Re: [PATCH v6 4/6] gpio: davinci: add OF support

[Grygorii Strashko <grygorii.strashko@ti.com>]

On 11/26/2013 07:12 PM, Sekhar Nori wrote:
> On Tuesday 26 November 2013 06:03 PM, Grygorii Strashko wrote:
>> On 11/25/2013 01:00 PM, Sekhar Nori wrote:
>>> On Thursday 21 November 2013 11:45 PM, Prabhakar Lad wrote:
>>>> From: KV Sujith <sujithkv@ti.com>
>>>>
>>>> This patch adds OF parser support for davinci gpio
>>>> driver and also appropriate documentation in gpio-davinci.txt
>>>> located at Documentation/devicetree/bindings/gpio/.
>>>>
>>>> Acked-by: Linus Walleij <linus.walleij@linaro.org>
>>>> Acked-by: Rob Herring <rob.herring@calxeda.com>
>>>> Signed-off-by: KV Sujith <sujithkv@ti.com>
>>>> Signed-off-by: Philip Avinash <avinashphilip@ti.com>
>>>> [prabhakar.csengg@gmail.com: simplified the OF code, removed
>>>> 		unnecessary DT property and also simplified
>>>> 		the commit message]
>>>> Signed-off-by: Lad, Prabhakar <prabhakar.csengg@gmail.com>
>>>> ---
>>>>    .../devicetree/bindings/gpio/gpio-davinci.txt      |   41 ++++++++++++++
>>>>    drivers/gpio/gpio-davinci.c                        |   57 ++++++++++++++++++--
>>>>    2 files changed, 95 insertions(+), 3 deletions(-)
>>>>    create mode 100644 Documentation/devicetree/bindings/gpio/gpio-davinci.txt
>>>>
>>>> diff --git a/Documentation/devicetree/bindings/gpio/gpio-davinci.txt b/Documentation/devicetree/bindings/gpio/gpio-davinci.txt
>>>> new file mode 100644
>>>> index 0000000..a2e839d
>>>> --- /dev/null
>>>> +++ b/Documentation/devicetree/bindings/gpio/gpio-davinci.txt
>>>> @@ -0,0 +1,41 @@
>>>> +Davinci GPIO controller bindings
>>>> +
>>>> +Required Properties:
>>>> +- compatible: should be "ti,dm6441-gpio"
>>>> +
>>>> +- reg: Physical base address of the controller and the size of memory mapped
>>>> +       registers.
>>>> +
>>>> +- gpio-controller : Marks the device node as a gpio controller.
>>>> +
>>>> +- interrupt-parent: phandle of the parent interrupt controller.
>>>> +
>>>> +- interrupts: Array of GPIO interrupt number. Only banked or unbanked IRQs are
>>>> +	      supported at a time.
>>>
>>> If this is true..
>>>
>>>> +
>>>> +- ti,ngpio: The number of GPIO pins supported.
>>>> +
>>>> +- ti,davinci-gpio-unbanked: The number of GPIOs that have an individual interrupt
>>>> +		             line to processor.
>>>
>>> .. then why do you need to maintain this separately? Number of elements
>>> in interrupts property should give you this answer, no?
>>>
>>> There can certainly be devices (past and future) which use a mixture of
>>> banked and unbanked IRQs. So a binding which does not take care of this
>>> is likely to change in future and that is a problem since it brings in
>>> backward compatibility of the binding into picture.
>>>
>>> The right thing would be to define the DT node per-bank similar to what
>>> is done on OMAP rather than for all banks together. That way there can
>>> be a separate property which determines whether that bank supports
>>> direct-mapped or banked IRQs (or that could be inferred if the number of
>>> tuples in the interrupts property is more than one).
>>
>> Number of IRQ can't be simply used to determine type of IRQ - need to handle IRQ names,
>> because each bank(32 gpios) may have up to 2 banked IRQs (one per 16 GPIO).
>
> Okay. That's why I inserted that comment in parenthesis :)
>
>>
>> Few things here:
>> - The mixed banked/unbanked functionality has never been supported before.
>
> True. I actually misread the driver before.
>
>> - The Davinci GPIO IP is different from OMAP and has common
>>    control registers for all banks.
>
> Well the only common register I can see is BINTEN - bank interrupt
> enable. This register can simply be initialized to enable interrupts
> from all banks possible as until the rising and falling edge triggers
> are programmed, there wont be any actual interrupts generated.
>
>> - The proposed approach is more less easy to implement for DT case, but for not-DT
>>    case - the platform data will need to be changed significantly (.
>>    So, from this point of view, that would be a big change (actually the total driver rewriting).
>
> Well, I certainly don't think its a complete driver re-write. It will
> take a bit of effort agreed, but I think the driver will also come out a
> lot cleaner.
>
> Honestly, I am not so much worried about the kernel code here. Its the
> bindings I am worried about. Once the bindings go in assuming there will
> never be banked and unbanked GPIO IRQs on the same SoC, changing them to
> do something else will be very painful with the need to keep backward
> compatibility and support for both semantics.
>
> That said, because there is no present hardware which needs both banked
> and unbanked at the same time, I wont press for this to be done endlessly.
>
>> Do you have any thoughts about how it can be done in a simpler way?
>
> I don't know if there is a "simpler" way, but I don't think there is too
> much effort too. I leave it to those implementing it though.

Oh. I see no problem to implement it for DT, but this change require to 
convert one device to the tree of devices:
  GPIO controller
  |- GPIO bank1
...
  |- GPIO bankX

And that's will need to be handled somehow from platform code (which is 
non-DT and which I can't verify and which I don't want to touch actually ;).

>
>>
>> Actually, the same was proposed by Linus, but we've tried avoid such huge rework -
>> by switching to one irq_domain per all banks for example.
>
> I didn't really read that proposal from Linus so if two people
> independently suggested the same thing, there must be something worth
> considering there :)

I'm thinking more and more about new DT only compatible driver, so there 
will be no problem with non-DT code ("no regression") and even about 
moving the old driver back to the platform. :) Just thinking aloud.

Regards,
-grygorii