Re: (EXT) Re: [PATCH 1/4] pwm: pca9685: remove unused duty_cycle struct element

[Thierry Reding]

On Sat, Apr 04, 2020 at 04:17:00PM -0400, Sven Van Asbroeck wrote:
> On Sat, Apr 4, 2020 at 1:35 PM Clemens Gruber
> <clemens.gruber@xxxxxxxxxxxx> wrote:
> >
> > As the user is setting the duty cycle in nanoseconds, it makes sense
> > that the relative duty cycle decreases in an absolute period increase.
> > As for the behavior that the other channels remain at the same relative
> > duty cycle: Not sure how we can avoid this, other than reprogramming all
> > 15 other channels if one of them is changed and that's not really
> > acceptable, I think.
> 
> Thank you for the explanation, Clemens.
> 
> Yes, it does make sense that the relative duty cycle changes when we change
> the period. A relative duty cycle of duty_cycle / period is what the user would
> expect to see.
> 
> It also kind-of makes sense that the relative duty cycles of the other
> pwm channels
> do not change: after all, the user is not touching them, so would not expect
> them to change.
> 
> However, the following does not make sense to me. Imagine pwm0 and pwm1
> are both active and at 50%: period=5000000, duty_cycle=2500000. Then, change
> the period on pwm0:
> 
> $ echo 10000000 > pwm0/period
> 
> Then pwm0 gets dimmer (makes sense) and pwm1 keeps the same relative duty
> cycle (makes sense). However, if we now read out sysfs for pwm1, we get:
> 
> $ echo pwm1/period
> 5000000 (wrong!)
> $ echo pwm1/duty_cycle
> 2500000 (wrong! although relative duty cycle is correct)
> 
> Although the pwm1 period has changed, the API calls do not reflect this.
> This makes it next to impossible for users to know what the current period
> is set to.
> 
> Moving to the atomic API won't help, because .get_state is called only
> once, when the chip is registered.

The .get_state() callback is actually called when the PWM is requested,
which could be much later than when the chip is registered. That doesn't
change the fact that it would be useless for this case, though.

> It does look like we have a square peg (this chip) in a round hole (the
> standard assumptions the pwm core makes) ?

There are other chips where a single period is shared across multiple
PWM channels. Typically what we do there is once a period is configured
for a given channel, all subsequent PWM channel configurations must use
the same period, or otherwise the driver will return an error code.

See for example:

  - stm32_pwm_config() in drivers/pwm/pwm-stm32.c
  - lpc18xx_pwm_config() in drivers/pwm/pwm-lpc18xx-sct.c
  - pwm_imx_tpm_apply_hw() in drivers/pwm/pwm-imx-tpm.c
  - fsl_pwm_apply_config() in drivers/pwm/pwm-fsl-ftm.c

The rationale behind that is that we must not change a PWM configuration
without a consumer having explicitly requested it.

It seems like PCA9685 is somewhere halfway between in that it will
actually keep the same duty-cycle/period ratio, but that doesn't mean it
is automatically okay to do this. The problem is that the duty-cycle to
period ratio is only relevant in some cases. If all you care about is
the power output of the PWM signal, which admittedly seems to be about
95% of all cases, then yes, this behaviour would be okay. But what if we
have a consumer that relies on a particular width of the PWM pulse in
absolute terms rather than relative to the period?

Thierry

Attachment: signature.asc
Description: PGP signature