Re: [PATCH v2 1/3] dt-bindings: pinctrl: Add bindings for pinctrl-mchp-sgpio driver

From: Lars Povlsen
Date: Mon Oct 05 2020 - 04:21:09 EST



Linus Walleij writes:

> Hi Lars,
>
> thanks for working on this!
>
> On Sun, Sep 13, 2020 at 9:11 PM Lars Povlsen <lars.povlsen@xxxxxxxxxxxxx> wrote:
>
>> > What I do not understand is why this GPIO controller is placed in the
>> > bindings of the pin controllers? Do you plan to add pin control
>> > properties to the bindings in the future?
>>
>> I have made provisions for some of the generic pinconf parameters, and
>> since the controller also has support for some alternate modes like
>> (syncronized) blink at various rates, I thought I better add it as
>> pinctrl straight away.
>
> OK fair enough let's keep the bindings here.
>
> BTW the latter function sounds like some kind of PWM?

Yes, it has PWM functionality as well.

>
>> >> + gpio-controller: true
>> >> +
>> >> + '#gpio-cells':
>> >> + description: GPIO consumers must specify four arguments, first the
>> >> + port number, then the bit number, then a input/output flag and
>> >> + finally the GPIO flags (from include/dt-bindings/gpio/gpio.h).
>> >> + The dt-bindings/gpio/mchp-sgpio.h file define manifest constants
>> >> + PIN_INPUT and PIN_OUTPUT.
>> >> + const: 4
>> >
>> > I do not follow this new third input/output flag at all.
>>
>> Its actually a sort of bank address, since the individual "pins" are
>> unidirectional.
>
> I'm a bit confused here...
> The standard advice for any "banked" GPIOs is to represent
> each "bank" as a separate node (with a corresponding gpio_chip
> in the Linux kernel). Then you can just use the standard
> bindings to pick a line from one of these nodes.

Yes, that seems to be a good model.

>
>> The PIN_INPUT/PIN_OUTPUT is defined in similar fashion in other pinctrl
>> binding header files... I can drop the define and use, but as it will be
>> used to address individual pins, I think it adds to readability.
>
> Hmmm. What makes these names expecially confusing is the
> Documentation/devicetree/bindings/pinctrl/pincfg-node.yaml defines:
> input-enable
> input-disable
> output-enable
> output-high
> output-low
>
> In the Linux kernel further there is:
> include/linux/pinctrl/pinconf-generic.h that defines:
> PIN_CONFIG_INPUT_ENABLE
> PIN_CONFIG_OUTPUT_ENABLE
> PIN_CONFIG_OUTPUT
>
> Since you are using the pin control framework this gets really
> hard to hash out.
>

Yes, as the pins are fixed-function, the "input-enable", "input-disable"
and "output-enable" are not really useful.

> I don't really understand why it is needed.
>
>> Like this (excerpts from a DT with a switchdev driver using SFP's and
>> LED's on sgpio):
>>
>> /{
>> leds {
>> compatible = "gpio-leds";
>> led@0 {
>> label = "eth60:yellow";
>> gpios = <&sgpio1 28 0 PIN_OUTPUT GPIO_ACTIVE_LOW>;
>> default-state = "off";
>> };
>> ...
>> };
>> };
>
> If what you intend to achieve is to make the GPIO come up in output mode,
> you can either just have the driver do that as needed by the consumer.
> If you absolutely have to do it in the device tree, then implement
> pin control (pin config) and have it something like this:
>
> leds {
> compatible = "gpio-leds";
> pinctrl-names = "default";
> pinctrl-0 = <&my_led_pinctrl>;
> led@0 {
> label = "eth60:yellow";
> gpios = <&sgpio1 28 GPIO_ACTIVE_LOW>;
> default-state = "off";
> };
> ...
>
> my_led_pinctrl: pinctrl-led {
> pins = "gpio95"; // Just an example way of referring to the pin
> bias-disable;
> output-enable;
> };
> };

No, the PIN_OUTPUT is purely for adressing. But as you suggested, I'll
split the into separate nodes. That will eliminate the "PIN_OUTPUT" and
the bindings header.

>
>> >> + microchip,sgpio-port-ranges:
>> >> + description: This is a sequence of tuples, defining intervals of
>> >> + enabled ports in the serial input stream. The enabled ports must
>> >> + match the hardware configuration in order for signals to be
>> >> + properly written/read to/from the controller holding
>> >> + registers. Being tuples, then number of arguments must be
>> >> + even. The tuples mast be ordered (low, high) and are
>> >> + inclusive. Arguments must be between 0 and 31.
>> >> + $ref: /schemas/types.yaml#/definitions/uint32-array
>> >> + minItems: 2
>> >> + maxItems: 64
>> >
>> > And you are *absolutely sure* that you can't just figure this out
>> > from the compatible string? Or add a few compatible strings for
>> > the existing variants?
>>
>> Yes, this really needs to be configured for each board individually -
>> and cant be probed. It defines how the bitstream to/from the shift
>> registers is constructed/demuxed.
>
> And you have considered the option of simply letting the driver
> check which board we are then? The property at the very
> top of the device tree.
>
> if (of_machine_is_compatible("my_board")) {
> ....
> } else if (of_machine_is_compatible("my_other_board")) {
> ....
> }

No, board-specific code is undesireable, as our customers should be able
to design own boards without driver changes.

>
> So that you simply use the board compatible string to determine
> this?
>
>> >> +/* mchp-sgpio specific pin type defines */
>> >> +#undef PIN_OUTPUT
>> >> +#undef PIN_INPUT
>> >> +#define PIN_OUTPUT 0
>> >> +#define PIN_INPUT 1
>> >
>> > I'm not a fan of this. It seems like something that should be set in
>> > response to the gpiochip callbacks .direction_input and
>> > .direction_output callbacks.
>> >
>>
>> As I tried to explain above, its a part of the pin address - aka bank
>> selector - whether your are accessing the input or the output side. And
>> since the directions have totally different - and concurrent - use, they
>> need to be individually addressed, not "configured".
>>
>> In the example presented, sgpio2-p28b0 IN is loss-of-signal, and the
>> OUT is the sfp tx-disable control.
>
> I suspect the proper way to do it is to create one node for
> the input side and one node for the output side and also create
> two different gpio chips in the kernel.
>
> my-device {
> compatible = "my-device";
> gpioin: input-gpio {
> ....
> };
> gpioout: output-gpio {
> ....
> };
> };
>
> Note: I didn't think over the naming in this example.
>
> You will need code in your driver to parse the subnodes and
> populate two gpio_chips.

Yes, I will modify the driver to use separate nodes for each direction.

Thank you for your comments, it is highly appreciated.

---Lars

>
> Yours,
> Linus Walleij

--
Lars Povlsen,
Microchip