server/ctdb_monitor.c

   1 /*
   2    monitoring links to all other nodes to detect dead nodes
   3
   4
   5    Copyright (C) Ronnie Sahlberg 2007
   6
   7    This program is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3 of the License, or
  10    (at your option) any later version.
  11
  12    This program is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with this program; if not, see <http://www.gnu.org/licenses/>.
  19 */
  20
  21 #include "includes.h"
  22 #include "lib/events/events.h"
  23 #include "system/filesys.h"
  24 #include "system/wait.h"
  25 #include "../include/ctdb_private.h"
  26
  27 /*
  28   see if any nodes are dead
  29  */
  30 static void ctdb_check_for_dead_nodes(struct event_context *ev, struct timed_event *te,
  31                                       struct timeval t, void *private_data)
  32 {
  33         struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
  34         int i;
  35
  36         /* send a keepalive to all other nodes, unless */
  37         for (i=0;i<ctdb->num_nodes;i++) {
  38                 struct ctdb_node *node = ctdb->nodes[i];
  39                 if (node->pnn == ctdb->pnn) {
  40                         continue;
  41                 }
  42
  43                 if (node->flags & NODE_FLAGS_DISCONNECTED) {
  44                         /* it might have come alive again */
  45                         if (node->rx_cnt != 0) {
  46                                 ctdb_node_connected(node);
  47                         }
  48                         continue;
  49                 }
  50
  51
  52                 if (node->rx_cnt == 0) {
  53                         node->dead_count++;
  54                 } else {
  55                         node->dead_count = 0;
  56                 }
  57
  58                 node->rx_cnt = 0;
  59
  60                 if (node->dead_count >= ctdb->tunable.keepalive_limit) {
  61                         DEBUG(0,("dead count reached for node %u\n", node->pnn));
  62                         ctdb_node_dead(node);
  63                         ctdb_send_keepalive(ctdb, node->pnn);
  64                         /* maybe tell the transport layer to kill the
  65                            sockets as well?
  66                         */
  67                         continue;
  68                 }
  69
  70                 if (node->tx_cnt == 0) {
  71                         DEBUG(5,("sending keepalive to %u\n", node->pnn));
  72                         ctdb_send_keepalive(ctdb, node->pnn);
  73                 }
  74
  75                 node->tx_cnt = 0;
  76         }
  77
  78         event_add_timed(ctdb->ev, ctdb->monitor_context,
  79                         timeval_current_ofs(ctdb->tunable.keepalive_interval, 0),
  80                         ctdb_check_for_dead_nodes, ctdb);
  81 }
  82
  83 static void ctdb_check_health(struct event_context *ev, struct timed_event *te,
  84                               struct timeval t, void *private_data);
  85
  86 /*
  87   called when a health monitoring event script finishes
  88  */
  89 static void ctdb_health_callback(struct ctdb_context *ctdb, int status, void *p)
  90 {
  91         struct ctdb_node *node = ctdb->nodes[ctdb->pnn];
  92         TDB_DATA data;
  93         struct ctdb_node_flag_change c;
  94         uint32_t next_interval;
  95
  96         c.pnn = ctdb->pnn;
  97         c.old_flags = node->flags;
  98
  99         if (status != 0 && !(node->flags & NODE_FLAGS_UNHEALTHY)) {
 100                 DEBUG(0,("monitor event failed - disabling node\n"));
 101                 node->flags |= NODE_FLAGS_UNHEALTHY;
 102         } else if (status == 0 && (node->flags & NODE_FLAGS_UNHEALTHY)) {
 103                 DEBUG(0,("monitor event OK - node re-enabled\n"));
 104                 node->flags &= ~NODE_FLAGS_UNHEALTHY;
 105         }
 106
 107         if (node->flags & NODE_FLAGS_UNHEALTHY) {
 108                 next_interval = ctdb->tunable.monitor_retry;
 109         } else {
 110                 next_interval = ctdb->tunable.monitor_interval;
 111         }
 112
 113         event_add_timed(ctdb->ev, ctdb->monitor_context,
 114                         timeval_current_ofs(next_interval, 0),
 115                         ctdb_check_health, ctdb);
 116
 117         if (c.old_flags == node->flags) {
 118                 return;
 119         }
 120
 121         c.new_flags = node->flags;
 122
 123         data.dptr = (uint8_t *)&c;
 124         data.dsize = sizeof(c);
 125
 126         /* tell the other nodes that something has changed */
 127         ctdb_daemon_send_message(ctdb, CTDB_BROADCAST_CONNECTED,
 128                                  CTDB_SRVID_NODE_FLAGS_CHANGED, data);
 129
 130 }
 131
 132
 133 /*
 134   called when the startup event script finishes
 135  */
 136 static void ctdb_startup_callback(struct ctdb_context *ctdb, int status, void *p)
 137 {
 138         if (status != 0) {
 139                 DEBUG(0,("startup event failed\n"));
 140         } else if (status == 0) {
 141                 DEBUG(0,("startup event OK - enabling monitoring\n"));
 142                 ctdb->done_startup = true;
 143         }
 144
 145         if (ctdb->done_startup) {
 146                 event_add_timed(ctdb->ev, ctdb->monitor_context,
 147                                 timeval_zero(),
 148                                 ctdb_check_health, ctdb);
 149         } else {
 150                 event_add_timed(ctdb->ev, ctdb->monitor_context,
 151                                 timeval_current_ofs(ctdb->tunable.monitor_interval, 0),
 152                                 ctdb_check_health, ctdb);
 153         }
 154
 155 }
 156
 157
 158 /*
 159   see if the event scripts think we are healthy
 160  */
 161 static void ctdb_check_health(struct event_context *ev, struct timed_event *te,
 162                               struct timeval t, void *private_data)
 163 {
 164         struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
 165         int ret;
 166
 167         if (ctdb->recovery_mode != CTDB_RECOVERY_NORMAL ||
 168             (ctdb->monitoring_mode == CTDB_MONITORING_DISABLED && ctdb->done_startup)) {
 169                 event_add_timed(ctdb->ev, ctdb->monitor_context,
 170                                 timeval_current_ofs(ctdb->tunable.monitor_interval, 0),
 171                                 ctdb_check_health, ctdb);
 172                 return;
 173         }
 174
 175         if (!ctdb->done_startup) {
 176                 ret = ctdb_event_script_callback(ctdb,
 177                                                  timeval_current_ofs(ctdb->tunable.script_timeout, 0),
 178                                                  ctdb->monitor_context, ctdb_startup_callback,
 179                                                  ctdb, "startup");
 180         } else {
 181                 ret = ctdb_event_script_callback(ctdb,
 182                                                  timeval_current_ofs(ctdb->tunable.script_timeout, 0),
 183                                                  ctdb->monitor_context, ctdb_health_callback,
 184                                                  ctdb, "monitor");
 185         }
 186
 187         if (ret != 0) {
 188                 DEBUG(0,("Unable to launch monitor event script\n"));
 189                 event_add_timed(ctdb->ev, ctdb->monitor_context,
 190                                 timeval_current_ofs(ctdb->tunable.monitor_retry, 0),
 191                                 ctdb_check_health, ctdb);
 192         }
 193 }
 194
 195 /* stop any monitoring */
 196 void ctdb_stop_monitoring(struct ctdb_context *ctdb)
 197 {
 198         talloc_free(ctdb->monitor_context);
 199         ctdb->monitor_context = NULL;
 200
 201         ctdb->monitoring_mode  = CTDB_MONITORING_DISABLED;
 202         DEBUG(0,("Monitoring has been stopped\n"));
 203 }
 204
 205 /*
 206   start watching for nodes that might be dead
 207  */
 208 void ctdb_start_monitoring(struct ctdb_context *ctdb)
 209 {
 210         struct timed_event *te;
 211
 212         if (ctdb->monitoring_mode == CTDB_MONITORING_ACTIVE) {
 213                 return;
 214         }
 215
 216         ctdb_stop_monitoring(ctdb);
 217
 218         ctdb->monitor_context = talloc_new(ctdb);
 219         CTDB_NO_MEMORY_FATAL(ctdb, ctdb->monitor_context);
 220
 221         te = event_add_timed(ctdb->ev, ctdb->monitor_context,
 222                              timeval_current_ofs(ctdb->tunable.keepalive_interval, 0),
 223                              ctdb_check_for_dead_nodes, ctdb);
 224         CTDB_NO_MEMORY_FATAL(ctdb, te);
 225
 226         te = event_add_timed(ctdb->ev, ctdb->monitor_context,
 227                              timeval_current_ofs(ctdb->tunable.monitor_retry, 0),
 228                              ctdb_check_health, ctdb);
 229         CTDB_NO_MEMORY_FATAL(ctdb, te);
 230
 231         ctdb->monitoring_mode  = CTDB_MONITORING_ACTIVE;
 232         DEBUG(0,("Monitoring has been started\n"));
 233 }
 234
 235
 236 /*
 237   modify flags on a node
 238  */
 239 int32_t ctdb_control_modflags(struct ctdb_context *ctdb, TDB_DATA indata)
 240 {
 241         struct ctdb_node_modflags *m = (struct ctdb_node_modflags *)indata.dptr;
 242         TDB_DATA data;
 243         struct ctdb_node_flag_change c;
 244         struct ctdb_node *node = ctdb->nodes[ctdb->pnn];
 245         uint32_t old_flags = node->flags;
 246
 247         node->flags |= m->set;
 248         node->flags &= ~m->clear;
 249
 250         if (node->flags == old_flags) {
 251                 DEBUG(0, ("Control modflags on node %u - Unchanged - flags 0x%x\n", ctdb->pnn, node->flags));
 252                 return 0;
 253         }
 254
 255         DEBUG(0, ("Control modflags on node %u - flags now 0x%x\n", ctdb->pnn, node->flags));
 256
 257         /* if we have been banned, go into recovery mode */
 258         c.pnn = ctdb->pnn;
 259         c.old_flags = old_flags;
 260         c.new_flags = node->flags;
 261
 262         data.dptr = (uint8_t *)&c;
 263         data.dsize = sizeof(c);
 264
 265         /* tell the other nodes that something has changed */
 266         ctdb_daemon_send_message(ctdb, CTDB_BROADCAST_CONNECTED,
 267                                  CTDB_SRVID_NODE_FLAGS_CHANGED, data);
 268
 269         if ((node->flags & NODE_FLAGS_BANNED) && !(old_flags & NODE_FLAGS_BANNED)) {
 270                 /* make sure we are frozen */
 271                 DEBUG(0,("This node has been banned - forcing freeze and recovery\n"));
 272                 /* Reset the generation id to 1 to make us ignore any
 273                    REQ/REPLY CALL/DMASTER someone sends to us.
 274                    We are now banned so we shouldnt service database calls
 275                    anymore.
 276                 */
 277                 ctdb->vnn_map->generation = INVALID_GENERATION;
 278
 279                 ctdb_start_freeze(ctdb);
 280                 ctdb_release_all_ips(ctdb);
 281                 ctdb->recovery_mode = CTDB_RECOVERY_ACTIVE;
 282         }
 283
 284         return 0;
 285 }