Re: [PATCH 2/2] io_uring: implement async hybrid mode for pollable requests

[Pavel Begunkov <asml.silence@gmail.com>]

On 10/18/21 11:34, Hao Xu wrote:
> 在 2021/10/18 下午7:29, Hao Xu 写道:
>> The current logic of requests with IOSQE_ASYNC is first queueing it to
>> io-worker, then execute it in a synchronous way. For unbound works like
>> pollable requests(e.g. read/write a socketfd), the io-worker may stuck
>> there waiting for events for a long time. And thus other works wait in
>> the list for a long time too.
>> Let's introduce a new way for unbound works (currently pollable
>> requests), with this a request will first be queued to io-worker, then
>> executed in a nonblock try rather than a synchronous way. Failure of
>> that leads it to arm poll stuff and then the worker can begin to handle
>> other works.
>> The detail process of this kind of requests is:
>>
>> step1: original context:
>>             queue it to io-worker
>> step2: io-worker context:
>>             nonblock try(the old logic is a synchronous try here)
>>                 |
>>                 |--fail--> arm poll
>>                              |
>>                              |--(fail/ready)-->synchronous issue
>>                              |
>>                              |--(succeed)-->worker finish it's job, tw
>>                                             take over the req
>>
>> This works much better than the old IOSQE_ASYNC logic in cases where
>> unbound max_worker is relatively small. In this case, number of
>> io-worker eazily increments to max_worker, new worker cannot be created
>> and running workers stuck there handling old works in IOSQE_ASYNC mode.
>>
>> In my 64-core machine, set unbound max_worker to 20, run echo-server,
>> turns out:
>> (arguments: register_file, connetion number is 1000, message size is 12
>> Byte)
>> original IOSQE_ASYNC: 76664.151 tps
>> after this patch: 166934.985 tps
>>
>> Suggested-by: Jens Axboe <axboe@kernel.dk>
>> Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
> An irrelevant question: why do we do linked timeout logic in
> io_wq_submit_work() again regarding that we've already done it in
> io_queue_async_work().

Because io_wq_free_work() may enqueue new work (by returning it)
without going through io_queue_async_work(), and we don't care
enough to split those cases.

-- 
Pavel Begunkov