Re: [PATCH 2/3] dt: rfkill-gpio: add bindings documentation

Am 13.02.2012 20:43, schrieb Olof Johansson:

On Mon, Feb 13, 2012 at 5:47 AM, Rob Herring[off-list ref]  wrote:

quoted

On 02/12/2012 02:21 PM, Simon Glass wrote:


On Sun, Feb 12, 2012 at 11:13 AM, Marc Dietrich[off-list ref]  wrote:

quoted

quoted

Add device tree bindings information for rfkill gpio switches.

Cc: linux-wireless@vger.kernel.org
Cc: "John W. Linville"<redacted>
Cc: Johannes Berg<johannes@sipsolutions.net>
Cc: Rhyland Klein<redacted>
Cc: Grant Likely<redacted>
Cc: devicetree-discuss@lists.ozlabs.org
Signed-off-by: Marc Dietrich<marvin24@gmx.de>
---
  Documentation/devicetree/bindings/gpio/rfkill.txt |   38 +++++++++++++++++++++
  1 files changed, 38 insertions(+), 0 deletions(-)
  create mode 100644 Documentation/devicetree/bindings/gpio/rfkill.txt

diff --git a/Documentation/devicetree/bindings/gpio/rfkill.txt b/Documentation/devicetree/bindings/gpio/rfkill.txt
new file mode 100644
index 0000000..22bf22a
--- /dev/null
+++ b/Documentation/devicetree/bindings/gpio/rfkill.txt

@@ -0,0 +1,38 @@
+RFKILL switches connected to GPIO lines
+
+Required properties:
+- compatible : should be "rfkill-gpio".
+
+Each rfkill switch is represented as a sub-node of the rfkill-gpio device.
+Each node has a label property which represents the name of the corresponding
+rfkill device.
+
+RFKILL sub-node properties:
+- label :  (optional) The label for this rfkill switch.  If omitted, the label is
+  taken from the node name (excluding the unit address).
+- reset-gpio, shutdown-gpio :  Should specify the rfkill gpios for reset and
+  shutdown (see "Specifying GPIO information for devices" in

Should that be reset-gpios, shutdown-gpios? Even though you have only
one it seems that people put an 's' on the end.

Agreed.

quoted

quoted

quoted

+  Documentation/devicetree/booting-without-of.txt).
+- type : enumerated type of the gpio (see include/linux/rfkill.h).

It would be better I think if this were explicit here. If you have a
number, then what values does it take and what do they mean?

This should most likely be moved to a set of properties instad of an
enumerated type, I agree. And/or use a string to encode the type
simiar to how powerpc does some of the USB interfaces.

quoted

quoted

quoted

+- clock : (optional) name of the clock name associated with the rfkill switch

Can this be a phandle instead of a string?

This seems to be in the wrong place altogether. The gpio controller
would have a clock, not particular gpio line.

And either way, this should conform to the standard clock binding, not
use something locally hacked up.

quoted

quoted

quoted

+  (see include/linux/rfkill-gpio.h)

IMO device tree bindings should be fully documented in this file,
rather than needing to look at a separate header. This is particularly
true if the binding is used in another project.

Correct. A binding should not be Linux specific. It should describe the h/w.

quoted

quoted

+
+Examples:
+
+rfkill-switches {
+       compatible = "rfkill-gpio";
+
+       wifi {
+               label = "wifi";
+               reset-gpio =<&gpio 25 0>; /* Active high */
+               shutdown-gpio =<&gpio 85 0>; /* Active high */
+               type =<1>;
+       };
+
+       bt {
+               label = "bluetooth";
+               reset-gpio =<&gpio 17 0>; /* Active high */
+               shutdown-gpio =<&gpio 35 0>; /* Active high */
+               type =<1>;
+       };

Why wouldn't the gpio lines just be part of the bt and wifi device nodes
themselves? The DT is supposed to describe h/w connections.

The thing is, that "rfkill" isn't a _device_, and Marc is trying to
describe it as one. It's really just a software abstraction of a
collection of power supplies and/or GPIO lines that are used to power
up/down specific peripherals.

Oh, this gets more complicated than I wanted. First I only translated 
the *existing* platform_data to devicetree format. I understand that 
devicetree needs some more careful thoughts. Let me shortly describe my 
hw setup and lets see how it can me mapped.

I know that the USB modem, for example, is probed through autoprobing
the USB bus. So there's no device to associate it with, per se. But
the USB slot that the modem is connected to, which is also the
connector that the GPIO controls the power supplies and reset line to,
are connected directly to one of the USB host controllers, right? So
maybe describing it there is a better option.

The wifi module (some models also include an additional bluetooth part) 
is plugged into a mini-pci-e like slot. This has pins for usb (the 
module uses usb), dap, sdio, uart, and some other pins which also 
include the gpio lines for rfkill. The other gpio end is directly 
connected to the tegra soc.

So there is no usb slot and a standard usb slot also does not contain 
any gpio lines (just 2 power and 2 data lines). The usb controller also 
sits in the tegra soc (well, more or less because it is the ulpi port 
connected to SMC3315 and later a SMC2512 4-port usb hub). So gpio and 
usb take a totally different path. IMHO, it makes no sense to add a gpio 
to the usb controller, but see below.

That still leaves the issue of actually having something to bind it
against. As I already said, rfkill isn't a device, so crafting one
just because linux wants one is the wrong way to go about. Maybe using
/chosen to refer to the device nodes for the GPIO lines under USB
instead, and have rfkill look for those and create a device if they're
found is a better way to go about it.

To me it looks more like how the LEDs are implemented but with a 
"linux,rfkill_type" property. To me this looks saner than adding a gpio 
to an usb controller, even if there if no physical rfkill device. This 
still leaves the problem of what to do with the clock (it is not used by 
any machine AFAIK).

The main problem here is that usb devices have no device tree entries 
(only the host controllers). Perhaps it is possible to a fake usb device 
as a workaround? This has the advantage that it exists physically, but 
the resources are not taken from the device tree. Instead all other non 
standard properties (gpios, clocks) are.

Marc