Re: [v2 3/4] platform: (TS-5500) add LED support

[Vivien Didelot <vivien.didelot@savoirfairelinux.com>]

On Tue, 30 Aug 2011 14:17:15 -0700,
"H. Peter Anvin" <hpa@zytor.com> wrote:

> On 08/30/2011 02:15 PM, Mark Brown wrote:
> > On Tue, Aug 30, 2011 at 05:14:24PM -0400, Vivien Didelot wrote:
> >> "H. Peter Anvin" <hpa@zytor.com> wrote:
> >>> On 08/29/2011 03:16 PM, Mark Brown wrote:
> > 
> >>>> Can you not do outb() from atomic context?  The reason lots of
> >>>> LED drivers update the hardware in a workqueue is that they
> >>>> communicate with the hardware over buses that can't be used in
> >>>> atomic context like I2C or SPI but if that's not an issue then
> >>>> the workqueue is not required and the code can be simplified.
> > 
> >>> outb() can definitely be executed from atomic context.
> > 
> >> Good to know, thanks. I removed the work_struct and instead lock a
> >> mutex before setting led->new_brightness and calling outb().
> > 
> > You can't take a mutex in atomic context...
> 
> OK, so what is the potential race that this mutex is called for?  If
> it just means that the brightness can be redundantly set to the same
> value more than once, no atomicity is needed.
> 
> 	-hpa
> 
I wrote the led_set function like:

  static void ts5500_led_set(struct led_classdev *led_cdev,
                             enum led_brightness value)
  {
          struct ts5500_led *led = container_of(led_cdev,
                                                struct ts5500_led,
                                                cdev);
          mutex_lock(&led->lock);
          led->new_brightness = (value == LED_OFF) ? LED_OFF : LED_FULL;
          outb(value, led->ioaddr);
          mutex_unlock(&led->lock);
  }

I guess the wrong value could be read if we get preempted just before
the outb() call, am I wrong?

	Vivien