Re: [RFC v2 3/5] clk: bcm2835: use firmware interface to update pllb

From: Stefan Wahren
Date: Tue May 21 2019 - 17:46:49 EST



> Nicolas Saenz Julienne <nsaenzjulienne@xxxxxxx> hat am 21. Mai 2019 um 17:47 geschrieben:
>
>
> Hi Stefan, thanks for your comments!
>
> On Tue, 2019-05-21 at 14:40 +0200, Stefan Wahren wrote:
> > Hi Nicolas,
> >
> > On 20.05.19 14:11, Stefan Wahren wrote:
> > > Hi Nicolas,
> > >
> > > the following comments applies only in case Eric is fine with the whole
> > > approach.
> > >
> > > On 20.05.19 12:47, Nicolas Saenz Julienne wrote:
> > > > Raspberry Pi's firmware, which runs in a dedicated processor, keeps
> > > maybe we should clarify that the firmware is running in the VPU
> > > > track of the board's temperature and voltage. It's resposible for
> > > > scaling the CPU frequency whenever it deems the device reached an unsafe
> > > > state. On top of that the firmware provides an interface which allows
> > > > Linux to to query the clock's state or change it's frequency.
> > > I think this requires a separate update of the devicetree binding.
> > > > Being the sole user of the bcm2835 clock driver, this integrates the
> > > > firmware interface into the clock driver and adds a first user: the CPU
> > > > pll, also known as 'pllb'.
> > > Please verify that the kernel still works (and this clock driver probe)
> > > under the following conditions:
> > >
> > > - CONFIG_RASPBERRYPI_FIRMWARE=n
> > > - CONFIG_RASPBERRYPI_FIRMWARE=m
> > > - older DTBs without patch #1
> > i thought about this and the case this driver would return
> > -EPROBE_DEFER. The clock driver is too essential for doing such a thing.
> > So i think the best solution would be to move these changes into a
> > separate driver which should be register by the clock driver (similiar
> > to vchiq). This also avoid the need of a new device tree binding.
>
> I understand your concerns.
>
> Wouldn't you prefer registering the device trough the device tree? I'd go with
> the same approach as the firmware touchscreen driver, which is registered after
> the firmware's probe trough dt's 'simple-bus'. That said, it's not a strongly
> held opinion, I'm happy with whatever solution as long as it works.

A devicetree binding always introduce some kind of inflexibility. In case someone finds a better solution later things can get really messy. A recent example is the clock handling for i2c-bcm2835.

>
> I get from your comments that you'd like the register based version of 'pllb'
> and 'pllb_arm' to be loaded if for some reason the firmware isn't available. Is
> that right?

This wasn't my intention. I would prefer a simple approch here (no handover).

> The main problem I see with this is the duplication of 'pllb' and
> 'pllb_arm'. Both drivers will create the same clock device through different
> interfaces. Any suggestions on how to deal with that? If not I can simply
> remove 'pllb' and 'pllb_arm' from clk-bcm2835.c.

Yes. So even if this driver is disabled, there shouldn't be a regression. Or did i miss something?

>
> Regards,
> Nicolas
>