Re: [PATCH] ACPI / GPIO: Pass index to acpi_get_gpiod_by_index() when using properties

From: Grant Likely
Date: Tue Nov 04 2014 - 09:49:13 EST


On Mon, Nov 3, 2014 at 9:52 PM, Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> On Monday, November 03, 2014 04:25:08 PM Rafael J. Wysocki wrote:
>> On Sunday, November 02, 2014 08:49:37 PM Darren Hart wrote:
>> >
>> > On 11/1/14 4:11, Grant Likely wrote:
>> > > On Tue, 28 Oct 2014 22:59:57 +0100
>> > > , "Rafael J. Wysocki" <rjw@xxxxxxxxxxxxx>
>> > > wrote:
>> > >> On Tuesday, October 28, 2014 01:15:27 PM Mika Westerberg wrote:
>> > >>> acpi_dev_add_driver_gpios() makes it possible to set up mapping between
>> > >>> properties and ACPI GpioIo resources in a driver, so we can take index
>> > >>> parameter in acpi_find_gpio() into use with _DSD device properties now.
>> > >>>
>> > >>> This index can be used to select a GPIO from a property with multiple
>> > >>> GPIOs:
>> > >>>
>> > >>> Package () {
>> > >>> "data-gpios",
>> > >>> Package () {
>> > >>> \_SB.GPIO, 0, 0, 0,
>> > >>> \_SB.GPIO, 1, 0, 0,
>> > >>> \_SB.GPIO, 2, 0, 1,
>> > >>> }
>> > >>> }
>> > >>>
>> > >>> In order to retrieve the last GPIO from a driver we can simply do:
>> > >>>
>> > >>> desc = devm_gpiod_get_index(dev, "data", 2);
>> > >>>
>> > >>> and so on.
>> > >>>
>> > >>> Signed-off-by: Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx>
>> > >>
>> > >> Cool. :-)
>> > >>
>> > >> Any objections anyone?
>> > >
>> > > Actually, I do. Not in the idea, but in the implementation. The way this gets encoded:
>> > >
>> > > Package () {
>> > > \_SB.GPIO, 0, 0, 0,
>> > > \_SB.GPIO, 1, 0, 0,
>> > > \_SB.GPIO, 2, 0, 1,
>> > > }
>> > >
>> > > Means that decoding each GPIO tuple requires the length of a tuple to be
>> > > fixed, or to implement a DT-like #gpio-cells. If it is fixed, then there
>> > > is no way to expand the binding later. Can this be done in the following
>> > > way instead?
>> > >
>> > > Package () {
>> > > Package () { \_SB.GPIO, 0, 0, 0 },
>> > > Package () { \_SB.GPIO, 1, 0, 0 },
>> > > Package () { \_SB.GPIO, 2, 0, 1 },
>> > > }
>> > >
>> > > This is one of the biggest pains in device tree. We don't have any way
>> > > to group tuples so it requires looking up stuff across the tree to
>> > > figure out how to parse each multi-item property.
>> > >
>> > > I know that last year we talked about how bios vendors would get
>> > > complicated properties wrong, but I think there is little risk in this
>> > > case. If the property is encoded wrong, the driver simply won't work and
>> > > it is unlikely to get shipped before being fixed.
>> >
>> > This particular nesting of Packages is expressly prohibited by the
>> > Device Properties UUID for the reasons you mention.
>> >
>> > http://www.uefi.org/sites/default/files/resources/_DSD-device-properties-UUID.pdf
>>
>> Also we don't use properties where single name is assigned to multiple GPIOs
>> anywhere in the current device-properties patchset, so this is not relevant at
>> the moment.
>>
>> Moreover, even if we were to use them, we would need to ensure that this:
>>
>> Package () {
>> \_SB.GPIO, 0, 0, 0
>> }
>>
>> was equivalent to
>>
>> Package () {
>> Package () { \_SB.GPIO, 0, 0, 0 }
>> }
>>
>> This is not impossible to do and I suppose we could even explain that in the
>> implementation guide document in a sensible way, but that would require the
>> document linked above to be changed first and *then* we can think about writing
>> kernel code to it. Not the other way around, please.
>>
>> So Grant, do you want us to proceed with that?
>
> Before you reply, one more observation that seems to be relevant.
>
> In ACPI, both this:
>
> Package () {
> \_SB.GPIO, 0, 0, 0,
> \_SB.GPIO, 1, 0, 0,
> \_SB.GPIO, 2, 0, 1,
> }
>
> and this:
>
> Package () {
> Package () { \_SB.GPIO, 0, 0, 0 },
> Package () { \_SB.GPIO, 1, 0, 0 },
> Package () { \_SB.GPIO, 2, 0, 1 },
> }
>
> carry the same information, because every element of a package has a type,
> so there is no danger of confusing an ACPI_TYPE_LOCAL_REFERENCE with
> ACPI_TYPE_INTEGER. Thus one can easily count the number of GPIOs represented
> by the first package by counting the number of reference elements in it.
> The second one has more structure which in this particular case is arguably
> redundant.

Okay, this make sense. I'm okay with this approach, and I would
recommend making that the only valid method for parsing in
acpi_dev_get_property_reference(). Get rid of the *size_prop argument
so that it always behaves the same way and users aren't tempted to do
something clever.

>
> Of course, that's not the case for list properties where each item consists
> of a bunch of integers, like
>
> Package () {
> Package () { 0, 0, 0 },
> Package () { 1, 0, 0 },
> Package () { 2, 0, 1 },
> }
>
> but I'm not sure if this is relevant at all.

Probably not. With a pure list it isn't implicitly referencing data in
another node. In the ref+args pattern the length of each tuple can
vary based on which node it references, but on a simple list the
parsing is going to be a lot simpler.

g.
--
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/