Re: [PATCH v3 1/4] pwm: imx27: fix race condition .apply,.get_state

[Leif Middelschulte <leif.middelschulte@gmail.com>]

> Am 06.09.2023 um 17:42 schrieb Uwe Kleine-König <u.kleine-koenig@pengutronix.de>:
> 
> On Tue, Aug 15, 2023 at 12:43:29PM +0200, Leif Middelschulte wrote:
>> From: Leif Middelschulte <Leif.Middelschulte@klsmartin.com>
>> 
>> With CONFIG_PWM_DEBUG=y after writing a value to the PWMSAR
>> register in .apply(), the register is read in .get_state().
>> Unless a period completed in the meantime, this read yields the
>> previously used duty cycle configuration. As the PWM_DEBUG code
>> applies the read out configuration for testing purposes this
>> effectively undoes the intended effect by rewriting the previous
>> hardware state.
>> 
>> Note that this change merely implements a sensible heuristic.
>> The i.MX has a 4 slot FIFO to configure the duty cycle. This FIFO
>> cannot be read back in its entirety. The "write x then read back
>> x from hw" semantics are therefore not easily applicable.
>> With this change, the .get_state() function tries to wait for some
>> stabilization in the FIFO (empty state). In this state it keeps
>> applying the last value written to the sample register.
>> 
>> Signed-off-by: Leif Middelschulte <Leif.Middelschulte@klsmartin.com>
>> ---
>> drivers/pwm/pwm-imx27.c | 50 ++++++++++++++++++++++++++++++++++++++---
>> 1 file changed, 47 insertions(+), 3 deletions(-)
>> 
>> diff --git a/drivers/pwm/pwm-imx27.c b/drivers/pwm/pwm-imx27.c
>> index 29a3089c534c..32389ca2da3e 100644
>> --- a/drivers/pwm/pwm-imx27.c
>> +++ b/drivers/pwm/pwm-imx27.c
>> @@ -75,6 +75,7 @@
>>   (x)) + 1)
>> 
>> #define MX3_PWM_SWR_LOOP 5
>> +#define MX3_PWM_FIFOAV_EMPTY_LOOP 4
>> 
>> /* PWMPR register value of 0xffff has the same effect as 0xfffe */
>> #define MX3_PWMPR_MAX 0xfffe
>> @@ -118,8 +119,28 @@ static void pwm_imx27_clk_disable_unprepare(struct pwm_imx27_chip *imx)
>> clk_disable_unprepare(imx->clk_ipg);
>> }
>> 
>> +static int pwm_imx27_wait_fifo_empty(struct pwm_chip *chip,
>> +     struct pwm_device *pwm)
>> +{
>> + struct pwm_imx27_chip *imx = to_pwm_imx27_chip(chip);
>> + struct device *dev = chip->dev;
>> + unsigned int period_ms = DIV_ROUND_UP_ULL(pwm_get_period(pwm), NSEC_PER_MSEC);
>> + int tries = MX3_PWM_FIFOAV_EMPTY_LOOP;
>> + int fifoav;
>> + u32 sr;
>> +
>> + while (tries--) {
>> + sr = readl(imx->mmio_base + MX3_PWMSR);
>> + fifoav = FIELD_GET(MX3_PWMSR_FIFOAV, sr);
>> + if (fifoav == MX3_PWMSR_FIFOAV_EMPTY)
>> + return;
>> + msleep(period_ms);
>> + }
>> + dev_warn(dev, "FIFO has been refilled concurrently\n");
>> +}
>> +
>> static int pwm_imx27_get_state(struct pwm_chip *chip,
>> -       struct pwm_device *pwm, struct pwm_state *state)
>> + struct pwm_device *pwm, struct pwm_state *state)
> 
> This looks wrong. Aligning at the opening ( was right.

Good catch. That change as unintentional.

> 
>> {
>> struct pwm_imx27_chip *imx = to_pwm_imx27_chip(chip);
>> u32 period, prescaler, pwm_clk, val;
>> @@ -161,10 +182,33 @@ static int pwm_imx27_get_state(struct pwm_chip *chip,
>> * PWMSAR can be read only if PWM is enabled. If the PWM is disabled,
>> * use the cached value.
>> */
>> - if (state->enabled)
>> + if (state->enabled) {
>> + /*
>> + * From the i.MX PWM reference manual:
>> + * "A read on the sample register yields the current FIFO value that
>> + *  is being used, or will be used, by the PWM for generation on the
>> + *  output signal. Therefore, a write and a subsequent read on the
>> + *  sample register may result in different values being obtained."
>> + * Furthermore:
>> + * "When a new value is written, the duty cycle changes after the
>> + *  current period is over."
>> + * Note "changes" vs. "changes to the given value"!
>> + * Finally:
>> + * "The PWM will run at the last set duty-cycle setting if all the
>> + *  values of the FIFO has been utilized, until the FIFO is reloaded
>> + *  or the PWM is disabled."
>> + * Try to be at least a bit more deterministic about which value is
>> + * read by waiting until the FIFO is empty. In this state the last/most
>> + * recently pushed sample (duty cycle) value is continuously applied.
>> + * Beware that this approach is still racy, as a new value could have
>> + * been supplied and a period expired between the call of the wait
>> + * function and the subsequent readl.
> 
> this would only happen if there are concurrent calls into the driver,
> wouldn't it? I think it's safe to assume this doesn't happen.

This assessment seems correct at them moment. This comment is supposed to be a source of explanation to future developers if - for some reason - the driver/subsystem is changed and they trip over this behaviour (again), just as I did.

> 
> Patch 3 of this series improves the function that is only introduced
> here. I suggest to squash these together.

I will provide v4 with the changes squashed into the initial commit, that introduced the function.

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

Thank you,

Leif