On Thu, Apr 06, 2023 at 04:30:23PM +0200, Thierry Reding wrote:
> On Thu, Apr 06, 2023 at 03:53:58PM +0200, Uwe Kleine-König wrote:
> > Hello Thierry,
> > 
> > On Thu, Apr 06, 2023 at 03:38:48PM +0200, Thierry Reding wrote:
> > > On Thu, Mar 09, 2023 at 02:04:10AM +0100, Lorenz Brun wrote:
> > > > +	 * appear to have the capability to invert the output.
> > > > +	 * This means that inverted mode can not be fully supported as the
> > > > +	 * waveform will always start with the low period and end with the high
> > > > +	 * period. Thus reject non-normal polarity if the shape of the waveform
> > > > +	 * matters, i.e. usage_power is not set.
> > > > +	 */
> > > > +	if (state->polarity != PWM_POLARITY_NORMAL && !state->usage_power)
> > > >  		return -EINVAL;
> > > >  
> > > >  	if (!state->enabled) {
> > > > @@ -213,7 +221,11 @@ static int pwm_mediatek_apply(struct pwm_chip *chip, struct pwm_device *pwm,
> > > >  		return 0;
> > > >  	}
> > > >  
> > > > -	err = pwm_mediatek_config(pwm->chip, pwm, state->duty_cycle, state->period);
> > > > +	duty_cycle = state->duty_cycle;
> > > > +	if (state->polarity == PWM_POLARITY_INVERSED)
> > > > +		duty_cycle = state->period - state->duty_cycle;
> > > 
> > > That's not really what state->usage_power was meant to address.
> > 
> > I don't understand your concern here. I don't like .usage_power, but
> > AFAICT this is a legitimite use. With .usage_power = true, the lowlevel
> > driver is free to shift the phase_offset and even modify the period size
> > and the goal is just that the average power-output matches.
> > 
> > Lorenz's patch does exactly this: It even keeps the period and only
> > shifts the phase (by period - duty_cycle). If you consider this not
> > legitmate, I think we have to improve the docs about .usage_power.
> 
> I realize that I'm being nitpicky here. Setting usage_power = true and
> duty = period - duty is a lazy way of achieving what you can easily do
> by adjusting the input duty cycle.
> 
> If you all really want this, then it should go into the core, because
> it's something that can be implemented on basically every single PWM
> controller.

You'd need something like:

diff --git a/drivers/pwm/core.c b/drivers/pwm/core.c
index e01147f66e15..6bb851c2e55e 100644
--- a/drivers/pwm/core.c
+++ b/drivers/pwm/core.c
@@ -556,6 +556,7 @@ static void pwm_apply_state_debug(struct pwm_device *pwm,
 int pwm_apply_state(struct pwm_device *pwm, const struct pwm_state *state)
 {
 	struct pwm_chip *chip;
+	bool retry_inverted = true;
 	int err;
 
 	/*
@@ -580,10 +581,19 @@ int pwm_apply_state(struct pwm_device *pwm, const struct pwm_state *state)
 	    state->usage_power == pwm->state.usage_power)
 		return 0;
 
+retry:
 	err = chip->ops->apply(chip, pwm, state);
 	trace_pwm_apply(pwm, state, err);
-	if (err)
+	if (err) {
+		if (err == -EINVAL && state->usage_power && retry_inverted) {
+			state->duty_cycle = state->period - state->duty_cycle;
+			state->polarity = 1 - state->polarity;
+			retry_inverted = false;
+			goto retry;
+		}
+
 		return err;
+	}
 
 	pwm->state = *state;
 
(Just to show the idea. It doesn't work like that, because *state is const.)

I don't like that .apply() is called twice and without having thought
much about it, I'd prefer explicit support in the lowlevel drivers over
this approach.

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |