Re: [PATCH v1] clk: Add devm_clk_{prepare,enable,prepare_enable}

[Geert Uytterhoeven]

Hi Marc,

On Mon, Nov 25, 2019 at 3:11 PM Marc Gonzalez <marc.w.gonzalez@xxxxxxx> wrote:
> On 25/11/2019 14:37, Russell King - ARM Linux admin wrote:
> > On Mon, Nov 25, 2019 at 02:10:21PM +0100, Marc Gonzalez wrote:
> >> On 25/11/2019 13:55, Russell King - ARM Linux admin wrote:
> >>> It's also worth reading https://lore.kernel.org/patchwork/patch/755667/
> >>> and considering whether you really are using the clk_prepare() and
> >>> clk_enable() APIs correctly.  Wanting these devm functions suggests
> >>> you aren't...
> >>
> >> In that older thread, you wrote:
> >>
> >>> If you take the view that trying to keep clocks disabled is a good way
> >>> to save power, then you'd have the clk_prepare() or maybe
> >>> clk_prepare_enable() in your run-time PM resume handler, or maybe even
> >>> deeper in the driver... the original design goal of the clk API was to
> >>> allow power saving and clock control.
> >>>
> >>> With that in mind, getting and enabling the clock together in the
> >>> probe function didn't make sense.
> >>>
> >>> I feel that aspect has been somewhat lost, and people now regard much
> >>> of the clk API as a bit of a probe-time nuisance.
> >>
> >> In the few drivers I've written, I call clk_prepare_enable() at probe.
> >
> > Right, so the clocks are enabled as soon as the device is probed,
> > in other words at boot time. It remains enabled for as long as the
> > device is bound to its driver, whether or not the device is actually
> > being used. Every switching edge causes heat to be generated. Every
> > switching edge causes energy to be wasted.
> >
> > That's fine if you have an infinite energy supply. That hasn't been
> > discovered yet.
> >
> > Given the prevalence of technology, don't you think we should be
> > doing as much as we possibly can to reduce the energy consumption
> > of the devices we use? It may be peanuts per device, but at scale
> > it all adds up.
>
> OK, I'm starting to see the bigger picture.
>
> (To provide some rationale for the patch, I think devm is a huge
> improvement for probe error-handling, and I did not understand
> why every driver must do manual error-handling when dealing with
> clocks in probe.)
>
> I did envision kernel modules being loaded on an as-needed basis,
> somewhat side-stepping the energy-waste issue you point out.
> But I realize that such a use-case may be uncommon. (Especially
> due to module auto-loading.)
>
> A few months ago, I was discussing a similar issue with GKH:
> Consider a device with a "START" register. Basically, if we write 0,
> the device turns itself off; if we write 1, it runs as configured.
>
> I was trying to start the device only when at least one user had
> it "open". So I used reference counting, and started the device
> on 0->1 open transitions, and stopped the device on 1->0 close
> transitions. GKH told me that was the wrong way to do it, and IIRC
> suggested to start the device in probe.
>
> I probably misunderstood Greg's suggestion. Where is the right place
> to start/stop a device (or gate its clocks)?

In the device driver's Runtime PM callbacks?
In the Power/Clock Domain Controller driver?

See drivers/base/power/domain.c:genpd_{start,stop}_dev(), and how/when
it's called.

Embedded device driver writers typically care.
Server device driver writes typically don't.

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds