Re: [PATCH v2 1/3] pwm: jz4740: Use clocks from TCU driver

From: Paul Cercueil
Date: Mon Nov 18 2019 - 08:43:06 EST

Hi Uwe,


Le lun., nov. 18, 2019 at 12:19, Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxxxx> a écrit :
Hello Paul,

On Mon, Nov 18, 2019 at 11:55:56AM +0100, Paul Cercueil wrote:
Le lun., nov. 18, 2019 at 08:15, Uwe Kleine-König
<u.kleine-koenig@xxxxxxxxxxxxxx> a écrit :
> On Sun, Nov 17, 2019 at 11:58:43PM +0100, Paul Cercueil wrote:
> > Le dim., nov. 17, 2019 at 21:20, Uwe Kleine-König
> > <u.kleine-koenig@xxxxxxxxxxxxxx> a écrit :
> > > On Sat, Nov 16, 2019 at 06:36:11PM +0100, Paul Cercueil wrote:
> > > > struct jz4740_pwm_chip {
> > > > struct pwm_chip chip;
> > > > - struct clk *clk;
> > >
> > > What is the motivation to go away from this approach to store the
> > clock?
> >
> > It's actually not the same clock. Instead of obtaining "ext" clock
> > from the
> > probe, we obtain "timerX" clocks (X being the PWM channel) from the
> > request
> > callback.
>
> Before you used driver data and container_of to get it, now you used
> pwm_set_chip_data. I wondered why you changed the approach to store
> data. That the actual data is different now is another thing (and
> obviously ok).

Thierry suggested it: https://lkml.org/lkml/2019/3/4/486

If you motivate that in the commit log (preferably with a better
rationale than "Thierry suggested it") that's fine for. (Do I claim now
without having read the rationale :-)

I don't really have a better rationale. The alternative was to have a "struct clk[NB_PWMS];" in the struct jz4740_pwm_chip, so this is arguably better. I'm not sure it's worth mentioning in the commit message, is it?


> > > > static void jz4740_pwm_free(struct pwm_chip *chip, struct pwm_device *pwm)
> > > > {
> > > > + struct clk *clk = pwm_get_chip_data(pwm);
> > > > +
> > > > jz4740_timer_set_ctrl(pwm->hwpwm, 0);
> > >
> > > What is the purpose of this call? I would have expected that all these
> > > would go away when converting to the clk stuff?!
> >
> > Some go away in patch [1/3] as they are clock-related, this one will go away
> > in patch [2/3] when the driver is converted to use regmap.
>
> I'd like to understand what it does. Judging from the name I expect this
> is somehow related to the clock stuff and so I wonder if the conversion
> to the clk API is as complete as it should be.

It clears the PWM channel's CTRL register. That's the register used for
instance to enable the PWM function of a TCU channel.

OK, so this is a register in a different register range than the PWM
related registers to set duty and period, right? Looking at the code,
this register has a bit to enable PWM mode and other than that bit
fields to tune the clock feeding the PWM counters, right?

They are actually all in the same register range. Each channel has 4 32-bit registers, the first one is the CTRL (aka. TCSR) register which is written to here. The following two configure the duty/period values, the last one is the counter. The 'timer enable' bit is however in the global TCU registers area.

The clock bits of the TCSRs registers (including the TCSR registers of the watchdog and 64-bit OS timer) are controlled by the clocks driver. All register accesses are properly handled thanks to regmap, that we add in patch [2/3].


This probably explains my resistance because such a setup if really hard
to map to nice code. At least the "PWM enable" bit doesn't fit the clk
abstraction, no good idea here. Maybe it's easier and more straight
forward to not wrap that register in a clock driver and only use a clk
for the parent? What is the motivation to convert this piece of hardware
to a clk driver? Or abstract it as a proper clk and provide a function
to enable PWM mode for channel X?

The motivation behind converting it to a clocks driver, is that it's not only used for PWM, but for system timers, the watchdog, and the 64-bit OS timer.

There is some overview of the TCU hardware here: linux/Documentation/mips/ingenic-tcu.rst (and yes, that hardware is a mess).

All the bits have been merged or accepted upstream minus the OST driver which is still under review, and this patchset.


> > > > - jz4740_timer_stop(pwm->hwpwm);
> > > > + clk_disable_unprepare(clk);
> > > > + clk_put(clk);
> > > > }
> > > >
> > > > static int jz4740_pwm_enable(struct pwm_chip *chip, struct
> > pwm_device *pwm)
> > > > @@ -91,17 +110,21 @@ static int jz4740_pwm_apply(struct
> > pwm_chip *chip, struct pwm_device *pwm,
> > > > const struct pwm_state *state)
> > > > {
> > > > struct jz4740_pwm_chip *jz4740 = to_jz4740(pwm->chip);
> > > > + struct clk *clk = pwm_get_chip_data(pwm),
> > > > + *parent_clk = clk_get_parent(clk);
> > > > + unsigned long rate, period, duty;
> > > > unsigned long long tmp;
> > > > - unsigned long period, duty;
> > > > unsigned int prescaler = 0;
> > > > uint16_t ctrl;
> > > >
> > > > - tmp = (unsigned long long)clk_get_rate(jz4740->clk) *
> > state->period;
> > > > + rate = clk_get_rate(parent_clk);
> > >
> > > Why is it the parent's rate that is relevant here?
> >
> > We calculate the divider to be used for the "timerX" clock, so we
> > need to
> > know the parent clock.
>
> Then the approach here is wrong. You should not assume anything about
> the internal details of the clock, that's the task of the clock driver.
> As a consumer of the clock just request a rate (or use clk_round_rate to
> find a good setting first) and use that.

Totally agreed. I wanted to do that, but you were fighting tooth and nails
against my patch "Improve algorithm of clock calculation", remember?

No, I don't, but I looked that up :-) And I fighted because I thought
the clk API isn't used properly (and I think your problem is that the
clk API as is today doesn't give you what you want, so there is more
work to do on the clk side of the problem).

That's a question of point of view, really. I need to specify the maximum clock rate that still gives me a valid value, the clk API gives me clk_set_max_rate(). Doesn't look like I'm doing something out of bounds.

The conceptual problem I see is that currently the code uses some
internal knowledge about how this timer clock works but as soon as you
use the clk abstraction it's wrong to use such internal knowledge.

Well, this patch is still a step in the right direction. I too would like to see a better solution, but I don't want to hold it until we fix the problem mentioned above. Right now jz4740-pwm is broken upstream (incompatible with the documented DT bindings) and that's something I want fixed ASAP, hence this reduced patchset.

Best regards,
-Paul