Re: [PATCH 2/2] io_uring: implementation of IOSQE_ASYNC_HYBRID logic

[Hao Xu <haoxu@linux.alibaba.com>]

在 2021/10/9 下午8:46, Pavel Begunkov 写道:
> On 10/8/21 13:36, Hao Xu wrote:
>> The 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 IOSQE_ASYNC in cases where cpu resources
>> are scarce or unbound max_worker is small. In these cases, 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 machine, set unbound max_worker to 20, run echo-server, turns out:
>> (arguments: register_file, connetion number is 1000, message size is 12
>> Byte)
>> IOSQE_ASYNC: 76664.151 tps
>> IOSQE_ASYNC_HYBRID: 166934.985 tps
>>
>> Suggested-by: Jens Axboe <axboe@kernel.dk>
>> Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
>> ---
>>   fs/io_uring.c | 42 ++++++++++++++++++++++++++++++++++++++----
>>   1 file changed, 38 insertions(+), 4 deletions(-)
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index a99f7f46e6d4..024cef09bc12 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -1409,7 +1409,7 @@ static void io_prep_async_work(struct io_kiocb 
>> *req)
>>       req->work.list.next = NULL;
>>       req->work.flags = 0;
>> -    if (req->flags & REQ_F_FORCE_ASYNC)
>> +    if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_ASYNC_HYBRID))
>>           req->work.flags |= IO_WQ_WORK_CONCURRENT;
>>       if (req->flags & REQ_F_ISREG) {
>> @@ -5575,7 +5575,13 @@ static int io_arm_poll_handler(struct io_kiocb 
>> *req)
>>       req->apoll = apoll;
>>       req->flags |= REQ_F_POLLED;
>>       ipt.pt._qproc = io_async_queue_proc;
>> -    io_req_set_refcount(req);
>> +    /*
>> +     * REQ_F_REFCOUNT set indicate we are in io-worker context, where we
> 
> Nope, it indicates that needs more complex refcounting. It includes linked
> timeouts but also poll because of req_ref_get for double poll. fwiw, with
> some work it can be removed for polls, harder (and IMHO not necessary) 
> to do
> for timeouts.Agree, I now realize that the explanation I put here is not good at all,
I actually want to say that the io-worker already set refs = 2 (also
possible that prep_link_out set 1, and io-worker adds the other 1,
previously I miss this situation). One will be put at completion time,
the other one will be put in io_wq_free_work(). So no need to set the
refcount here again. I looked into io_req_set_refcount(), since it does
nothing if refcount is already not zero, I should be ok to keep this one
as it was.
> 
>> +     * already explicitly set the submittion and completion ref. So no
> 
> I'd say there is no notion of submission vs completion refs anymore.
> 
>> +     * need to set refcount here if that is the case.
>> +     */
>> +    if (!(req->flags & REQ_F_REFCOUNT))
> 
> Compare it with io_req_set_refcount(), that "if" is a a no-op
> 
>> +        io_req_set_refcount(req);
>>       ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
>>                       io_async_wake);
>> @@ -6704,8 +6710,11 @@ static void io_wq_submit_work(struct io_wq_work 
>> *work)
>>           ret = -ECANCELED;
>>       if (!ret) {
>> +        bool need_poll = req->flags & REQ_F_ASYNC_HYBRID;
>> +
>>           do {
>> -            ret = io_issue_sqe(req, 0);
>> +issue_sqe:
>> +            ret = io_issue_sqe(req, need_poll ? IO_URING_F_NONBLOCK : 
>> 0);
> 
> It's buggy, you will get all kinds of kernel crashes and leaks.
> Currently IO_URING_F_NONBLOCK has dual meaning: obvious nonblock but
> also whether we hold uring_lock or not. You'd need to split the flag
> into two, i.e. IO_URING_F_LOCKED
I'll look into it. I was thinking about to do the first nowait try in
the original context, but then I thought it doesn't make sense to bring
up a worker just for poll infra arming since thread creating and
scheduling has its overhead.
>