Re: [PATCH 2/2] PM: domain: use per-genpd lockdep class

From: Ulf Hansson
Date: Thu Jul 01 2021 - 11:14:49 EST


On Thu, 1 Jul 2021 at 13:01, Dmitry Baryshkov
<dmitry.baryshkov@xxxxxxxxxx> wrote:
>
> On Thu, 1 Jul 2021 at 13:07, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> >
> > On Tue, 29 Jun 2021 at 17:09, Bjorn Andersson
> > <bjorn.andersson@xxxxxxxxxx> wrote:
> > >
> > > On Mon 28 Jun 14:55 CDT 2021, Dmitry Baryshkov wrote:
> > >
> > > > Hi,
> > > >
> > > > On 17/06/2021 12:07, Ulf Hansson wrote:
> > > > > + Rajendra
> > > > >
> > > > > On Tue, 15 Jun 2021 at 17:55, Bjorn Andersson
> > > > > <bjorn.andersson@xxxxxxxxxx> wrote:
> > > [..]
> > > > > > But I am unable to find a way for the gdsc driver to get hold of the
> > > > > > struct generic_pm_domain of the resources exposed by the rpmhpd driver.
> > > > >
> > > > > You don't need a handle to the struct generic_pm_domain, to assign a
> > > > > parent/child domain. Instead you can use of_genpd_add_subdomain(),
> > > > > which takes two "struct of_phandle_args*" corresponding to the
> > > > > parent/child device nodes of the genpd providers and then let genpd
> > > > > internally do the look up.
> > > >
> > > [..]
> > > >
> > > > I think I'd need this function anyway for the gdsc code. During gdsc_init()
> > > > we check gdsc status and this requires register access (and thus powering on
> > > > the parent domain) before the gdsc is registered itself as a power domain.
> > > >
> > >
> > > But this is a register access in the dispcc block, which is the context
> > > that our gdsc_init() operates. So describing that MMCX is the
> > > power-domain for dispcc should ensure that the power-domain is enabled.
> >
> > Right.
> >
> > As a note, when we assign a child domain to a parent domain, via
> > of_genpd_add_subdomain() for example - and the child domain has been
> > powered on, this requires the parent domain to be turned on as well.
>
> Most probably we should also teach genpd code to call pm_runtime
> functions on respective devices when the genpd is powered on or off.
> For now I had to do this manually.

No, that's not the way it works or should work for that matter.

It's the runtime PM status of the devices that are attached to a
genpd, that controls whether a genpd should be powered on/off.
Additionally, if there is a child domain powered on, then its parent
needs to be powered on too and so forth.

>
> >
> > >
> > > We do however need to make sure that dispcc doesn't hog its
> > > power-domain, and that any register accesses in runtime is done with the
> > > parenting power-domain enabled. E.g. the clock framework wraps all
> > > operations in pm_runtime_get/put(), but I don't see anything in the
> > > gnepd code for this.
> > >
> > >
> > > And for gcc I'm worried that we might have some GDSCs that are parented
> > > by CX and some by MX, but I do still think that the register accesses
> > > are only related to one of these.
> > >
> > > But this seems like a continuation of the special case in dispcc, so I
> > > think we should be able to focus on getting that right before we attempt
> > > the general case (and I don't know if we have a need for this today).
> > >
> > > Regards,
> > > Bjorn
> >
> > Unfortunately, I didn't understand all the above things.
> >
> > In any case, please tell me if there is anything else that blocks you
> > from moving forward with the power domain conversion? I am happy to
> > help.
>
> Patch series was submitted:
> https://lore.kernel.org/linux-arm-msm/20210630133149.3204290-1-dmitry.baryshkov@xxxxxxxxxx/

Okay, I will have a look over there. Thanks!

Kind regards
Uffe