Re: [PATCH 1/4] dt-bindings: memory: additional compatible strings for Broadcom DPFE

[Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>]

On 06/12/2023 18:36, Florian Fainelli wrote:
> On 12/6/23 09:29, Krzysztof Kozlowski wrote:
>> On 06/12/2023 17:32, Florian Fainelli wrote:
>>>
>>>
>>> On 12/6/2023 3:09 AM, Krzysztof Kozlowski wrote:
>>>> On 05/12/2023 19:47, Markus Mayer wrote:
>>>>> Add versioned compatible strings for Broadcom DPFE. These take the form
>>>>> brcm,dpfe-cpu-v<N> where <N> is a number from 1 to 4.
>>>>>
>>>>> These API version related compatible strings are more specific than the
>>>>> catch-all "brcm,dpfe-cpu" and more generic than chip-specific compatible
>>>>> strings.
>>>>
>>>> None of this explains: Why? I don't see any point in this and commit
>>>> does not explain.
>>>>
>>>>>
>>>>> Signed-off-by: Markus Mayer <mmayer@broadcom.com>
>>>>> ---
>>>>>    .../bindings/memory-controllers/brcm,dpfe-cpu.yaml        | 8 +++++++-
>>>>>    1 file changed, 7 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/Documentation/devicetree/bindings/memory-controllers/brcm,dpfe-cpu.yaml b/Documentation/devicetree/bindings/memory-controllers/brcm,dpfe-cpu.yaml
>>>>> index 08cbdcddfead..6dffa7b62baf 100644
>>>>> --- a/Documentation/devicetree/bindings/memory-controllers/brcm,dpfe-cpu.yaml
>>>>> +++ b/Documentation/devicetree/bindings/memory-controllers/brcm,dpfe-cpu.yaml
>>>>> @@ -16,6 +16,11 @@ properties:
>>>>>          - enum:
>>>>>              - brcm,bcm7271-dpfe-cpu
>>>>>              - brcm,bcm7268-dpfe-cpu
>>>>> +      - enum:
>>>>> +          - brcm,dpfe-cpu-v1
>>>>> +          - brcm,dpfe-cpu-v2
>>>>> +          - brcm,dpfe-cpu-v3
>>>>> +          - brcm,dpfe-cpu-v4
>>>>
>>>> No, that's just wrong. So you want to say bcm7271 is brcm,dpfe-cpu-v4?
>>>
>>> No as the example shows it "speaks" API v1.
>>
>> Example is not a binding. It does not matter except of validating the
>> binding. This is just incorrect.
>>
>>>
>>> I would be inclined to completely remove the chip specific compatible
>>> strings from the binding because they are not sufficient or descriptive
>>> enough to determine which API version is being spoken, since the
>>> firmware is unfortunately allowed to change major APIs (and the
>>> messaging format, because why not?) at a moments notice.
>>
>> Then versions do not give you anything more.
> 
> The versions indicate exactly which API to be spoken to with the 
> firmware. The firmware API was not properly designed, it should have had 
> a way to indicate which API it has, regardless of the messaging format 
> it implements, but for reasons unknown that is not how it was implemented.
> 
> Essentially we need to know right away and ahead of time which API to be 
> used, otherwise that means doing runtime detection like what patch 4 
> does which you do not want to see.

Yeah, I see, you explained this deeper in response to 3/4, which I read
after this one.

Deprecating specific compatibles makes sense. If you have subset of FW
per given SoC, you could keep the specific compatible followed by subset
of version-compatibles (e.g. bcm7271 + v1 + generic fallback). However
then generic fallback is useless and you should actually drop it. The
only, *ONLY* point of generic fallback is to be used by OS alone. In
that case it cannot be used alone, so it is useless.

We do not use generic compatibles in a way of "I want to call all of
these devices a DPFE" or "I want to call it a default".

Now, if you do not have subset of FW per given SoC, so anything can
match with anything, then in one commit:
1. Deprecate specific compatible followed by useless generic fallback
2. Add versioned-compatibles alone, since generic fallback gives nothing.

Best regards,
Krzysztof