Re: [PATCH v2 Resend] pwm: samsung: Fix output race on disabling

[Sjoerd Simons <sjoerd.simons@collabora.co.uk>]

On Wed, 2015-03-18 at 09:08 +0100, Lukasz Majewski wrote:
> Hi Sjoerd,
> 
> > Hey Jingoo, Kukjijn, Lukasz,
> > 
> > Pinging on this one again, could you please review this patch so it
> > can be merged through the PWM tree? 
> 
> As fair as I remember, I've already acked the patch :-)

I don't think you did, but i might have missed it ofcourse. Seems
patchwork also missed it though:   
  https://patchwork.ozlabs.org/patch/446643/

Mind redoing you're acked-by so it gets picked up by patchwork ? :)

> > 
> > On Thu, 2015-03-05 at 09:14 +0100, Sjoerd Simons wrote:
> > > When disabling the samsung PWM the output state remains at the
> > > level it was in the end of a pwm cycle. In other words, calling
> > > pwm_disable when at 100% duty will keep the output active, while at
> > > all other setting the output will go/stay inactive. On top of that
> > > the samsung PWM settings are double-buffered, which means the new
> > > settings only get applied at the start of a new PWM cycle.
> > > 
> > > This results in a race if the PWM is at 100% duty and a driver
> > > calls: pwm_config (pwm, 0, period);
> > >   pwm_disable (pwm);
> > > 
> > > In this case the PWMs output will unexpectedly stay active, unless
> > > a new PWM cycle happened to start between the register writes in
> > > _config and _disable. As far as i can tell this is a regression
> > > introduced by 3bdf878, before that a call to pwm_config would call
> > > pwm_samsung_enable which, while heavy-handed, made sure the
> > > expected settings were live.
> > > 
> > > To resolve this, while not re-introducing the issues 3bdf878
> > > (flickering as the PWM got reset while in a PWM cycle). Only force
> > > an update of the settings when at 100% duty, which shouldn't have a
> > > noticeable effect on the output but is enough to ensure the
> > > behaviour is as expected on disable.
> > > 
> > > Signed-off-by: Sjoerd Simons <sjoerd.simons@collabora.co.uk>
> > > ---
> > > Changes since v1:
> > >   Fix small issues pointed out by Tomasz Figa
> > >   - Correct various coding style issues
> > >   - Read the current value of the tcmp register for comparison
> > > rather then using a non-trivial comparison to decide whether the
> > > current state was 100% duty
> > >   - Move the code to force manual update out into its own function
> > >   - Clarify the comment indicating why a manual update is sometimes
> > > required
> > > 
> > >  drivers/pwm/pwm-samsung.c | 31 ++++++++++++++++++++++++++++++-
> > >  1 file changed, 30 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/pwm/pwm-samsung.c b/drivers/pwm/pwm-samsung.c
> > > index 3e9b583..649f6c4 100644
> > > --- a/drivers/pwm/pwm-samsung.c
> > > +++ b/drivers/pwm/pwm-samsung.c
> > > @@ -269,12 +269,31 @@ static void pwm_samsung_disable(struct
> > > pwm_chip *chip, struct pwm_device *pwm)
> > > spin_unlock_irqrestore(&samsung_pwm_lock, flags); }
> > >  
> > > +static void pwm_samsung_manual_update(struct samsung_pwm_chip
> > > *chip,
> > > +				      struct pwm_device *pwm)
> > > +{
> > > +	unsigned int tcon_chan = to_tcon_channel(pwm->hwpwm);
> > > +	u32 tcon;
> > > +	unsigned long flags;
> > > +
> > > +	spin_lock_irqsave(&samsung_pwm_lock, flags);
> > > +
> > > +	tcon = readl(chip->base + REG_TCON);
> > > +	tcon |= TCON_MANUALUPDATE(tcon_chan);
> > > +	writel(tcon, chip->base + REG_TCON);
> > > +
> > > +	tcon &= ~TCON_MANUALUPDATE(tcon_chan);
> > > +	writel(tcon, chip->base + REG_TCON);
> > > +
> > > +	spin_unlock_irqrestore(&samsung_pwm_lock, flags);
> > > +}
> > > +
> > >  static int pwm_samsung_config(struct pwm_chip *chip, struct
> > > pwm_device *pwm, int duty_ns, int period_ns)
> > >  {
> > >  	struct samsung_pwm_chip *our_chip =
> > > to_samsung_pwm_chip(chip); struct samsung_pwm_channel *chan =
> > > pwm_get_chip_data(pwm);
> > > -	u32 tin_ns = chan->tin_ns, tcnt, tcmp;
> > > +	u32 tin_ns = chan->tin_ns, tcnt, tcmp, oldtcmp;
> > >  
> > >  	/*
> > >  	 * We currently avoid using 64bit arithmetic by using the
> > > @@ -288,6 +307,7 @@ static int pwm_samsung_config(struct pwm_chip
> > > *chip, struct pwm_device *pwm, return 0;
> > >  
> > >  	tcnt = readl(our_chip->base + REG_TCNTB(pwm->hwpwm));
> > > +	oldtcmp = readl(our_chip->base + REG_TCMPB(pwm->hwpwm));
> > >  
> > >  	/* We need tick count for calculation, not last tick. */
> > >  	++tcnt;
> > > @@ -335,6 +355,15 @@ static int pwm_samsung_config(struct pwm_chip
> > > *chip, struct pwm_device *pwm, writel(tcnt, our_chip->base +
> > > REG_TCNTB(pwm->hwpwm)); writel(tcmp, our_chip->base +
> > > REG_TCMPB(pwm->hwpwm)); 
> > > +	/* In case the PWM is currently at 100% duty, force a
> > > manual update
> > > +	 * to prevent the signal staying high in the pwm is
> > > disabled shortly
> > > +	 * afer this update (before it autoreloaded the new
> > > values) .
> > > +	 */
> > > +	if (oldtcmp == (u32) -1) {
> > > +		dev_dbg(our_chip->chip.dev, "Forcing manual
> > > update");
> > > +		pwm_samsung_manual_update(our_chip, pwm);
> > > +	}
> > > +
> > >  	chan->period_ns = period_ns;
> > >  	chan->tin_ns = tin_ns;
> > >  	chan->duty_ns = duty_ns;
> > 
> > 
> 
> 
> 


-- 
Sjoerd Simons <sjoerd.simons@collabora.co.uk>
Collabora Ltd.