Re: [PATCH v2 2/6] dt-bindings: pwm: amlogic: add new compatible for meson8 pwm type

[Krzysztof Kozlowski]

On 22/11/2023 16:23, Jerome Brunet wrote:
> 
> On Wed 22 Nov 2023 at 16:04, Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx> wrote:
> 
>> On 22/11/2023 15:34, Jerome Brunet wrote:
>>>
>>> On Wed 22 Nov 2023 at 09:37, Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx> wrote:
>>>
>>>> On 20/11/2023 11:04, Jerome Brunet wrote:
>>>>>>>>>>    .../devicetree/bindings/pwm/pwm-amlogic.yaml  | 36 +++++++++++++++++--
>>>>>>>>>>    1 file changed, 34 insertions(+), 2 deletions(-)
>>>>>>>>>>
>>>>>>>>> Reviewed-by: Rob Herring <robh@xxxxxxxxxx>
>>>>>>>>>
>>>>>>>>
>>>>>>>> I'm puzzled, isn't it recommended to have a per-soc compatible now ?
>>>>
>>>> Yes, it is.
>>>>
>>>>>>> I have specifically addressed this matter in the description,
>>>>>>> haven't I ? What good would it do in this case ?
>>>>
>>>> There is nothing about compatible naming in commit msg.
>>>
>>> Krzysztof, the whole commit desciption is explanation about why a new
>>> compatible is introduced. I don't understand this comment.
>>>
>>>>
>>>>>>
>>>>>> Yes you did but I was asked for the last year+ that all new compatible
>>>>>> should be soc specific (while imprecise, in our care soc family should be ok),
>>>>>> with a possible semi-generic callback with an IP version or a first soc
>>>>>> implementing the IP.
>>>>>>
>>>>>>> Plus the definition of a SoC is very vague. One could argue that
>>>>>>> the content of the list bellow are vaguely defined families. Should we
>>>>>>> add meson8b, gxl, gxm, sm1 ? ... or even the actual SoC reference ?
>>>>>>> This list gets huge for no reason.
>>>>>>
>>>>>> I think in our case soc family is reasonable since they share same silicon
>>>>>> design.
>>>>>>
>>>>>>> We know all existing PWM of this type are the same. We have been using
>>>>>>> them for years. It is not a new support we know nothing about.
>>>>>>>
>>>>>>>>
>>>>>>>> I thought something like:
>>>>>>>> - items:
>>>>>>>>      - enum:
>>>>>>>>          - amlogic,gxbb-pwm
>>>>>>>>          - amlogic,axg-pwm
>>>>>>>>          - amlogic,g12a-pwm
>>>>>>>>      - const: amlogic,pwm-v1
>>>>>>> I'm not sure I understand what you are suggesting here.
>>>>>>> Adding a "amlogic,pwm-v1" for the obsolete compatible ? No amlogic DT
>>>>>>> has that and I'm working to remove this type, so I don't get the point.
>>>>>>>
>>>>>>>>
>>>>>>>> should be preferred instead of a single amlogic,meson8-pwm-v2 ?
>>>>>>> This is named after the first SoC supporting the type.
>>>>>>> Naming it amlogic,pwm-v2 would feel weird with the s4 coming after.
>>>>>>> Plus the doc specifically advise against this type of names.
>>>>>>
>>>>>> The -v2 refers to a pure software/dt implementation versioning and not
>>>>>> an HW version, so I'm puzzled and I requires DT maintainers advice here.
>>>>>>
>>>>>> Yes meson8b is the first "known" platform, even if I'm pretty sure meson6 has
>>>>
>>>> Yes, this should be SoC-based compatible, unless you have clear
>>>> versioning scheme by SoC/IP block vendor. You named it not a HW version,
>>>> which kind of answers to the "unless" case - that's not hardware version.
>>>>
>>>
>>> This is specifically the point of the comment in commit description.
>>> We know all the PWMs compatible are the same HW (version) as one found
>>> in the meson8b.
>>>
>>> It is certain that adding more compatible, listing all the SoC, will be
>>> useless. I can do it if you insist.
>>
>> The docs you references insist on that, so yeah, I insist as well.
>>
>>>
>>>>>
>>>>> This is not my point. I picked this name because I have to pick a
>>>>> specific device based one. Not because it is actually the first or
>>>>> not. I don't see a problem with meson6 being compatible with
>>>>> meson8-pwm-v2, if that ever comes along.
>>>>
>>>> No, the point is not to use "v2". Use SoC compatibles.
>>>
>>> It is a SoC compatible. The second one.
>>
>> "v2" is not the soc. I assume meson8 is one specific SoC, right? Because
>> elinux says it is a *family*:
>> https://elinux.org/Amlogic/SoCs
>>
> 
> It is a family. yes
> 
>>
>>>
>>> The first one, as explained in the description was describing the driver
>>> more that the HW.
>>>
>>> Changing the way clock are passed from DT to the driver would be break
>>> user of the old compatible. So a new compatible is introduced. I believe
>>> this is recommended way to introduce incompatible binding changes.
>>
>> The way is not to introduce incompatible changes. Your way is also not
>> good as it breaks other users of DTS.
>>
>>>
>>> v2 here denote a new interface version, nothing to do with HW
>>> versioning. I happy to pick something else to denote this.
>>
>> Sorry, then it is not a SoC-based compatible.
>>
>>>
>>>>
>>>>>
>>>>> I think the binding here satisfy the rule that it should be specific,
>>>>> and the intent that goes with it:
>>>>>
>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings/writing-bindings.rst?h=v6.7-rc2#n42
>>>>>
>>>>>> the same pwm architecture, this is why "amlogic,pwm-v1" as fallback seems more
>>>>>> reasonable and s4 and later pwm could use the "amlogic,pwm-v2"
>>>>>> fallback.
>>>>>
>>>>> That is not how understand this:
>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings/writing-bindings.rst?h=v6.7-rc2#n82
>>>>>
>>>>
>>>> Again, where the "v2" is defined? Where is any document explaining the
>>>> mapping between version blocks and SoC parts? Why do you list here only
>>>> major version? Blocks almost always have also minor (e.g. v2.0).
>>>
>>> Again, v2 does has nothing to do with the HW. Never wrote it was.
>>> The HW remains the same.
>>
>> Don't add compatibles which are not related to HW, but represent
>> software versioning. Software does not matter for the bindings.
> 
> What I did I explicitly what is recommended in Grant's presentation from
> 2013. 10y old, but I assume slide 10 "Making an incompatible update" is
> still valid.
> 
> https://elinux.org/images/1/1e/DT_Binding_Process_glikely_ksummit_2013_10_28.pdf
> 
> Breaking the ABI of the old compatible would break all boards which use
> u-boot DT and pass it to the kernel, because the meaning of the clock
> property would change.

You broke U-Boot now as well - it will get your new DTS from the kernel
and stop working.

> 
> Doing things has suggested in this slide, and this patch, allows every
> device to continue to work properly, whether the DT given is the one
> shipped with u-boot (using the old compatible for now) or the kernel.

OK, that explains the reasons. I read your commit msg and nothing like
this was mentioned there. What's more, you did not deprecate the old
binding, thus the confusion - it looked like you add entirely new
hardware (although you put "deprecated" but in some unrelated place, not
next to the compatibles).

Anyway, the main point of Neil was that you started using generic
compatible for all SoCs, which is wrong as well. I guess this was the
original discussion.

Best regards,
Krzysztof