On Tue, Mar 24, 2015 at 04:22:16PM +0100, Lars-Peter Clausen wrote:
Add Alexandre and linux-gpio to Cc.
On 03/24/2015 04:06 PM, Mika Westerberg wrote:
On Tue, Mar 24, 2015 at 02:57:49PM +0100, Linus Walleij wrote:
On Tue, Mar 24, 2015 at 2:38 PM, Lars-Peter Clausen <lars@xxxxxxxxxx> wrote:
On 03/24/2015 02:26 PM, Robert Dolca wrote:
On Tue, Mar 24, 2015 at 2:17 PM, Lars-Peter Clausen <lars@xxxxxxxxxx>
wrote:
In the ACPI description you specify one or more IRQ GPIO pins. In the
driver you request the GPIO pin using the index. In the ACPI 5.1
specification you can use named GPIOs instead of index.
But is there a way to distinguish between IRQ GPIOs and non IRQ GPIOs? If it
is clear that a certain GPIO is the IRQ for the device the I2C framework
should take care of assigning the client->irq field, instead of doing it
manually in each and every device driver.
In the device tree case we have a mechanism where each
GPIO chip implements two API:s, one gpio_chip API and
one irqchip API.
Then in the tree both the GPIO and IRQs can be assigned as
resources to clients, orthogonally. Usually this will only work
if there is a 1-to-1 correspondence between the GPIO lines
and available IRQ line triggers on the GPIO chip, but that is
indeed the most common. They will then usually also have
the same line offset numbers. In some odd cases I guess it
won't work this way.
The I2C subsystem does this for the device tree case in
i2c_device_probe() like this:
if (!client->irq && dev->of_node) {
int irq = of_irq_get(dev->of_node, 0);
if (irq == -EPROBE_DEFER)
return irq;
if (irq < 0)
irq = 0;
client->irq = irq;
}
This is why the code does not contain any OF/DT
IRQ assignment code.
However in the ACPI probe path I guess it doesn't
happen then?
In ACPI we have two kind of GPIOs: GpioIo and GpioInt. The latter is
used to describe GPIOs that can be used as interrupts.
In order to translate a GpioInt to an interrupt number we would need to
request the GPIO first here (in the I2C core), call gpiod_to_irq() to it
and assign that to the client->irq.
Maybe the API can be extended to support to translate a GPIO to a IRQ
without actually requesting the GPIO first.
We still need to take care the the GPIO is properly requested and locked
as IRQ. Otherwise something else (userspace for example) can mess this
up.
This has few problems that I have not yet figured out. Maybe someone
here can suggest what to do:
1) Who is responsible in releasing the GPIO?
2) What if the driver wants to use that pin as a GPIO instead? The GPIO
is already requested by the I2C core.
3) We may have multiple GpioInts for devices like GPIO button array.
Which one we should pick, or should we let the driver to handle this
separetely?
Well, we have the same issue with devicetree already. I'd say use the first
IRQ for client->irq and ignore the other ones for now.
For devices like the button array above doing that leaves the driver
wondering where the heck is one of my GPIOs :-) Perhaps we could to
automatic translation if we find out that there is only one GpioInt for
this device.