On Thu, Jun 18, 2015 at 02:33:48PM +0200, Ludovic Desroches wrote:

On Wed, Jun 17, 2015 at 09:55:56AM -0600, Stephen Warren wrote:

quoted

On 06/17/2015 06:38 AM, Ludovic Desroches wrote:

quoted

Hi Stephen,

On Mon, Jun 15, 2015 at 09:58:05AM -0600, Stephen Warren wrote:

quoted

On 06/10/2015 09:04 AM, Ludovic Desroches wrote:

quoted

When having a controller which allows per pin muxing, declaring with
which groups a function can be used is a useless constraint since groups
are something virtual.

This isn't true.

Irrespective of whether a particular piece of pinmux HW can control the mux
function for each pin individually, or only in groups, it's quite likely
that each function can only be selected onto a subset of those pins or
groups. Requiring the pinctrl driver to inform the core which set of
pins/groups particular functions can be selected onto seems quite
reasonable.

In my opinion at least, for HW that can select the mux function at the
per-pin level, the only sensible set of groups is one group per pin with
each group containing a single pin. Any other use of groups is a
SW/user-level construct, and is something unrelated to why the pinctrl
subsystem supports groups. If we want to represent those groups in pinctrl,
there should be two separate sets of groups; one to represent the actual HW
capabilities, and one to represent the SW/user-level convenience
abstractions.

Groups that I would like to use are not something related to the user or
software. It's an hardware reality but they would be more flexibles.

Usually the muxing is done by selecting a function (which seems to be
device related: usart, spi, etc.), then you select on which pins you
want this function.

In my case, I can't select a function only by choosing a mux. It is a
combination of the mux and the pin on which it is applied. So my
functions will be GPIO, A, B, C, etc. If I use function A on pin 12, I
will have my i2c clock signal but I can have this signal on pin 58 if I
use function C. Do you understand what I mean? It's not very easy to
explain... So here is a real example:

 --------------------------------------------------
|              | pio peripheral                    |
 --------------------------------------------------
| signal | dir | func | signal       | dir | ioset |
 --------------------------------------------------
| PA8    | I/O | A    | SDMMC0_DAT6  | I/O | 1     |
|        |     | B    | QSPI1_IO1    | I/O | 1     |
|        |     | D    | TCLK5        | I   | 1     |
|        |     | E    | FLEXCOM2_IO2 | I/O | 1     |
|        |     | F    | NWE/NANDWE   | O   | 2     |
 --------------------------------------------------
| PD28   | I/O | A    | SPI1_NPCS0   | I/O | 3     |
|        |     | B    | TDI          | I/O | 3     |
|        |     | C    | FLEXCOM2_IO2 | I   | 2     |
 --------------------------------------------------


You are right that having a group per pin is a solution.

If I follow your proposal (tell me if I'm wrong):
- I will have 128 groups since I have 128 pins.

Yes.

quoted

- I will have functions GPIO, A, B, C, D, E, F.

You could have functions A..F, and require the user to determine what option
they want for each pin, find the pin-specific mux function value for each
pin, and put that into the DT (or other pinmux data source). For example,
the bcm2835 pinctrl driver works this way.

In that case, each of the functions A..F could be selected on each pin, so
you'd have a very simple "get pins for function" implementation.

Alternatively, you could define a logic function per IO controller or signal
that gets pinmuxed. In the example above, FLEXCOM2_IO2 is one such example.
Given that set of functions, you'd need a mapping table to convert from the
logical functions seen by the pinctrl subsystem to the HW values that need
to be written into registers. For example, the Tegra pinctrl drivers work
this way.

In that case, each (pinctrl) function could only be selected on a specific
subset of all pins/groups, and so you'd probably need a table-based
implementation of "get pins for function".

Thanks for giving me some examples. Let's take a look at these drivers.
My concern is that I didnn't want to have many and/or big tables in my
driver. I will have to update it for a new SoC even if the pin
controller is the same. I prefer to have it in the device as we did
before and as some drivers continue to do.

quoted

quoted

- I have to give the groups which can achieve a function so I will have
128 groups for each function. It means 128 x 7 = 896 groups.

I don't think so no. I'm not sure why you'd consider multiplying 128 and 7
here. I'd expect 128 groups since you have 128 pins[1].

Sorry my mistake, I mean 896 possibilites since I have 128 groups for a
function.

quoted

Well, it's possible to have slightly more groups if, say, mux function is
selectable per pin, whereas something else like drive strength is configured
by a register that affects multiple pins at once. You'd need separate sets
of groups for muxing and for drive strength configuration. Some Tegra SoCs
are like this. Still, we just add the different sets of groups together
here, not multiply. The overall set of groups is not that much larger than
the set of pins.

quoted

Does it seems to be something reasonable and intelligible? From my point
of view no. And what about the sysfs readability?

With my current implementation, I have something quite understandable
for the user if he needs to check the pinmuxing:

  # cat /sys/kernel/debug/pinctrl/ahb\:apb\:pinctrl at fc038000/pinmux-pins
  pin 17 (PA17): (MUX UNCLAIMED) (GPIO UNCLAIMED)
  pin 18 (PA18): b0000000.sdio-host (GPIO UNCLAIMED) function E group sdmmc1_0
  pin 19 (PA19): b0000000.sdio-host (GPIO UNCLAIMED) function E group sdmmc1_0

  # cat /sys/kernel/debug/pinctrl/pinctrl-maps
  Pinctrl maps:

  device b0000000.sdio-host
  state default
  type MUX_GROUP (2)
  controlling device ahb:apb:pinctrl at fc038000
  group sdmmc1_0
  function E

  device b0000000.sdio-host
  state default
  type CONFIGS_PIN (3)
  controlling device ahb:apb:pinctrl at fc038000
  pin PA28
  config 00010003

  device b0000000.sdio-host
  state default
  type CONFIGS_PIN (3)
  controlling device ahb:apb:pinctrl at fc038000
  pin PA18
  config 00010003


Doing as you propose, I assume the result should be:

  # cat /sys/kernel/debug/pinctrl/ahb\:apb\:pinctrl at fc038000/pinmux-pins
  pin 17 (PA17): (MUX UNCLAIMED) (GPIO UNCLAIMED)
  pin 18 (PA18): b0000000.sdio-host (GPIO UNCLAIMED) function E group PA18
  pin 19 (PA19): b0000000.sdio-host (GPIO UNCLAIMED) function E group PA19

  # cat /sys/kernel/debug/pinctrl/pinctrl-maps
  Pinctrl maps:

  device b0000000.sdio-host
  state default
  type MUX_GROUP (2)
  controlling device ahb:apb:pinctrl at fc038000
  group PA28
  function E

  device b0000000.sdio-host
  state default
  type CONFIGS_PIN (3)
  controlling device ahb:apb:pinctrl at fc038000
  pin PA28
  config 00010003

  device b0000000.sdio-host
  state default
  type MUX_GROUP (2)
  controlling device ahb:apb:pinctrl at fc038000
  group PA18
  function E

  device b0000000.sdio-host
  state default
  type CONFIGS_PIN (3)
  controlling device ahb:apb:pinctrl at fc038000
  pin PA18
  config 00010003

I think it is more difficult to understand what is done here.

I don't think I agree. The HW level groups are the individual pins, so I
think the second option is clearer and more correct. What is the "sdmmc1_0"
group in the first example? Does any such thing even exist in HW?

If a user see this, I will have to take the datasheet to undersand what
using function E on this pin really means.

sdmmc1_0 group doesn't really exists as a group but a collection of
pins. The minimum requirement is to have CK, CMD and DAT0 signals. The
user can choose if he wants to use DAT1,2,3,4,5,6,7 and CD. Maybe he
would like to keep this pins available for another function.

This is what I have in my device tree.

pinctrl at fc038000 {
	group_defs {
		sdmmc1_0 {
			pins = <PIN_PA22__SDMMC1_CK>,
			       <PIN_PA28__SDMMC1_CMD>,
			       <PIN_PA18__SDMMC1_DAT0>,
			       <PIN_PA19__SDMMC1_DAT1>,
			       <PIN_PA20__SDMMC1_DAT2>,
			       <PIN_PA21__SDMMC1_DAT3>,
			       <PIN_PA30__SDMMC1_CD>;
		};
	};

	pinctrl_sdmmc1_default: sdmmc1_default {
		mux {
			function = "E";
			groups = "sdmmc1_0";
		};

		conf-cmd_data {
			pins = <PIN_PA28__SDMMC1_CMD>,
			       <PIN_PA18__SDMMC1_DAT0>,
			       <PIN_PA19__SDMMC1_DAT1>,
			       <PIN_PA20__SDMMC1_DAT2>,
			       <PIN_PA21__SDMMC1_DAT3>;
			bias-pull-up;
		};

		conf-ck_cd {
			pins = <PIN_PA22__SDMMC1_CK>,
			       <PIN_PA30__SDMMC1_CD>;
			bias-disable;
		};
	};
}

From my point of view, it is not so far from what generic pinconf does.
The only addition is the group_defs node because I don't want SoC
dependant tables in my driver.

+1 for not having SoC dependent tables in the driver. This additional
indirection makes the pinctrl drivers quite big and hard to follow.

However, I don't understand why you need the group_defs node. I can
understand why having this may be helpful in the code, but otherwise it
doesn't contain any new information. It basically collects all pins also
contained in the sdmmc1_default node, so this information can be
retrieved there. The "Function E" seems rather uninteresting since not
all pins have to be configured in function E. I bet it's possible to
implement card detection as gpio and then you'll have PIN_PA30__GPIO_X_Y
which is not function E.

So my take is: remove the group_defs node.

Note that in the recently introduced Mediatek pinctrl driver we used
'pinmux' for the property that you name 'pins' here. We probably want to
use the same name.

Sascha

-- 
Pengutronix e.K.                           |                             |
Industrial Linux Solutions                 | http://www.pengutronix.de/  |
Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |