Re: [PATCH 1/5] pwm: jz4740: Fix pin level of disabled TCU2 channels, part 1

["Uwe Kleine-König" <u.kleine-koenig@pengutronix.de>]

[-- Attachment #1: Type: text/plain, Size: 5179 bytes --]

Hello Paul,

On Fri, Nov 18, 2022 at 09:55:40AM +0000, Paul Cercueil wrote:
> Le jeu. 17 nov. 2022 à 14:29:27 +0100, Uwe Kleine-König
> <u.kleine-koenig@pengutronix.de> a écrit :
> > Hello Paul,
> > 
> > On Tue, Oct 25, 2022 at 11:02:00AM +0100, Paul Cercueil wrote:
> > >  Le mar. 25 oct. 2022 à 08:21:29 +0200, Uwe Kleine-König
> > >  <u.kleine-koenig@pengutronix.de> a écrit :
> > >  > Hello,
> > >  >
> > >  > On Mon, Oct 24, 2022 at 09:52:09PM +0100, Paul Cercueil wrote:
> > >  > >  The "duty > cycle" trick to force the pin level of a disabled
> > > TCU2
> > >  > >  channel would only work when the channel had been enabled
> > >  > > previously.
> > >  > >
> > >  > >  Address this issue by enabling the PWM mode in
> > > jz4740_pwm_disable
> > >  > >  (I know, right) so that the "duty > cycle" trick works before
> > >  > > disabling
> > >  > >  the PWM channel right after.
> > >  > >
> > >  > >  This issue went unnoticed, as the PWM pins on the majority of
> > > the
> > >  > > boards
> > >  > >  tested would default to the inactive level once the
> > > corresponding
> > >  > > TCU
> > >  > >  clock was enabled, so the first call to jz4740_pwm_disable()
> > > would
> > >  > > not
> > >  > >  actually change the pin levels.
> > >  > >
> > >  > >  On the GCW Zero however, the PWM pin for the backlight (PWM1,
> > > which
> > >  > > is
> > >  > >  a TCU2 channel) goes active as soon as the timer1 clock is
> > > enabled.
> > >  > >  Since the jz4740_pwm_disable() function did not work on
> > > channels not
> > >  > >  previously enabled, the backlight would shine at full
> > > brightness
> > >  > > from
> > >  > >  the moment the backlight driver would probe, until the
> > > backlight
> > >  > > driver
> > >  > >  tried to *enable* the PWM output.
> > >  > >
> > >  > >  With this fix, the PWM pins will be forced inactive as soon as
> > >  > >  jz4740_pwm_apply() is called (and might be reconfigured to
> > > active if
> > >  > >  dictated by the pwm_state). This means that there is still a
> > > tiny
> > >  > > time
> > >  > >  frame between the .request() and .apply() callbacks where the
> > > PWM
> > >  > > pin
> > >  > >  might be active. Sadly, there is no way to fix this issue: it
> > > is
> > >  > >  impossible to write a PWM channel's registers if the
> > > corresponding
> > >  > > clock
> > >  > >  is not enabled, and enabling the clock is what causes the PWM
> > > pin
> > >  > > to go
> > >  > >  active.
> > >  > >
> > >  > >  There is a workaround, though, which complements this fix:
> > > simply
> > >  > >  starting the backlight driver (or any PWM client driver) with a
> > >  > > "init"
> > >  > >  pinctrl state that sets the pin as an inactive GPIO. Once the
> > >  > > driver is
> > >  > >  probed and the pinctrl state switches to "default", the
> > > regular PWM
> > >  > > pin
> > >  > >  configuration can be used as it will be properly driven.
> > >  > >
> > >  > >  Fixes: c2693514a0a1 ("pwm: jz4740: Obtain regmap from parent
> > > node")
> > >  > >  Signed-off-by: Paul Cercueil <paul@crapouillou.net>
> > >  > >  Cc: stable@vger.kernel.org
> > >  >
> > >  > OK, understood the issue. I think there is another similar issue:
> > > The
> > >  > clk is get and enabled only in the .request() callback. The
> > > result is (I
> > >  > think---depends on a few further conditions) that if you have the
> > >  > backlight driver as a module and the bootloader enables the
> > > backlight to
> > >  > show a splash screen, the backlight goes off because of the
> > >  > clk_disable_unused initcall.
> > > 
> > >  I will have to verify, but I'm pretty sure disabling the clock
> > > doesn't
> > >  change the pin level back to inactive.
> > 
> > Given that you set the clk's rate depending on the period to apply, I'd
> > claim that you need to keep the clk on. Maybe it doesn't hurt, because
> > another component of the system keeps the clk running, but it's wrong
> > anyhow. Assumptions like these tend to break on new chip revisions.
> 
> If the backlight driver is a module then it will probe before the
> clk_disable_unused initcall, unless something is really wrong.

I'd claim the clk_disable_unused initcall is called before userspace
starts and so before the module can be loaded. Who is wrong here?

> So the backlight would stay ON if it was enabled by the bootloader,
> unless the DTB decides it doesn't have to be.

Don't understand that. How could hte DTB decide the backlight can be
disabled?
 
> Anyway, I can try your suggestion, and move the trick to force-disable PWM
> pins in the probe(). After that, the clocks can be safely disabled, so I can
> disable them (for the disabled PWMs) at the end of the probe and re-enable
> them again in their respective .request() callback.

I really lost track of the problem here and would appreciate a new
submission of the remaining (and improved?) patches.

Best regards
Uwe

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

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]