io_uring: don't count rqs failed after current one

When checking for draining with __req_need_defer(), it tries to match
how many requests were sent before a current one with number of already
completed. Dropped SQEs are included in req->sequence, and they won't
ever appear in CQ. To compensate for that, __req_need_defer() substracts
ctx->cached_sq_dropped.
However, what it should really use is number of SQEs dropped __before__
the current one. In other words, any submitted request shouldn't
shouldn't affect dequeueing from the drain queue of previously submitted
ones.

Instead of saving proper ctx->cached_sq_dropped in each request,
substract from req->sequence it at initialisation, so it includes number
of properly submitted requests.

note: it also changes behaviour of timeouts, but
1. it's already diverge from the description because of using SQ
2. the description is ambiguous regarding dropped SQEs

Signed-off-by: Pavel Begunkov <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>

Author: isilence

fs/io_uring.c

@@ -957,8 +957,8 @@ static inline bool __req_need_defer(struct io_kiocb *req)
 {
 	struct io_ring_ctx *ctx = req->ctx;
 
-	return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
-					+ atomic_read(&ctx->cached_cq_overflow);
+	return req->sequence != ctx->cached_cq_tail
+				+ atomic_read(&ctx->cached_cq_overflow);
 }
 
 static inline bool req_need_defer(struct io_kiocb *req)
@@ -5801,7 +5801,7 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	 * it can be used to mark the position of the first IO in the
 	 * link list.
 	 */
-	req->sequence = ctx->cached_sq_head;
+	req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
 	req->opcode = READ_ONCE(sqe->opcode);
 	req->user_data = READ_ONCE(sqe->user_data);
 	req->io = NULL;