Message-ID: <CAK1VsR2e9oHCE1dBD0db0B5TTb6Q4hd4AnB6AYt=4CiaGBuRgA@mail.gmail.com>
Date: Mon, 15 Apr 2024 16:51:28 -0700
From: Oliver Crumrine <ozlinuxc@...il.com>
To: Pavel Begunkov <asml.silence@...il.com>, Oliver Crumrine <ozlinuxc@...il.com>, axboe@...nel.dk
Cc: io-uring@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 1/3] io_uring: Add REQ_F_CQE_SKIP support for io_uring zerocopy

Pavel Begunkov wrote:
> On 4/11/24 01:52, Oliver Crumrine wrote:
> > Pavel Begunkov wrote:
> >> On 4/9/24 02:33, Oliver Crumrine wrote:
> >>> Pavel Begunkov wrote:
> >>>> On 4/7/24 20:14, Oliver Crumrine wrote:
> >>>>> Oliver Crumrine wrote:
> >>>>>> Pavel Begunkov wrote:
> >>>>>>> On 4/5/24 21:04, Oliver Crumrine wrote:
> >>>>>>>> Pavel Begunkov wrote:
> >>>>>>>>> On 4/4/24 23:17, Oliver Crumrine wrote:
> >>>>>>>>>> In his patch to enable zerocopy networking for io_uring, Pavel Begunkov
> >>>>>>>>>> specifically disabled REQ_F_CQE_SKIP, as (at least from my
> >>>>>>>>>> understanding) the userspace program wouldn't receive the
> >>>>>>>>>> IORING_CQE_F_MORE flag in the result value.
> >>>>>>>>>
> >>>>>>>>> No. IORING_CQE_F_MORE means there will be another CQE from this
> >>>>>>>>> request, so a single CQE without IORING_CQE_F_MORE is trivially
> >>>>>>>>> fine.
> >>>>>>>>>
> >>>>>>>>> The problem is the semantics, because by suppressing the first
> >>>>>>>>> CQE you're loosing the result value. You might rely on WAITALL
> >>>>>>>> That's already happening with io_send.
> >>>>>>>
> >>>>>>> Right, and it's still annoying and hard to use
> >>>>>> Another solution might be something where there is a counter that stores
> >>>>>> how many CQEs with REQ_F_CQE_SKIP have been processed. Before exiting,
> >>>>>> userspace could call a function like: io_wait_completions(int completions)
> >>>>>> which would wait until everything is done, and then userspace could peek
> >>>>>> the completion ring.
> >>>>>>>
> >>>>>>>>> as other sends and "fail" (in terms of io_uring) the request
> >>>>>>>>> in case of a partial send posting 2 CQEs, but that's not a great
> >>>>>>>>> way and it's getting userspace complicated pretty easily.
> >>>>>>>>>
> >>>>>>>>> In short, it was left out for later because there is a
> >>>>>>>>> better way to implement it, but it should be done carefully
> >>>>>>>> Maybe we could put the return values in the notifs? That would be a
> >>>>>>>> discrepancy between io_send and io_send_zc, though.
> >>>>>>>
> >>>>>>> Yes. And yes, having a custom flavour is not good. It'd only
> >>>>>>> be well usable if apart from returning the actual result
> >>>>>>> it also guarantees there will be one and only one CQE, then
> >>>>>>> the userspace doesn't have to do the dancing with counting
> >>>>>>> and checking F_MORE. In fact, I outlined before how a generic
> >>>>>>> solution may looks like:
> >>>>>>>
> >>>>>>> https://github.com/axboe/liburing/issues/824
> >>>>>>>
> >>>>>>> The only interesting part, IMHO, is to be able to merge the
> >>>>>>> main completion with its notification. Below is an old stash
> >>>>>>> rebased onto for-6.10. The only thing missing is relinking,
> >>>>>>> but maybe we don't even care about it. I need to cover it
> >>>>>>> well with tests.
> >>>>>> The patch looks pretty good. The only potential issue is that you store
> >>>>>> the res of the normal CQE into the notif CQE. This overwrites the
> >>>>>> IORING_CQE_F_NOTIF with IORING_CQE_F_MORE. This means that the notif would
> >>>>>> indicate to userspace that there will be another CQE, of which there
> >>>>>> won't.
> >>>>> I was wrong here; Mixed up flags and result value.
> >>>>
> >>>> Right, it's fine. And it's synchronised by the ubuf refcounting,
> >>>> though it might get more complicated if I'd try out some counting
> >>>> optimisations.
> >>>>
> >>>> FWIW, it shouldn't give any performance wins. The heavy stuff is
> >>>> notifications waking the task, which is still there. I can even
> >>>> imagine that having separate CQEs might be more flexible and would
> >>>> allow more efficient CQ batching.
> >>> I've actaully been working on this issue for a little while now. My current
> >>> idea is that an id is put into the optval section of the SQE, and then it
> >>> can be used to tag that req with a certain group. When a req has a flag
> >>> set on it, it can request for all of group's notifs to be "flushed" in one
> >>> notif that encompasses that entire group. If the id is zero, it won't be
> >>> associated with a group and will generate a notif. LMK if you see anything
> >>> in here that could overcomplicate userspace. I think it's pretty simple,
> >>> but you've had a crack at this before so I'd like to hear your opinion.
> >>
> >> You can take a look at early versions of the IORING_OP_SEND_ZC, e.g.
> >> patchset v1, probably even later ones. It was basically doing what
> >> you've described with minor uapi changes, like you had to declare groups
> >> (slots) in advance, i.e. register them.
> > My idea is that insead of allocating slots before making requests, "slots"
> > will be allocated as the group ids show up. Instead of an array of slots, a
> > linked list can be used so things can be kmalloc'ed on the fly to make
> > the uapi simpler.
> >>
> >> More flexible and so performant in some circumstances, but the overall
> >> feedback from people trying it is that it's complicated. The user should
> >> allocate group ids, track bound requests / buffers, do other management.
> >> The next question is how the user should decide what bind to what. There
> >> is some nastiness in using the same group for multiple sockets, and then
> > Then maybe we find a way to prevent multiple sockets on one group.
>
> You don't have to explicitly prevent it unless there are other reasons,
> it's just not given a real app would be able to use it this way.
>
> >> what's the cut line to flush the previous notif? I probably forgot a
> > I'd make it the max for a u32 -- I'm (probably) going to use an atomic_t
> > to store the counter of how many reqs have been completed, so a u32 max
> > would make sense.
>
> To be clear, the question raised is entirely for userspace to decide
> if we're talking about the design when the user has to flush a group
> notificaiton via flag or so. Atomics or not is a performance side,
> that's separate.
>
> >> couple more complaints.
> >>
> >> TL;DR;
> >>
> >> The performance is a bit of a longer story, problems are mostly coming
> >> from the async nature of io_uring, and it'd be nice to solve at least a
> >> part of it generically, not only for sendzc. The expensive stuff is
> >> waking up the task, it's not unique to notifications, recv will trigger
> >> it with polling as well as other opcodes. Then the key is completion
> >> batching.
> > Maybe the interface is made for sendzc first, and people could test it
> > there. Then if it is considered beneficial to other places, it could be
> > implemented there.
> >>
> >> What's interesting, take for example some tx only toy benchmark with
> >> DEFER_TASKRUN (recommended to use in any case). If you always wait for
> >> sends without notifications and add eventual *_get_events(), that would
> >> completely avoid the wake up overhead if there are enough buffers,
> >> and if it's not it can 1:1 replace tx polling.
> > Seems like an interesting way to eliminate waiting overhead.
> >>
> >> Try groups, see if numbers are good. And a heads up, I'm looking at
> > I will. Working hard to have the code done by Sunday.
>
> Good, and here is the patchset I mentioned:
>
> https://lore.kernel.org/io-uring/cover.1712923998.git.asml.silence@gmail.com/T/
Wow! 6x improvment is crazy. I just finished the code for notif grouping, and
will be benchmarking it in the upcoming hours/days. It's still in a pre-alpha
state, so I'll have to put a little more work into it. (pre-alpha means leaking
memory. I have 32 gigs in my system. Assuming I don't go too crazy on the
benchmarking I should be fine) Either way, my patch will need a little bit of
work to be compatible with yours, as it modifies the ubuf callback, and yours
does too.
>
> >> improving it a little bit for TCP because of a report, not changing
> >> uapi but might change performance math.
>
> --
> Pavel Begunkov

Powered by blists - more mailing lists