nbd: clear last_sqe's SQE_LINK after the staggered ios are queued

If staggered ios in next sqe chain are just queued inside
nbd_handle_io_bg() because current sqe chain are completed, we have
to clear the last sqe's SQE_LINK after the whole next sqe chain is
queued.

Fix one io hang issue, which is often triggered when running 512k bs
IO with localhost nbd server. And not easy to reproduce with remote
nbd server.

Signed-off-by: Ming Lei <tom.leiming@gmail.com>

nbd/tgt_nbd.cpp

@@ -626,11 +626,6 @@ static void nbd_handle_send_bg(const struct ublksrv_queue *q,
 	if (q_data->send_sqe_chain_busy && !q_data->chained_send_ios)
 		q_data->send_sqe_chain_busy = 0;
 
-	if (q_data->last_send_sqe) {
-		q_data->last_send_sqe->flags &= ~IOSQE_IO_LINK;
-		q_data->last_send_sqe = NULL;
-	}
-
 	if (!q_data->send_sqe_chain_busy) {
 		std::vector<const struct ublk_io_data *> &ios =
 			q_data->next_chain;
@@ -648,6 +643,10 @@ static void nbd_handle_send_bg(const struct ublksrv_queue *q,
 		if (q_data->chained_send_ios && !q_data->send_sqe_chain_busy)
 			q_data->send_sqe_chain_busy = 1;
 	}
+	if (q_data->last_send_sqe) {
+		q_data->last_send_sqe->flags &= ~IOSQE_IO_LINK;
+		q_data->last_send_sqe = NULL;
+	}
 }
 
 static void nbd_recv_reply(const struct ublksrv_queue *q)