Re: [PATCH]input: Change timer function to workqueue for gpio_keys driver

[Alek Du <alek.du@intel.com>]

On Thu, 25 Jun 2009 21:31:33 +0800
Jani Nikula <ext-jani.1.nikula@nokia.com> wrote:

> On Thu, 2009-06-25 at 15:06 +0200, ext Alek Du wrote:
> > On Thu, 25 Jun 2009 18:29:25 +0800
> > Jani Nikula <ext-jani.1.nikula@nokia.com> wrote:
> > 
> > > On Fri, Jun 12, 2009 at 8:40 PM, Trilok Soni<soni.trilok@gmail.com> wrote:
> > > >>  static irqreturn_t gpio_keys_isr(int irq, void *dev_id)
> > > >>  {
> > > >>        struct gpio_button_data *bdata = dev_id;
> > > >> @@ -62,10 +61,10 @@ static irqreturn_t gpio_keys_isr(int irq, void *dev_id)
> > > >>        BUG_ON(irq != gpio_to_irq(button->gpio));
> > > >>
> > > >>        if (button->debounce_interval)
> > > >> -               mod_timer(&bdata->timer,
> > > >> -                       jiffies + msecs_to_jiffies(button->debounce_interval));
> > > >> +               schedule_delayed_work(&bdata->work,
> > > >> +                       msecs_to_jiffies(button->debounce_interval));
> > > >>        else
> > > >> -               gpio_keys_report_event(bdata);
> > > >> +               schedule_work(&bdata->work.work);
> > > >>
> > > >>        return IRQ_HANDLED;
> > > >>  }
> > > 
> > > Correct me if I'm wrong, but as far as I can tell,
> > > schedule_delayed_work doesn't modify the timer if the work was already
> > > pending. The result is not the same as with the timer. This breaks the
> > > debouncing.
> > 
> > No. The workqueue is per button, if the work is already pending, then last
> > key press is not handled yet. That keeps the debouncing. Why you want the second
> > key press to break the first one? The second key press should be ignored, that's
> > the meaning of debouncing right?
> 
> No, debouncing is supposed to let the gpio line stabilize to either
> state before doing *anything*. You only want to schedule the work (and
> send the input event) once the line has been in the same state for
> debounce_interval ms. This is what the original code did, by kicking the
> timer further at each state change.
> 
If you schedule the timer when you decide it "stabilized", the final gpio_get_value()
could still return 0 in the timer handler, if the key released at that time. So your previous
"stabilized" state is useless.
Isn't the delay work itself the mechanism to decide the "stabilized" ?

The work will finally call gpio_get_value to determine the state to be sent
to input layer. I don't think there is any defect here. 
 
> IMHO it should be either fixed or reverted.
> 

No, the original timer handler will crash kernel if you are using a I2C GPIO or SPI GPIO expander
Since it try to call sleep-able gpio_get_value in atomic context.