Re: [PATCH v2 1/4] fsstress: add IO_URING read and write operations

[Jens Axboe <axboe@kernel.dk>]

On 8/22/20 12:14 PM, Zorro Lang wrote:
>>> +	if ((e = io_uring_submit(&ring)) != 1) {
>>> +		if (v)
>>> +			printf("%d/%d: %s - io_uring_submit failed %d\n", procid, opno,
>>> +			       iswrite ? "uring_write" : "uring_read", e);
>>> +		goto uring_out1;
>>> +	}
>>> +	if ((e = io_uring_wait_cqe(&ring, &cqe)) < 0) {
>>> +		if (v)
>>> +			printf("%d/%d: %s - io_uring_wait_cqe failed %d\n", procid, opno,
>>> +			       iswrite ? "uring_write" : "uring_read", e);
>>> +		goto uring_out1;
>>> +	}
>>
>> You could use io_uring_submit_and_wait() here, that'll save a system
>> call for sync IO. Same comment goes for 4/4.
> 
> Hi Jens,
> 
> Sorry I think I haven't learned about io_uring enough, why the
> io_uring_submit_and_wait can save a system call? Is it same with
> io_uring_submit(), except a wait_nr ? The io_uring_wait_cqe() and
> io_uring_cqe_seen() are still needed, right?

If you just call io_uring_submit(), it'll enter the kernel and submit
that IO. Then right after that you're saying "I want to wait for
completion of a request", which is then another system call. If you do
io_uring_submit_and_wait() you're entering the kernel with the intent of
"submit my request(s), and wait for N requests" hence only doing a
single system call even though it's an async interface.

Nothing else changes, io_uring_wait_cqe() will not enter the kernel if a
cqe is available in the ring already.

-- 
Jens Axboe