[patch 3/3] OCFS2 Configurable timeouts - Protocol changes
authorAndrew Beekhof <abeekhof@suse.de>
Mon, 4 Dec 2006 13:04:55 +0000 (14:04 +0100)
committerMark Fasheh <mark.fasheh@oracle.com>
Mon, 11 Dec 2006 22:26:44 +0000 (14:26 -0800)
Modify the OCFS2 handshake to ensure essential timeouts are configured
identically on all nodes.

Only allow changes when there are no connected peers

Improves the logic in o2net_advance_rx() which broke now that
sizeof(struct o2net_handshake) is greater than sizeof(struct o2net_msg)

Included is the field for userspace-heartbeat timeout to avoid the need for
further protocol changes.

Uses a global spinlock to ensure the decisions to update configfs entries
are made on the correct value.  The region covered by the spinlock when
incrementing the counter is much larger as this is the more critical case.

Small cleanup contributed by Adrian Bunk <bunk@stusta.de>

Signed-off-by: Andrew Beekhof <abeekhof@suse.de>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
fs/ocfs2/cluster/nodemanager.c
fs/ocfs2/cluster/tcp.c
fs/ocfs2/cluster/tcp.h
fs/ocfs2/cluster/tcp_internal.h

index 234f83f2897fd0cc120419512617e2c78c93be2e..357f1d551771201f4b9587186fc01a97809ad06b 100644 (file)
@@ -573,12 +573,21 @@ static ssize_t o2nm_cluster_attr_idle_timeout_ms_write(
        ret =  o2nm_cluster_attr_write(page, count, &val);
 
        if (ret > 0) {
-               if (val <= cluster->cl_keepalive_delay_ms) {
+               if (cluster->cl_idle_timeout_ms != val
+                       && o2net_num_connected_peers()) {
+                       mlog(ML_NOTICE,
+                            "o2net: cannot change idle timeout after "
+                            "the first peer has agreed to it."
+                            "  %d connected peers\n",
+                            o2net_num_connected_peers());
+                       ret = -EINVAL;
+               } else if (val <= cluster->cl_keepalive_delay_ms) {
                        mlog(ML_NOTICE, "o2net: idle timeout must be larger "
                             "than keepalive delay\n");
-                       return -EINVAL;
+                       ret = -EINVAL;
+               } else {
+                       cluster->cl_idle_timeout_ms = val;
                }
-               cluster->cl_idle_timeout_ms = val;
        }
 
        return ret;
@@ -599,12 +608,21 @@ static ssize_t o2nm_cluster_attr_keepalive_delay_ms_write(
        ret =  o2nm_cluster_attr_write(page, count, &val);
 
        if (ret > 0) {
-               if (val >= cluster->cl_idle_timeout_ms) {
+               if (cluster->cl_keepalive_delay_ms != val
+                   && o2net_num_connected_peers()) {
+                       mlog(ML_NOTICE,
+                            "o2net: cannot change keepalive delay after"
+                            " the first peer has agreed to it."
+                            "  %d connected peers\n",
+                            o2net_num_connected_peers());
+                       ret = -EINVAL;
+               } else if (val >= cluster->cl_idle_timeout_ms) {
                        mlog(ML_NOTICE, "o2net: keepalive delay must be "
                             "smaller than idle timeout\n");
-                       return -EINVAL;
+                       ret = -EINVAL;
+               } else {
+                       cluster->cl_keepalive_delay_ms = val;
                }
-               cluster->cl_keepalive_delay_ms = val;
        }
 
        return ret;
index ebbaee664c667d26c4e9562c49697c45885a0fe0..457753df1ae76719933463a268e7fd44a7dd9fc1 100644 (file)
@@ -380,6 +380,13 @@ static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
                sc_put(sc);
 }
 
+static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
+
+int o2net_num_connected_peers(void)
+{
+       return atomic_read(&o2net_connected_peers);
+}
+
 static void o2net_set_nn_state(struct o2net_node *nn,
                               struct o2net_sock_container *sc,
                               unsigned valid, int err)
@@ -390,6 +397,11 @@ static void o2net_set_nn_state(struct o2net_node *nn,
 
        assert_spin_locked(&nn->nn_lock);
 
+       if (old_sc && !sc)
+               atomic_dec(&o2net_connected_peers);
+       else if (!old_sc && sc)
+               atomic_inc(&o2net_connected_peers);
+
        /* the node num comparison and single connect/accept path should stop
         * an non-null sc from being overwritten with another */
        BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
@@ -1123,6 +1135,44 @@ static int o2net_check_handshake(struct o2net_sock_container *sc)
                return -1;
        }
 
+       /*
+        * Ensure timeouts are consistent with other nodes, otherwise
+        * we can end up with one node thinking that the other must be down,
+        * but isn't. This can ultimately cause corruption.
+        */
+       if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
+                               o2net_idle_timeout(sc->sc_node)) {
+               mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
+                    "%u ms, but we use %u ms locally.  disconnecting\n",
+                    SC_NODEF_ARGS(sc),
+                    be32_to_cpu(hand->o2net_idle_timeout_ms),
+                    o2net_idle_timeout(sc->sc_node));
+               o2net_ensure_shutdown(nn, sc, -ENOTCONN);
+               return -1;
+       }
+
+       if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
+                       o2net_keepalive_delay(sc->sc_node)) {
+               mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
+                    "%u ms, but we use %u ms locally.  disconnecting\n",
+                    SC_NODEF_ARGS(sc),
+                    be32_to_cpu(hand->o2net_keepalive_delay_ms),
+                    o2net_keepalive_delay(sc->sc_node));
+               o2net_ensure_shutdown(nn, sc, -ENOTCONN);
+               return -1;
+       }
+
+       if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
+                       O2HB_MAX_WRITE_TIMEOUT_MS) {
+               mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
+                    "%u ms, but we use %u ms locally.  disconnecting\n",
+                    SC_NODEF_ARGS(sc),
+                    be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
+                    O2HB_MAX_WRITE_TIMEOUT_MS);
+               o2net_ensure_shutdown(nn, sc, -ENOTCONN);
+               return -1;
+       }
+
        sc->sc_handshake_ok = 1;
 
        spin_lock(&nn->nn_lock);
@@ -1155,6 +1205,23 @@ static int o2net_advance_rx(struct o2net_sock_container *sc)
        sclog(sc, "receiving\n");
        do_gettimeofday(&sc->sc_tv_advance_start);
 
+       if (unlikely(sc->sc_handshake_ok == 0)) {
+               if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
+                       data = page_address(sc->sc_page) + sc->sc_page_off;
+                       datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
+                       ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
+                       if (ret > 0)
+                               sc->sc_page_off += ret;
+               }
+
+               if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
+                       o2net_check_handshake(sc);
+                       if (unlikely(sc->sc_handshake_ok == 0))
+                               ret = -EPROTO;
+               }
+               goto out;
+       }
+
        /* do we need more header? */
        if (sc->sc_page_off < sizeof(struct o2net_msg)) {
                data = page_address(sc->sc_page) + sc->sc_page_off;
@@ -1162,15 +1229,6 @@ static int o2net_advance_rx(struct o2net_sock_container *sc)
                ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
                if (ret > 0) {
                        sc->sc_page_off += ret;
-
-                       /* this working relies on the handshake being
-                        * smaller than the normal message header */
-                       if (sc->sc_page_off >= sizeof(struct o2net_handshake)&&
-                           !sc->sc_handshake_ok && o2net_check_handshake(sc)) {
-                               ret = -EPROTO;
-                               goto out;
-                       }
-
                        /* only swab incoming here.. we can
                         * only get here once as we cross from
                         * being under to over */
@@ -1272,6 +1330,18 @@ static int o2net_set_nodelay(struct socket *sock)
        return ret;
 }
 
+static void o2net_initialize_handshake(void)
+{
+       o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
+               O2HB_MAX_WRITE_TIMEOUT_MS);
+       o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(
+               o2net_idle_timeout(NULL));
+       o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
+               o2net_keepalive_delay(NULL));
+       o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
+               o2net_reconnect_delay(NULL));
+}
+
 /* ------------------------------------------------------------ */
 
 /* called when a connect completes and after a sock is accepted.  the
@@ -1286,6 +1356,7 @@ static void o2net_sc_connect_completed(struct work_struct *work)
               (unsigned long long)O2NET_PROTOCOL_VERSION,
              (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
 
+       o2net_initialize_handshake();
        o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
        sc_put(sc);
 }
@@ -1514,6 +1585,8 @@ static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
 
        if (node_num != o2nm_this_node())
                o2net_disconnect_node(node);
+
+       BUG_ON(atomic_read(&o2net_connected_peers) < 0);
 }
 
 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
@@ -1677,6 +1750,7 @@ static int o2net_accept_one(struct socket *sock)
        o2net_register_callbacks(sc->sc_sock->sk, sc);
        o2net_sc_queue_work(sc, &sc->sc_rx_work);
 
+       o2net_initialize_handshake();
        o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
 
 out:
index 2e08976050fb0d448cd4e2968cf0672af3e4caa4..21a4e43df836159c177771974ffc215957d26bf3 100644 (file)
@@ -108,6 +108,7 @@ void o2net_unregister_hb_callbacks(void);
 int o2net_start_listening(struct o2nm_node *node);
 void o2net_stop_listening(struct o2nm_node *node);
 void o2net_disconnect_node(struct o2nm_node *node);
+int o2net_num_connected_peers(void);
 
 int o2net_init(void);
 void o2net_exit(void);
index 56f7ee1d254737a917cfa1ef175d2b93cac719e3..b700dc9624d13774cd693ab3bde426829fc9e79a 100644 (file)
@@ -38,6 +38,9 @@
  * locking semantics of the file system using the protocol.  It should 
  * be somewhere else, I'm sure, but right now it isn't.
  *
+ * New in version 5:
+ *     - Network timeout checking protocol
+ *
  * New in version 4:
  *     - Remove i_generation from lock names for better stat performance.
  *
  *     - full 64 bit i_size in the metadata lock lvbs
  *     - introduction of "rw" lock and pushing meta/data locking down
  */
-#define O2NET_PROTOCOL_VERSION 4ULL
+#define O2NET_PROTOCOL_VERSION 5ULL
 struct o2net_handshake {
        __be64  protocol_version;
        __be64  connector_id;
+       __be32  o2hb_heartbeat_timeout_ms;
+       __be32  o2net_idle_timeout_ms;
+       __be32  o2net_keepalive_delay_ms;
+       __be32  o2net_reconnect_delay_ms;
 };
 
 struct o2net_node {