Re: [PATCH v2 2/3] cpufreq: qcom-nvmem: Enable virtual power domain devices

[Stephan Gerhold]

On Thu, Oct 19, 2023 at 01:26:19PM +0200, Ulf Hansson wrote:
> On Thu, 19 Oct 2023 at 12:24, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> > On Wed, 18 Oct 2023 at 10:06, Stephan Gerhold
> > <stephan.gerhold@xxxxxxxxxxxxxxx> wrote:
> > >
> > > The genpd core caches performance state votes from devices that are
> > > runtime suspended as of commit 3c5a272202c2 ("PM: domains: Improve
> > > runtime PM performance state handling"). They get applied once the
> > > device becomes active again.
> > >
> > > To attach the power domains needed by qcom-cpufreq-nvmem the OPP core
> > > calls genpd_dev_pm_attach_by_id(). This results in "virtual" dummy
> > > devices that use runtime PM only to control the enable and performance
> > > state for the attached power domain.
> > >
> > > However, at the moment nothing ever resumes the virtual devices created
> > > for qcom-cpufreq-nvmem. They remain permanently runtime suspended. This
> > > means that performance state votes made during cpufreq scaling get
> > > always cached and never applied to the hardware.
> > >
> > > Fix this by enabling the devices after attaching them and use
> > > dev_pm_syscore_device() to ensure the power domains also stay on when
> > > going to suspend. Since it supplies the CPU we can never turn it off
> > > from Linux. There are other mechanisms to turn it off when needed,
> > > usually in the RPM firmware (RPMPD) or the cpuidle path (CPR genpd).
> >
> > I believe we discussed using dev_pm_syscore_device() for the previous
> > version. It's not intended to be used for things like the above.
> >

Sorry, looks like we still had a misunderstanding in the conclusion of
the previous discussion. :')

> > Moreover, I was under the impression that it wasn't really needed. In
> > fact, I would think that this actually breaks things for system
> > suspend/resume, as in this case the cpr driver's genpd
> > ->power_on|off() callbacks are no longer getting called due this,
> > which means that the cpr state machine isn't going to be restored
> > properly. Or did I get this wrong?
> 

We strictly need the RPMPDs to be always-on, also across system suspend
[1]. The RPM firmware will drop the votes internally as soon as the
CPU(s) have entered deep cpuidle. We can't do this from Linux, because
we need the CPU to continue running until it was shut down cleanly.

For CPR, we strictly need the backing regulator to be always-on, also
across system suspend. Typically the hardware will turn off the
regulator as soon as the CPU(s) enter deep cpuidle. Similarly, we can't
do this from Linux, because we need the CPU to continue running until it
was shut down cleanly.

My understanding was that we're going to pause the CPR state machine
using the system suspend/resume callbacks on the driver, instead of
using the genpd->power_on|off() callbacks [2]. I can submit a separate
patch for this.

I didn't prioritize this because QCS404 (as the only current user of
CPR) doesn't have proper deep cpuidle/power management set up yet. It's
not entirely clear to me if there is any advantage (or perhaps even
disadvantage) if we pause the CPR state machine while the shared L2
cache is still being actively powered by the CPR power rail during
system suspend. I suspect this is a configuration that was never
considered in the hardware design.

Given the strict requirement for the RPMPDs, I only see two options:

 1. Have an always-on consumer that prevents the power domains to be
    powered off during system suspend. This is what this patch tries to
    achieve.

Or:

 2. Come up with a way to register the RPMPDs used by the CPU with
    GENPD_FLAG_ALWAYS_ON. This would also be doable, but isn't as
    straightfoward as "regulator-always-on" in the DT because the rpmpd
    DT node represents multiple genpds in a single DT node [3].

What do you think? Do you see some other solution perhaps? I hope we can
clear up the misunderstanding. :-)

[1]: https://lore.kernel.org/linux-arm-msm/ZQGqfMigCFZP_HLA@xxxxxxxxxxx/
[2]: https://lore.kernel.org/linux-arm-msm/CAPDyKFoiup8KNv=1LFGKDdDLA1pHsdJUgTTWMdgxnikEmReXzg@xxxxxxxxxxxxxx/
[3]: https://lore.kernel.org/linux-arm-msm/ZSg-XtwMxg3_fWxc@xxxxxxxxxxx/

> BTW, if you really need something like the above, the proper way to do
> it would instead be to call device_set_awake_path() for the device.
> 
> This informs genpd that the device needs to stay powered-on during
> system suspend (assuming that GENPD_FLAG_ACTIVE_WAKEUP has been set
> for it), hence it will keep the corresponding PM domain powered-on
> too.
> 

Thanks, I can try if this works as alternative to the
dev_pm_syscore_device()!

I will wait for your thoughts on the above before accidentally going
into the wrong direction again. :-)

Thanks!
Stephan