Re: [PATCH v2 06/11] dt-bindings: timer: Add Sophgo sg2042 clint

[Inochi Amaoto]

>On Wed, Sep 20, 2023 at 07:24:21PM +0800, Inochi Amaoto wrote:
>>>
>>> Yo,
>>>
>>> On Wed, Sep 20, 2023 at 05:08:41PM +0800, Inochi Amaoto wrote:
>>>>> On Wed, Sep 20, 2023 at 02:39:39PM +0800, Chen Wang wrote:
>>>>>> From: Inochi Amaoto <inochiama@xxxxxxxxxxx>
>>>>>>
>>>>>> Add two new compatible string formatted like `C9xx-clint-xxx` to identify
>>>>>> the timer and ipi device separately, and do not allow c900-clint as the
>>>>>> fallback to avoid conflict.
>>>>>>
>>>>>> Signed-off-by: Inochi Amaoto <inochiama@xxxxxxxxxxx>
>>>>>> Signed-off-by: Chen Wang <wangchen20@xxxxxxxxxxx>
>>>>>
>>>>> Have you ignored Krzysztof's comments on this? I don't see a response or
>>>>> a reaction to his comments about the compatibles on the last version.
>>>>> Additionally, where is the user for these? I don't see any drivers that
>>>>> actually make use of these.
>>>>>
>>>>
>>>> Sorry for late reply and wrong message-id.
>>>>
>>>> The clint is parsed by sbi.
>>>
>>> That needs to go in the commit message.
>>
>> Yes, it will.
>
>Thanks.
>
>>>> As use the same compatible, the opensbi will
>>>> parse the device twice. This will cause a fault.
>>>
>>> Then only have one compatible with 2 register ranges? Then your SBI
>>> implementation can use those two register ranges to find out the base
>>> address for the mtimer bits and for the mswi bits.
>>> I don't understand why this cannot be done, could you please explain.
>>
>> That is a good idea, but now SBI use the second register ranges as
>> mtimecmp address for aclint. And there is a aclint-mswi in the SBI.
>> Maybe a change is needed?
>
>Yeah, I don't think the model for this in OpenSBI at the moment (and
>since I checked, in QEMU too) is correct. I think we should re-do things
>correctly and it'd be great if things didn't get merged to those
>projects that end up being objected to by dt-binding people.
>I've started keeping a closer eye on QEMU recently in that regard, but I
>am not super attentive. I'll try to be better at that going forward!
>
>>
>>> I also don't see anything in the opensbi repo right now that is using
>>> these (nor could I easily see any patches for opensbi adding this).
>>> Is there another SBI implementation that you are using that I can take
>>> a look at to try and understand this better?
>>>
>>
>> This will be sumbit in a short time.
>> Now we only use it is sophgo vendor SBI, which url is [1].
>>
>> [1] https://github.com/sophgo/opensbi
>
>Thanks.
>
>>>>> Why do you need to have 2 compatibles (and therefore 2 devices) for the
>>>>> clint? I thought the clint was a single device, of which the mtimer and
>>>>> mswi bits were just "features"? Having split register ranges isn't a
>>>>> reason to have two compatibles, so I must be missing something here...
>>>
>>>> Sorry for late reply, The clint consists of mtimer and ipi devices, which
>>>> is defined in [1].
>>>
>>> Yes, I have looked at the spec. I went to check it again before replying
>>> here in case there was something immediately obvious that I was missing.
>>>
>>
>> I think nothing missed.
>>
>>>> This standard shows clint(or the aclint) has two device,
>>>
>>> The wording used here doesn't really matter. It's one interrupt
>>> controller that does mtimer and mswi.
>>>
>>>> but not one. In another word, there is no need to defined mtimer and ipi
>>>> device on the same base address.
>>>
>>> There's also no need to have two compatibles for the same interrupt
>>> controller, so I do not get this reasoning. What actually _requires_
>>> them to be split?
>>>
>>
>> Yes, it is one, but can be mapped into different address. So I think we
>> need two.
>
>Not two compatibles though, just two memory addresses that you need to
>locate (or maybe even 3, for SSWI?)
>

We may need four (mtime, mtimecmp, mswi, sswi) if use register range.

Anyway, I will use a vendor spec implementation as a temporary solution.
I hope this will be corrected in a predictable future, and we can use a
standard way to resolve this at that time. :)

>>
>>>> So we need two compatibles to allow sbi to identify them correctly.
>>>
>>> Why is it not sufficient to identify the individual memory regions?
>>>
>>
>> FYI, Anup. As I have no idea for aclint implementation.
>>
>>> Thanks,
>>> Conor.
>>>
>