[linux-sunxi] Re: [PATCH 7/7] ARM: sun7i: cubietruck: enable bluetooth module

[Chen-Yu Tsai]

Hi,

On Wed, Apr 16, 2014 at 5:44 PM, Maxime Ripard
<maxime.ripard at free-electrons.com> wrote:
> Hi,
>
> Please try to keep me in CC, even though the ML doesn't make it easy..

Sorry about that.

> On Wed, Apr 16, 2014 at 12:06:59AM +0800, Chen-Yu Tsai wrote:
>> >> @@ -139,4 +152,16 @@
>> >>       reg_usb2_vbus: usb2-vbus {
>> >>               status = "okay";
>> >>       };
>> >> +
>> >> +     rfkill_bt {
>> >> +             compatible = "rfkill-gpio";
>> >> +             pinctrl-names = "default";
>> >> +             pinctrl-0 = <&bt_pwr_pin_cubietruck>, <&clk_out_a_pins_a>;
>> >> +             clocks = <&clk_out_a>;
>> >> +             clock-frequency = <32768>;
>> >> +             gpios = <&pio 7 18 0>; /* PH18 */
>> >> +             gpio-names = "reset";
>> >> +             rfkill-name = "bt";
>> >> +             rfkill-type = <2>;
>> >> +     };
>> >
>> > Hmmm, I don't think that's actually right.
>> >
>> > If you have such a device, then I'd expect it to be represented as a
>> > full device in the DT, probably with one part for the WiFi, one part
>> > for the Bluetooth, and here the definition of the rfkill device that
>> > controls it.
>>
>> The AP6210 is not one device, but 2 separate chips in one module. Each
>> chip has its own controls and interface. They just so happen to share
>> the same enclosure. Even 2-in-1 chips by Broadcom have separate controls
>> and interfaces. The WiFi side is most likely connected via SDIO, while
>> the Bluetooth side is connected to a UART, and optionally I2S for sound.
>
> It's even easier to represent then.
>
>> > But tying parts of the device to the rfkill that controls it, such as
>> > the clocks, or the frequency it runs at seems just wrong.
>>
>> I understand where you're coming from. For devices on buses that require
>> drivers (such as USB, SDIO) these properties probably should be tied to
>> the device node.
>>
>> For our use case here, which is a bluetooth chip connected on the UART,
>> there is no in kernel representation or driver to tie them to. Same goes
>> for UART based GPS chips. They just so happen to require toggling a GPIO,
>> and maybe enabling a specific clock, to get it running. Afterwards,
>> accessing it is done solely from userspace. For our Broadcom chips, the
>> user has to upload its firmware first, then designate the tty as a Bluetooth
>> HCI using hciattach.
>>
>> We are using the rfkill device as a on-off switch.
>
> I understand your point, but the fact that it's implemented in
> user-space, or that UART is not a bus (which probably should be), is
> only a Linux specific story, and how it's implemented in Linux (even
> if the whole rfkill node is another one, but let's stay on topic).

I gave it some thought last night. You are right. My whole approach
is wrong. But let's try to make it right.

So considering the fact that it's primarily connected to a UART,
maybe I should make it a sub-node to the UART node it's actually
connected to? Something like:

        uart2: serial at 01c28800 {
                pinctrl-names = "default";
                pinctrl-0 = <&uart2_pins_a>;
                status = "okay";

                bt: bt_hci {
                        compatible = "brcm,bcm20710";
                        /* maybe add some generic compatible */
                        pinctrl-names = "default";
                        pinctrl-0 = <&clk_out_a_pins_a>,
<&bt_pwr_pin_cubietruck>;
                        clocks = <&clk_out_a>;
                        clock-frequency = <32768>;
                        gpios = <&pio 7 18 0>; /* PH18 */
                };
        };

And let the uart core handle power sequencing for sub-nodes.

The rfkill node would still have the gpios and clocks, but not the
clock-frequency property. It's sole purpose would be to toggle the
controls. But I think the placement is still odd. Perhaps these
virtual devices shouldn't live in the DT at all.

> This is a huge abstraction leak.
>
> Let's say you need the I2S stream you mentionned for some
> reason. Would you tie the audio stream to the rfkill node as well?
> I'm sorry, but from an hardware description perspective, it makes no
> sense.

The above revision should be better, from a hardware perspective. I'm
not sure how to tie in the I2S stream, and there I haven't found any
examples in the DT tree.

> What's the feeling of the DT maintainers?


Cheers

ChenYu