Re: [PATCH v2 4/4] io_uring: get next req on subm ref drop

[Jens Axboe <axboe@kernel.dk>]

On 3/3/20 9:04 AM, Jens Axboe wrote:
> On 3/3/20 3:46 AM, Pavel Begunkov wrote:
>> On 3/3/2020 9:54 AM, Pavel Begunkov wrote:
>>> On 03/03/2020 07:26, Jens Axboe wrote:
>>>> On 3/2/20 1:45 PM, Pavel Begunkov wrote:
>>>>> Get next request when dropping the submission reference. However, if
>>>>> there is an asynchronous counterpart (i.e. read/write, timeout, etc),
>>>>> that would be dangerous to do, so ignore them using new
>>>>> REQ_F_DONT_STEAL_NEXT flag.
>>>>
>>>> Hmm, not so sure I like this one. It's not quite clear to me where we
>>>> need REQ_F_DONT_STEAL_NEXT. If we have an async component, then we set
>>>> REQ_F_DONT_STEAL_NEXT. So this is generally the case where our
>>>> io_put_req() for submit is not the last drop. And for the other case,
>>>> the put is generally in the caller anyway. So I don't really see what
>>>> this extra flag buys us?
>>>
>>> Because io_put_work() holds a reference, no async handler can achive req->refs
>>> == 0, so it won't return next upon dropping the submission ref (i.e. by
>>> put_find_nxt()). And I want to have next before io_put_work(), to, instead of as
>>> currently:
>>>
>>> run_work(work);
>>> assign_cur_work(NULL); // spinlock + unlock worker->lock
>>> new_work = put_work(work);
>>> assign_cur_work(new_work); // the second time
>>>
>>> do:
>>>
>>> new_work = run_work(work);
>>> assign_cur_work(new_work); // need new_work here
>>> put_work(work);
>>>
>>>
>>> The other way:
>>>
>>> io_wq_submit_work() // for all async handlers
>>> {
>>> 	...
>>> 	// Drop submission reference.
>>> 	// One extra ref will be put in io_put_work() right
>>> 	// after return, and it'll be done in the same thread
>>> 	if (atomic_dec_and_get(req) == 1)
>>> 		steal_next(req);
>>> }
>>>
>>> Maybe cleaner, but looks fragile as well. Would you prefer it?
>>
>> Any chance you've measured your next-work fix? I wonder how much does it
>> hurt performance, and whether we need a terse patch for 5.6.
> 
> Unless I'm missing something, the worker will pick up the next work
> without sleeping, since the request will have finished. So it really
> should not add any extra overhead, except you'll do an extra wqe lock
> roundtrip.
> 
> But I'll run some testing to be totally sure.

Testing with link-cp, not seeing much if anything of a difference. Not
in wqe load either.

-- 
Jens Axboe