dlm: fix connection close handling
authorLars Marowsky-Bree <lmb@suse.de>
Tue, 11 Aug 2009 21:18:23 +0000 (16:18 -0500)
committerDavid Teigland <teigland@redhat.com>
Mon, 24 Aug 2009 18:13:56 +0000 (13:13 -0500)
Closing a connection to a node can create problems if there are
outstanding messages for that node.  The problems include dlm_send
spinning attempting to reconnect, or BUG from tcp_connect_to_sock()
attempting to use a partially closed connection.

To cleanly close a connection, we now first attempt to send any pending
messages, cancel any remaining workqueue work, and flag the connection
as closed to avoid reconnect attempts.

Signed-off-by: Lars Marowsky-Bree <lmb@suse.de>
Signed-off-by: Christine Caulfield <ccaulfie@redhat.com>
Signed-off-by: David Teigland <teigland@redhat.com>
fs/dlm/lowcomms.c

index 210d52c48808d77f8b2fa4139675062aab309091..bda690cd3640e53fc5b91fe17a81f2b82b2bb9c7 100644 (file)
@@ -106,6 +106,7 @@ struct connection {
 #define CF_CONNECT_PENDING 3
 #define CF_INIT_PENDING 4
 #define CF_IS_OTHERCON 5
+#define CF_CLOSE 6
        struct list_head writequeue;  /* List of outgoing writequeue_entries */
        spinlock_t writequeue_lock;
        int (*rx_action) (struct connection *); /* What to do when active */
@@ -299,6 +300,8 @@ static void lowcomms_write_space(struct sock *sk)
 
 static inline void lowcomms_connect_sock(struct connection *con)
 {
+       if (test_bit(CF_CLOSE, &con->flags))
+               return;
        if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags))
                queue_work(send_workqueue, &con->swork);
 }
@@ -1368,6 +1371,13 @@ int dlm_lowcomms_close(int nodeid)
        log_print("closing connection to node %d", nodeid);
        con = nodeid2con(nodeid, 0);
        if (con) {
+               clear_bit(CF_CONNECT_PENDING, &con->flags);
+               clear_bit(CF_WRITE_PENDING, &con->flags);
+               set_bit(CF_CLOSE, &con->flags);
+               if (cancel_work_sync(&con->swork))
+                       log_print("canceled swork for node %d", nodeid);
+               if (cancel_work_sync(&con->rwork))
+                       log_print("canceled rwork for node %d", nodeid);
                clean_one_writequeue(con);
                close_connection(con, true);
        }
@@ -1393,9 +1403,10 @@ static void process_send_sockets(struct work_struct *work)
 
        if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
                con->connect_action(con);
+               set_bit(CF_WRITE_PENDING, &con->flags);
        }
-       clear_bit(CF_WRITE_PENDING, &con->flags);
-       send_to_sock(con);
+       if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags))
+               send_to_sock(con);
 }