[PATCH 14/14] io_uring: save atomic dec for inline executed reqs

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

When a request is completed with comp_state, its completion reference
put is deferred to io_submit_flush_completions(), but the submission
is put not far from there, so do it together to save one atomic dec per
request. That targets requests that complete inline, e.g. buffered rw,
send/recv.

Proper benchmarking haven't been conducted but for nops(batch=32) it was
around 7901 vs 8117 KIOPS (~2.7%), or ~4% per perf profiling.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 fs/io_uring.c | 17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 1e46d471aa76..fb4e2a97e4f3 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -618,6 +618,7 @@ enum {
 	REQ_F_NO_FILE_TABLE_BIT,
 	REQ_F_WORK_INITIALIZED_BIT,
 	REQ_F_LTIMEOUT_ACTIVE_BIT,
+	REQ_F_COMPLETE_INLINE_BIT,
 
 	/* not a real bit, just to check we're not overflowing the space */
 	__REQ_F_LAST_BIT,
@@ -661,6 +662,8 @@ enum {
 	REQ_F_WORK_INITIALIZED	= BIT(REQ_F_WORK_INITIALIZED_BIT),
 	/* linked timeout is active, i.e. prepared by link's head */
 	REQ_F_LTIMEOUT_ACTIVE	= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
+	/* completion is deferred through io_comp_state */
+	REQ_F_COMPLETE_INLINE	= BIT(REQ_F_COMPLETE_INLINE_BIT),
 };
 
 struct async_poll {
@@ -1899,14 +1902,15 @@ static void io_submit_flush_completions(struct io_comp_state *cs)
 		 * io_free_req() doesn't care about completion_lock unless one
 		 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
 		 * because of a potential deadlock with req->work.fs->lock
+		 * We defer both, completion and submission refs.
 		 */
 		if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
 				 |REQ_F_WORK_INITIALIZED)) {
 			spin_unlock_irq(&ctx->completion_lock);
-			io_put_req(req);
+			io_double_put_req(req);
 			spin_lock_irq(&ctx->completion_lock);
 		} else {
-			io_put_req(req);
+			io_double_put_req(req);
 		}
 	}
 	io_commit_cqring(ctx);
@@ -1922,8 +1926,7 @@ static void io_req_complete_state(struct io_kiocb *req, long res,
 	io_clean_op(req);
 	req->result = res;
 	req->compl.cflags = cflags;
-	list_add_tail(&req->compl.list, &cs->list);
-	cs->nr++;
+	req->flags |= REQ_F_COMPLETE_INLINE;
 }
 
 static inline void __io_req_complete(struct io_kiocb *req, long res,
@@ -6537,9 +6540,9 @@ static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
 			io_queue_linked_timeout(linked_timeout);
 	} else if (likely(!ret)) {
 		/* drop submission reference */
-		if (cs) {
-			io_put_req(req);
-			if (cs->nr >= 32)
+		if (req->flags & REQ_F_COMPLETE_INLINE) {
+			list_add_tail(&req->compl.list, &cs->list);
+			if (++cs->nr >= 32)
 				io_submit_flush_completions(cs);
 			req = NULL;
 		} else {
-- 
2.24.0