Re: [PATCH v9 4/6] pwm: dwc: use clock rate in hz to avoid rounding issues

["Uwe Kleine-König" <u.kleine-koenig@pengutronix.de>]

[-- Attachment #1: Type: text/plain, Size: 4989 bytes --]

Hello,

[Dropped William Salmon and Jude Onyenegecha from the list of recipents,
their email addresses don't seem to work any more.]

On Thu, Sep 07, 2023 at 05:12:40PM +0100, Ben Dooks wrote:
> As noted, the clock-rate when not a nice multiple of ns is probably
> going to end up with inacurate calculations, as well as on a non pci
> system the rate may change (although we've not put a clock rate
> change notifier in this code yet) so we also add some quick checks
> of the rate when we do any calculations with it.
> 
> Signed-off-by; Ben Dooks <ben.dooks@codethink.co.uk>
> Reported-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>
> ---
> v9:
>  - fixed commit spelling
>  - changed to use codethink email instead of sifive
> v8:
>  - fixup post rename
>  - move to earlier in series
> ---
>  drivers/pwm/pwm-dwc-core.c | 24 +++++++++++++++---------
>  drivers/pwm/pwm-dwc.h      |  2 +-
>  2 files changed, 16 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/pwm/pwm-dwc-core.c b/drivers/pwm/pwm-dwc-core.c
> index 3fc281a78c9a..3b856685029d 100644
> --- a/drivers/pwm/pwm-dwc-core.c
> +++ b/drivers/pwm/pwm-dwc-core.c
> @@ -49,13 +49,14 @@ static int __dwc_pwm_configure_timer(struct dwc_pwm *dwc,
>  	 * periods and check are the result within HW limits between 1 and
>  	 * 2^32 periods.
>  	 */
> -	tmp = DIV_ROUND_CLOSEST_ULL(state->duty_cycle, dwc->clk_ns);
> +	tmp = state->duty_cycle * dwc->clk_rate;

This might overflow. You can prevent this by asserting that clk_rate is
<= NSEC_PER_SEC and using mul_u64_u64_div_u64.

> +	tmp = DIV_ROUND_CLOSEST_ULL(tmp, NSEC_PER_SEC);
>  	if (tmp < 1 || tmp > (1ULL << 32))
>  		return -ERANGE;
>  	low = tmp - 1;
>  
> -	tmp = DIV_ROUND_CLOSEST_ULL(state->period - state->duty_cycle,
> -				    dwc->clk_ns);
> +	tmp = (state->period - state->duty_cycle) * dwc->clk_rate;
> +	tmp = DIV_ROUND_CLOSEST_ULL(tmp, NSEC_PER_SEC);
>  	if (tmp < 1 || tmp > (1ULL << 32))
>  		return -ERANGE;
>  	high = tmp - 1;
> @@ -121,11 +122,14 @@ static int dwc_pwm_get_state(struct pwm_chip *chip, struct pwm_device *pwm,
>  			     struct pwm_state *state)
>  {
>  	struct dwc_pwm *dwc = to_dwc_pwm(chip);
> +	unsigned long clk_rate;
>  	u64 duty, period;
>  	u32 ctrl, ld, ld2;
>  
>  	pm_runtime_get_sync(chip->dev);
>  
> +	clk_rate = dwc->clk_rate;
> +
>  	ctrl = dwc_pwm_readl(dwc, DWC_TIM_CTRL(pwm->hwpwm));
>  	ld = dwc_pwm_readl(dwc, DWC_TIM_LD_CNT(pwm->hwpwm));
>  	ld2 = dwc_pwm_readl(dwc, DWC_TIM_LD_CNT2(pwm->hwpwm));
> @@ -137,17 +141,19 @@ static int dwc_pwm_get_state(struct pwm_chip *chip, struct pwm_device *pwm,
>  	 * based on the timer load-count only.
>  	 */
>  	if (ctrl & DWC_TIM_CTRL_PWM) {
> -		duty = (ld + 1) * dwc->clk_ns;
> -		period = (ld2 + 1)  * dwc->clk_ns;
> +		duty = ld + 1;
> +		period = ld2 + 1;
>  		period += duty;
>  	} else {
> -		duty = (ld + 1) * dwc->clk_ns;
> +		duty = ld + 1;
>  		period = duty * 2;
>  	}
>  
> +	duty *= NSEC_PER_SEC;
> +	period *= NSEC_PER_SEC;

A comment that/why this cannot overflow would be nice. (I didn't check,
maybe it can?)

> +	state->period = DIV_ROUND_CLOSEST_ULL(period, clk_rate);
> +	state->duty_cycle = DIV_ROUND_CLOSEST_ULL(duty, clk_rate);

Without having thought deeply about this, I think you need to round up
here. Hmm, but given that .apply() uses round-closest, too, this needs
to be addressed separately.

(The ugly thing about round-closest is that .apply(mypwm,
.get_state(mypwm)) isn't idempotent in general. Consider a PWM that can
implement period = 41.7ns and period = 42.4 ns. If it's configured with
42.4, .get_state will return period = 42. Reapplying this will configure
for 41.7ns. This won't happen with the PCI clkrate, but it might in the
of case. Another reason to use rounding-down in .apply is that
mul_u64_u64_div_u64 doesn't have a round-nearest variant.)

>  	state->polarity = PWM_POLARITY_INVERSED;
> -	state->period = period;
> -	state->duty_cycle = duty;
>  
>  	pm_runtime_put_sync(chip->dev);
>  
> @@ -168,7 +174,7 @@ struct dwc_pwm *dwc_pwm_alloc(struct device *dev)
>  	if (!dwc)
>  		return NULL;
>  
> -	dwc->clk_ns = 10;
> +	dwc->clk_rate = NSEC_PER_SEC / 10;
>  	dwc->chip.dev = dev;
>  	dwc->chip.ops = &dwc_pwm_ops;
>  	dwc->chip.npwm = DWC_TIMERS_TOTAL;
> diff --git a/drivers/pwm/pwm-dwc.h b/drivers/pwm/pwm-dwc.h
> index 64795247c54c..e0a940fd6e87 100644
> --- a/drivers/pwm/pwm-dwc.h
> +++ b/drivers/pwm/pwm-dwc.h
> @@ -42,7 +42,7 @@ struct dwc_pwm_ctx {
>  struct dwc_pwm {
>  	struct pwm_chip chip;
>  	void __iomem *base;
> -	unsigned int clk_ns;
> +	unsigned long clk_rate;

Given that clk_ns was only introduced in patch #2 I think it would be
cleaner to squash these two patches together.

Best regards
Uwe

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

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]