On 4/22/24 10:57, Jerome Brunet wrote:
>
> On Mon 22 Apr 2024 at 09:16, Jerome Brunet <jbrunet@xxxxxxxxxxxx> wrote:
>
>> On Sun 21 Apr 2024 at 20:14, Krzysztof Kozlowski <krzk@xxxxxxxxxx> wrote:
>>
>>> On 20/04/2024 18:15, Jan Dakinevich wrote:
>>>>
>>>>
>>>> On 4/20/24 00:09, Rob Herring wrote:
>>>>> On Fri, Apr 19, 2024 at 03:58:10PM +0300, Jan Dakinevich wrote:
>>>>>> Add device tree bindings for A1 SoC audio clock and reset controllers.
>>>>>>
>>>>>> Signed-off-by: Jan Dakinevich <jan.dakinevich@xxxxxxxxxxxxxxxxx>
>>>>>> ---
>>>>>>
>>>>>> This controller has 6 mandatory and up to 20 optional clocks. To describe
>>>>>> this, I use 'additionalItems'. It produces correct processed-schema.json:
>>>>>>
>>>>>> "clock-names": {
>>>>>> "maxItems": 26,
>>>>>> "items": [
>>>>>> {
>>>>>> "const": "pclk"
>>>>>> },
>>>>>> {
>>>>>> "const": "dds_in"
>>>>>> },
>>>>>> {
>>>>>> "const": "fclk_div2"
>>>>>> },
>>>>>> {
>>>>>> "const": "fclk_div3"
>>>>>> },
>>>>>> {
>>>>>> "const": "hifi_pll"
>>>>>> },
>>>>>> {
>>>>>> "const": "xtal"
>>>>>> }
>>>>>> ],
>>>>>> "additionalItems": {
>>>>>> "oneOf": [
>>>>>> {
>>>>>> "pattern": "^slv_sclk[0-9]$"
>>>>>> },
>>>>>> {
>>>>>> "pattern": "^slv_lrclk[0-9]$"
>>>>>> }
>>>>>> ]
>>>>>> },
>>>>>> "type": "array",
>>>>>> "minItems": 6
>>>>>> },
>>>>>>
>>>>>> and it behaves as expected. However, the checking is followed by
>>>>>> complaints like this:
>>>>>>
>>>>>> Documentation/devicetree/bindings/clock/amlogic,a1-audio-clkc.yaml: properties:clock-names:additionalItems: {'oneOf': [{'pattern': '^slv_sclk[0-9]$'}, {'pattern': '^slv_lrclk[0-9]$'}]} is not of type 'boolean'
>>>>>>
>>>>>> And indeed, 'additionalItems' has boolean type in meta-schema. So, how to
>>>>>> do it right?
>>>>>
>>>>> The meta-schemas are written both to prevent nonsense that json-schema
>>>>> allows by default (e.g additionalitems (wrong case)) and constraints to
>>>>> follow the patterns we expect. I'm happy to loosen the latter case if
>>>>> there's really a need.
>>>>>
>>>>> Generally, most bindings shouldn't be using 'additionalItems' at all as
>>>>> all entries should be defined, but there's a few exceptions. Here, the
>>>>> only reasoning I see is 26 entries is a lot to write out, but that
>>>>> wouldn't really justify it.
>>>>
>>>> Writing a lot of entries don't scary me too much, but the reason is that
>>>> the existence of optional clock sources depends on schematics. Also, we
>>>
>>> Aren't you documenting SoC component, not a board? So how exactly it
>>> depends on schematics? SoC is done or not done...
>>>
>>>> unable to declare dt-nodes for 'clocks' array in any generic way,
>>>> because their declaration would depends on that what is actually
>>>> connected to the SoC (dt-node could be "fixed-clock" with specific rate
>>>> or something else).
>>>
>>> So these are clock inputs to the SoC?
>>>
>>
>> Yes, possibly.
>> Like an external crystal or a set clocks provided by an external codec
>> where the codec is the clock master of the link.
>>
>> This is same case as the AXG that was discussed here:
>> https://lore.kernel.org/linux-devicetree/20230808194811.113087-1-alexander.stein@xxxxxxxxxxx/
>>
>> IMO, like the AXG, only the pclk is a required clock.
>> All the others - master and slave clocks - are optional.
>> The controller is designed to operate with grounded inputs
>
> Looking again at the implementation of the controller, there is a clear
> indication in patch 3 that the controller interface is the same as the
> AXG and that the above statement is true.
> > The AXG had 8 master clocks wired in. The A1 just has 5 - and 3 grounded
> master clocks. This is why you to had to provide a mux input table to
> skip the grounded inputs. You would not have to do so if the controller was
> properly declared with the 8 master clock input, as it actually is.
>
For simplicity, I could make something like this in device tree:
clocks = <&clk0,
&clk1,
&clk2,
&clk3,
&clk4,
0,
0,
0>
clock-names = <"mst_in0",
"mst_in1",
"mst_in2"
"mst_in2"
"mst_in3"
"mst_in4"
"mst_in5"
"mst_in6"
"mst_in7">
But I don't see in the doc that the last 3 clocks are grounded to
anywhere. It will be just community's assumption about internals of the
controller.
Anyway, I still don't understand what to do with external slv_* clocks.
I can do the same as in example above: list slv_(s|lr)clk[0-9] in
"clock-names" and fill the rest if "clocks" by "0" phandles.
> It also shows that it is a bad idea to name input after what is coming
> in (like you do with "dds_in" or "fclk_div2") instead of what they
> actually are like in the AXG (mst0, mst1, etc ...)
>
I agree, these are not the best names.
>>
>>>>
>>>> By the way, I don't know any example (neither for A1 SoC nor for other
>>>> Amlogic's SoCs) where these optional clocks are used, but they are
>>>> allowed by hw.
>>
>> Those scenario exists and have been tested. There is just no dts using
>> that upstream because they are all mostly copy of the AML ref design.
>>
>>>>
>>>> This is my understanding of this controller. I hope, Jerome Brunet will
>>>> clarify how it actually works.
>>>
>>
>> I think the simpliest way to deal with this to just list all the clocks
>> with 'minItems = 1'. It is going be hard to read with a lot of '<0>,' in
>> the DTS when do need those slave clocks but at least the binding doc
>> will be simple.
>>
>>> Best regards,
>>> Krzysztof
>>
>> If you are going ahead with this, please name the file
>> amlogic,axg-audio-clkc.yaml because this is really the first controller
>> of the type and is meant to be documented in the same file.
>>
>> You are free to handle the conversion of the AXG at the same time if
>> you'd like. It would be much appreciated if you do.
>
>
--
Best regards
Jan Dakinevich
