Re: [PATCH 11/11] io_uring: get rid of intermediate aux cqe caches

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 3/15/24 11:26 AM, Pavel Begunkov wrote:
> On 3/15/24 16:49, Jens Axboe wrote:
>> On 3/15/24 10:44 AM, Pavel Begunkov wrote:
>>> On 3/15/24 16:27, Jens Axboe wrote:
>>>> On 3/15/24 10:25 AM, Jens Axboe wrote:
>>>>> On 3/15/24 10:23 AM, Pavel Begunkov wrote:
>>>>>> On 3/15/24 16:20, Jens Axboe wrote:
>>>>>>> On 3/15/24 9:30 AM, Pavel Begunkov wrote:
>>>>>>>> io_post_aux_cqe(), which is used for multishot requests, delays
>>>>>>>> completions by putting CQEs into a temporary array for the purpose
>>>>>>>> completion lock/flush batching.
>>>>>>>>
>>>>>>>> DEFER_TASKRUN doesn't need any locking, so for it we can put completions
>>>>>>>> directly into the CQ and defer post completion handling with a flag.
>>>>>>>> That leaves !DEFER_TASKRUN, which is not that interesting / hot for
>>>>>>>> multishot requests, so have conditional locking with deferred flush
>>>>>>>> for them.
>>>>>>>
>>>>>>> This breaks the read-mshot test case, looking into what is going on
>>>>>>> there.
>>>>>>
>>>>>> I forgot to mention, yes it does, the test makes odd assumptions about
>>>>>> overflows, IIRC it expects that the kernel allows one and only one aux
>>>>>> CQE to be overflown. Let me double check
>>>>>
>>>>> Yeah this is very possible, the overflow checking could be broken in
>>>>> there. I'll poke at it and report back.
>>>>
>>>> It does, this should fix it:
>>>>
>>>>
>>>> diff --git a/test/read-mshot.c b/test/read-mshot.c
>>>> index 8fcb79857bf0..501ca69a98dc 100644
>>>> --- a/test/read-mshot.c
>>>> +++ b/test/read-mshot.c
>>>> @@ -236,7 +236,7 @@ static int test(int first_good, int async, int overflow)
>>>>            }
>>>>            if (!(cqe->flags & IORING_CQE_F_MORE)) {
>>>>                /* we expect this on overflow */
>>>> -            if (overflow && (i - 1 == NR_OVERFLOW))
>>>> +            if (overflow && i >= NR_OVERFLOW)
>>>
>>> Which is not ideal either, e.g. I wouldn't mind if the kernel stops
>>> one entry before CQ is full, so that the request can complete w/o
>>> overflowing. Not supposing the change because it's a marginal
>>> case, but we shouldn't limit ourselves.
>>
>> But if the event keeps triggering we have to keep posting CQEs,
>> otherwise we could get stuck. 
> 
> Or we can complete the request, then the user consumes CQEs
> and restarts as usual

So you'd want to track if we'd overflow, wait for overflow to clear, and
then restart that request? I think that sounds a bit involved, no?
Particularly for a case like overflow, which generally should not occur.
If it does, just terminate it, and have the user re-issue it. That seems
like the simpler and better solution to me.

>> As far as I'm concerned, the behavior with
>> the patch looks correct. The last CQE is overflown, and that terminates
>> it, and it doesn't have MORE set. The one before that has MORE set, but
>> it has to, unless you aborted it early. But that seems impossible,
>> because what if that was indeed the last current CQE, and we reap CQEs
>> before the next one is posted.
>>
>> So unless I'm missing something, I don't think we can be doing any
>> better.
> 
> You can opportunistically try to avoid overflows, unreliably
> 
> bool io_post_cqe() {
>     // Not enough space in the CQ left, so if there is a next
>     // completion pending we'd have to overflow. Avoid that by
>     // terminating it now.
>     //
>     // If there are no more CQEs after this one, we might
>     // terminate a bit earlier, but that better because
>     // overflows are so expensive and unhandy and so on.
>     if (cq_space_left() <= 1)
>         return false;
>     fill_cqe();
>     return true;
> }
> 
> some_multishot_function(req) {
>     if (!io_post_cqe(res))
>         complete_req(req, res);
> }
> 
> Again, not suggesting the change for all the obvious reasons, but
> I think semantically we should be able to do it.

Yeah not convinced this is worth looking at. If it was the case that the
hot path would often see overflows and it'd help to avoid it, then
probably it'd make sense. But I don't think that's the case.

-- 
Jens Axboe





[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux