Re: [PATCH v4] clk: add duty cycle support

From: Geert Uytterhoeven
Date: Tue Jul 03 2018 - 06:03:11 EST


Hi Jerome,

On Tue, Jul 3, 2018 at 11:58 AM Jerome Brunet <jbrunet@xxxxxxxxxxxx> wrote:
> On Tue, 2018-07-03 at 11:27 +0200, Geert Uytterhoeven wrote:
> > On Tue, Jun 19, 2018 at 4:42 PM Jerome Brunet <jbrunet@xxxxxxxxxxxx> wrote:
> > > Add the possibility to apply and query the clock signal duty cycle ratio.
> > >
> > > This is useful when the duty cycle of the clock signal depends on some
> > > other parameters controlled by the clock framework.
> > >
> > > For example, the duty cycle of a divider may depends on the raw divider
> > > setting (ratio = N / div) , which is controlled by the CCF. In such case,
> > > going through the pwm framework to control the duty cycle ratio of this
> > > clock would be a burden.
> > >
> > > A clock provider is not required to implement the operation to set and get
> > > the duty cycle. If it does not implement .get_duty_cycle(), the ratio is
> > > assumed to be 50%.
> > >
> > > This change also adds a new flag, CLK_DUTY_CYCLE_PARENT. This flag should
> > > be used to indicate that a clock, such as gates and muxes, may inherit
> > > the duty cycle ratio of its parent clock. If a clock does not provide a
> > > get_duty_cycle() callback and has CLK_DUTY_CYCLE_PARENT, then the call
> > > will be directly forwarded to its parent clock, if any. For
> > > set_duty_cycle(), the clock should also have CLK_SET_RATE_PARENT for the
> > > call to be forwarded
> > >
> > > Signed-off-by: Jerome Brunet <jbrunet@xxxxxxxxxxxx>
> >
> > Thanks for your patch!
> >
> > > ---
> > > The series has been developed to handled the sample clocks provided by
> > > audio clock controller of amlogic's A113 SoC. To support i2s modes, this
> > > clock need to have a 50% duty cycle ratio, while it should be just one
> > > pulse of the parent clock in dsp modes.
> >
> > "one pulse" means num = 1, den = the clock rate, right?
>
> No, it would be num = 1, den = divider

Right, thanks for correcting me!

> > > --- a/include/linux/clk-provider.h
> > > +++ b/include/linux/clk-provider.h
> > > @@ -66,6 +68,17 @@ struct clk_rate_request {
> > > struct clk_hw *best_parent_hw;
> > > };
> > >
> > > +/**
> > > + * struct clk_duty - Struture encoding the duty cycle ratio of a clock
> > > + *
> > > + * @num: Numerator of the duty cycle ratio
> > > + * @den: Denominator of the duty cycle ratio
> > > + */
> > > +struct clk_duty {
> > > + unsigned int num;
> > > + unsigned int den;
> >
> > So shouldn't both fields be "unsigned long" instead, to match clock rates?
> > (Yes, I do know we don't support +4.3 GHz clock rates on 32-bit yet ;-)
>
> Not sure we need to match clock rates, long seems a bit too much.
> In the end, all we want a ratio, so a [0 - 1] number. Fraction using unsigned
> int already provide a pretty good precision (around 0.0002 ppm with 32bit)
>
> Do you have a use case where you need more than that ?

No, if den = divider, "unsigned int" is fine.
I wrongly assumed it could be equal to the clock rate.

> > Also, you may want to have a higher precision than degrees for the
> > phase property when handling pulses.
>
> Is this comment related to this patch ?

It's something to consider for the future, in case den > 360,
Probably its rare for divider values to be that large, though.
Again, I wrongly assumed den could be equal to the clock rate.

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