Re: [PATCH v8 06/34] dt-bindings: clock: tegra-car: Document new tegra-clocks sub-node

[Dmitry Osipenko <digetx@xxxxxxxxx>]

19.08.2021 19:31, Thierry Reding пишет:
>> The "device" representation is internal to the kernel. It's okay to me
>> to have PLLs represented by a device, it's a distinct h/w by itself.
>>
>> CCF supports managing of clock's RPM and it requires to have clock to be
>> backed by a device. That's what we are using here.
>>
>> Please see
>> https://elixir.bootlin.com/linux/v5.14-rc6/source/drivers/clk/clk.c#L109
> Looking at the implementation of __clk_register() and where that device
> pointer typically comes from, I don't think the way this is used here is
> what was intended. The way I interpret the code is that a clock is
> registered with a parent device (i.e. its provider) and
> clk_pm_runtime_get() is then used internally as a way to make sure that
> when a clock is prepared, it's parent device is runtime resumed. This is
> presumably to ensure that any registers that the driver might need to
> access in order to prepare and enable the clock are accessible (i.e. the
> CAR is not powered off or in reset).
> 
> So the struct device that is passed to __clk_register() (or its callers)
> should be that of the CAR rather than virtual struct devices created by
> the CAR.
> 
> And it's a bit debatable whether or not PLLs represent distinct
> hardware. Ultimately every transistor on a chip could be considered
> distinct hardware. But a platform device is a device on a platform bus,
> which is really just another way of saying it's a hardware block that's
> accessible from the CPU via a memory-mapped address. A PLL (just like
> other clocks) is merely a resource exposed by means of access to these
> registers. So I don't think they should be platform devices. Even making
> them struct device:s seems a bit of a stretch.
> 
> Is there any reason why struct clk can't be used for this? I mean, the
> whole purpose of that structure is to represent clocks. Why do we need
> to make them special?

Because we need to perform DVFS for PLLs. The only way to do it without
having to reinvent existing frameworks is to use these frameworks and
they require a device.