+* Sat May 21 2005 Brendan O'Dea <bod@optusnet.com.au> 2.0.21
+- Cluster changes from Michael, intended to prevent a stray master
+ from trashing a cluster:
+ + Ignore heartbeats from peers claiming to be the master before the
+ timeout on the old master has expired.
+ + A master receiving a stray heartbeat sends a unicast HB back, which
+ should cause the rogue to die due to the tie-breaker code.
+ + Keep probing the master for late heartbeats.
+ + Drop BGP as soon as we become master with peers (TODO: pre-emptively
+ drop routes if the master is late and we are the best candidate).
+ + A slave (presumably a recently restarted master) receiving a
+ LASTSEEN message sends a zero basetime ping (as sent by the master
+ on clean shutdown).
+
* Mon May 16 2005 Brendan O'Dea <bod@optusnet.com.au> 2.0.20
- Add handling of "throttle=N" RADIUS attributes.
- Fix RADIUS indexing (should have 16K entries with 64 sockets).
- Revise BGP config to use "router bgp AS" syntax (requires libcli >= 1.8.2)
* Tue Nov 9 2004 Brendan O'Dea <bod@optusnet.com.au> 2.0.5
-- Handle routing properly in lone-master case
+- Handle routing properly in lone-master case
- Fix intercepts: don't double-snoop throttled customers, ensure
byte/packet counts are only updated once
- Add a callback to allow plugins to fetch values from the running config
// L2TPNS Clustering Stuff
-char const *cvs_id_cluster = "$Id: cluster.c,v 1.26.2.5 2005/05/02 09:03:27 bodea Exp $";
+char const *cvs_id_cluster = "$Id: cluster.c,v 1.26.2.6 2005/05/21 13:05:36 bodea Exp $";
#include <stdio.h>
#include <stdlib.h>
// 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.
- 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)));
if (!num_peers) // lone master
advertise();
+#ifdef BGP
+ else if (bgp_configured)
+ bgp_enable_routing(0);
+#endif /* BGP */
+
// FIXME. We need to fix up the tunnel control message
// queue here! There's a number of other variables we
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! Sending a null PING.\n");
+ // Send a ping to the stray slave saying that we're a master that's
+ // just shutdown. This should force the slave to listen for the real
+ // master.
+ peer_send_message(slave, C_PING, 0, NULL, 0);
+ return 0;
+ }
diff = config->cluster_seq_number - seq; // How many packet do we need to send?
if (diff < 0)
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) {
- 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);
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.
}
+ //
+ // 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
+ && (config->cluster_last_hb + config->cluster_hb_timeout - 11) > TIME) {
+ 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;