Re: [PATCH v4 1/6] pwm: core: add pwm_get_waveform_step_might_sleep()
From: Jonathan Cameron
Date: Sat Nov 21 2026 - 18:35:52 EST
On Thu, 19 Nov 2026 13:24:15 +0200
Nicolae-Daniel Deaconescu <Nicolae-daniel.Deaconescu@xxxxxxxxxx> wrote:
> A consumer that must meet minimum times, or place an edge relative to the
> cycles of a clock a PWM gates, can't do that with the waveform API alone:
> rounding goes down, and the margin a minimum needs is the PWM step, known
> only to the provider.
>
> Let the provider report the rate of the clock its timing derives from and
> the clock cycles per step. A consumer then works in whole steps and rounds
> up when converting to nanoseconds, instead of reading the rate from a clock
> it gets on its own, which breaks if the PWM divides its clock.
>
> Signed-off-by: Nicolae-Daniel Deaconescu <Nicolae-daniel.Deaconescu@xxxxxxxxxx>
> ---
> drivers/pwm/core.c | 69 +++++++++++++++++++++++++++++++++++++++++++++
> include/linux/pwm.h | 34 ++++++++++++++++++++++
> 2 files changed, 103 insertions(+)
>
> diff --git a/drivers/pwm/core.c b/drivers/pwm/core.c
> index 2a050ff4608b..a63b8f65a43d 100644
> --- a/drivers/pwm/core.c
> +++ b/drivers/pwm/core.c
> @@ -14,6 +14,7 @@
> #include <linux/of.h>
> #include <linux/pwm.h>
> #include <linux/list.h>
> +#include <linux/math64.h>
> #include <linux/mutex.h>
> #include <linux/err.h>
> #include <linux/slab.h>
> @@ -325,6 +326,68 @@ int pwm_get_waveform_might_sleep(struct pwm_device *pwm, struct pwm_waveform *wf
> }
> EXPORT_SYMBOL_GPL(pwm_get_waveform_might_sleep);
>
> +/**
> + * pwm_get_waveform_step_might_sleep - Query the granularity of the waveforms
> + * Cannot be used in atomic context.
Odd to have that second bit here in the kernel-doc. Particularly as you have
a Context entry below saying the same thing below and a name of _might_sleep.
> + * @pwm: PWM device
> + * @step: output parameter
> + *
> + * Stores in @step the clock rate the timing of @pwm is derived from and how
> + * many cycles of that clock one step lasts. Period, duty length and duty
Can this be a cross reference to the docs in for the structure?
Nice to not duplicate stuff where possible.
> + * offset are multiples of that step, so a waveform that is asked for in whole
> + * steps, and is within the range of the hardware, is not changed by rounding.
> + * Compute the length of n steps in nanoseconds as
> + * n * @step->clk_cycles_per_step * NSEC_PER_SEC / @step->clk_rate_Hz and round
I'd tweak the wrap to just have the formula on a line on its own.
> + * it up, because pwm_round_waveform_might_sleep() and
> + * pwm_set_waveform_might_sleep() round down. The product can exceed 64 bits, so
> + * compute it with mul_u64_u64_div_u64_roundup().
I don't understand this. I was thinking it was an implementation detail that
should not be in these docs anyway, but I can't immediately spot it at all.
So rewrite or drop.
> + *
> + * The values can change when a clock changes its rate, so they are only valid
> + * as long as the provider's input clocks are not reconfigured.
> + *
> + * Context: May sleep.
> + * Returns: 0 on success, -EOPNOTSUPP if the provider doesn't report it or a
> + * negative errno.
> + */
> +int pwm_get_waveform_step_might_sleep(struct pwm_device *pwm,
> + struct pwm_waveform_step *step)
> +{
> + struct pwm_chip *chip = pwm->chip;
> + const struct pwm_ops *ops = chip->ops;
> + int ret;
> +
> + might_sleep();
> +
> + if (!pwmchip_supports_waveform(chip) || !ops->get_waveform_step)
> + return -EOPNOTSUPP;
> +
> + guard(pwmchip)(chip);
> +
> + if (!chip->operational)
> + return -ENODEV;
> +
> + if (IS_ENABLED(CONFIG_PWM_DEBUG) && chip->atomic) {
> + /*
> + * Catch any drivers that have been marked as atomic but
> + * that will sleep anyway.
Wrap comments to 80 chars unless it helps readability to not do so.
> + */
> + non_block_start();
> + ret = ops->get_waveform_step(chip, pwm, step);
> + non_block_end();
> + } else {
> + ret = ops->get_waveform_step(chip, pwm, step);
> + }
> + if (ret)
> + return ret;
> +
> + if (!step->clk_rate_Hz ||
> + step->clk_rate_Hz > mul_u32_u32(step->clk_cycles_per_step, NSEC_PER_SEC))
> + return -EINVAL;
> +
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(pwm_get_waveform_step_might_sleep);
> +
> /* Called with the pwmchip lock held */
> static int __pwm_set_waveform(struct pwm_device *pwm,
> const struct pwm_waveform *wf,
> @@ -1710,6 +1773,12 @@ static bool pwm_ops_check(const struct pwm_chip *chip)
> if (!ops->apply)
> return false;
>
> + if (ops->get_waveform_step) {
> + dev_warn(pwmchip_parent(chip),
> + ".get_waveform_step() needs .write_waveform()\n");
> + return false;
> + }
> +
> if (IS_ENABLED(CONFIG_PWM_DEBUG) && !ops->get_state)
> dev_warn(pwmchip_parent(chip),
> "Please implement the .get_state() callback\n");
> diff --git a/include/linux/pwm.h b/include/linux/pwm.h
> index b11ae91723f8..c52f4c9153c2 100644
> --- a/include/linux/pwm.h
> +++ b/include/linux/pwm.h
> @@ -76,6 +76,34 @@ struct pwm_waveform {
> u64 duty_offset_ns;
> };
>
> +/**
> + * struct pwm_waveform_step - granularity of the waveforms a PWM can generate
> + * @clk_rate_Hz: rate of the clock the PWM's timing is derived from
> + * @clk_cycles_per_step: number of those clock cycles one step lasts
> + *
> + * The period, the duty length and the duty offset of a waveform are multiples
> + * of one step, which lasts @clk_cycles_per_step / @clk_rate_Hz seconds. A PWM
> + * that divides its input clock in hardware has @clk_cycles_per_step > 1. A
> + * waveform that is asked for in whole steps, and is within the range of the
> + * hardware, is not changed by rounding.
> + *
> + * The step is the same for every waveform and for the period, the duty length
> + * and the duty offset, so a provider whose granularity depends on the
> + * requested period (for example through a prescaler) can't describe itself
> + * with it. One step also has to last at least 1 ns, otherwise rounding a
> + * length in nanoseconds up can add a whole step.
> + *
> + * A consumer that needs a minimum time, or that has to place an edge relative
> + * to the cycles of the clock a PWM channel gates, can use this to ask for
> + * the next possible value instead of one that is rounded down, and to compute
> + * the frequency a rounded period results in without the error of the
> + * nanosecond resolution of struct pwm_waveform.
> + */
> +struct pwm_waveform_step {
> + unsigned long clk_rate_Hz;
> + unsigned int clk_cycles_per_step;
> +};
For this to go through IIO (or an immutable branch if needed) I'll obviously
need PWM tags.
Thanks,
Jonathan