On Thu, 3 Oct 2024 at 12:10, Bernd Schubert <bernd.schubert@xxxxxxxxxxx> wrote: > What I mean is that you wanted to get rid of the 'tag' - using any kind of > search means we still need it. I.e. we cannot just take last list head > or tail and use that. > The array is only dynamic at initialization time. And why spending O(logN) > to search instead of O(1)? Because for sane queue depths they are essentially the same. This is not where we can gain or lose any significant performance. > And I know that it is an implementation detail, I just would like to avoid > many rebasing rounds on these details. I think the logical interface would be: - pass a userspace buffer to FETCH (you told me, but I don't remember why sqe->addr isn't suitable) - set sqe->user_data to an implementation dependent value, this could be just the userspace buffer, but it could be a request object - kernel allocates an idle request and queues it. - request comes in, kernel takes a request from the idle queue and fills it - cqe->user_data is returned with the original sqe->user_data, which should be sufficient for the server to identify the request - process request, send COMMIT_AND_FETCH with the userspace buffer and user data - the kernel reads the header from the userspace buffer, finds outh->unique, finds and completes the request - then queues the request on the idle queue ... What's wrong with that? Thanks, Miklos
