[PATCH] io_uring: Optimization of buffered random write

[luhongfei]

The buffered random write performance of io_uring is poor
due to the following reason:
By default, when performing buffered random writes, io_sq_thread
will call io_issue_sqe writes req, but due to the setting of
IO_URING_F_NONBLOCK, req is executed asynchronously in iou-wrk,
where io_wq_submit_work calls io_issue_sqe completes the write req,
with issue_flag as IO_URING_F_UNLOCKED | IO_URING_F_IOWQ,
which will reduce performance.
This patch will determine whether this req is a buffered random write,
and if so, io_sq_thread directly calls io_issue_sqe(req, 0)
completes req instead of completing it asynchronously in iou wrk.

Performance results:
  For fio the following results have been obtained with a queue depth of
  8 and 4k block size:

             random writes:
             without patch          with patch      libaio      psync
iops:           287k                   560k          248K        324K
bw:            1123MB/s               2188MB/s       970MB/s    1267MB/s
clat:         52760ns                69918ns       28405ns      2109ns

Signed-off-by: luhongfei <luhongfei@xxxxxxxx>
---
 io_uring/io_uring.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
 mode change 100644 => 100755 io_uring/io_uring.c

diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 4a865f0e85d0..64bb91beb4d6
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -2075,8 +2075,23 @@ static inline void io_queue_sqe(struct io_kiocb *req)
 	__must_hold(&req->ctx->uring_lock)
 {
 	int ret;
+	bool is_write;
 
-	ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
+	switch (req->opcode) {
+	case IORING_OP_WRITEV:
+	case IORING_OP_WRITE_FIXED:
+	case IORING_OP_WRITE:
+		is_write = true;
+		break;
+	default:
+		is_write = false;
+		break;
+	}
+
+	if (!is_write || (req->rw.kiocb.ki_flags & IOCB_DIRECT))
+		ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
+	else
+		ret = io_issue_sqe(req, 0);
 
 	/*
 	 * We async punt it if the file wasn't marked NOWAIT, or if the file
-- 
2.39.0