On 2/3/22 2:47 PM, Pavel Begunkov wrote: > On 2/3/22 19:54, Usama Arif wrote: >> On 03/02/2022 19:06, Jens Axboe wrote: >>> On 2/3/22 12:00 PM, Pavel Begunkov wrote: >>>> On 2/3/22 18:29, Jens Axboe wrote: >>>>> On 2/3/22 11:26 AM, Usama Arif wrote: >>>>>> Hmm, maybe i didn't understand you and Pavel correctly. Are you >>>>>> suggesting to do the below diff over patch 3? I dont think that would be >>>>>> correct, as it is possible that just after checking if ctx->io_ev_fd is >>>>>> present unregister can be called by another thread and set ctx->io_ev_fd >>>>>> to NULL that would cause a NULL pointer exception later? In the current >>>>>> patch, the check of whether ev_fd exists happens as the first thing >>>>>> after rcu_read_lock and the rcu_read_lock are extremely cheap i believe. >>>>> >>>>> They are cheap, but they are still noticeable at high requests/sec >>>>> rates. So would be best to avoid them. >>>>> >>>>> And yes it's obviously racy, there's the potential to miss an eventfd >>>>> notification if it races with registering an eventfd descriptor. But >>>>> that's not really a concern, as if you register with inflight IO >>>>> pending, then that always exists just depending on timing. The only >>>>> thing I care about here is that it's always _safe_. Hence something ala >>>>> what you did below is totally fine, as we're re-evaluating under rcu >>>>> protection. >>>> >>>> Indeed, the patch doesn't have any formal guarantees for propagation >>>> to already inflight requests, so this extra unsynchronised check >>>> doesn't change anything. >>>> >>>> I'm still more сurious why we need RCU and extra complexity when >>>> apparently there is no use case for that. If it's only about >>>> initial initialisation, then as I described there is a much >>>> simpler approach. >>> >>> Would be nice if we could get rid of the quiesce code in general, but I >>> haven't done a check to see what'd be missing after this... >>> >> >> I had checked! I had posted below in in reply to v1 (https://lore.kernel.org/io-uring/02fb0bc3-fc38-b8f0-3067-edd2a525ef29@xxxxxxxxx/T/#m5ac7867ac61d86fe62c099be793ffe5a9a334976), but i think it got missed! Copy-pasting here for reference: > > May have missed it then, apologies > >> " >> I see that if we remove ring quiesce from the the above 3 opcodes, then >> only IORING_REGISTER_ENABLE_RINGS and IORING_REGISTER_RESTRICTIONS is >> left for ring quiesce. I just had a quick look at those, and from what i >> see we might not need to enter ring quiesce in >> IORING_REGISTER_ENABLE_RINGS as the ring is already disabled at that point? >> And for IORING_REGISTER_RESTRICTIONS if we do a similar approach to >> IORING_REGISTER_EVENTFD, i.e. wrap ctx->restrictions inside an RCU >> protected data structure, use spin_lock to prevent multiple >> io_register_restrictions calls at the same time, and use read_rcu_lock >> in io_check_restriction, then we can remove ring quiesce from >> io_uring_register altogether? >> >> My usecase only uses IORING_REGISTER_EVENTFD, but i think entering ring >> quiesce costs similar in other opcodes. If the above sounds reasonable, >> please let me know and i can send patches for removing ring quiesce for >> io_uring_register. >> " >> >> Let me know if above makes sense, i can add patches on top of the current patchset, or we can do it after they get merged. >> >> As for why, quiesce state is very expensive. its making io_uring_register the most expensive syscall in my usecase (~15ms) compared to ~0.1ms now with RCU, which is why i started investigating this. And this patchset avoids ring quiesce for 3 of the opcodes, so it would generally be quite helpful if someone does registers and unregisters eventfd multiple times. > > I agree that 15ms for initial setup is silly and it has to be > reduced. However, I'm trying weight the extra complexity against > potential benefits of _also_ optimising [de,re]-registration > > Considering that you only register it one time at the beginning, > we risk adding a yet another feature that nobody is going to ever > use. This doesn't give me a nice feeling, well, unless you do > have a use case. It's not really a new feature, it's just making the existing one not suck quite as much... > To emphasise, I'm comparing 15->0.1 improvement for only initial > registration (which is simpler) vs 15->0.1 for both registration > and unregistration. reg+unreg should be way faster too, if done properly with the assignment tricks. > fwiw, it alters userpace visible behaviour in either case, shouldn't > be as important here but there is always a chance to break userspace It doesn't alter userspace behavior, if the registration works like I described with being able to assign a new one while the old one is being torn down. Or do you mean wrt inflight IO? I don't think the risk is very high there, to be honest. -- Jens Axboe
