Re: [PATCH v6 1/4] gpio: mvebu: fix pwm get_state period calculation

On Sun, Jan 10, 2021 at 07:14:17PM +0200, Baruch Siach wrote:

Hi Uwe,

Thanks for your review comments.

On Thu, Jan 07 2021, Uwe Kleine-König wrote:

quoted

On Wed, Jan 06, 2021 at 09:37:37AM +0200, Baruch Siach wrote:

quoted

The period is the sum of on and off values.

Reported-by: Russell King <linux@armlinux.org.uk>
Fixes: 757642f9a584e ("gpio: mvebu: Add limited PWM support")
Signed-off-by: Baruch Siach <baruch@tkos.co.il>
---
v6: divide (on + off) sum to reduce rounding error (RMK)
---
 drivers/gpio/gpio-mvebu.c | 19 ++++++++-----------
 1 file changed, 8 insertions(+), 11 deletions(-)

diff --git a/drivers/gpio/gpio-mvebu.c b/drivers/gpio/gpio-mvebu.c
index 672681a976f5..a912a8fed197 100644
--- a/drivers/gpio/gpio-mvebu.c
+++ b/drivers/gpio/gpio-mvebu.c

@@ -676,20 +676,17 @@ static void mvebu_pwm_get_state(struct pwm_chip *chip,
 	else
 		state->duty_cycle = 1;
 
+	val = (unsigned long long) u; /* on duration */
 	regmap_read(mvpwm->regs, mvebu_pwmreg_blink_off_duration(mvpwm), &u);
-	val = (unsigned long long) u * NSEC_PER_SEC;
+	val += (unsigned long long) u; /* period = on + off duration */
+	val *= NSEC_PER_SEC;
 	do_div(val, mvpwm->clk_rate);
-	if (val < state->duty_cycle) {
+	if (val > UINT_MAX)
+		state->period = UINT_MAX;

state->period is an u64, so there is no reason to not use values greater
than UINT_MAX.

I'll post a patch for that.

quoted

quoted

+	else if (val)
+		state->period = val;
+	else
 		state->period = 1;

This case assigning 1 looks strange. An explanation in a comment would
be great. I wonder if this is a hardware property or if it is only used
to not report 0 in case that mvpwm->clk_rate is high.

I guess that this is because 0 period does not make sense for pwm. It is
like a zero divisor. What is the common behavior?

It depends on how the hardware behaves in this case. One measure is to
use round-up (in .get_state) for the divisions -- as I already pointed
out. Then the problem only triggers if both on and off registers are
zero.

quoted

I found a few further shortcommings in the mvebu_pwm implementation while
looking through it:

 a) The rounding problem that RMK found is also present in .apply

    There we have:

    	val = clk_rate * (period - duty_cycle) / NSEC_PER_SEC

    while

    	val = clk_rate * period / NSEC_PER_SEC - on

    would be more exact.

I'll post a patch for that.

quoted

 b) To make

 	pwm_get_state(pwm, &state);
	pwm_apply_state(pwm, &state);

    idempotent .get_state should round up the division results.

I'll post a patch for that as well.

quoted

 c) .apply also has a check for val being zero and configures at least 1
    cycle for the on and off intervals. Is this a hardware imposed
    limitation? 

Not sure what was the original intention. Maybe Andrew can explain.

On my hardware (Armada 8040), zero 'on' value does not work as
expected. There is a blink once in a long while. Maybe this is the
reason?

It would be good to understand and document this. Hardware PWMs not
supporting a zero duty cycle are always a surprise for me and probably
also others.

As I understand, all these issues should not block this patch, right?

Yes, having this patch in a series fixing all the stuff I pointed out
would still be very welcome :-)

BTW, the key you used to sign your message is expired since 2020-01-10
on the key server I use (keys.gnupg.net). Where can I find your updated
key?

Hmm, I thought keyservers are out of fashion. Anyhow, I updated my key
there. There is no WKD for @pengutronix.de addresses, but

	gpg --locate-external-keys uwe@kleine-koenig.org

should work just fine, too.

Best regards
Uwe

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