GPIO bindings guidelines (Was: Re: [PATCH v5 10/12] gpio: Support for unified device properties interface)

[Alexandre Courbot]

On Mon, Oct 20, 2014 at 11:26 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Monday 20 October 2014 15:12:50 Alexandre Courbot wrote:
>> On Sat, Oct 18, 2014 at 6:47 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>> > On Friday 17 October 2014 20:09:51 Arnd Bergmann wrote:
>> >> On October 17, 2014 2:16:00 PM CEST, "Rafael J. Wysocki" <rjw@xxxxxxxxxxxxx> wrote:
>> >> >From: Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx>
>> >> >
>> >> >Some drivers need to deal with only firmware representation of its
>> >> >GPIOs. An example would be a GPIO button array driver where each button
>> >> >is described as a separate firmware node in device tree. Typically
>> >> >these
>> >> >child nodes do not have physical representation in the Linux device
>> >> >model.
>> >> >
>> >> >In order to help device drivers to handle such firmware child nodes we
>> >> >add dev[m]_get_named_gpiod_from_child() that takes a child firmware
>> >> >node pointer as its second argument (the first one is the parent device
>> >> >itself), finds the GPIO using whatever is the underlying firmware
>> >> >method, and requests the GPIO properly.
>> >>
>> >> Could we also have a wrapper around this function without a "name" argument,
>> >> using just the index?
>> >
>> > Expanding on this thought: I think we should mandate for new bindings
>> > that they use either a name and no index, or an index but not name,
>>
>> I'm afraid this could forbid some useful use-cases, namely the ones
>> where several GPIOs serve the same function (and are typically set
>> together). We had a few patch proposals to handle such GPIO groups,
>> and even though one was in pretty good shape the submitter did not
>> push it until the end. :/
>>
>> But my concern is that instead of having this:
>>
>> enable-gpio = <&gpio 0 GPIO_ACTIVE_HIGH>;
>> value-gpios = <&gpio 1 GPIO_ACTIVE_HIGH ... &gpio 8 GPIO_ACTIVE_HIGH>;
>>
>> We would force this:
>>
>> enable-gpio = <&gpio 0 GPIO_ACTIVE_HIGH>;
>> value0-gpio = <&gpio 1 GPIO_ACTIVE_HIGH>;
>> ...
>> value7-gpio = <&gpio 8 GPIO_ACTIVE_HIGH>;
>>
>> Or this:
>>
>> // First GPIO is enable, other GPIOs are value
>> gpios = <&gpio 0 GPIO_ACTIVE_HIGH &gpio 1 GPIO_ACTIVE_HIGH ... &gpio 8
>> GPIO_ACTIVE_HIGH>;
>>
>> Most bindings don't need that much sophistication, and for these we
>> should indeed make sure that they stick to using either the names or
>> index (and in a consistent manner), but closing the possibility to use
>> both together may bite us in the end.
>
> I would actually prefer the single-property case here, but I see your
> point. Could we make it a strong suggestion rather than a mandatory
> requirement for new bindings then?

Definitely, and a very strong suggestion even. Having to use both
names *and* index in GPIO properties should not be needed 99% of the
time.

>
>> > and I also think that for named gpios, we should try to converge on a
>> > common naming scheme. As discussed, we will probably want to support all
>> > the existing ways to do this even with ACPI and with the unified
>> > interface, but it doesn't have to be the obvious way.
>>
>> Personally, I like the idea that each GPIO has a function, so now that
>> ACPI fully supports this I'd suggest the policy of using names for
>> each GPIO (e.g. never use the fallback "gpios" or "gpio" property),
>> and only ressort to indexes if several GPIOs happen to serve the same
>> function. I know we haven't reached consensus about this so far, but
>> it would be nice it we could discuss this point again in the light of
>> the new ACPI capabilities and come with something to write as a
>> guideline in the GPIO documentation.
>
> We have enforced naming things for the dmaengine binding, which has
> just led to everyone calling things "rx" and "tx". My fear is that
> if we start to enforce giving a name, we'd end up with lots of
> drivers that use a "gpio-gpios" property or something silly.

Checking the bindings is also part of the review process. :) Things
like "gpio-gpios" should simply not be accepted to begin with.

This sounds like a good chance to finally land some guidelines
regarding GPIO bindings. Let's summarize the situation:
- GPIO bindings can be defined using both DT and ACPI (both interfaces
nicely abstracted by the interface introduced by this series)
- Both firmware interfaces support indexed GPIOs
- Both firmware interfaces support named GPIO properties, with an
optional index (can we absolutely take this for granted on ACPI now?)
- For DT bindings, both foo-gpio and foo-gpios are valid properties
for the GPIO "foo".

That's what we have now and need to maintain. However for new drivers
we want to come with guidelines that will hopefully make things easier
to review. Here are my thoughts on the topic:

- GPIOs have a function, and this function should be the GPIO name.
Thus now that ACPI supports named properties, all new GPIO properties
*must* have an accurate, explicit name.
- Indexes are only used if several GPIOs fulfill the same function
(like parallel data lines). This situation should be exceptional.
GPIOs not fulfilling the same function are in no way allowed to reside
under the same property.
- The DT property suffix should be "-gpio" for properties expecting
exactly one GPIO and "-gpios" for properties specifying one or more
GPIOs, to make it clear whether one should bother about the number of
GPIOs present of not.

Anything against this, especially from the ACPI side? If not I will
come with a patch to amend the GPIO documentation.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/