Hello Sean,

On Fri, Dec 04, 2020 at 11:38:46AM +0000, Sean Young wrote:
> On Fri, Dec 04, 2020 at 12:13:26PM +0100, Uwe Kleine-König wrote:
> > On Fri, Dec 04, 2020 at 08:44:17AM +0000, Sean Young wrote:
> > > On Fri, Dec 04, 2020 at 12:42:15AM +0100, Lino Sanfilippo wrote:
> > > > > You're storing an unsigned long long (i.e. 64 bits) in an u32. If
> > > > > you are sure that this won't discard relevant bits, please explain
> > > > > this in a comment for the cursory reader.
> > > > 
> > > > What about an extra check then to make sure that the period has not been truncated,
> > > > e.g:
> > > > 
> > > > 	value = DIV_ROUND_CLOSEST_ULL(state->period, scaler);
> > > > 
> > > > 	/* dont accept a period that is too small or has been truncated */
> > > > 	if ((value < PERIOD_MIN) ||
> > > > 	    (value != DIV_ROUND_CLOSEST_ULL(state->period, scaler)))
> > > > 		return -EINVAL;
> > > 
> > > Rather than doing another 64 bit division which is expensive (esp on 32 bit
> > > kernels), you could assign to u64 and check:
> > > 
> > > 	if (value < PERIOD_MIN || value > U32_MAX)
> > > 		return -EINVAL;
> > 
> > Given that value is a u32, value > U32_MAX will never trigger.
> 
> I meant that value is declared u64 as well ("assign to u64").
> 
> > Maybe checking period before doing the division is more sensible.
> 
> That could introduce rounding errors, exactly why PERIOD_MIN was introduced.

If done correctly it doesn't introduce rounding errors.

> > > > > Also note that round_closed is probably wrong, as .apply() is
> > > > > supposed to round down the period to the next achievable period. (But
> > > > > fixing this has to do done in a separate patch.)
> > > > 
> > > > According to commit 11fc4edc4 rounding to the closest integer has been introduced
> > > > to improve precision in case that the pwm controller is used by the pwm-ir-tx driver.
> > > > I dont know how strong the requirement is to round down the period in apply(), but I
> > > > can imagine that this may be a good reason to deviate from this rule.
> > > > (CCing Sean Young who introduced DIV_ROUND_CLOSEST)
> > > 
> > > There was a problem where the carrier is incorrect for some IR hardware
> > > which uses a carrier of 455kHz. With periods that small, rounding errors
> > > do really matter and rounding down might cause problems.
> > > 
> > > A policy of rounding down the carrier is not the right thing to do
> > > for pwm-ir-tx, and such a change will probably break pwm-ir-tx in some
> > > edge cases.
> > 
> > IMO it's not an option to say: pwm-driver A is used for IR, so A's
> > .apply uses round-nearest and pwm-driver B is used for $somethingelse,
> > so B's .apply uses round-down.
> 
> I'm not saying that one driver should have one it one way and another driver
> another way.

I read between your lines that you think that round-nearest is the
single best strategy, is that right?

> > To be a sensible API pwm_apply_state
> > should have a fixed behaviour. I consider round-down the sensible
> > choice (because it is easier to implmement the other options with this)
> 
> It's not sensible when it's wrong about half the time.

So round-nearest which is wrong about the other half is better?
If you have two consumer drivers and one requires round-nearest and the
other requires round-down, how would you suggest to implement these two?
Always adapting the low-level driver depending on which consumer is in
use sounds wrong. So I conclude that the expectation about the
implemented rounding behaviour should be the same for all drivers. And
if your consumer happens to require a different strategy you're either
out of luck (bad), or we need to expand the PWM API to make this
possible, probably by implementing a round_state callback that tells the
caller the resulting state if the given state is applied.

> Why is is easier to implement?

If pwm_apply_state (and so pwm_round_state) rounds down, you can achieve
round-nearest (simplified: Ignoring polarity, just looking for period) using:

	lower_state = pwm_round_state(pwm, target_state);
	upper_state = {
		.period = 2 * target_state.period - lower_state.period,
		...
	}
	tmp = pwm_round_state(pwm, upper)

	if tmp.period < target_state.period:
		# tmp == lower_state
		return lower_state

	else while tmp.period > target_state.period:
		upper = tmp;
		tmp.period -= 1
		tmp = pwm_round_state(pwm, tmp)

I admit it is not pretty. But please try to implement it the other way
around (i.e. pwm_round_state rounding to nearest and search for a
setting that yields the biggest period not above target.period without
just trying all steps). I spend a few brain cycles and the corner cases
are harder. (But maybe I'm not smart enough, so please convince me.)

Note that with round-nearest there is another complication: Assume a PWM
that can implement period = 500 µs and period = 1000 µs (and nothing
inbetween). That corresponds to the frequencies 2000 Hz and 1000 Hz.
round_nearest for state with period = 700 µs (corresponding to 1428.5714
Hz) would then pick 500 µs (corresponding to 2000 Hz), right? So is
round-nearest really what you prefer?

> > and for consumers like the IR stuff we need to provide some more
> > functions to allow it selecting a better suited state. Something like:
> > 
> > 	pwm_round_state_nearest(pwm, { .period = 2198, .. }, &state)
> > 
> > which queries the hardwares capabilities and then assigns state.period =
> > 2200 instead of 2100.
> 
> This is very elaborate and surely not "easier to implement". Why not just
> do the right thing in the first place and round-closest?

I looked through the history of drivers/pwm for commits changing the
rounding behaviour. I found:

 - 11fc4edc483 which changes bcm2835 from round-down to round-closest
   (I didn't check but given that the driver divides by the result of a
   division the rounding might not always be round-closest.)
 - 12f9ce4a519 which changes pwm-rockchip from round-down to
   round-closest
   (The motivation described in the commit log is wrong today as
   pwm_get_state() gives the last set value, not the result of the
   lowlevel driver's .get_state callback. Also this problem can be fixed
   with drivers implementing round-down by just letting .get_state round
   up. (Which by the way is how I recommend how to implement it when
   reviewing new drivers.))

Did I miss something?

For a quick (and maybe unreliable) overview:

	$ git grep -l _CLOSEST drivers/pwm/ | wc -l
	15

so we might have 15 drivers that round to nearest and the remaining 40
round down. (I checked a few and didn't find a false diagnose.)

For me this isn't a clear indication that round-nearest is
unconditionally better. What is the fact that convinces you that
round-nearest is better in general?

> > Where can I find the affected (consumer) driver?
> 
> So there is the pwm-ir-tx driver. The infrared led is directly connected
> to the pwm output pin, so that's all there is.

Ah, found it, drivers/media/rc/pwm-ir-tx.c, thanks.

Best regards
Uwe

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