[PATCH liburing 1/1] src/queue: simplify IORING_SETUP_SQ_REWIND handling

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

io_uring_load_sq_head() shouldn't need a IORING_SETUP_SQ_REWIND check
as the SQ head should already be zero. Also, add a couple of words about
the tail resetting logic.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 src/include/liburing.h | 3 ---
 src/queue.c            | 7 +++++++
 2 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/src/include/liburing.h b/src/include/liburing.h
index b41958ce..8d45d400 100644
--- a/src/include/liburing.h
+++ b/src/include/liburing.h
@@ -1684,9 +1684,6 @@ IOURINGINLINE void io_uring_prep_pipe_direct(struct io_uring_sqe *sqe, int *fds,
 IOURINGINLINE unsigned io_uring_load_sq_head(const struct io_uring *ring)
 	LIBURING_NOEXCEPT
 {
-	if (ring->flags & IORING_SETUP_SQ_REWIND)
-		return 0;
-
 	/*
 	 * Without acquire ordering, we could overwrite a SQE before the kernel
 	 * finished reading it. We don't need the acquire ordering for
diff --git a/src/queue.c b/src/queue.c
index 00995ace..87b4dbbe 100644
--- a/src/queue.c
+++ b/src/queue.c
@@ -205,6 +205,13 @@ static unsigned __io_uring_flush_sq(struct io_uring *ring)
 	struct io_uring_sq *sq = &ring->sq;
 	unsigned tail = sq->sqe_tail;
 
+	/*
+	 * With IORING_SETUP_SQ_REWIND, the kernel ignores the SQ head / tail
+	 * and submits entries from the beginning of the queue. Continue using
+	 * the indices as before but reset the tail on submission. With the
+	 * head kept to be zero, io_uring_get_sqe() / etc. will work without
+	 * any extra changes.
+	 */
 	if (ring->flags & IORING_SETUP_SQ_REWIND) {
 		sq->sqe_tail = 0;
 		return tail;
-- 
2.52.0