Re: [PATCH] PM: domains: Don't attach a device to genpd that corresponds to a provider

[Bjorn Andersson]

On Fri 09 Jul 08:58 CDT 2021, Ulf Hansson wrote:

> On Fri, 9 Jul 2021 at 15:47, Bjorn Andersson <bjorn.andersson@xxxxxxxxxx> wrote:
> >
> > On Fri 09 Jul 08:22 CDT 2021, Ulf Hansson wrote:
> >
> > > On Fri, 9 Jul 2021 at 15:07, Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > > >
> > > > Hi Ulf,
> > > >
> > > > Thanks for your patch!
> > > >
> > > > On Fri, Jul 9, 2021 at 2:56 PM Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> > > > > According to the common power domain DT bindings, a power domain provider
> > > > > must have a "#power-domain-cells" property in its OF node. Additionally, if
> > > > > a provider has a "power-domains" property, it means that it has a parent
> > > > > domain.
> > > >
> > > > OK.
> > > >
> > > > > It has turned out that some OF nodes that represents a genpd provider may
> > > > > also be compatible with a regular platform device. This leads to, during
> > > > > probe, genpd_dev_pm_attach(), genpd_dev_pm_attach_by_name() and
> > > > > genpd_dev_pm_attach_by_id() tries to attach the corresponding struct device
> > > > > to the genpd provider's parent domain, which is wrong. Instead the genpd
> > > >
> > > > Why is that wrong?
> > >
> > > It may lead to that the struct device that corresponds to a genpd
> > > provider may be attached to the parent domain. In other words, the
> > > parent domain will not only be controlled by a child domain
> > > (corresponding to the provider), but also through the provider's
> > > struct device. As far as I can tell, this has never been the intent
> > > for how things should work in genpd.
> > >
> > > So wrong or not, I guess it depends on what you expect to happen.
> > >
> > > Do you see an issue with changing this?
> > >
> >
> > But this exactly what we have in the case of the "dispcc" in the
> > Qualcomm platform that Dmitry is working on.
> >
> > The provider driver needs the parent power-domain to be powered in order
> > to poke the registers and then it is the parent of the power-domains
> > exposed.
> >
> > If I understand your proposed patch we'll have to manually attach the
> > parent domain to the struct device of the controller with this patch?
> 
> Not even that would work after $subject patch, as it prevents
> providers from being attached to a domain.
> 

That's definitely going to be a problem.

> It sure sounds like you need to control power for the parent domain,
> not only by registering a child domain to it.
> 

Yes, we certainly need power to the genpd provider.

> >
> > Is the Qualcomm case unique or will this change cut power do other genpd
> > providers assuming the same?
> 
> I think the Qualcomm case is a bit unique or at least the first I
> heard of. However, this change would affect all and of course we must
> not break things.
> 

I'm surprised that we'd be alone one that needs power to our genpd
provider. Does everyone else have their genpd providers in some
always-on power domain?

> >
> >
> >
> > Worth mentioning as we discuss this is that we have another genpd
> > provider, where I think the exposed genpds are parented by a few
> > different (each one with a specific) parent domains. In this case we'd
> > be forced to manually attach the genpd provider to the parent domain
> > that it actually is powered by (as no automatic attachment happens when
> > multiple domains are specified).
> 
> Yes, that's correct (assuming we don't apply $subject patch).
> 

Afaict this patch wouldn't change the case where the genpd provider has
multiple power-domains, as it wouldn't automatically attach the device
to any one of them anyways.

Regards,
Bjorn

> To sum up:
> 
> Rafael I am withdrawing the $subject patch, it seems like it may break
> existing expectations of what will happen during attach.
> 
> Moreover, it may actually be beneficial to allow the attach to succeed
> for the Qcom case, so let's leave this as is.
> 
> Kind regards
> Uffe