Russell King - ARM Linux wrote at Friday, August 05, 2011 3:40 AM:
On Thu, Aug 04, 2011 at 05:00:17PM -0600, Stephen Warren wrote:
In http://www.spinics.net/lists/linux-tegra/msg01731.html, Mark Brown pointed out that it was a little silly forcing every board or driver to gpio_request() a GPIO that is later converted to an IRQ, and passed to request_irq. The first patch in this series instead makes the core IRQ code perform these calls when appropriate, to avoid duplicating it everywhere.
Trying to go from IRQ to GPIO is not a good idea - most of the IRQ <-> GPIO macros we have today are just plain broken. Many of them just add or subtract a constant, which means non-GPIO IRQs have an apparant GPIO number too. Couple this with the fact that all positive GPIO numbers are valid, and this is a recipe for wrong GPIOs getting used and GPIOs being requested for non-GPIO IRQs.
I think this was also discussed in the past, and the conclusion was that IRQs should be kept separate from GPIOs. Maybe views have changed since then...
However, if we do want to do this, then it would be much better to provide a new API for requesting GPIO IRQs, eg:
gpio_request_irq()
which would wrap around request_threaded_irq(), takes a GPIO number, does the GPIO->IRQ conversion internally, and whatever GPIO setup is required. Something like this:
With that approach, drivers need to explicitly know whether they're passed a GPIO or an IRQ, and do something different, or they need to choose to only accept a GPIO or IRQ.
Take the case of the WM8903 audio codec, which is the case I was primarily motivated by.
The struct i2c_board_info for the WM8903 contains a .irq field, which is already an IRQ number. This IRQ might be a plain IRQ (i.e. a dedicated interrupt pin on the SoC's package), or a pin that's a GPIO with interrupt capabilities.
It'd be best if the WM8903 driver could handle either case, and be Completely unaware of which case was used in practice; it simply wants to request the interrupt, and have that work appropriately.
Having a separate API for this means pushing this GPIO-vs-IRQ knowledge into all drivers, which doesn't seem like a good thing.
So it seems like, as was mentioned elsewhere in this thread, the upshot of this conversation is that interrupt chip drivers should do this internally, both to avoid requiring a general irq_to_gpio function, and because calling gpio_direction_input for GPIOs-used-as-IRQs isn't appropriate for all hardware.
int gpio_request_threaded_irq(int gpio, irq_handler_t handler, irq_handler_t thread_fn, unsigned long flags, const char *name, void *dev)
...