Re: [PATCH v4 3/3] gpio: add support for the Diolan DLN-2 USB GPIO driver

[Johan Hovold <johan@kernel.org>]

On Thu, Sep 18, 2014 at 06:54:34PM +0300, Octavian Purdila wrote:
> On Thu, Sep 18, 2014 at 3:46 PM, Johan Hovold <johan@kernel.org> wrote:
> > On Thu, Sep 18, 2014 at 03:43:07PM +0300, Octavian Purdila wrote:
> >> On Thu, Sep 18, 2014 at 1:54 PM, Johan Hovold <johan@kernel.org> wrote:
> >> > On Tue, Sep 09, 2014 at 10:24:46PM +0300, Octavian Purdila wrote:
> >
> >> > Either way, it looks like this could race with get_direction() if you
> >> > get a set_direction() while get_direction() is retrieving the direction
> >> > from the device.
> >> >
> >> > This would break gpio_get().
> >> >
> >> I don't think gpio_set_direction() and gpio_get() are allowed to race.
> >
> > I wrote that set_direction() and get_direction() could race, which in
> > turn would break gpio_get() as you would be caching the wrong
> > direction setting.
> >
> 
> OK, I now see the problem. I think doing this in get_direction() will
> fix the issue:
> 
>                 if (!test_and_set_bit(offset, dln2->pin_dir_set))
>                         set/clear_bit(offset, dln2->pin_dir);
> 
> because gpiolib calls get_direction() while requesting a pin and
> request cannot race with itself. Which means that get_direction() can
> not race with itself the first time it is called, when the set/clear
> operation will be run.
> 
> And because we know that get_direction() is called first, we can even
> remove the set/clear operation from set_direction().

Why not simply fetch the direction in request() and get rid of the
pin_dir_set bitmask?

Johan