clear cluster_master on election so that slaves will accept a new master
[l2tpns.git] / cluster.c
index 808ab84..8f9c01e 100644 (file)
--- a/cluster.c
+++ b/cluster.c
@@ -1,6 +1,6 @@
 // L2TPNS Clustering Stuff
 
 // L2TPNS Clustering Stuff
 
-char const *cvs_id_cluster = "$Id: cluster.c,v 1.26 2004/12/16 23:40:31 bodea Exp $";
+char const *cvs_id_cluster = "$Id: cluster.c,v 1.26.2.12 2005/07/03 02:58:06 bodea Exp $";
 
 #include <stdio.h>
 #include <stdlib.h>
 
 #include <stdio.h>
 #include <stdlib.h>
@@ -192,7 +192,7 @@ static void add_type(char **p, int type, int more, char *data, int size)
 }
 
 // advertise our presence via BGP or gratuitous ARP
 }
 
 // advertise our presence via BGP or gratuitous ARP
-static void advertise(void)
+static void advertise_routes(void)
 {
 #ifdef BGP
        if (bgp_configured)
 {
 #ifdef BGP
        if (bgp_configured)
@@ -203,6 +203,15 @@ static void advertise(void)
                        send_garp(config->bind_address);        // Start taking traffic.
 }
 
                        send_garp(config->bind_address);        // Start taking traffic.
 }
 
+// withdraw our routes (BGP only)
+static void withdraw_routes(void)
+{
+#ifdef BGP
+       if (bgp_configured)
+               bgp_enable_routing(0);
+#endif /* BGP */
+}
+
 static void cluster_uptodate(void)
 {
        if (config->cluster_iam_uptodate)
 static void cluster_uptodate(void)
 {
        if (config->cluster_iam_uptodate)
@@ -214,7 +223,7 @@ static void cluster_uptodate(void)
        config->cluster_iam_uptodate = 1;
 
        LOG(0, 0, 0, "Now uptodate with master.\n");
        config->cluster_iam_uptodate = 1;
 
        LOG(0, 0, 0, "Now uptodate with master.\n");
-       advertise();
+       advertise_routes();
 }
 
 //
 }
 
 //
@@ -405,18 +414,18 @@ void master_update_counts(void)
                if ( walk_session_number > config->cluster_highest_sessionid)
                        walk_session_number = 1;
 
                if ( walk_session_number > config->cluster_highest_sessionid)
                        walk_session_number = 1;
 
-               if (!sess_count[walk_session_number].cin && !sess_count[walk_session_number].cout)
+               if (!sess_local[walk_session_number].cin && !sess_local[walk_session_number].cout)
                        continue; // Unused. Skip it.
 
                b[c].sid = walk_session_number;
                        continue; // Unused. Skip it.
 
                b[c].sid = walk_session_number;
-               b[c].in = sess_count[walk_session_number].cin;
-               b[c].out = sess_count[walk_session_number].cout;
+               b[c].in = sess_local[walk_session_number].cin;
+               b[c].out = sess_local[walk_session_number].cout;
 
                if (++c > MAX_B_RECS)   // Send a max of 400 elements in a packet.
                        break;
 
                        // Reset counters.
 
                if (++c > MAX_B_RECS)   // Send a max of 400 elements in a packet.
                        break;
 
                        // Reset counters.
-               sess_count[walk_session_number].cin = sess_count[walk_session_number].cout = 0;
+               sess_local[walk_session_number].cin = sess_local[walk_session_number].cout = 0;
        }
 
        if (!c)         // Didn't find any that changes. Get out of here!
        }
 
        if (!c)         // Didn't find any that changes. Get out of here!
@@ -456,17 +465,22 @@ void cluster_check_slaves(void)
                        continue;       // Shutdown peer! Skip them.
 
                if (peers[i].uptodate)
                        continue;       // Shutdown peer! Skip them.
 
                if (peers[i].uptodate)
-                       have_peers = 1;
-
-               if (!peers[i].uptodate)
+                       have_peers++;
+               else
                        config->cluster_iam_uptodate = 0; // Start fast heartbeats
        }
 
                        config->cluster_iam_uptodate = 0; // Start fast heartbeats
        }
 
-#ifdef BGP
-       // in a cluster, withdraw/add routes when we get a peer/lose all peers
-       if (bgp_configured && have_peers != had_peers)
-               bgp_enable_routing(!have_peers);
-#endif /* BGP */
+       // in a cluster, withdraw/add routes when we get a peer/lose peers
+       if (have_peers != had_peers)
+       {
+               if (had_peers < config->cluster_master_min_adv &&
+                   have_peers >= config->cluster_master_min_adv)
+                       withdraw_routes();
+
+               else if (had_peers >= config->cluster_master_min_adv &&
+                   have_peers < config->cluster_master_min_adv)
+                       advertise_routes();
+       }
 }
 
 //
 }
 
 //
@@ -479,20 +493,22 @@ void cluster_check_master(void)
        int last_free = 0;
        clockt t = TIME;
        static int probed = 0;
        int last_free = 0;
        clockt t = TIME;
        static int probed = 0;
+       int have_peers;
 
        if (config->cluster_iam_master)
                return;         // Only runs on the slaves...
 
        // If the master is late (missed 2 hearbeats by a second and a
        // hair) it may be that the switch has dropped us from the
 
        if (config->cluster_iam_master)
                return;         // Only runs on the slaves...
 
        // If the master is late (missed 2 hearbeats by a second and a
        // hair) it may be that the switch has dropped us from the
-       // multicast group, try unicasting one probe to the master
+       // multicast group, try unicasting probes to the master
        // which will hopefully respond with a unicast heartbeat that
        // will allow us to limp along until the querier next runs.
        // which will hopefully respond with a unicast heartbeat that
        // will allow us to limp along until the querier next runs.
-       if (TIME > (config->cluster_last_hb + 2 * config->cluster_hb_interval + 11))
+       if (config->cluster_master_address
+           && TIME > (config->cluster_last_hb + 2 * config->cluster_hb_interval + 11))
        {
        {
-               if (!probed && config->cluster_master_address)
+               if (!probed || (TIME > (probed + 2 * config->cluster_hb_interval)))
                {
                {
-                       probed = 1;
+                       probed = TIME;
                        LOG(1, 0, 0, "Heartbeat from master %.1fs late, probing...\n",
                                0.1 * (TIME - (config->cluster_last_hb + config->cluster_hb_interval)));
 
                        LOG(1, 0, 0, "Heartbeat from master %.1fs late, probing...\n",
                                0.1 * (TIME - (config->cluster_last_hb + config->cluster_hb_interval)));
 
@@ -507,10 +523,15 @@ void cluster_check_master(void)
                return; // Everything's ok!
 
        config->cluster_last_hb = TIME + 1;     // Just the one election thanks.
                return; // Everything's ok!
 
        config->cluster_last_hb = TIME + 1;     // Just the one election thanks.
+       config->cluster_master_address = 0;
 
        LOG(0, 0, 0, "Master timed out! Holding election...\n");
 
 
        LOG(0, 0, 0, "Master timed out! Holding election...\n");
 
-       for (i = 0; i < num_peers; i++)
+       // In the process of shutting down, can't be master
+       if (main_quit)
+               return;
+
+       for (i = have_peers = 0; i < num_peers; i++)
        {
                if ((peers[i].timestamp + config->cluster_hb_timeout) < t)
                        continue;       // Stale peer! Skip them.
        {
                if ((peers[i].timestamp + config->cluster_hb_timeout) < t)
                        continue;       // Stale peer! Skip them.
@@ -528,6 +549,9 @@ void cluster_check_master(void)
                        LOG(1, 0, 0, "Expecting %s to become master\n", fmtaddr(peers[i].peer, 0));
                        return;         // They'll win the election. Wait for them to come up.
                }
                        LOG(1, 0, 0, "Expecting %s to become master\n", fmtaddr(peers[i].peer, 0));
                        return;         // They'll win the election. Wait for them to come up.
                }
+
+               if (peers[i].uptodate)
+                       have_peers++;
        }
 
                // Wow. it's been ages since I last heard a heartbeat
        }
 
                // Wow. it's been ages since I last heard a heartbeat
@@ -535,10 +559,14 @@ void cluster_check_master(void)
                // to become a master!!!
 
        config->cluster_iam_master = 1;
                // to become a master!!!
 
        config->cluster_iam_master = 1;
-       config->cluster_master_address = 0;
 
        LOG(0, 0, 0, "I am declaring myself the master!\n");
 
 
        LOG(0, 0, 0, "I am declaring myself the master!\n");
 
+       if (have_peers < config->cluster_master_min_adv)
+               advertise_routes();
+       else
+               withdraw_routes();
+
        if (config->cluster_seq_number == -1)
                config->cluster_seq_number = 0;
 
        if (config->cluster_seq_number == -1)
                config->cluster_seq_number = 0;
 
@@ -570,38 +598,42 @@ void cluster_check_master(void)
                        ++count;
                }
 
                        ++count;
                }
 
-               if (session[i].tunnel == T_FREE) { // Unused session. Add to free list.
+               if (!session[i].opened) { // Unused session. Add to free list.
+                       memset(&session[i], 0, sizeof(session[i]));
+                       session[i].tunnel = T_FREE;
                        session[last_free].next = i;
                        session[i].next = 0;
                        last_free = i;
                        session[last_free].next = i;
                        session[i].next = 0;
                        last_free = i;
+                       continue;
                }
 
                }
 
-                       // Reset all the idle timeouts..
+                       // Reset idle timeouts..
                session[i].last_packet = time_now;
 
                session[i].last_packet = time_now;
 
+                       // Reset die relative to our uptime rather than the old master's
+               if (session[i].die) session[i].die = TIME;
+
                        // Accumulate un-sent byte counters.
                        // Accumulate un-sent byte counters.
-               session[i].cin += sess_count[i].cin;
-               session[i].cout += sess_count[i].cout;
-               session[i].total_cin += sess_count[i].cin;
-               session[i].total_cout += sess_count[i].cout;
+               session[i].cin += sess_local[i].cin;
+               session[i].cout += sess_local[i].cout;
+               session[i].total_cin += sess_local[i].cin;
+               session[i].total_cout += sess_local[i].cout;
 
 
-               sess_count[i].cin = sess_count[i].cout = 0;
+               sess_local[i].cin = sess_local[i].cout = 0;
 
                session[i].radius = 0;  // Reset authentication as the radius blocks aren't up to date.
 
                if (session[i].unique_id >= high_unique_id)     // This is different to the index into the session table!!!
                        high_unique_id = session[i].unique_id+1;
 
 
                session[i].radius = 0;  // Reset authentication as the radius blocks aren't up to date.
 
                if (session[i].unique_id >= high_unique_id)     // This is different to the index into the session table!!!
                        high_unique_id = session[i].unique_id+1;
 
-
                session[i].tbf_in = session[i].tbf_out = 0; // Remove stale pointers from old master.
                throttle_session(i, session[i].throttle_in, session[i].throttle_out);
 
                session[i].tbf_in = session[i].tbf_out = 0; // Remove stale pointers from old master.
                throttle_session(i, session[i].throttle_in, session[i].throttle_out);
 
-               if (session[i].tunnel != T_FREE && i > config->cluster_highest_sessionid)
-                       config->cluster_highest_sessionid = i;
+               config->cluster_highest_sessionid = i;
        }
 
        session[last_free].next = 0;    // End of chain.
        }
 
        session[last_free].next = 0;    // End of chain.
-       last_id = high_unique_id;               // Keep track of the highest used session ID.
+       last_id = high_unique_id;       // Keep track of the highest used session ID.
 
        become_master();
 
 
        become_master();
 
@@ -615,9 +647,6 @@ void cluster_check_master(void)
        config->cluster_undefined_tunnels = 0;
        config->cluster_iam_uptodate = 1; // assume all peers are up-to-date
 
        config->cluster_undefined_tunnels = 0;
        config->cluster_iam_uptodate = 1; // assume all peers are up-to-date
 
-       if (!num_peers) // lone master
-               advertise();
-
        // FIXME. We need to fix up the tunnel control message
        // queue here! There's a number of other variables we
        // should also update.
        // FIXME. We need to fix up the tunnel control message
        // queue here! There's a number of other variables we
        // should also update.
@@ -649,12 +678,12 @@ static void cluster_check_sessions(int highsession, int freesession_ptr, int hig
        config->cluster_undefined_sessions = 0;
        for (i = 1 ; i < MAXSESSION; ++i) {
                if (i > highsession) {
        config->cluster_undefined_sessions = 0;
        for (i = 1 ; i < MAXSESSION; ++i) {
                if (i > highsession) {
-                       session[i].tunnel = 0; // Defined.
+                       if (session[i].tunnel == T_UNDEF) session[i].tunnel = T_FREE; // Defined.
                        continue;
                }
                        continue;
                }
-               if (session[i].tunnel != T_UNDEF)
-                       continue;
-               ++config->cluster_undefined_sessions;
+
+               if (session[i].tunnel == T_UNDEF)
+                       ++config->cluster_undefined_sessions;
        }
 
                // Clear out defined tunnels, counting the number of
        }
 
                // Clear out defined tunnels, counting the number of
@@ -662,12 +691,12 @@ static void cluster_check_sessions(int highsession, int freesession_ptr, int hig
        config->cluster_undefined_tunnels = 0;
        for (i = 1 ; i < MAXTUNNEL; ++i) {
                if (i > hightunnel) {
        config->cluster_undefined_tunnels = 0;
        for (i = 1 ; i < MAXTUNNEL; ++i) {
                if (i > hightunnel) {
-                       tunnel[i].state = TUNNELFREE; // Defined.
+                       if (tunnel[i].state == TUNNELUNDEF) tunnel[i].state = TUNNELFREE; // Defined.
                        continue;
                }
                        continue;
                }
-               if (tunnel[i].state != TUNNELUNDEF)
-                       continue;
-               ++config->cluster_undefined_tunnels;
+
+               if (tunnel[i].state == TUNNELUNDEF)
+                       ++config->cluster_undefined_tunnels;
        }
 
 
        }
 
 
@@ -892,6 +921,13 @@ static int cluster_catchup_slave(int seq, in_addr_t slave)
        int diff;
 
        LOG(1, 0, 0, "Slave %s sent LASTSEEN with seq %d\n", fmtaddr(slave, 0), seq);
        int diff;
 
        LOG(1, 0, 0, "Slave %s sent LASTSEEN with seq %d\n", fmtaddr(slave, 0), seq);
+       if (!config->cluster_iam_master) {
+               LOG(1, 0, 0, "Got LASTSEEN but I'm not a master! Redirecting it to %s.\n",
+                       fmtaddr(config->cluster_master_address, 0));
+
+               peer_send_message(slave, C_MASTER, config->cluster_master_address, NULL, 0);
+               return 0;
+       }
 
        diff = config->cluster_seq_number - seq;        // How many packet do we need to send?
        if (diff < 0)
 
        diff = config->cluster_seq_number - seq;        // How many packet do we need to send?
        if (diff < 0)
@@ -903,9 +939,11 @@ static int cluster_catchup_slave(int seq, in_addr_t slave)
                return peer_send_message(slave, C_KILL, seq, NULL, 0);// Kill the slave. Nothing else to do.
        }
 
                return peer_send_message(slave, C_KILL, seq, NULL, 0);// Kill the slave. Nothing else to do.
        }
 
+       LOG(1, 0, 0, "Sending %d catchup packets to slave %s\n", diff, fmtaddr(slave, 0) );
+
                // Now resend every packet that it missed, in order.
        while (seq != config->cluster_seq_number) {
                // Now resend every packet that it missed, in order.
        while (seq != config->cluster_seq_number) {
-               s = seq%HB_HISTORY_SIZE;
+               s = seq % HB_HISTORY_SIZE;
                if (seq != past_hearts[s].seq) {
                        LOG(0, 0, 0, "Tried to re-send heartbeat for %s but %d doesn't match %d! (%d,%d)\n",
                                fmtaddr(slave, 0), seq, past_hearts[s].seq, s, config->cluster_seq_number);
                if (seq != past_hearts[s].seq) {
                        LOG(0, 0, 0, "Tried to re-send heartbeat for %s but %d doesn't match %d! (%d,%d)\n",
                                fmtaddr(slave, 0), seq, past_hearts[s].seq, s, config->cluster_seq_number);
@@ -958,8 +996,10 @@ static int cluster_add_peer(in_addr_t peer, time_t basetime, pingt *pp, int size
        }
 
        // Is this the master shutting down??
        }
 
        // Is this the master shutting down??
-       if (peer == config->cluster_master_address && !basetime) {
-               LOG(3, 0, 0, "Master %s shutting down...\n", fmtaddr(config->cluster_master_address, 0));
+       if (peer == config->cluster_master_address) {
+               LOG(3, 0, 0, "Master %s %s\n", fmtaddr(config->cluster_master_address, 0),
+                       basetime ? "has restarted!" : "shutting down...");
+
                config->cluster_master_address = 0;
                config->cluster_last_hb = 0; // Force an election.
                cluster_check_master();
                config->cluster_master_address = 0;
                config->cluster_last_hb = 0; // Force an election.
                cluster_check_master();
@@ -1000,6 +1040,20 @@ static int cluster_add_peer(in_addr_t peer, time_t basetime, pingt *pp, int size
        return 1;
 }
 
        return 1;
 }
 
+// A slave responds with C_MASTER when it gets a message which should have gone to a master.
+static int cluster_set_master(in_addr_t peer, in_addr_t master)
+{
+       if (config->cluster_iam_master) // Sanity...
+               return 0;
+
+       LOG(3, 0, 0, "Peer %s set the master to %s...\n", fmtaddr(peer, 0),
+               fmtaddr(master, 1));
+
+       config->cluster_master_address = master;
+       cluster_check_master();
+       return 0;
+}
+
 /* Handle the slave updating the byte counters for the master. */
 //
 // Note that we don't mark the session as dirty; We rely on
 /* Handle the slave updating the byte counters for the master. */
 //
 // Note that we don't mark the session as dirty; We rely on
@@ -1027,7 +1081,9 @@ static int cluster_handle_bytes(char *data, int size)
 
                session[b->sid].cin += b->in;
                session[b->sid].cout += b->out;
 
                session[b->sid].cin += b->in;
                session[b->sid].cout += b->out;
-               session[b->sid].last_packet = time_now; // Reset idle timer!
+
+               if (b->in)
+                       session[b->sid].last_packet = time_now; // Reset idle timer!
 
                size -= sizeof(*b);
                ++b;
 
                size -= sizeof(*b);
                ++b;
@@ -1168,18 +1224,54 @@ static int cluster_process_heartbeat(uint8_t *data, int size, int more, uint8_t
                        exit(1);
                }
 
                        exit(1);
                }
 
+                       //
+                       // Send it a unicast heartbeat to see give it a chance to die.
+                       // NOTE: It's actually safe to do seq-number - 1 without checking
+                       // for wrap around.
+                       //
+               cluster_catchup_slave(config->cluster_seq_number - 1, addr);
+
                return -1; // Skip it.
        }
 
                return -1; // Skip it.
        }
 
+               //
+               // Try and guard against a stray master appearing.
+               //
+               // Ignore heartbeats received from another master before the
+               // timeout (less a smidgen) for the old master has elapsed.
+               //
+               // Note that after a clean failover, the cluster_master_address
+               // is cleared, so this doesn't run. 
+               //
+       if (config->cluster_master_address && addr != config->cluster_master_address) {
+                   LOG(0, 0, 0, "Ignoring stray heartbeat from %s, current master %s has not yet timed out (last heartbeat %.1f seconds ago).\n",
+                           fmtaddr(addr, 0), fmtaddr(config->cluster_master_address, 1),
+                           0.1 * (TIME - config->cluster_last_hb));
+                   return -1; // ignore
+       }
+
        if (config->cluster_seq_number == -1)   // Don't have one. Just align to the master...
                config->cluster_seq_number = h->seq;
 
        config->cluster_last_hb = TIME; // Reset to ensure that we don't become master!!
 
        if (config->cluster_seq_number != h->seq) {     // Out of sequence heartbeat!
        if (config->cluster_seq_number == -1)   // Don't have one. Just align to the master...
                config->cluster_seq_number = h->seq;
 
        config->cluster_last_hb = TIME; // Reset to ensure that we don't become master!!
 
        if (config->cluster_seq_number != h->seq) {     // Out of sequence heartbeat!
-               LOG(1, 0, 0, "HB: Got seq# %d but was expecting %d. asking for resend.\n", h->seq, config->cluster_seq_number);
+               static int lastseen_seq = 0;
+               static time_t lastseen_time = 0;
 
 
-               peer_send_message(addr, C_LASTSEEN, config->cluster_seq_number, NULL, 0);
+               // limit to once per second for a particular seq#
+               int ask = (config->cluster_seq_number != lastseen_seq || time_now != lastseen_time);
+
+               LOG(1, 0, 0, "HB: Got seq# %d but was expecting %d.  %s.\n",
+                       h->seq, config->cluster_seq_number,
+                       ask ? "Asking for resend" : "Ignoring");
+
+               if (ask)
+               {
+                       lastseen_seq = config->cluster_seq_number;
+                       lastseen_time = time_now;
+                       peer_send_message(addr, C_LASTSEEN, config->cluster_seq_number, NULL, 0);
+               }
 
                config->cluster_last_hb = TIME; // Reset to ensure that we don't become master!!
 
 
                config->cluster_last_hb = TIME; // Reset to ensure that we don't become master!!
 
@@ -1335,10 +1427,13 @@ int processcluster(char *data, int size, in_addr_t addr)
        s -= sizeof(uint32_t);
 
        switch (type) {
        s -= sizeof(uint32_t);
 
        switch (type) {
-       case C_PING:    // Update the peers table.
+       case C_PING: // Update the peers table.
                return cluster_add_peer(addr, more, (pingt *) p, s);
 
                return cluster_add_peer(addr, more, (pingt *) p, s);
 
-       case C_LASTSEEN:        // Catch up a slave (slave missed a packet).
+       case C_MASTER: // Our master is wrong
+               return cluster_set_master(addr, more);
+
+       case C_LASTSEEN: // Catch up a slave (slave missed a packet).
                return cluster_catchup_slave(more, addr);
 
        case C_FORWARD: { // Forwarded control packet. pass off to processudp.
                return cluster_catchup_slave(more, addr);
 
        case C_FORWARD: { // Forwarded control packet. pass off to processudp.
@@ -1379,6 +1474,11 @@ int processcluster(char *data, int size, in_addr_t addr)
                return 0;
 
        case C_BYTES:
                return 0;
 
        case C_BYTES:
+               if (!config->cluster_iam_master) {
+                       LOG(0, 0, 0, "I'm not the master, but I got a C_BYTES from %s?\n", fmtaddr(addr, 0));
+                       return -1;
+               }
+
                return cluster_handle_bytes(p, s);
 
        case C_KILL:    // The master asked us to die!? (usually because we're too out of date).
                return cluster_handle_bytes(p, s);
 
        case C_KILL:    // The master asked us to die!? (usually because we're too out of date).