// L2TPNS Clustering Stuff
-char const *cvs_id_cluster = "$Id: cluster.c,v 1.37 2005/05/08 08:00:49 bodea Exp $";
+char const *cvs_id_cluster = "$Id: cluster.c,v 1.43 2005/06/27 04:52:54 bodea Exp $";
#include <stdio.h>
#include <stdlib.h>
*/
// Module variables.
-int cluster_sockfd = 0; // The filedescriptor for the cluster communications port.
+extern int cluster_sockfd; // The filedescriptor for the cluster communications port.
in_addr_t my_address = 0; // The network address of my ethernet port.
static int walk_session_number = 0; // The next session to send when doing the slow table walk.
}
// advertise our presence via BGP or gratuitous ARP
-static void advertise(void)
+static void advertise_routes(void)
{
#ifdef BGP
if (bgp_configured)
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)
config->cluster_iam_uptodate = 1;
LOG(0, 0, 0, "Now uptodate with master.\n");
- advertise();
+ advertise_routes();
}
//
LOG(4, 0, 0, "Forwarding packet from %s to master (size %d)\n", fmtaddr(addr, 0), size);
STAT(c_forwarded);
- add_type(&p, C_FORWARD, addr, (char *) &port, sizeof(port));
+ add_type(&p, C_FORWARD, addr, (char *) &port, sizeof(port)); // ick. should be uint16_t
memcpy(p, data, size);
p += size;
if (config->cluster_iam_master) // Only happens on the slaves.
return;
- if (!config->cluster_master_address) // If we don't have a master, skip it for a while.
+ if (!config->cluster_master_address) // If we don't have a master, skip it for a while.
+ return;
+
+ // C_BYTES format changed in 2.1.0 (cluster version 5)
+ // during upgrade from previous versions, hang onto our counters
+ // for a bit until the new master comes up
+ if (config->cluster_last_hb_ver < 5)
return;
- i = MAX_B_RECS * 5; // Examine max 2000 sessions;
+ i = MAX_B_RECS * 5; // Examine max 3000 sessions;
if (config->cluster_highest_sessionid > i)
i = config->cluster_highest_sessionid;
walk_session_number = 1;
if (!sess_local[walk_session_number].cin && !sess_local[walk_session_number].cout)
- continue; // Unused. Skip it.
+ continue; // Unchanged. Skip it.
b[c].sid = walk_session_number;
- 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;
+ b[c].pin = sess_local[walk_session_number].pin;
+ b[c].pout = sess_local[walk_session_number].pout;
+ b[c].cin = sess_local[walk_session_number].cin;
+ b[c].cout = sess_local[walk_session_number].cout;
// Reset counters.
+ sess_local[walk_session_number].pin = sess_local[walk_session_number].pout = 0;
sess_local[walk_session_number].cin = sess_local[walk_session_number].cout = 0;
+
+ if (++c > MAX_B_RECS) // Send a max of 600 elements in a packet.
+ break;
}
if (!c) // Didn't find any that changes. Get out of here!
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
}
-#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();
+ }
}
//
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
- // 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)));
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");
- 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.
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
// 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");
+ if (have_peers < config->cluster_master_min_adv)
+ advertise_routes();
+ else
+ withdraw_routes();
+
if (config->cluster_seq_number == -1)
config->cluster_seq_number = 0;
// 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.
- 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;
+ // Accumulate un-sent byte/packet counters.
+ increment_counter(&session[i].cin, &session[i].cin_wrap, sess_local[i].cin);
+ increment_counter(&session[i].cout, &session[i].cout_wrap, sess_local[i].cout);
+ session[i].cin_delta += sess_local[i].cin;
+ session[i].cout_delta += sess_local[i].cout;
+
+ session[i].pin += sess_local[i].pin;
+ session[i].pout += sess_local[i].pout;
sess_local[i].cin = sess_local[i].cout = 0;
+ sess_local[i].pin = sess_local[i].pout = 0;
sess_local[i].radius = 0; // Reset authentication as the radius blocks aren't 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.
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)
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);
}
// 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();
- return 0;
}
if (i >= num_peers)
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;
+ if (master)
+ {
+ // catchup with new master
+ peer_send_message(master, C_LASTSEEN, config->cluster_seq_number, NULL, 0);
+
+ // delay next election
+ config->cluster_last_hb = TIME;
+ }
+
+ // run election (or reset "probed" if master was set)
+ 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
return -1; /* Abort processing */
}
- session[b->sid].total_cin += b->in;
- session[b->sid].total_cout += b->out;
+ session[b->sid].pin += b->pin;
+ session[b->sid].pout += b->pout;
- session[b->sid].cin += b->in;
- session[b->sid].cout += b->out;
+ increment_counter(&session[b->sid].cin, &session[b->sid].cin_wrap, b->cin);
+ increment_counter(&session[b->sid].cout, &session[b->sid].cout_wrap, b->cout);
- if (b->in)
+ session[b->sid].cin_delta += b->cin;
+ session[b->sid].cout_delta += b->cout;
+
+ if (b->cin)
session[b->sid].last_packet = time_now; // Reset idle timer!
size -= sizeof(*b);
new.nr = old->nr;
new.ns = old->ns;
new.magic = old->magic;
- new.cin = old->cin;
- new.cout = old->cout;
new.pin = old->pin;
new.pout = old->pout;
- new.total_cin = old->total_cin;
- new.total_cout = old->total_cout;
+ new.cin = old->total_cin;
+ new.cout = old->total_cout;
+ new.cin_delta = old->cin;
+ new.cout_delta = old->cout;
new.throttle_in = old->throttle_in;
new.throttle_out = old->throttle_out;
new.filter_in = old->filter_in;
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) {
+ 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!!
+ config->cluster_last_hb_ver = hb_ver; // remember what cluster version the master is using
if (config->cluster_seq_number != h->seq) { // Out of sequence heartbeat!
static int lastseen_seq = 0;
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);
- case C_LASTSEEN: // Catch up a slave (slave missed a packet).
- if (!config->cluster_iam_master) { // huh?
- LOG(0, 0, 0, "I'm not the master, but I got a C_LASTSEEN from %s?\n", fmtaddr(addr, 0));
- return -1;
- }
+ 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 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).