libceph: fix messenger retry
Sage Weil [Tue, 10 Jul 2012 18:53:34 +0000 (11:53 -0700)]
(cherry picked from commit 5bdca4e0768d3e0f4efa43d9a2cc8210aeb91ab9)

In ancient times, the messenger could both initiate and accept connections.
An artifact if that was data structures to store/process an incoming
ceph_msg_connect request and send an outgoing ceph_msg_connect_reply.
Sadly, the negotiation code was referencing those structures and ignoring
important information (like the peer's connect_seq) from the correct ones.

Among other things, this fixes tight reconnect loops where the server sends
RETRY_SESSION and we (the client) retries with the same connect_seq as last
time.  This bug pretty easily triggered by injecting socket failures on the
MDS and running some fs workload like workunits/direct_io/test_sync_io.

Signed-off-by: Sage Weil <sage@inktank.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

include/linux/ceph/messenger.h
net/ceph/messenger.c

index 2521a95..44c87e7 100644 (file)
@@ -163,16 +163,8 @@ struct ceph_connection {
 
        /* connection negotiation temps */
        char in_banner[CEPH_BANNER_MAX_LEN];
-       union {
-               struct {  /* outgoing connection */
-                       struct ceph_msg_connect out_connect;
-                       struct ceph_msg_connect_reply in_reply;
-               };
-               struct {  /* incoming */
-                       struct ceph_msg_connect in_connect;
-                       struct ceph_msg_connect_reply out_reply;
-               };
-       };
+       struct ceph_msg_connect out_connect;
+       struct ceph_msg_connect_reply in_reply;
        struct ceph_entity_addr actual_peer_addr;
 
        /* message out temps */
index 375ae39..0e27016 100644 (file)
@@ -1423,7 +1423,7 @@ static int process_connect(struct ceph_connection *con)
                 * dropped messages.
                 */
                dout("process_connect got RESET peer seq %u\n",
-                    le32_to_cpu(con->in_connect.connect_seq));
+                    le32_to_cpu(con->in_reply.connect_seq));
                pr_err("%s%lld %s connection reset\n",
                       ENTITY_NAME(con->peer_name),
                       ceph_pr_addr(&con->peer_addr.in_addr));
@@ -1450,10 +1450,10 @@ static int process_connect(struct ceph_connection *con)
                 * If we sent a smaller connect_seq than the peer has, try
                 * again with a larger value.
                 */
-               dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
+               dout("process_connect got RETRY_SESSION my seq %u, peer %u\n",
                     le32_to_cpu(con->out_connect.connect_seq),
-                    le32_to_cpu(con->in_connect.connect_seq));
-               con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
+                    le32_to_cpu(con->in_reply.connect_seq));
+               con->connect_seq = le32_to_cpu(con->in_reply.connect_seq);
                ceph_con_out_kvec_reset(con);
                ret = prepare_write_connect(con);
                if (ret < 0)
@@ -1468,9 +1468,9 @@ static int process_connect(struct ceph_connection *con)
                 */
                dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
                     con->peer_global_seq,
-                    le32_to_cpu(con->in_connect.global_seq));
+                    le32_to_cpu(con->in_reply.global_seq));
                get_global_seq(con->msgr,
-                              le32_to_cpu(con->in_connect.global_seq));
+                              le32_to_cpu(con->in_reply.global_seq));
                ceph_con_out_kvec_reset(con);
                ret = prepare_write_connect(con);
                if (ret < 0)