On 12/17/20 11:09 AM, Linus Torvalds wrote: > On Thu, Dec 17, 2020 at 8:19 AM Jens Axboe <axboe@xxxxxxxxx> wrote: >> [..] >> which shows a failed nonblock lookup, then punt to worker, and then we >> complete with fd == 4. This takes 65 usec in total. Re-running the same >> test case again: >> [..] >> shows the same request completing inline, also returning fd == 4. This >> takes 6 usec. > > I think this example needs to be part of the git history - either move > it into the individual commits, or we just need to make sure it > doesn't get lost as a cover letter and ends up part of the merge > message. I agree, and I think the best approach will depend a bit on whether Al takes the core bits, and then I'm left with just the io_uring part, or if the series goes in as a whole. In any case, I'll make sure it gets in as either the merge message, or (at the least) in the io_uring enablement of it. > Despite having seen the patch series so many times now, I'm still just > really impressed by how small and neat it is, considering the above > numbers, and considering just how problematic this case was > historically (ie I remember all the discussions we had about > nonblocking opens back in the days). > > So I continue to go "this is the RightWay(tm)" just from that. Well, that's largely thanks to your and Al's feedback. I did attempt this one time earlier, and the break through just wasn't there in terms of condensing it properly. I'm very happy with it now. -- Jens Axboe
