Date: Fri, 12 Apr 2024 14:20:34 +0100
From: Pavel Begunkov <asml.silence@...il.com>
To: 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

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/

>> 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