---

yaml
---
r: 193593
b: refs/heads/master
c: 550a800
h: refs/heads/master
i:
  193591: 890bae9
v: v3

[refs]

@@ -1,2 +1,2 @@
 ---
-refs/heads/master: 97069788d6784ac92d050557a02f6e7bf4d1f53d
+refs/heads/master: 550a8002e4340eaf3bc333e33b59427e9c20272d

trunk/net/rds/send.c

@@ -632,9 +632,6 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 		list_move(&rm->m_sock_item, &list);
 		rds_send_sndbuf_remove(rs, rm);
 		clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
-
-		/* If this is a RDMA operation, notify the app. */
-		__rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
 	}
 
 	/* order flag updates with the rs lock */
@@ -643,16 +640,17 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 
 	spin_unlock_irqrestore(&rs->rs_lock, flags);
 
-	if (wake)
-		rds_wake_sk_sleep(rs);
-
 	conn = NULL;
 
 	/* now remove the messages from the conn list as needed */
 	list_for_each_entry(rm, &list, m_sock_item) {
 		/* We do this here rather than in the loop above, so that
 		 * we don't have to nest m_rs_lock under rs->rs_lock */
 		spin_lock_irqsave(&rm->m_rs_lock, flags2);
+		/* If this is a RDMA operation, notify the app. */
+		spin_lock(&rs->rs_lock);
+		__rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
+		spin_unlock(&rs->rs_lock);
 		rm->m_rs = NULL;
 		spin_unlock_irqrestore(&rm->m_rs_lock, flags2);
 
@@ -681,6 +679,9 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
 	if (conn)
 		spin_unlock_irqrestore(&conn->c_lock, flags);
 
+	if (wake)
+		rds_wake_sk_sleep(rs);
+
 	while (!list_empty(&list)) {
 		rm = list_entry(list.next, struct rds_message, m_sock_item);
 		list_del_init(&rm->m_sock_item);