[alsa-devel] [PATCH 2/4] ASoC: s3c64xx/smartq: use dynamic registration

[Alexandre Courbot]

On Wed, Jul 16, 2014 at 4:28 PM, Alexandre Courbot <gnurou at gmail.com> wrote:
> On Wed, Jul 16, 2014 at 4:12 PM, Thierry Reding
> <thierry.reding at gmail.com> wrote:
>> On Wed, Jul 16, 2014 at 12:00:45PM +0900, Alexandre Courbot wrote:
>>> On Tue, Jul 15, 2014 at 6:14 PM, Alexandre Courbot <gnurou at gmail.com> wrote:
>>> > On Tue, Jul 15, 2014 at 4:58 PM, Lars-Peter Clausen <lars at metafoo.de> wrote:
>>> >> On 07/15/2014 09:36 AM, Alexandre Courbot wrote:
>>> >>>
>>> >>> On Tue, Jul 15, 2014 at 4:19 PM, Arnd Bergmann <arnd at arndb.de> wrote:
>>> >>>>
>>> >>>> On Monday 14 July 2014 19:36:24 Mark Brown wrote:
>>> >>>>>
>>> >>>>> On Mon, Jul 14, 2014 at 08:23:55PM +0200, Arnd Bergmann wrote:
>>> >>>>>>
>>> >>>>>> On Monday 14 July 2014 18:18:12 Lars-Peter Clausen wrote:
>>> >>>>>
>>> >>>>>
>>> >>>>>>> Yes. But now that you say it the gpiod_direction_output() call is
>>> >>>>>>> missing
>>> >>>>>>> from this patch.
>>> >>>>>
>>> >>>>>
>>> >>>>>> I'm lost now. The GPIO_ACTIVE_HIGH I added comes from
>>> >>>>>> Documentation/gpio/board.txt
>>> >>>>>> and as Linus Walleij explained to me the other day, the lookup is
>>> >>>>>> supposed
>>> >>>>>> to replace devm_gpio_request_one(), which in turn replaced both the
>>> >>>>>> gpio_request and the gpio_direction_output(). Do I need to put the
>>> >>>>>> gpiod_direction_output() back or is there another interface for that
>>> >>>>>> when
>>> >>>>>> registering the board gpios?
>>> >>>>>
>>> >>>>>
>>> >>>>> Indeed.  If you *do* need an explicit _output() then that sounds to me
>>> >>>>> like we either need a gpiod_get_one() or an extension to the table,
>>> >>>>> looking at the code it seems like this is indeed the case.  We can set
>>> >>>>> if the GPIO is active high/low, or open source/drain but there's no flag
>>> >>>>> for the initial state.
>>> >>>>
>>> >>>>
>>> >>>> (adding Alexandre and the gpio list)
>>> >>>>
>>> >>>> GPIO people: any guidance on how a board file should set a gpio to
>>> >>>> output/default-high in a GPIO_LOOKUP() table to replace a
>>> >>>> devm_gpio_request_one() call in a device driver with devm_gpiod_get()?
>>> >>>> Do we need to add an interface extension to do this, e.g. passing
>>> >>>> GPIOF_OUT_INIT_HIGH as the flags rather than GPIO_ACTIVE_HIGH?
>>> >>>
>>> >>>
>>> >>> The way I see it, GPIO mappings (whether they are done using the
>>> >>> lookup tables, DT, or ACPI) should only care about details that are
>>> >>> relevant to the device layout and that should be abstracted to the
>>> >>> driver (e.g. whether the GPIO is active low or open drain) so drivers
>>> >>> do not need to check X conditions every time they want to drive the
>>> >>> GPIO.
>>> >>>
>>> >>> Direction and initial value, on the other hand, are clearly properties
>>> >>> that ought to be set by the driver itself. Thus my expectation here
>>> >>> would be that the driver sets the GPIO direction and initial value as
>>> >>> soon as it gets it using gpiod_direction_output(). In other words,
>>> >>> there is no replacement for gpio_request_one() with the gpiod
>>> >>> interface. Is there any use-case that cannot be covered by calling
>>> >>> gpiod_direction_output() right after gpiod_get()? AFAICT this is what
>>> >>> gpio_request_one() was doing anyway.
>>> >>
>>> >>
>>> >> I agree with you that this is something that should be done in the driver
>>> >> and not in the lookup table. I think that it is still a good idea to have a
>>> >> replacement for gpio_request_one with the new GPIO descriptor API. A large
>>> >> share of the drivers want to call either gpio_direction_input() or
>>> >> gpio_direction_output() right after requesting the GPIO. Combining both the
>>> >> requesting and the configuration of the GPIO into one function call makes
>>> >> the code a bit shorter and also simplifies the error handling. Even more so
>>> >> if e.g. the GPIO is optional. This was one of the main reasons why
>>> >> gpio_request_one was introduced, see the commit[1] that added it.
>>> >
>>> > I am not opposed to it as a convenience function. Note that since the
>>> > open-source and open-drain flags are already handled by the lookup
>>> > table, the only flags it should handle are those related to direction,
>>> > value, and (maybe) sysfs export.
>>>
>>> Problem is, too much convenience functions seems to ultimately kill convenience.
>>>
>>> The canonical way to request a GPIO is by providing a (device,
>>> function, index) triplet to gpiod_get_index(). Since most functions
>>> only need one GPIO, we have gpiod_get(device, function) which is
>>> basically an alias to gpiod_get_index(device, function, 0) (note to
>>> self: we should probably inline it).
>>>
>>> On top of these comes another set of convenience functions,
>>> gpiod_get_optional() and gpiod_get_index_optional(), which return NULL
>>> instead of -ENOENT if the requested GPIO mapping does not exist. This
>>> is useful for the common case where a driver can work without a GPIO.
>>>
>>> Of course these functions all have devm counterparts, so we currently
>>> have 8 (devm_)gpiod_get(_index)(_optional) functions.
>>>
>>> If we are to add functions with an init flags parameter, we will end
>>> with 16 functions. That starts to be a bit too much to my taste, and
>>> maybe that's where GPIO consumers should sacrifice some convenience to
>>> preserve a comprehensible GPIO API.
>>>
>>> There might be other ways to work around this though. For instance, we
>>> could replace the _optional functions by a GPIOF_OPTIONAL flag to be
>>> passed to a more generic function that would also accept direction and
>>> init value flags. Actually I am not seeing any user of the _optional
>>> variant in -next, so maybe we should just do this. Thierry, since you
>>> introduced the _optional functions, can we get your thoughts about
>>> this?
>>
>> I personally prefer explicit naming of the functions rather than putting
>> a bunch of flags into some parameter. If you're overly concerned about
>> the amount of convenience functions, perhaps the _index variants can be
>> left out for gpiod_get_one(). I'd argue that if drivers want to deal
>> with that level of detail anyway, they may just as well add the index
>> explicitly when calling the function.
>>
>> While we're at it, gpiod_get_one() doesn't sound like a very good name.
>> All other variants only request "one" as well. Perhaps something like
>> gpiod_get_with_flags() would be a better name.
>>
>> Then again, maybe rather than add a new set of functions we should bite
>> the bullet and change gpiod_get() (and variants) to take an additional
>> flags parameter. There aren't all that many users yet (I count 26
>> outside of drivers/gpio), so maybe now would still be a good time to do
>> that.
>
> That sounds reasonable indeed. And preferable to getting an aneurysm
> after trying to spell devm_gpiod_get_index_optional_with_flags().
>
> This also makes the most sense since most GPIO users will want to set
> a direction and value right after obtaining one. So if there is no
> objection I will probably start refactoring gpiod_get() this week.

Spammed half of the kernel developers with a patch adding a flags
argument to the gpiod getters and updating all gpiod users (there were
more than I thought!). I'm not sure how such a cross-subsystem patch
is supposed to be applied ; suggestions are welcome!