Re: [RFC PATCH v2 1/2] PM: domains: Skip disabling unused domains if provider has sync_state

[Matthias Kaehlcke <mka@chromium.org>]

On Mon, Feb 06, 2023 at 06:31:21PM +0200, Abel Vesa wrote:
> On 23-02-02 18:24:15, Matthias Kaehlcke wrote:
> > Hi Abel,
> > 
> > On Fri, Jan 27, 2023 at 12:40:53PM +0200, Abel Vesa wrote:
> > > Currently, there are cases when a domain needs to remain enabled until
> > > the consumer driver probes. Sometimes such consumer drivers may be built
> > > as modules. Since the genpd_power_off_unused is called too early for
> > > such consumer driver modules to get a chance to probe, the domain, since
> > > it is unused, will get disabled. On the other hand, the best time for
> > > an unused domain to be disabled is on the provider's sync_state
> > > callback. So, if the provider has registered a sync_state callback,
> > > assume the unused domains for that provider will be disabled on its
> > > sync_state callback. Also provide a generic sync_state callback which
> > > disables all the domains unused for the provider that registers it.
> > > 
> > > Signed-off-by: Abel Vesa <abel.vesa@linaro.org>
> > > ---
> > > 
> > > This approach has been applied for unused clocks as well.
> > > With this patch merged in, all the providers that have sync_state
> > > callback registered will leave the domains enabled unless the provider's
> > > sync_state callback explicitly disables them. So those providers will
> > > need to add the disabling part to their sync_state callback. On the
> > > other hand, the platforms that have cases where domains need to remain
> > > enabled (even if unused) until the consumer driver probes, will be able,
> > > with this patch in, to run without the pd_ignore_unused kernel argument,
> > > which seems to be the case for most Qualcomm platforms, at this moment.
> > 
> > I recently encountered a related issue on a Qualcomm platform with a
> > v6.2-rc kernel, which includes 3a39049f88e4 ("soc: qcom: rpmhpd: Use
> > highest corner until sync_state"). The issue involves a DT node with a
> > rpmhpd, the DT node is enabled, however the corresponding device driver
> > is not enabled in the kernel. In such a scenario the sync_state callback
> > is never called, because the genpd consumer never probes. As a result
> > the Always-on subsystem (AOSS) of the SoC doesn't enter sleep mode during
> > system suspend, which results in a substantially higher power consumption
> > in S3.
> 
> If I get this correctly, one of the providers is missing (doesn't matter
> the reason), in which case, your kernel needs that driver, period. There
> is no reason why you would expect the consumer to work without the
> provider. Or, you could just remove the property in the devicetree node,
> the property that makes the consumer wait for that provider. Anyway, you
> should never end up with a consumer provider relationship in devicetree
> without providing the provider driver.

I would agree if it was actually a provider that's missing, however it's a
'missing' consumer that prevents the sync_state() call.

> > I wonder if genpd (and some other frameworks) needs something like
> > regulator_init_complete(), which turns off unused regulators 30s after
> > system boot. That's conceptually similar to the current
> > genpd_power_off_unused(), but would provide time for modules being loaded.
> 
> NACK, timeouts are just another hack in this case, specially when we
> have a pretty reliable mechanism like sync_state.

It does not work properly unless all consumers are probed successfully. It
makes sense to wait some time for the consumers to probe, but not eternally,
it's perfectly valid that a driver for a (potential) consumer is not enabled.