Re: [PATCH v6 00/12] Unregister critical branch clocks + some RPM

[Konrad Dybcio]

On 1/24/24 08:41, Abel Vesa wrote:
> On 24-01-13 15:50:49, Konrad Dybcio wrote:
> > On Qualcomm SoCs, certain branch clocks either need to be always-on, or
> > should be if you're interested in touching some part of the hardware.
> >
> > Using CLK_IS_CRITICAL for this purpose sounds like a genius idea,
> > however that messes with the runtime pm handling - if a clock is
> > marked as such, the clock controller device will never enter the
> > "suspended" state, leaving the associated resources online, which in
> > turn breaks SoC-wide suspend.
>
> Generally speaking, HW-wise, if the power domain of a clock controller
> is being disabled, all clocks that it provides are being disabled.

Generally speaking, if that's the case, that's true.

> Are you saying that is not the case ?

Dragons however, are peculiar creatures and it seems like the clock
controllers are not *really* disabled when we think they are,
e.g. due to RPM(h) owning a share of GCC clocks, or due to the
MX rail being always-on. It would indeed be an issue with
hibernation where the registers would need to be reprogrammed
after battery power is removed.

As we spoke off-list, I'll split this series into two: adding
common helpers and then taking care of 2290/6375/6115.

I'm not yet sure how far we can go with converting existing clock
drivers to use pm_clk_add so that the _AHB, _XO, and _SLEEP clocks
for a given subsystem are only enabled when necessary - if we
require an entry in clock-names, backwards compatibility goes away,
and if we don't - we potentially miss out on a devlink between X_CC
and GCC, plus the name needs to be hardcoded for global parent lookup.

For new drivers, we'll likely just require a clean solution (runtime
PM enabled + subsys clocks gotten as pm_clk through a dt entry on
the consumer).


Konrad