On Fri, Nov 20, 2015 at 08:27:35PM +0200, Vladimir Zapolskiy wrote:

On 20.11.2015 16:13, Rob Herring wrote:

quoted

On Fri, Nov 20, 2015 at 03:29:52AM +0200, Vladimir Zapolskiy wrote:

quoted

For the purpose of better description of NXP LPC32xx GPIO controller
hardware in device tree format, extend the existing description with
device tree subnodes, which represent 6 GPIO banks within the
controller.

Note, client interface to the GPIO controller is untouched.

Signed-off-by: Vladimir Zapolskiy <vz@mleia.com>
---
 .../devicetree/bindings/gpio/gpio_lpc32xx.txt      | 121 ++++++++++++++++++++-
 1 file changed, 120 insertions(+), 1 deletion(-)

diff --git a/Documentation/devicetree/bindings/gpio/gpio_lpc32xx.txt b/Documentation/devicetree/bindings/gpio/gpio_lpc32xx.txt
index 4981936..d2da63c 100644
--- a/Documentation/devicetree/bindings/gpio/gpio_lpc32xx.txt
+++ b/Documentation/devicetree/bindings/gpio/gpio_lpc32xx.txt

@@ -15,7 +15,43 @@ Required properties:
    2) pin number
    3) optional parameters:
       - bit 0 specifies polarity (0 for normal, 1 for inverted)
-- reg: Index of the GPIO group
+- #address-cells: should be 2, which stands for GPIO bank id and
+  physical base address of this GPIO bank.

Now you need special code to do address translation. I'd really think 
twice about doing this.

Correct, address translation code is needed here...

quoted

Why do you need the bank number?

Only one reason -- backward compatibility in sense of referencing a GPIO
line on client's side. This API design is broken, I agree.

What client exactly? Between the dtb and kernel or kernel and userspace 
or ...?

Honestly I would prefer to get rid of this "feature", new code allows to
reference on client's side either a parent GPIO controller device node,
or bank nodes, probably the improvement can be done in a few steps?

  - this change,
  - convert clients to reference a GPIO bank directly,
  - remove root GPIO controller (e.g. make it "simple-bus") and convert
GPIO banks to "gpio-controller"s.

Can an evolution like this happen?

You generally don't want bindings to evolve.

quoted

quoted

+		gpio_p2: gpio-controller at 2 {
+			reg = <2 0x10 0x18>;
+			gpio-bank-name = "p2";
+			gpios = <13>;
+			gpio-no-output-state;
+		};
+
+		gpio_gpio: gpio-controller at 3 {
+			reg = <3 0x00 0x1C>;

This overlaps with bank 2.

Yes, it is. Thousand thanks to hardware designers.

Then you might want to split these into 2 regions. The problem is 
request_resource does not work with overlapping resources. Or just don't 
do subnodes. If there is not a lot of variations in the subnode data, 
then just leave that information in the kernel.

Rob