Re: [PATCH v11 2/2] leds: Add driver for Qualcomm LPG

[Uwe Kleine-König]

Hello Bjorn,

On Wed, Feb 02, 2022 at 01:40:21PM -0800, Bjorn Andersson wrote:
> On Wed 02 Feb 08:29 PST 2022, Uwe Kleine-K?nig wrote:
> > did you consider my earlier feedback "It would also be good if the PWM
> > code could live in drivers/pwm"?
> > (https://lore.kernel.org/r/20210505051958.e5lvwfxuo2skdu2q@xxxxxxxxxxxxxx)
> 
> Yes, I did consider this. Because the downstream driver is (at least was
> when I looked at it originally) split like that.
> 
> We have a number of different Qualcomm PMICs containing the LPG modules,
> which consists of N PWM channels, a pattern lookup table and a set of
> current sinks.
> 
> Each PWM channel can either be used as a traditional PWM, a LED or be
> grouped together with other channels to form a multicolor LED. So we
> need a design that allows different boards using a particular PMIC to
> freely use the N channels according to one of these three operational
> modes.
> 
> The pattern lookup table is a shared resource containing duty cycle
> values and each of the PWM channels can be configured to have their duty
> cycle modified from the lookup table on some configured cadence.
> 
> In the even that multiple PWM channels are ganged together to form a
> multicolor LED, which is driven by a pattern, the pattern generator for
> the relevant channels needs to be synchronized.

Is this some material for the commit log to motivate the design
decision?

> If we consider the PWM channel to be the basic primitive we need some
> mechanism to configure the pattern properties for each of the channels
> and we need some mechanism to synchronize the pattern generators for
> some subset of the PWM channels.
> 
> 
> In other words we need some custom API between the LED driver part and
> the PWM driver, to configure these properties. This was the design
> of the downstream driver when I started looking at this driver.
> 
> 
> Another alternative that has been considered is to create two
> independent drivers - for the same hardware. This would allow the system
> integrator to pick the right driver for each of the channels.
> 
> One problem with this strategy is that the DeviceTree description of the
> LPG hardware will have to be modified depending on the use case. In
> particular this prevents me from writing a platform dtsi describing the
> LPG hardware and then describe the LEDs and pwm channels in a board dts.
> 
> And we can't express the individual channels, because the multicolor
> definition needs to span multiple channels.
> 
> 
> So among all the options, implementing the pwm_chip in the LED driver
> makes it possible for us to describe the LPG as one entity, with
> board-specific LEDs and a set of PWM channels.

ok.

> > At least splitting in two patches would be good IMHO.
> 
> I guess I can split out the parts related to the pwmchip in a separate
> patch. Seems to be a rather small portion of the code though. Is that
> what you have in mind?

I didn't try to understand the pattern part. I know that for PWMs there
is no pattern support, wasn't aware it's a thing for LEDs.

Anyhow, not a hard requirement to split from my side.

> > > +/*
> > > + * Limitations:
> > > + * - Updating both duty and period is not done atomically, so the output signal
> > > + *   will momentarily be a mix of the settings.
> > 
> > Is the PWM well-behaved? (i.e. does it emit the inactive level when
> > disabled?)
> 
> Yes, a disabled channel outputs a logical 0.

Please add this to the Limitations section. It's not actually a
limitation, but still this is a good place to put this information.

> > Does it complete a period before switching to the new
> > setting?
> 
> I see nothing indicating the answer to this, in either direction...

Can you test that? It's as easy as configuring a long period with 0%
relative duty cycle and then immediately a 100% relative duty cycle.

> > > +static void lpg_pwm_get_state(struct pwm_chip *chip, struct pwm_device *pwm,
> > > +			      struct pwm_state *state)
> > > +{
> > > +	struct lpg *lpg = container_of(chip, struct lpg, pwm);
> > > +	struct lpg_channel *chan = &lpg->channels[pwm->hwpwm];
> > > +	u64 duty = DIV_ROUND_UP_ULL(chan->pwm_value * chan->period, LPG_RESOLUTION - 1);
> > > +
> > > +	state->period = chan->period;
> > > +	state->duty_cycle = duty;
> > > +	state->polarity = PWM_POLARITY_NORMAL;
> > > +	state->enabled = chan->enabled;
> > 
> > This doesn't work if .get_state() is called before .apply() was called,
> > does it?
> > 
> 
> You mean that I would return some bogus state and not the actual
> hardware state?

Yes. At least I only found lpg_calc_freq() assigning chan->period and
chan->enabled. And unless I missed something this isn't called before
the pwm core calls .get_state().

> > > +}

Best regards
Uwe

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

Attachment: signature.asc
Description: PGP signature