On 07/16/2013 02:47 AM, Christian Ruppert wrote:

On Wed, Jul 10, 2013 at 01:27:52PM -0600, Stephen Warren wrote:

quoted

On 07/08/2013 07:02 AM, Christian Ruppert wrote:
...

quoted

OK, a small drawing of our hardware should make this clear, let's take
an imaginary example of one port with 10 pins, one i2c interface, one
spi interface and one GPIO bank:

              | mux N-1|
              +........+
              |        |  2
              |        +--/-- i2c
              |        |
           10 |        |  4
   Pins  --/--+ mux N  +--/-- spi
              |        |
              |        |  10
              |        +--/-- GPIO
              |        |
              +........+
              | mux N+1|

This example shows the mux N inside the pin controller. It controls
all pins associated to port N through a single register value M. Let's
assume the pins are configured as follows in function of the register
value:

 pin      M=0       M=1     M=2      M=3
  0      GPIO0   SPI_MISO  GPIO0   SPI_MISO
  1      GPIO1   SPI_MOSI  GPIO1   SPI_MOSI
  2      GPIO2    SPI_CK   GPIO2    SPI_CK
  3      GPIO3    SPI_CS   GPIO3    SPI_CS
  4      GPIO4    GPIO4    GPIO4    GPIO4
  5      GPIO5    GPIO5    GPIO5    GPIO5
  6      GPIO6    GPIO6    GPIO6    GPIO6
  7      GPIO7    GPIO7    GPIO7    GPIO7
  8      GPIO8    GPIO8   I2C_SDA  I2C_SDA
  9      GPIO9    GPIO9   I2C_SCL  I2C_SCL

In that scenario, in the language of Linux's pinctrl subsystem, what you
have is:

10 pins, named 0..9
1 pin group, named perhaps "mux N".
4 different functions; values M==0, 1, 2, 3.

quoted

We now have three pin groups defined, corresponding to the chip-side
ports of the pin controller:
GPIO = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}
SPI  = {0, 1, 2, 3}
I2C  = {8, 9}

You would usually only define pin groups for the pin/ball/package side
of the pinmux HW. IIRC, you're also wanting to define pin groups for the
intra-chip side of the pinmux HW. However, you're not muxing functions
onto those pingroups; they're just there to help with naming the
GPIO<->pinmux mapping. You only mux functions onto the pin/ball/package
side pins/pingroups.

Well, the GPIO<->pinmux mapping is not the only reason for defining
these groups wrt. the chip-side of the pin controller. The other reasons
are:
  - Make different interfaces on the same MUX orthogonal wrt. each
    other, i.e. make it possible to request one independently of the
    other. In the example above, SPI and I2C can be requested completely
    independently and the pin controller driver decides which mode to
    use.

But the pinctrl subsystem and bindings don't have any concept of that;
what gets requested is the pins on the chip, or the "outer" side of the
pin controller. There's no concept of resource management for the
"inside" of the pin controller.

  - Make pin allocation more fine-grained (in the example above, only
    pins 0-4 are "allocated" in case SPI is requested). This makes
    GPIO<->interface pin conflict management more natural.

I think you'd want to either:

a) Just deal with this in the driver; it knows the HW, and it knows
which mux function is selected for each mux, and hence knows exactly
which pins can be requested as GPIOs for each combination, and can
therefore allow/disallow any GPIO request or mux function change.

b) Extend the pinctrl core to know about this explicitly, and pass
information to the pinctrl core. Presumably, for each combination of
(pingroup, mux function), you'd need a list or bitmask indicating which
pins within the pingroup are actually used. Then, the pinctrl core can
perform all the validation. If you do this, you don't need to invent new
pinctrl groups in order to try and shoe-horn this into pinctrl.