gdsc genpd would be powered on (keeping parent
domains in the on state), but the provider device itself would be
runtime-suspended (neat idea by Bjorn). However this relied on changes
in rpmhpd behaviour (which still did not make it to linux-next).

In v7 we have to keep the provider device in resumed state while the
gdsc genpd is powered on (to keep the required-opps vote in place).

I suppose that 'child requires minimum parent's performance state'
might become common property at some point, allowing us to drop this
pm_runtime handling.

> Another option would simply be to manage this solely in the
> platform/soc specific genpd provider. Would that work?

Yes, I've had this in the very old iteration of mmcx fixup patchset
(even before mmcx-regulator came into play). It ended up with quite an
ugly piece of code.

>
> > Until that time I have changed code to
> > enforce having clock controller in pm resume state when gdsc is
> > enabled, thus CC itself votes on parent's (rpmhpd) performance state.
> >
> >
> > --
> > With best wishes
> > Dmitry
>
> Kind regards
> Uffe



--
With best wishes
Dmitry