[PATCH v1] drm/amdgpu: give each kernel job a unique id

[Christian König]

On 6/11/25 17:11, Danilo Krummrich wrote:
>>>> Mhm, reiterating our internal discussion on the mailing list.
>>>>
>>>> I think it would be nicer if we could use negative values for the kernel submissions and positive for userspace. But as discussed internally we would need to adjust the scheduler trace points for that once more.
>>>>
>>>> @Philip and @Danilo any opinion on that?
>>>
>>> Both, the U64_MAX and the positive-negative approach, are a bit hacky. I wonder
>>> why we need client_id to be a u64, wouldn't a u32 not be enough?
>>
>> That can trivially overflow on long running boxes.
> 
> I don't know if "trivially" is the word of choice given that the number is
> 4,294,967,295.
> 
> But I did indeed miss that this is a for ever increasing atomic. Why is it an
> atomic? Why is it not an IDA?

Well IDA has some extra overhead compared to an ever increasing atomic, additional to that it might not be the best choice to re-use numbers for clients in a trace log.

On the other hand using smaller numbers is usually nicer for manual inspection.

Regards,
Christian.