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

[Jens Axboe <axboe@kernel.dk>]

On 3/4/24 3:18 PM, Pavel Begunkov wrote:
> On 3/4/24 22:07, Pavel Begunkov wrote:
>> On 3/1/24 03:45, Jens Axboe wrote:
>>> On 2/29/24 9:36 AM, Pavel Begunkov wrote:
>>>> With defer taskrun we store aux cqes into a cache array and then flush
>>>> into the CQ, and we also maintain the ordering so aux cqes are flushed
>>>> before request completions. Why do we need the cache instead of pushing
>>>> them directly? We acutally don't, so let's kill it.
>>>>
>>>> One nuance is synchronisation -- the path we touch here is only for
>>>> DEFER_TASKRUN and guaranteed to be executed in the task context, and
>>>> all cqe posting is serialised by that. We also don't need locks because
>>>> of that, see __io_cq_lock().
>>>
>>> Nicely spotted! Looks good to me.
>>
>> Apparently I'm wrong as "defer" in that function is not about
>> defer taskrun, but rather IO_URING_F_COMPLETE_DEFER. Jens, can
>> you drop it for now?

Dropped

> One way forward would be to limit the aux fast path to defer
> taskrun, I don't really care about other cases, and aux is
> mostly useful for multishots hence net, which should preferably
> be DEFER_TASKRUN.

I think that'd be fine.

> In any case the code is a bit unhandy if not smelly, can use
> some refactoring.

For sure.

-- 
Jens Axboe