Mark Brown wrote at Thursday, August 04, 2011 11:35 PM:
On Thu, Aug 04, 2011 at 08:53:34PM -0700, Stephen Warren wrote:
Well, things break. This is essentially the problem I was describing in the PATCH 0 email, just with a slightly different motivation.
There's a bunch of existing code using that idiom.
Certainly.
I suppose that an alternative here would be to simply ignore any errors from gpio_request. This might have the benefit of removing the need for the other two patches I posted in the series. However, it seems a little dirty; one benefit of the IRQ code calling gpio_request and honoring errors would be to detect when some completely unrelated code had a bug and had called gpio_request on the GPIO before. Such detection would be non-existent if we don't error out on gpio_request. Perhaps some mechanism is needed to indicate that the driver has explicitly already called gpio_request for a legitimate shared purpose, and only then ignore errors?
But it's not a bug to use a GPIO as an IRQ source, otherwise we wouldn't have gpio_to_irq() in the first place.
True, but I think there are two cases:
1) Some code legitimately uses a pin as both a GPIO and IRQ, and is fully cognizant of the fact, just like in your example -> no bug.
2) Two pieces of unrelated code somehow accidentally get a GPIO and IRQ number that map to the same resource, e.g. due to incorrect board files or Device Tree content. This is probably a bug, but ends up looking exactly the same as far as the IRQ code's gpio_request call failing in the patch I posted.
Feels like we need a backchannel between gpiolib and the IRQ code to do this. Or perhaps the drivers that implement this should be taking care of setting up the GPIO mode?