[U-Boot] [PATCH 1/4] dm: gpio: extend gpio api by dm_gpio_set_pull()

[Simon Glass]

Hi,

On 23 February 2015 at 03:51, Przemyslaw Marczak <p.marczak at samsung.com> wrote:
>
> Hello Simon,
>
>
> On 02/20/2015 08:29 PM, Simon Glass wrote:
>>
>> Hi,
>>
>> On 20 February 2015 at 10:50, Stephen Warren <swarren at wwwdotorg.org> wrote:
>>>
>>> On 02/20/2015 02:34 AM, Przemyslaw Marczak wrote:
>>>>
>>>>
>>>> Hello,
>>>>
>>>> On 02/19/2015 06:09 PM, Stephen Warren wrote:
>>>>>
>>>>>
>>>>> On 02/19/2015 05:11 AM, Przemyslaw Marczak wrote:
>>>>>>
>>>>>>
>>>>>> Hello,
>>>>>>
>>>>>> On 02/18/2015 05:39 PM, Stephen Warren wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 02/17/2015 10:01 PM, Simon Glass wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> +Stephen who might have an opinion on this.
>>>>>>>>
>>>>>>>> Hi Przemyslaw,
>>>>>>>>
>>>>>>>> On 17 February 2015 at 06:09, Przemyslaw Marczak
>>>>>>>> <p.marczak at samsung.com> wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> This commits extends:
>>>>>>>>> - dm gpio ops by: 'set_pull' call
>>>>>>>>> - dm gpio uclass by: dm_gpio_set_pull() function
>>>>>>>>>
>>>>>>>>> The pull mode is not defined so should be driver specific.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> It's good to implement this, but I think you should try to have a
>>>>>>>> standard interface. You could define the options you want to support
>>>>>>>> and pass in a standard value.
>>>>>>>>
>>>>>>>> Otherwise we are not really providing a driver abstraction, only an
>>>>>>>> interface.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I don't think that pull is a GPIO-related function/property. At
>>>>>>> least on
>>>>>>> Tegra, the GPIO controller allows you to set the pin direction and the
>>>>>>> output value and that's it. Configuring pull-up/down and other IO
>>>>>>> related properties is done in the pinmux controller instead. I don't
>>>>>>> think we want a standard API that has to touch both HW modules at once.
>>>>>>> What common code needs to manipulate a GPIO's pull-up/down setting? As
>>>>>>> precedent observe that pull-up/down isn't part of the Linux kernel's
>>>>>>> GPIO API, but rather that's part of the SoC-specific pinctrl driver,
>>>>>>> which controls pinmuxing etc.
>>>>>>>
>>>>>>
>>>>>> This is a quite different than in the Exynos, where all the gpio
>>>>>> functions and properties can be set by few registers within range of
>>>>>> each gpio port base address. So in this case we don't touch another
>>>>>> hardware module, we modify one of available gpio related registers.
>>>>>>
>>>>>> Anyway, if we want to have a single and common gpio API in the future,
>>>>>> then I think it is better to add pull option.
>>>>>
>>>>>
>>>>>
>>>>> Why? I'll ask again: What common driver code needs to manipulate
>>>>> pull-ups?
>>>>
>>>>
>>>>
>>>> Please look at driver: drivers/gpio/s5p_gpio.c
>>>>
>>>> It's one driver related to one gpio hardware submodule and it takes care
>>>> about standard gpio properties and also mux/pull/drv/rate.
>>>>
>>>> And the exynos pinmux code is only a software abstraction:
>>>> arch/arm/cpu/armv7/exynos/pinmux.c
>>>
>>>
>>>
>>> I didn't want to ask which driver implements the control of pullups, but
>>> rather which other driver needs to turn pullups on/off in a standard way
>>> across multiple SoCs.
>>>
>>> In other words, do you expect code in common/ to need to call a "set pin
>>> pullup" function? If so, then we certainly need a standard API to manipulate
>>> pullups. However if no common code needs to manipulate pullups, then I'd
>>> argue we don't actually need a common API to do this, since there's no code
>>> that would call that common API.
>>
>>
>> We do currently use the GPIO to handle pullup/pulldown for some boards
>> so until we have a pinmux API (which might be a long while) it seems
>> reasonable for it to live there.
>>
>> If not, does anyone plan to write such an API?
>>
>
> Right, we uses this in most Exynos boards. But the boards uses direct calls to s5p gpio driver, without uclass.
> I wonder if wouldn't it be better and faster to leave the board low-level init routines as they are now.
>
>
>>>
>>>
>>>>>   > And the driver will
>>>>>>
>>>>>>
>>>>>> implement what is required, instead of provide common and private api
>>>>>> for each driver.
>>>>>
>>>>>
>>>>>
>>>>> I'm not proposing driver-specific APIs, but rather having a common GPIO
>>>>> API and a common pinmux API. They need to be different since different
>>>>> HW modules may implement the functionality.
>>>>>
>>>>
>>>> As in the above example, for the Exynos it's the one hw module, so it's
>>>> simply.
>>>>
>>>>>> For the various devices it is unclear, what should be pinmux and what
>>>>>> should be gpio driver.
>>>>>
>>>>>
>>>>>
>>>>> How about the following are GPIO:
>>>>> * Set GPIO pin direction
>>>>> * Read GPIO input
>>>>> * Set GPIO output value
>>>>>
>>>>> ... and anything else is pinmux. That's the split in Linux and AFAIK it
>>>>> works out fine.
>>>>>
>>>>> It'd be perfectly fine for the same driver code to implement both a GPIO
>>>>> and a pinmux driver, if the HW supports it.
>>>>>
>>>>
>>>> Ok, I can drop this commit, since the current code works fine.
>>>>
>>>>>> Moreover in my opinion from the single external pin point of view the
>>>>>> pull up/down is the property of that pin.
>>>>>
>>>>>
>>>>>
>>>>> It's a property of the same pin, but semantically it's not manipulating
>>>>> a GPIO-related function.
>>>>>
>>>>>> Actually for Exynos, the pinmux is an abstraction and uses only GPIO
>>>>>> driver api in U-Boot.
>>>>>>
>>>>>> Do we need pinmux class?
>>>>>>
>>>>>> Best regards,
>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>> As I wrote in one of my previous e-mail, I was testing eMMC detect.
>>>> And setting the pull was required for this, before call the pinmux for
>>>> eMMC pins.
>>>> But finally the eMMC detect seem to be not useful in case of the present
>>>> 'mmc rescan' command.
>>>
>>>
>>>
>>> Why wouldn't the pinmux driver for the whole system simply apply the board's
>>> whole pinmux configuration before initializing any IO controller drivers? IO
>>> controller drivers shouldn't have to initialize board-/SoC-specific pinmux,
>>> but the board-/SoC-specfic code should do so.
>>>
>>> At most, the eMMC driver should call a function such as pinmux_emmc(), and
>>> the board/SoC code should implement that as appropriate for that board. The
>>> eMMC driver shouldn't have to know about applying specific pullups/downs to
>>> specific pins (since those settings and pins are likely board-/SoC-specific,
>>> and drivers shouldn't know about board-/SoC-specific details). The only
>>
>>
>> No this way lies madness. It is how things work on Jetson and Nyan.
>> Loads of opaque tables and no idea what the pins are connected to. It
>> has some value for pins that U-Boot doesn't use (so we are just
>> setting them up for Linux) but even then it is really opaque.
>>
>> We can't even sent patches to the file because it is auto-generated
>> from a tool in another repo. Tiny differences between boards are
>> hidden because we repeat all the information again with just a line or
>> two of changes. I really don't want exynos to go that way.
>>
>>> exception would be if the standard IO protocol requires pullups to be
>>> changed during regular operation. In which case, a specific callback from
>>> the driver could be added for each protocol-mandated configuration change,
>>> thus keeping the IO controller driver still completely isolated from details
>>> of the pins and pinmux APIs etc.
>>
>>
>> This is like the 'funcmux' in Tegra I think. I think this is more
>> useful and we should use it to set up all peripheral pins. We can
>> review the code, see changes, understand what they relate to, etc.
>>
>> Anyway this all seems off-topic from this patch.
>>
>> Unless someone plans to write a pinmux subsystem for U-Boot (which I
>> agree would be better) I think the general approach of this patch is
>> good.
>>
>> Regards,
>> Simon
>>
>
> Ok, so there are two next versions of this patch-set.
> Please decide, which one is better.
>
> For me, at present, the current s5p_gpio api works fine for all the exynos based boards.
> Introducing the pinmux uclass is not a quick task, now I'm trying to focus on pmic.

OK, then I think we should probably leave it as it is. If we add
pull-ups to driver model it should be done with pinctl as Stephen
says. I doubt this is a huge task, since we can likely port over the
code from Linux. But for now I think we should keep with the s5p API
until someone takes on pinctl.

Regards,
Simon