[PATCH v2 3/6] dt/bindings: Add bindings for Tegra GMI controller

[Mirza Krak]

2016-08-30 19:06 GMT+02:00 Rob Herring <robh at kernel.org>:
> On Wed, Aug 24, 2016 at 09:54:47PM +0200, Mirza Krak wrote:
>> 2016-08-24 17:56 GMT+02:00 Jon Hunter <jonathanh at nvidia.com>:
>> +
>> >> +Example with two SJA1000 CAN controllers connected to the GMI bus. We wrap the
>> >> +controllers with a simple-bus node since they are all connected to the same
>> >> +chip-select (CS4), in this example external address decoding is provided:
>> >> +
>> >> +gmi at 70090000 {
>> >> +     compatible = "nvidia,tegra20-gmi";
>> >> +     reg = <0x70009000 0x1000>;
>> >> +     #address-cells = <1>;
>> >> +     #size-cells = <1>;
>> >> +     clocks = <&tegra_car TEGRA20_CLK_NOR>;
>> >> +     clock-names = "gmi";
>> >> +     resets = <&tegra_car 42>;
>> >> +     reset-names = "gmi";
>> >> +     ranges = <4 0x48000000 0x7ffffff>;
>> >> +
>> >> +     status = "disabled";
>> >> +
>> >> +     bus at 4 {
>> >> +             compatible = "simple-bus";
>> >> +             reg = <4>;
>> >> +             #address-cells = <1>;
>> >> +             #size-cells = <1>;
>> >> +             ranges = <0 4 0x40100>;
>> >
>> > Does this work? I tried to add an example like this and I got ...
>> >
>> > Warning (reg_format): "reg" property in /gmi at 70009000/bus at 4 has invalid
>> > length (4 bytes) (#address-cells == 1, #size-cells == 1)
>>
>> Shoot, to get rid of the warning it should be
>>
>> reg = <4 0 >;
>>
>> But it works either way.
>
> The CS node should have #address-cells=2 with the first being CS# and
> the second being the offset (often 0).
>
>>
>> >
>> > I am wondering if we should just following the arm,pl172 example and
>> > have ...
>> >
>> >         cs4 {
>> >                 compatible = "simple-bus";
>> >                 #address-cells = <1>;
>> >                 #size-cells = <1>;
>> >                 ranges;
>
> Empty ranges is typically wrong and due to laziness...
>
> This should have the CS# in it.
>
>> >
>> >                 nvidia,snor-cs = <4>;
>> >                 nvidia,snor-mux-mode;
>> >                 nvidia,snor-adv-inv;
>> >
>> >                 can at 0 {
>> >                         reg = <0 0x100>;
>
> This can be 1 cell with just the offset.
>
>> >                         ...
>> >                 };
>> >
>> >                 ...
>> >         };
>> >
>>
>> That means to go back to V1 really (almost :)). Which I do not mind.
>> Will give it a test run.
>>
>> But I am a little hesitant if will be any better/cleaner. In your example above:
>>
>> can at 0 {
>>          reg = <0 0x100>;
>>          ...
>> };
>>
>> Would this really translate correctly? In the pl172 example they have
>> multiple ranges and address with "flash at 0,0" which a range defined in
>> parent node. "can at 0" does not have valid match in parent node in our
>> example. So I probably need add some more logic for it to properly
>> translate.
>
> pl172 has several things I don't like, so don't follow it. Mainly those
> are custom CS property and 3 levels of nodes. I'm fine with 3 levels if
> there is more than one device, but otherwise 2 levels with timing
> properties in the child device node.
>
>
>>
>> I have an idea which is following:
>>
>> gmi at 70090000 {
>>          status = "okay";
>>          #address-cells = <2>;
>>          #size-cells = <1>;
>>          ranges = <4 0 0x48000000 0x00040000>;
>>
>>          cs4 {
>
> cs at 4,0
>
>>                  compatible = "simple-bus";
>>                  #address-cells = <2>;
>
> 1 cell here.
>
>>                  #size-cells = <1>;
>>                  ranges;
>
> Fill this in to drop the 2nd cell on child addresses and just have the
> offset.
>
>>
>>                  nvidia,snor-cs = <4>;
>
> NAK, no custom CS properties.
>
>>                  nvidia,snor-mux-mode;
>>                  nvidia,snor-adv-inv;
>>
>>                  can at 0 {
>>                          compatible = "nxp,sja1000";
>>                          reg = <4 0 0x100>;
>>                          ...
>>                  };
>>
>>
>>                  can at 40000 {
>>                          compatible = "nxp,sja1000";
>>                          reg = <4 0x40000 0x100>;
>>                          ...
>>                  };
>>          };
>> };
>>

Thank you for your review Rob.

Taking your comments in to account I end up with this:

gmi at 70090000 {
        compatible = "nvidia,tegra20-gmi";
        reg = <0x70009000 0x1000>;
        #address-cells = <2>;
        #size-cells = <1>;
        clocks = <&tegra_car TEGRA20_CLK_NOR>;
        clock-names = "gmi";
        resets = <&tegra_car 42>;
        reset-names = "gmi";
        ranges = <4 0 0xd0000000 0xfffffff>;

        status = "okay";

        bus at 4,0 {
                compatible = "simple-bus";
                #address-cells = <1>;
                #size-cells = <1>;
                ranges = <0 4 0 0x40000>;

                nvidia,snor-mux-mode;
                nvidia,snor-adv-inv;

                can at 0 {
                        reg = <0 0x100>;
                        ...
                };

                can at 40000 {
                        reg = <0x40000 0x100>;
                        ...
                };
        };
};

Have I understood you correct?

Also wanted to verify the example case where you only have on device
connected to one CS#, from what I see in other implementations it
seems OK to put the CS# in the reg property in that case. Is this
correct?

Example with one SJA1000 CAN controller connected to the GMI bus
on CS4:

gmi at 70090000 {
        compatible = "nvidia,tegra20-gmi";
        reg = <0x70009000 0x1000>;
        #address-cells = <2>;
        #size-cells = <1>;
        clocks = <&tegra_car TEGRA20_CLK_NOR>;
        clock-names = "gmi";
        resets = <&tegra_car 42>;
        reset-names = "gmi";
        ranges = <4 0 0xd0000000 0xfffffff>;

        status = "okay";

        can at 4,0 {
                reg = <4 0 0x100>;
                nvidia,snor-mux-mode;
                nvidia,snor-adv-inv;
                ...
        };
};

Jon, to be able to handle both cases in the driver we would first
attempt to decode the CS# from the ranges property, and fallback to
reg property if no ranges are defined. Does that sound reasonable?

Best Regards
Mirza