On Tue, Jan 27, 2015 at 03:21:27PM +0000, Mika Westerberg wrote:

On Tue, Jan 27, 2015 at 03:06:06PM +0000, Mark Rutland wrote:

quoted

On Tue, Jan 27, 2015 at 02:41:27PM +0000, Mika Westerberg wrote:

quoted

On Tue, Jan 27, 2015 at 02:33:34PM +0000, Mark Rutland wrote:

quoted

Ok, that allays my fear w.r.t. ordering of the resources.

As I see it, the fact that we convert GpioInt entries to GPIOs rather
than irqs when parsing _CRS is the issue here, and to me it makes no
sense that we do so. Were we to treat them as interrupts, the binding is
fine as-is, and we'd do the same thing in DT and ACPI.

The reason GpioInt is separate from GpioIo is that a GpioInt _is_ an
interrupt (which happens to be backed by a GPIO), and is not something
that necessarily makes sense as a GPIO.

I would rather say that GpioInt *is* a GPIO. That can then used as an
interrupt but it should not prevent you from using it as GPIO instead.
For example if you just want to poll that something is 0 or 1. That
should be possible as well and nothing say that you cannot do that for
GpioInt().

quoted

From my POV a GpioInt is logically an interrupt, or it would be a

GpioIo. That doesn't necessarily mean it's invalid to try to query its
state as a GPIO, but I do not think that it makes sense to handle it by
default as a GPIO given that it was handed to us as a GpioInt so that it
can be used as an interrupt.

If it's just the case that ACPI and DT differ w.r.t. how this case (an
interrupt line wired to a GPIO) is described, that in itself is fine;
different standards have different models.

However, I do not think we must change the DT binding and violate
established DT practice simply becuase on the ACPI side things are
different (nor would it make sense to do things the other way around).
If the two have different rules, then we should handle those rules
separately rather than trying to force the two together when they
clearly don't fit.

I agree.

quoted

In the driver that can easily be achieved with separate probe paths.

Yes, that's what this patch is doing. It has different paths for
interrupt and GPIO cases. If we find that there is an interrupt number
already, then we use that directly.

If not we try to look for a GPIO which we could use as an interrupt.

My comment about separate probe paths was w.r.t. ACPI and DT. In DT,
acquiring a GPIO here shouldn't be necessary (and ideally shouldn't be
allowed) because the GPIO should be described as an interrupt (we need
this to have the correct flags).

The reason why this is not done only in the ACPI probe path is that
there is nothing ACPI specific in gpiod_get() and friends.

While the high-level APIs are not specific to ACPI or DT, the behaviour
we want for DT is subtly different top the behaviour you want for ACPI.

In DT we don't necessarily have the relevant flags (e.g. edge vs level,
active high vs active low) unless the GPIO is described as an interrupt,
so permitting it to be described as a GPIO permits situations which we
cannot handle correctly. In the DT case therefore we must not acquire
the interrupt GPIO as a GPIO; it must be described as an interrupt.

So if we need to acquire a GPIO in the ACPI case, this should be limited
to the ACPI case.

Looking further at ACPI, the flags issue also seems to be a problem
there. If the interrupt GPIO were described as a GpioIo, then you could
acquire it as a GPIO with gpiod_get(), but wouldn't have the EdgeLevel
and ActiveLevel properties, and therefore cannot necessarily correctly
convert the GPIO to an irq, no?

Thanks,
Mark.