Re: [PATCH v3] gpio: interrupt consistency check for OF GPIO IRQs

[Stephen Warren <swarren@wwwdotorg.org>]

On 09/02/2013 03:25 AM, Lars Poeschel wrote:
> Am Freitag, 30. August 2013, 13:55:26 schrieb Stephen Warren:
>> On 08/29/2013 06:24 PM, Javier Martinez Canillas wrote:
>> ...
>>
>>> We have been trying to solve this issue for a few months by now and Linus'
>>> approach seems to be the most sensible solution to me.
>>>
>>> Drivers that request an IRQ and assume that platform code will request and
>>> setup the GPIO have been broken since the boards using these drivers were
>>> migrated to DT (e.g: smsc911x on OMAP2+ boards).
>>
>> That's only true if the driver for the GPIO controller is buggy.
>> Whatever request_irq() maps down to in the GPIO/IRQ controller driver
>> simply needs to set up the pin as an interrupt input, then it doesn't
>> matter which order the driver does things.
> 
> Is it really that easy?

Yes.

> request_irq() should request the gpio and set it to input that it needs to 
> fulfill the irq request. That would then be the way to go for new drivers and 
> in the DT case.

Either explicitly request the GPIO, or simply directly program the HW in
whatever way is required for the pin to operate as an IRQ.

> Some leagcy drivers currently do this:
> 
> request_gpio(gpio);
> gpio_direction_input(gpio);
> request_irq(gpio_to_irq(gpio));
> 
> In that case request_irq should not fail because the driver is already the 
> correct owner of this gpio. But if some other entity owns this gpio it should 
> fail.

Yes.

> Also if I understand you correct the other way round should also possible:
> 
> request_irq(gpio_to_irq(gpio));
> request_gpio(gpio);
> gpio_direction_input(gpio);
> 
> request_irq() also requests the gpio then but the following request_gpio() 
> should also not fail.

I don't believe that code sequence is currently banned. If we want to
ban it, we should document that. Until this is documented as being
banned, I think we must fully support that code sequence.

> To have it work that way we have to track the owners of all requested gpios 
> somewhere. Or am I wrong here?
> Where and how to record these owners? Can gpio system achieve this? gpios are 
> requested without an owning device.

Yes. But, I believe we need to fully track every GPIO/IRQ owner already
right now; if a driver (not the GPIO/IRQ chip driver, but a driver that
uses the GPIOs/IRQs) calls gpio_request()/request_irq(), and this patch
has already requested them, then we already need to fully track GPIO/IRQ
ownership to make sure that the driver's own requests aren't failed
because the DT/GPIO core has already requested them on its behalf.