[PATCH v4 2/8] at91: pinctrl: don't request GPIOs used for interrupts but lock them as IRQ

[Linus Walleij]

On Wed, Feb 12, 2014 at 11:06 AM, Jean-Jacques Hiblot
<jjhiblot at traphandler.com> wrote:

> During the xlate stage of the DT interrupt parsing, the at91 pinctrl driver
> requests the GPIOs that are described as interrupt sources. This prevents a
> driver to request the gpio later to get its electrical value.
> This patch replaces the gpio_request with a gpio_lock_as_irq to prevent the
> gpio to be set as an ouput while allowing a subsequent gpio_request to succeed
>
> Signed-off-by: Jean-Jacques Hiblot <jjhiblot at traphandler.com>

OK, but is this really correct:

> @@ -1478,18 +1478,17 @@ static int at91_gpio_irq_domain_xlate(struct irq_domain *d,
>  {
>         struct at91_gpio_chip *at91_gpio = d->host_data;
>         int ret;
> -       int pin = at91_gpio->chip.base + intspec[0];
>
>         if (WARN_ON(intsize < 2))
>                 return -EINVAL;
>         *out_hwirq = intspec[0];
>         *out_type = intspec[1] & IRQ_TYPE_SENSE_MASK;
>
> -       ret = gpio_request(pin, ctrlr->full_name);
> +       ret = gpio_lock_as_irq(&at91_gpio->chip, intspec[0]);

So when resolving an IRQ resource, we take for granted that it will be used
for IRQs and IRQs only? Is it not possible that this resolution is done
and then the driver using it unloads or whatever and it is still marked
as IRQ?

I don't think the xlate function should have such side effects on
the gpio_chip internal state. I think it should just translate.

The line is locked for IRQ the moment its startup() callback is
called, is it not?

> -       ret = gpio_direction_input(pin);
> +       ret = at91_gpio_direction_input(&at91_gpio->chip, intspec[0]);

I actually don't like this either. This kind of thing was causing
problems in the OMAP driver like hell.

I think this should be deleted from xlate and at91_gpio_direction_input()
be called from the irqchip's .startup() or even .unmask() function
instead.

Yours,
Linus Walleij