On 1/23/25 19:32, Joanne Koong wrote: > On Thu, Jan 23, 2025 at 10:06 AM Bernd Schubert > <bernd.schubert@xxxxxxxxxxx> wrote: >> >> >> >> On 1/23/25 18:48, Joanne Koong wrote: >>> On Thu, Jan 23, 2025 at 9:19 AM Bernd Schubert >>> <bernd.schubert@xxxxxxxxxxx> wrote: >>>> >>>> Hi Joanne, >>>> >>>>>>> Thanks, applied and pushed with some cleanups including Luis's clamp idea. >>>>>> >>>>>> Hi Miklos, >>>>>> >>>>>> I don't think the timeouts do work with io-uring yet, I'm not sure >>>>>> yet if I have time to work on that today or tomorrow (on something >>>>>> else right now, I can try, but no promises). >>>>> >>>>> Hi Bernd, >>>>> >>>>> What are your thoughts on what is missing on the io-uring side for >>>>> timeouts? If a request times out, it will abort the connection and >>>>> AFAICT, the abort logic should already be fine for io-uring, as users >>>>> can currently abort the connection through the sysfs interface and >>>>> there's no internal difference in aborting through sysfs vs timeouts. >>>>> >>>> >>>> in fuse_check_timeout() it iterates over each fud and then fpq. >>>> In dev_uring.c fpq is is per queue but unrelated to fud. In current >>>> fuse-io-uring fud is not cloned anymore - using fud won't work. >>>> And Requests are also not queued at all on the other list >>>> fuse_check_timeout() is currently checking. >>> >>> In the io-uring case, there still can be fuds and their associated >>> fpqs given that /dev/fuse can be used still, no? So wouldn't the >>> io-uring case still need this logic in fuse_check_timeout() for >>> checking requests going through /dev/fuse? >> >> Yes, these need to be additionally checked. >> >>> >>>> >>>> Also, with a ring per core, maybe better to use >>>> a per queue check that is core bound? I.e. zero locking overhead? >>> >>> How do you envision a queue check that bypasses grabbing the >>> queue->lock? The timeout handler could be triggered on any core, so >>> I'm not seeing how it could be core bound. >> >> I don't want to bypass it, but maybe each queue could have its own >> workq and timeout checker? And then use queue_delayed_work_on()? >> >> >>> >>>> And I think we can also avoid iterating over hash lists (queue->fpq), >>>> but can use the 'ent_in_userspace' list. >>>> >>>> We need to iterate over these other entry queues anyway: >>>> >>>> ent_w_req_queue >>>> fuse_req_bg_queue >>>> ent_commit_queue >>>> >>> >>> Why do we need to iterate through the ent lists (ent_w_req_queue and >>> ent_commit_queue)? AFAICT, in io-uring a request is either on the >>> fuse_req_queue/fuse_req_bg_queue or on the fpq->processing list. Even >>> when an entry has been queued to ent_w_req_queue or ent_commit_queue, >>> the request itself is still queued on >>> fuse_req_queue/fuse_req_bg_queue/fpq->processing. I'm not sure I >>> understand why we still need to look at the ent lists? >> >> Yeah you are right, we could avoid ent_w_req_queue and ent_commit_queue >> if we use fpq->processing, but processing consists of 256 lists - >> overhead is much smaller by using the entry lists? >> >> >>> >>>> >>>> And we also need to iterate over >>>> >>>> fuse_req_queue >>>> fuse_req_bg_queue >>> >>> Why do we need to iterate through fuse_req_queue and >>> fuse_req_bg_queue? fuse_uring_request_expired() checks the head of >>> fuse_req_queue and fuse_req_bg_queue and given that requests are added >>> to fuse_req_queue/fuse_req_bg_queue sequentially (eg added to the tail >>> of these lists), why isn't this enough? >> >> I admit I'm a bit lost with that question. Aren't you pointing out >> the same lists as I do? >> > > Oh, I thought your comment was saying that we need to "iterate" over > it (eg go through every request on the lists)? It currently already > checks the fuse_req_queue and fuse_req_bg_queue lists (in > fuse_uring_request_expired() which gets invoked in the > fuse_check_timeout() timeout handler). > > Maybe the fuse_uring_request_expired() addition got missed - I tried > to call this out in the cover letter changelog, as I had to rebase > this patchset on top of the io-uring patches, so I added this function > in to make it work for io-uring. I believe this suffices for now for > the io uring case (with future optimizations that can be added)? Ah sorry, that is me, I had missed you had already rebased it to io-uring. So we are good to land this version. Just would be good if we could optimize this soon - our test systems have 96 cores - 24576 list heads to check... I won't manage to work on it today and probably also not tomorrow, but by Monday I should have an optimized version ready. Thanks, Bernd
