On Fri, Nov 27, 2020 at 12:08:51PM +0100, Thierry Reding wrote:
> On Fri, Nov 27, 2020 at 08:11:05AM +0100, Uwe Kleine-König wrote:
> > Hello Jonathan,
> >
> > On Fri, Nov 27, 2020 at 12:19:31AM +0100, Jonathan Neuschäfer wrote:
> > > On Tue, Nov 24, 2020 at 09:20:19AM +0100, Uwe Kleine-König wrote:
> > > > On Sun, Nov 22, 2020 at 11:27:36PM +0100, Jonathan Neuschäfer wrote:
> > > [...]
> > > > > +/*
> > > > > + * The time base used in the EC is 8MHz, or 125ns. Period and duty cycle are
> > > > > + * measured in this unit.
> > > > > + */
> > > > > +#define TIME_BASE_NS 125
> > > > > +
> > > > > +/*
> > > > > + * The maximum input value (in nanoseconds) is determined by the time base and
> > > > > + * the range of the hardware registers that hold the converted value.
> > > > > + * It fits into 32 bits, so we can do our calculations in 32 bits as well.
> > > > > + */
> > > > > +#define MAX_PERIOD_NS (TIME_BASE_NS * 0xffff)
> > > > > +
> > > > > +static int ntxec_pwm_apply(struct pwm_chip *chip, struct pwm_device *pwm_dev,
> > > > > + const struct pwm_state *state)
> > > > > +{
> > > > > + struct ntxec_pwm *priv = pwmchip_to_priv(pwm_dev->chip);
> > > > > + unsigned int duty = state->duty_cycle;
> > > > > + unsigned int period = state->period;
> > > >
> > > > state->duty_cycle and state->period are u64, so you're losing
> > > > information here. Consider state->duty_cycle = 0x100000001 and
> > > > state->period = 0x200000001.
> > >
> > > Oh, good point, I didn't notice the truncation.
> > >
> > > The reason I picked unsigned int was to avoid a 64-bit division;
> > > I suppose I can do something like this:
> > >
> > > period = (u32)period / TIME_BASE_NS;
> > > duty = (u32)duty / TIME_BASE_NS;
> >
> > You can do that after you checked period > MAX_PERIOD_NS below, yes.
> > Something like:
> >
> > if (state->polarity != PWM_POLARITY_NORMAL)
> > return -EINVAL;
> >
> > if (state->period > MAX_PERIOD_NS) {
> > period = MAX_PERIOD_NS;
> > else
> > period = state->period;
> >
> > if (state->duty_cycle > period)
> > duty_cycle = period;
> > else
> > duty_cycle = state->duty_cycle;
> >
> > should work with even keeping the local variables as unsigned int.
> >
> > > > > + int res = 0;
> > > > > +
> > > > > + if (state->polarity != PWM_POLARITY_NORMAL)
> > > > > + return -EINVAL;
> > > > > +
> > > > > + if (period > MAX_PERIOD_NS) {
> > > > > + period = MAX_PERIOD_NS;
> > > > > +
> > > > > + if (duty > period)
> > > > > + duty = period;
> > > > > + }
> > > > > +
> > > > > + period /= TIME_BASE_NS;
> > > > > + duty /= TIME_BASE_NS;
> > > > > +
> > > > > + res = regmap_write(priv->ec->regmap, NTXEC_REG_PERIOD_HIGH, ntxec_reg8(period >> 8));
> > > > > + if (res)
> > > > > + return res;
> > > >
> > > > I wonder if you can add some logic to the regmap in the mfd driver such
> > > > that ntxec_reg8 isn't necessary for all users.
> > >
> > > I think that would involve:
> > >
> > > 1. adding custom register access functions to the regmap, which decide
> > > based on the register number whether a register needs 8-bit or 16-bit
> > > access. So far I have avoided information about registers into the
> > > main driver, when the registers are only used in the sub-drivers.
> > >
> > > or
> > >
> > > 2. switching the regmap configuration to little endian, which would be
> > > advantageous for 8-bit registers, inconsequential for 16-bit
> > > registers that consist of independent high and low halves, and wrong
> > > for the 16-bit registers 0x41, which reads the battery voltage ADC
> > > value. It is also different from how the vendor kernel treats 16-bit
> > > registers.
> > >
> > > Perhaps there is another option that I haven't considered yet.
> >
> > I don't know enough about regmap to teach you something here. But maybe
> > Mark has an idea. (I promoted him from Cc: to To:, maybe he will
> > notice.)
> >
> > > > > + res = regmap_write(priv->ec->regmap, NTXEC_REG_PERIOD_LOW, ntxec_reg8(period));
> > > > > + if (res)
> > > > > + return res;
> > > > > +
> > > > > + res = regmap_write(priv->ec->regmap, NTXEC_REG_DUTY_HIGH, ntxec_reg8(duty >> 8));
> > > > > + if (res)
> > > > > + return res;
> > > > > +
> > > > > + res = regmap_write(priv->ec->regmap, NTXEC_REG_DUTY_LOW, ntxec_reg8(duty));
> > > > > + if (res)
> > > > > + return res;
> > > >
> > > > I think I already asked, but I don't remember the reply: What happens to
> > > > the output between these writes? A comment here about this would be
> > > > suitable.
> > >
> > > I will add something like the following:
> > >
> > > /*
> > > * Changes to the period and duty cycle take effect as soon as the
> > > * corresponding low byte is written, so the hardware may be configured
> > > * to an inconsistent state after the period is written and before the
> > > * duty cycle is fully written. If, in such a case, the old duty cycle
> > > * is longer than the new period, the EC will output 100% for a moment.
> > > */
> >
> > Is the value pair taken over by hardware atomically? That is, is it
> > really "will" in your last line, or only "might". (E.g. when changing
> > from duty_cycle, period = 1000, 2000 to 500, 800 and a new cycle begins
> > after reducing period, the new duty_cycle is probably written before the
> > counter reaches 500. Do we get a 100% cycle here?)
> >
> > Other than that the info is fine. Make sure to point this out in the
> > Limitations paragraph at the top of the driver please, too.
>
> Perhaps also use something like regmap_bulk_write() to make sure the
> time between these writes is a short as possible.
>
> >
> > > > > + .apply = ntxec_pwm_apply,
> > > >
> > > > /*
> > > > * The current state cannot be read out, so there is no .get_state
> > > > * callback.
> > > > */
> > > >
> > > > Hmm, at least you could provice a .get_state() callback that reports the
> > > > setting that was actually implemented for in the last call to .apply()?
> > >
> > > Yes... I see two options:
> > >
> > > 1. Caching the state in the driver's private struct. I'm not completely
> > > convinced of the value, given that the information is mostly
> > > available in the PWM core already (except for the adjustments that
> > > the driver makes).
> > >
> > > 2. Writing the adjusted state back into pwm_dev->state (via pwm_set_*).
> > > This seems a bit dirty.
> >
> > 2. isn't a good option. Maybe regmap caches this stuff anyhow for 1. (or
> > can be made doing that)?
> >
> > > > @Thierry: Do you have concerns here? Actually it would be more effective
> > > > to have a callback (like .apply()) that modfies its pwm_state
> > > > accordingly. (Some drivers did that in the past, but I changed that to
> > > > get an uniform behaviour in 71523d1812aca61e32e742e87ec064e3d8c615e1.)
> > > > The downside is that people have to understand that concept to properly
> > > > use it. I'm torn about the right approach.
> > >
> > > General guidance for such cases when the state can't be read back from
> > > the hardware would be appreciated.
> >
> > Yes, improving the documentation would be great here. Thierry, can you
> > please comment on
> > https://lore.kernel.org/r/20191209213233.29574-2-u.kleine-koenig@xxxxxxxxxxxxxx
> > which I'm waiting on before describing our understanding in more detail.
>
> Hm... that link gives me a "Not Found" message. Anyway, I think perhaps
I thought I tested that before sending the link, but it gives Not Found
for me, too.
https://patchwork.ozlabs.org/project/linux-pwm/patch/20191209213233.29574-2-u.kleine-koenig@xxxxxxxxxxxxxx/
is the patchwork link.
> the best compromise would be for the core to provide an implementation
> of ->get_state() that drivers can use if they can't read out hardware
> state. This generic implementation would then just copy over the
> internal state and we have to trust that that's really what was applied.
>
> One drawback of that is that we don't factor in things like rounding
> errors and other limitations. So a better alternative may be to require
> drivers to store a cached version of the state and return that in their
> ->get_state() implementation.
Yes. That's what Jonathan called 1.
> Or perhaps a hybrid of the above would work where the core provides the
> helper that copies cached state and a cached state structure for storage
> and then the drivers that can't properly read back hardware state just
> need to update the cached state during ->apply().
I thought about doing
diff --git a/include/linux/pwm.h b/include/linux/pwm.h
index a13ff383fa1d..c60a638ebfad 100644
--- a/include/linux/pwm.h
+++ b/include/linux/pwm.h
@@ -261,7 +261,7 @@ struct pwm_ops {
int (*capture)(struct pwm_chip *chip, struct pwm_device *pwm,
struct pwm_capture *result, unsigned long timeout);
int (*apply)(struct pwm_chip *chip, struct pwm_device *pwm,
- const struct pwm_state *state);
+ struct pwm_state *state);
void (*get_state)(struct pwm_chip *chip, struct pwm_device *pwm,
struct pwm_state *state);
struct module *owner;
and adapting the callbacks to provide the actually implemented setting
in *state. But I don't like this because providing this is (for probably
most drivers) extra effort. So the drivers should only cache the written
register values and calculate the actual timing from these values if and
when required. Providing help from the framework for this is more
complicated because different drivers have different needs here. (Two
u64 for duty_cycle and period is probably enough for everybody?)
> I slightly prefer variant 2 because it's not clear to me how often we'll
> need this and we can always easily convert to variant 3 if this becomes
> a more common thing to do.
I cannot follow because I don't know what variant is "variant 2" and
"variant 3" for you.
I suggest for now we go for "don't provide a .get_state() callback"
because in the only case where it is currently called there is no cached
data to rely on anyhow.
Best regards
Uwe
--
Pengutronix e.K. | Uwe Kleine-König |
Industrial Linux Solutions | https://www.pengutronix.de/ |
Attachment:
signature.asc
Description: PGP signature
