Re: [PATCHv2] pwm: avoid holding mutex in interrupt context

[Thierry Reding]

On Wed, Jan 20, 2016 at 08:29:58AM +0900, Krzysztof Kozlowski wrote:
> On 20.01.2016 00:04, Anand Moon wrote:
> > Hi Krzysztof,
> > 
> > On 18 January 2016 at 09:58, Krzysztof Kozlowski
> >>> Already within function pwm_samsung_set_invert is protected by
> >>> spin_lock_irqsave(&samsung_pwm_lock, flags);
> >>>
> >>> So no need to introduce another lock to control pwm_samsung_set_polarity.
> >>>
> >>> Best Regards.
> >>> -Anand Moon
> >>
> >> I don't have any clue what is your point here. I don't get what
> >> pwm_samsung_set_polarity has to do with main pwm core...
> >>
> >> Sorry, you need to be more specific.
> >>
> >> Best regards,
> >> Krzysztof
> >>
> >>
> > 
> > Below is the mapping of calls from pwm driver.
> > I have tried to map the functionality and I am trying to understand
> > the flow of the driver.
> > 
> > Also looking in document
> > 
> > https://www.kernel.org/doc/Documentation/pwm.txt
> > 
> > pwm-samsung driver controls the LEDS, fans...etc
> > 
> > Form the dts modes pwmleds
> > 
> >         pwmleds {
> >                 compatible = "pwm-leds";
> > 
> >                 blueled {
> >                         label = "blue:heartbeat";
> >                         pwms = <&pwm 2 2000000 0>;
> >                         pwm-names = "pwm2";
> >                         max_brightness = <255>;
> >                         linux,default-trigger = "heartbeat";
> >                 };
> >         };
> > 
> > Following is the map out from the device tree.
> > 
> > pwms = <&pwm 2 2000000 0>;
> > 
> > &pwm       ->  pwm: pwm@12dd0000 --->samsung,exynos4210-pwm
> > 2              ->  period
> > 2000000    ->  duty_cycle
> > 0              ->  polarity
> 
> I do not see any relations between DTS and the problem.
> 
> > 
> > And here is the mapping of the call of function
> > Note: This function call are as per my understanding of the flow in
> > the driver. I might be wrong.
> > 
> > pwm_samsung_set_polarity(struct pwm_chip *chip, struct pwm_device
> > *pwm, enum pwm_polarity polarity)
> > \
> >  pwm_samsung_set_invert(our_chip, pwm->hwpwm, invert);
> >  \
> >   pwm_set_polarity(struct pwm_device *pwm, enum pwm_polarity polarity)
> 
> No, pwm_samsung_set_invert does not call pwm_set_polarity(). This would
> result in a circular call - back to pwm_samsung_set_polarity().
> 
> >   \
> >    pwm->chip->ops->set_polarity(pwm->chip, pwm, polarity);
> >     \
> >      pwm_enable(struct pwm_device *pwm) or pwm_disable(struct pwm_device *pwm)
> > 
> > pwm_enable or pwm_disable will be triggered on change in pwm->flags by
> > the pwm core.
> > before pwm_set_polarity is called form the Samsung driver it hold with
> > following locks
> > 
> > Here is the locking
> > 
> > pwm_samsung_set_polarity(struct pwm_chip *chip, struct pwm_device
> > *pwm, enum pwm_polarity polarity)
> >  \
> >   pwm_samsung_set_invert(struct samsung_pwm_chip *chip, unsigned int
> > channel, bool invert)
> >     \
> >      spin_lock_irqsave(&samsung_pwm_lock, flags);
> >       \
> >        pwm_set_polarity(struct pwm_device *pwm, enum pwm_polarity polarity)
> >        \
> >         mutex_lock(&pwm->lock)
> > 
> >           pwm_enable(struct pwm_device *pwm) or pwm_disable(struct
> > pwm_device *pwm)
> >           \
> >            mutex_lock(&pwm->lock);
> > 
> > Problem I see that we are holding the lock in interrupt context.
> > I don't know how the this triggers this bug.
> > 
> > BUG: sleeping function called from invalid context at kernel/locking/mutex.c:97
> 
> So leave it. If your flow of calls was correct, you would spot the
> problem. But actually it does not matter - I think the flow is not correct.

The reason for the BUG that you're seeing is that the leds-pwm driver
differentiates between PWMs that can sleep and those that can't. This
used to be limited to some PWMs that were attached to a slow bus like
I2C, or that called functions which might sleep (like clk_prepare()).
With commit d1cd21427747 ("pwm: Set enable state properly on failed
call to enable"), effectively all PWM drivers may sleep. The lock
introduced in that commit must also be a mutex because it protects
sections which may sleep themselves (->enable() and ->set_polarity())
so turning it into a spinlock won't work for the general case.

Given that this is currently broken and we're quite close to -rc1 I
suggest the following fix for now:

--- >8 ---
diff --git a/drivers/pwm/core.c b/drivers/pwm/core.c
index d24ca5f281b4..7831bc6b51dd 100644
--- a/drivers/pwm/core.c
+++ b/drivers/pwm/core.c
@@ -889,7 +889,7 @@ EXPORT_SYMBOL_GPL(devm_pwm_put);
   */
 bool pwm_can_sleep(struct pwm_device *pwm)
 {
-	return pwm->chip->can_sleep;
+	return true;
 }
 EXPORT_SYMBOL_GPL(pwm_can_sleep);
 
--- >8 ---

For v4.6 I can remove all usage of the ->can_sleep and pwm_can_sleep()
because they're effectively useless now.

Does that sound reasonable to everyone?

Anand, the above should fix the issue for you. Can you give it a try
and report if it doesn't?

Thanks,
Thierry

Attachment: signature.asc
Description: PGP signature