server/ctdb_recoverd.c

   1 /*
   2    ctdb recovery daemon
   3
   4    Copyright (C) Ronnie Sahlberg  2007
   5
   6    This program is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3 of the License, or
   9    (at your option) any later version.
  10
  11    This program is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    You should have received a copy of the GNU General Public License
  17    along with this program; if not, see <http://www.gnu.org/licenses/>.
  18 */
  19
  20 #include "includes.h"
  21 #include "system/filesys.h"
  22 #include "system/time.h"
  23 #include "system/network.h"
  24 #include "system/wait.h"
  25 #include "popt.h"
  26 #include "cmdline.h"
  27 #include "../include/ctdb_client.h"
  28 #include "../include/ctdb_private.h"
  29 #include "db_wrap.h"
  30 #include "dlinklist.h"
  31
  32
  33 /* most recent reload all ips request we need to perform during the
  34    next monitoring loop
  35 */
  36 struct reloadips_all_reply *reload_all_ips_request = NULL;
  37
  38 /* list of "ctdb ipreallocate" processes to call back when we have
  39    finished the takeover run.
  40 */
  41 struct ip_reallocate_list {
  42         struct ip_reallocate_list *next;
  43         struct rd_memdump_reply *rd;
  44 };
  45
  46 struct ctdb_banning_state {
  47         uint32_t count;
  48         struct timeval last_reported_time;
  49 };
  50
  51 /*
  52   private state of recovery daemon
  53  */
  54 struct ctdb_recoverd {
  55         struct ctdb_context *ctdb;
  56         uint32_t recmaster;
  57         uint32_t num_active;
  58         uint32_t num_connected;
  59         uint32_t last_culprit_node;
  60         struct ctdb_node_map *nodemap;
  61         struct timeval priority_time;
  62         bool need_takeover_run;
  63         bool need_recovery;
  64         uint32_t node_flags;
  65         struct timed_event *send_election_te;
  66         struct timed_event *election_timeout;
  67         struct vacuum_info *vacuum_info;
  68         TALLOC_CTX *ip_reallocate_ctx;
  69         struct ip_reallocate_list *reallocate_callers;
  70         TALLOC_CTX *ip_check_disable_ctx;
  71         struct ctdb_control_get_ifaces *ifaces;
  72         TALLOC_CTX *deferred_rebalance_ctx;
  73 };
  74
  75 #define CONTROL_TIMEOUT() timeval_current_ofs(ctdb->tunable.recover_timeout, 0)
  76 #define MONITOR_TIMEOUT() timeval_current_ofs(ctdb->tunable.recover_interval, 0)
  77
  78 static void ctdb_restart_recd(struct event_context *ev, struct timed_event *te, struct timeval t, void *private_data);
  79
  80 /*
  81   ban a node for a period of time
  82  */
  83 static void ctdb_ban_node(struct ctdb_recoverd *rec, uint32_t pnn, uint32_t ban_time)
  84 {
  85         int ret;
  86         struct ctdb_context *ctdb = rec->ctdb;
  87         struct ctdb_ban_time bantime;
  88
  89         if (!ctdb_validate_pnn(ctdb, pnn)) {
  90                 DEBUG(DEBUG_ERR,("Bad pnn %u in ctdb_ban_node\n", pnn));
  91                 return;
  92         }
  93
  94         DEBUG(DEBUG_NOTICE,("Banning node %u for %u seconds\n", pnn, ban_time));
  95
  96         bantime.pnn  = pnn;
  97         bantime.time = ban_time;
  98
  99         ret = ctdb_ctrl_set_ban(ctdb, CONTROL_TIMEOUT(), pnn, &bantime);
 100         if (ret != 0) {
 101                 DEBUG(DEBUG_ERR,(__location__ " Failed to ban node %d\n", pnn));
 102                 return;
 103         }
 104
 105 }
 106
 107 enum monitor_result { MONITOR_OK, MONITOR_RECOVERY_NEEDED, MONITOR_ELECTION_NEEDED, MONITOR_FAILED};
 108
 109
 110 /*
 111   remember the trouble maker
 112  */
 113 static void ctdb_set_culprit_count(struct ctdb_recoverd *rec, uint32_t culprit, uint32_t count)
 114 {
 115         struct ctdb_context *ctdb = talloc_get_type(rec->ctdb, struct ctdb_context);
 116         struct ctdb_banning_state *ban_state;
 117
 118         if (culprit > ctdb->num_nodes) {
 119                 DEBUG(DEBUG_ERR,("Trying to set culprit %d but num_nodes is %d\n", culprit, ctdb->num_nodes));
 120                 return;
 121         }
 122
 123         /* If we are banned or stopped, do not set other nodes as culprits */
 124         if (rec->node_flags & NODE_FLAGS_INACTIVE) {
 125                 DEBUG(DEBUG_NOTICE, ("This node is INACTIVE, cannot set culprit node %d\n", culprit));
 126                 return;
 127         }
 128
 129         if (ctdb->nodes[culprit]->ban_state == NULL) {
 130                 ctdb->nodes[culprit]->ban_state = talloc_zero(ctdb->nodes[culprit], struct ctdb_banning_state);
 131                 CTDB_NO_MEMORY_VOID(ctdb, ctdb->nodes[culprit]->ban_state);
 132
 133
 134         }
 135         ban_state = ctdb->nodes[culprit]->ban_state;
 136         if (timeval_elapsed(&ban_state->last_reported_time) > ctdb->tunable.recovery_grace_period) {
 137                 /* this was the first time in a long while this node
 138                    misbehaved so we will forgive any old transgressions.
 139                 */
 140                 ban_state->count = 0;
 141         }
 142
 143         ban_state->count += count;
 144         ban_state->last_reported_time = timeval_current();
 145         rec->last_culprit_node = culprit;
 146 }
 147
 148 /*
 149   remember the trouble maker
 150  */
 151 static void ctdb_set_culprit(struct ctdb_recoverd *rec, uint32_t culprit)
 152 {
 153         ctdb_set_culprit_count(rec, culprit, 1);
 154 }
 155
 156
 157 /* this callback is called for every node that failed to execute the
 158    recovered event
 159 */
 160 static void recovered_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
 161 {
 162         struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
 163
 164         DEBUG(DEBUG_ERR, (__location__ " Node %u failed the recovered event. Setting it as recovery fail culprit\n", node_pnn));
 165
 166         ctdb_set_culprit(rec, node_pnn);
 167 }
 168
 169 /*
 170   run the "recovered" eventscript on all nodes
 171  */
 172 static int run_recovered_eventscript(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, const char *caller)
 173 {
 174         TALLOC_CTX *tmp_ctx;
 175         uint32_t *nodes;
 176         struct ctdb_context *ctdb = rec->ctdb;
 177
 178         tmp_ctx = talloc_new(ctdb);
 179         CTDB_NO_MEMORY(ctdb, tmp_ctx);
 180
 181         nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
 182         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_END_RECOVERY,
 183                                         nodes, 0,
 184                                         CONTROL_TIMEOUT(), false, tdb_null,
 185                                         NULL, recovered_fail_callback,
 186                                         rec) != 0) {
 187                 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event when called from %s\n", caller));
 188
 189                 talloc_free(tmp_ctx);
 190                 return -1;
 191         }
 192
 193         talloc_free(tmp_ctx);
 194         return 0;
 195 }
 196
 197 /* this callback is called for every node that failed to execute the
 198    start recovery event
 199 */
 200 static void startrecovery_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
 201 {
 202         struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
 203
 204         DEBUG(DEBUG_ERR, (__location__ " Node %u failed the startrecovery event. Setting it as recovery fail culprit\n", node_pnn));
 205
 206         ctdb_set_culprit(rec, node_pnn);
 207 }
 208
 209 /*
 210   run the "startrecovery" eventscript on all nodes
 211  */
 212 static int run_startrecovery_eventscript(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap)
 213 {
 214         TALLOC_CTX *tmp_ctx;
 215         uint32_t *nodes;
 216         struct ctdb_context *ctdb = rec->ctdb;
 217
 218         tmp_ctx = talloc_new(ctdb);
 219         CTDB_NO_MEMORY(ctdb, tmp_ctx);
 220
 221         nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
 222         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_START_RECOVERY,
 223                                         nodes, 0,
 224                                         CONTROL_TIMEOUT(), false, tdb_null,
 225                                         NULL,
 226                                         startrecovery_fail_callback,
 227                                         rec) != 0) {
 228                 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event. Recovery failed.\n"));
 229                 talloc_free(tmp_ctx);
 230                 return -1;
 231         }
 232
 233         talloc_free(tmp_ctx);
 234         return 0;
 235 }
 236
 237 static void async_getcap_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
 238 {
 239         if ( (outdata.dsize != sizeof(uint32_t)) || (outdata.dptr == NULL) ) {
 240                 DEBUG(DEBUG_ERR, (__location__ " Invalid length/pointer for getcap callback : %u %p\n",  (unsigned)outdata.dsize, outdata.dptr));
 241                 return;
 242         }
 243         if (node_pnn < ctdb->num_nodes) {
 244                 ctdb->nodes[node_pnn]->capabilities = *((uint32_t *)outdata.dptr);
 245         }
 246
 247         if (node_pnn == ctdb->pnn) {
 248                 ctdb->capabilities = ctdb->nodes[node_pnn]->capabilities;
 249         }
 250 }
 251
 252 /*
 253   update the node capabilities for all connected nodes
 254  */
 255 static int update_capabilities(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
 256 {
 257         uint32_t *nodes;
 258         TALLOC_CTX *tmp_ctx;
 259
 260         tmp_ctx = talloc_new(ctdb);
 261         CTDB_NO_MEMORY(ctdb, tmp_ctx);
 262
 263         nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
 264         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_CAPABILITIES,
 265                                         nodes, 0,
 266                                         CONTROL_TIMEOUT(),
 267                                         false, tdb_null,
 268                                         async_getcap_callback, NULL,
 269                                         NULL) != 0) {
 270                 DEBUG(DEBUG_ERR, (__location__ " Failed to read node capabilities.\n"));
 271                 talloc_free(tmp_ctx);
 272                 return -1;
 273         }
 274
 275         talloc_free(tmp_ctx);
 276         return 0;
 277 }
 278
 279 static void set_recmode_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
 280 {
 281         struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
 282
 283         DEBUG(DEBUG_ERR,("Failed to freeze node %u during recovery. Set it as ban culprit for %d credits\n", node_pnn, rec->nodemap->num));
 284         ctdb_set_culprit_count(rec, node_pnn, rec->nodemap->num);
 285 }
 286
 287 static void transaction_start_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
 288 {
 289         struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
 290
 291         DEBUG(DEBUG_ERR,("Failed to start recovery transaction on node %u. Set it as ban culprit for %d credits\n", node_pnn, rec->nodemap->num));
 292         ctdb_set_culprit_count(rec, node_pnn, rec->nodemap->num);
 293 }
 294
 295 /*
 296   change recovery mode on all nodes
 297  */
 298 static int set_recovery_mode(struct ctdb_context *ctdb, struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, uint32_t rec_mode)
 299 {
 300         TDB_DATA data;
 301         uint32_t *nodes;
 302         TALLOC_CTX *tmp_ctx;
 303
 304         tmp_ctx = talloc_new(ctdb);
 305         CTDB_NO_MEMORY(ctdb, tmp_ctx);
 306
 307         /* freeze all nodes */
 308         nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
 309         if (rec_mode == CTDB_RECOVERY_ACTIVE) {
 310                 int i;
 311
 312                 for (i=1; i<=NUM_DB_PRIORITIES; i++) {
 313                         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_FREEZE,
 314                                                 nodes, i,
 315                                                 CONTROL_TIMEOUT(),
 316                                                 false, tdb_null,
 317                                                 NULL,
 318                                                 set_recmode_fail_callback,
 319                                                 rec) != 0) {
 320                                 DEBUG(DEBUG_ERR, (__location__ " Unable to freeze nodes. Recovery failed.\n"));
 321                                 talloc_free(tmp_ctx);
 322                                 return -1;
 323                         }
 324                 }
 325         }
 326
 327
 328         data.dsize = sizeof(uint32_t);
 329         data.dptr = (unsigned char *)&rec_mode;
 330
 331         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECMODE,
 332                                         nodes, 0,
 333                                         CONTROL_TIMEOUT(),
 334                                         false, data,
 335                                         NULL, NULL,
 336                                         NULL) != 0) {
 337                 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode. Recovery failed.\n"));
 338                 talloc_free(tmp_ctx);
 339                 return -1;
 340         }
 341
 342         talloc_free(tmp_ctx);
 343         return 0;
 344 }
 345
 346 /*
 347   change recovery master on all node
 348  */
 349 static int set_recovery_master(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, uint32_t pnn)
 350 {
 351         TDB_DATA data;
 352         TALLOC_CTX *tmp_ctx;
 353         uint32_t *nodes;
 354
 355         tmp_ctx = talloc_new(ctdb);
 356         CTDB_NO_MEMORY(ctdb, tmp_ctx);
 357
 358         data.dsize = sizeof(uint32_t);
 359         data.dptr = (unsigned char *)&pnn;
 360
 361         nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
 362         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECMASTER,
 363                                         nodes, 0,
 364                                         CONTROL_TIMEOUT(), false, data,
 365                                         NULL, NULL,
 366                                         NULL) != 0) {
 367                 DEBUG(DEBUG_ERR, (__location__ " Unable to set recmaster. Recovery failed.\n"));
 368                 talloc_free(tmp_ctx);
 369                 return -1;
 370         }
 371
 372         talloc_free(tmp_ctx);
 373         return 0;
 374 }
 375
 376 /* update all remote nodes to use the same db priority that we have
 377    this can fail if the remove node has not yet been upgraded to
 378    support this function, so we always return success and never fail
 379    a recovery if this call fails.
 380 */
 381 static int update_db_priority_on_remote_nodes(struct ctdb_context *ctdb,
 382         struct ctdb_node_map *nodemap,
 383         uint32_t pnn, struct ctdb_dbid_map *dbmap, TALLOC_CTX *mem_ctx)
 384 {
 385         int db;
 386         uint32_t *nodes;
 387
 388         nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
 389
 390         /* step through all local databases */
 391         for (db=0; db<dbmap->num;db++) {
 392                 TDB_DATA data;
 393                 struct ctdb_db_priority db_prio;
 394                 int ret;
 395
 396                 db_prio.db_id     = dbmap->dbs[db].dbid;
 397                 ret = ctdb_ctrl_get_db_priority(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, dbmap->dbs[db].dbid, &db_prio.priority);
 398                 if (ret != 0) {
 399                         DEBUG(DEBUG_ERR,(__location__ " Failed to read database priority from local node for db 0x%08x\n", dbmap->dbs[db].dbid));
 400                         continue;
 401                 }
 402
 403                 DEBUG(DEBUG_INFO,("Update DB priority for db 0x%08x to %u\n", dbmap->dbs[db].dbid, db_prio.priority));
 404
 405                 data.dptr  = (uint8_t *)&db_prio;
 406                 data.dsize = sizeof(db_prio);
 407
 408                 if (ctdb_client_async_control(ctdb,
 409                                         CTDB_CONTROL_SET_DB_PRIORITY,
 410                                         nodes, 0,
 411                                         CONTROL_TIMEOUT(), false, data,
 412                                         NULL, NULL,
 413                                         NULL) != 0) {
 414                         DEBUG(DEBUG_ERR,(__location__ " Failed to set DB priority for 0x%08x\n", db_prio.db_id));
 415                 }
 416         }
 417
 418         return 0;
 419 }
 420
 421 /*
 422   ensure all other nodes have attached to any databases that we have
 423  */
 424 static int create_missing_remote_databases(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
 425                                            uint32_t pnn, struct ctdb_dbid_map *dbmap, TALLOC_CTX *mem_ctx)
 426 {
 427         int i, j, db, ret;
 428         struct ctdb_dbid_map *remote_dbmap;
 429
 430         /* verify that all other nodes have all our databases */
 431         for (j=0; j<nodemap->num; j++) {
 432                 /* we dont need to ourself ourselves */
 433                 if (nodemap->nodes[j].pnn == pnn) {
 434                         continue;
 435                 }
 436                 /* dont check nodes that are unavailable */
 437                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
 438                         continue;
 439                 }
 440
 441                 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
 442                                          mem_ctx, &remote_dbmap);
 443                 if (ret != 0) {
 444                         DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node %u\n", pnn));
 445                         return -1;
 446                 }
 447
 448                 /* step through all local databases */
 449                 for (db=0; db<dbmap->num;db++) {
 450                         const char *name;
 451
 452
 453                         for (i=0;i<remote_dbmap->num;i++) {
 454                                 if (dbmap->dbs[db].dbid == remote_dbmap->dbs[i].dbid) {
 455                                         break;
 456                                 }
 457                         }
 458                         /* the remote node already have this database */
 459                         if (i!=remote_dbmap->num) {
 460                                 continue;
 461                         }
 462                         /* ok so we need to create this database */
 463                         ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), pnn, dbmap->dbs[db].dbid,
 464                                             mem_ctx, &name);
 465                         if (ret != 0) {
 466                                 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbname from node %u\n", pnn));
 467                                 return -1;
 468                         }
 469                         ctdb_ctrl_createdb(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
 470                                            mem_ctx, name,
 471                                            dbmap->dbs[db].flags & CTDB_DB_FLAGS_PERSISTENT);
 472                         if (ret != 0) {
 473                                 DEBUG(DEBUG_ERR, (__location__ " Unable to create remote db:%s\n", name));
 474                                 return -1;
 475                         }
 476                 }
 477         }
 478
 479         return 0;
 480 }
 481
 482
 483 /*
 484   ensure we are attached to any databases that anyone else is attached to
 485  */
 486 static int create_missing_local_databases(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
 487                                           uint32_t pnn, struct ctdb_dbid_map **dbmap, TALLOC_CTX *mem_ctx)
 488 {
 489         int i, j, db, ret;
 490         struct ctdb_dbid_map *remote_dbmap;
 491
 492         /* verify that we have all database any other node has */
 493         for (j=0; j<nodemap->num; j++) {
 494                 /* we dont need to ourself ourselves */
 495                 if (nodemap->nodes[j].pnn == pnn) {
 496                         continue;
 497                 }
 498                 /* dont check nodes that are unavailable */
 499                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
 500                         continue;
 501                 }
 502
 503                 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
 504                                          mem_ctx, &remote_dbmap);
 505                 if (ret != 0) {
 506                         DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node %u\n", pnn));
 507                         return -1;
 508                 }
 509
 510                 /* step through all databases on the remote node */
 511                 for (db=0; db<remote_dbmap->num;db++) {
 512                         const char *name;
 513
 514                         for (i=0;i<(*dbmap)->num;i++) {
 515                                 if (remote_dbmap->dbs[db].dbid == (*dbmap)->dbs[i].dbid) {
 516                                         break;
 517                                 }
 518                         }
 519                         /* we already have this db locally */
 520                         if (i!=(*dbmap)->num) {
 521                                 continue;
 522                         }
 523                         /* ok so we need to create this database and
 524                            rebuild dbmap
 525                          */
 526                         ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
 527                                             remote_dbmap->dbs[db].dbid, mem_ctx, &name);
 528                         if (ret != 0) {
 529                                 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbname from node %u\n",
 530                                           nodemap->nodes[j].pnn));
 531                                 return -1;
 532                         }
 533                         ctdb_ctrl_createdb(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, name,
 534                                            remote_dbmap->dbs[db].flags & CTDB_DB_FLAGS_PERSISTENT);
 535                         if (ret != 0) {
 536                                 DEBUG(DEBUG_ERR, (__location__ " Unable to create local db:%s\n", name));
 537                                 return -1;
 538                         }
 539                         ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, dbmap);
 540                         if (ret != 0) {
 541                                 DEBUG(DEBUG_ERR, (__location__ " Unable to reread dbmap on node %u\n", pnn));
 542                                 return -1;
 543                         }
 544                 }
 545         }
 546
 547         return 0;
 548 }
 549
 550
 551 /*
 552   pull the remote database contents from one node into the recdb
 553  */
 554 static int pull_one_remote_database(struct ctdb_context *ctdb, uint32_t srcnode,
 555                                     struct tdb_wrap *recdb, uint32_t dbid)
 556 {
 557         int ret;
 558         TDB_DATA outdata;
 559         struct ctdb_marshall_buffer *reply;
 560         struct ctdb_rec_data *rec;
 561         int i;
 562         TALLOC_CTX *tmp_ctx = talloc_new(recdb);
 563
 564         ret = ctdb_ctrl_pulldb(ctdb, srcnode, dbid, CTDB_LMASTER_ANY, tmp_ctx,
 565                                CONTROL_TIMEOUT(), &outdata);
 566         if (ret != 0) {
 567                 DEBUG(DEBUG_ERR,(__location__ " Unable to copy db from node %u\n", srcnode));
 568                 talloc_free(tmp_ctx);
 569                 return -1;
 570         }
 571
 572         reply = (struct ctdb_marshall_buffer *)outdata.dptr;
 573
 574         if (outdata.dsize < offsetof(struct ctdb_marshall_buffer, data)) {
 575                 DEBUG(DEBUG_ERR,(__location__ " invalid data in pulldb reply\n"));
 576                 talloc_free(tmp_ctx);
 577                 return -1;
 578         }
 579
 580         rec = (struct ctdb_rec_data *)&reply->data[0];
 581
 582         for (i=0;
 583              i<reply->count;
 584              rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec), i++) {
 585                 TDB_DATA key, data;
 586                 struct ctdb_ltdb_header *hdr;
 587                 TDB_DATA existing;
 588
 589                 key.dptr = &rec->data[0];
 590                 key.dsize = rec->keylen;
 591                 data.dptr = &rec->data[key.dsize];
 592                 data.dsize = rec->datalen;
 593
 594                 hdr = (struct ctdb_ltdb_header *)data.dptr;
 595
 596                 if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
 597                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 598                         talloc_free(tmp_ctx);
 599                         return -1;
 600                 }
 601
 602                 /* fetch the existing record, if any */
 603                 existing = tdb_fetch(recdb->tdb, key);
 604
 605                 if (existing.dptr != NULL) {
 606                         struct ctdb_ltdb_header header;
 607                         if (existing.dsize < sizeof(struct ctdb_ltdb_header)) {
 608                                 DEBUG(DEBUG_CRIT,(__location__ " Bad record size %u from node %u\n",
 609                                          (unsigned)existing.dsize, srcnode));
 610                                 free(existing.dptr);
 611                                 talloc_free(tmp_ctx);
 612                                 return -1;
 613                         }
 614                         header = *(struct ctdb_ltdb_header *)existing.dptr;
 615                         free(existing.dptr);
 616                         if (!(header.rsn < hdr->rsn ||
 617                               (header.dmaster != ctdb->recovery_master && header.rsn == hdr->rsn))) {
 618                                 continue;
 619                         }
 620                 }
 621
 622                 if (tdb_store(recdb->tdb, key, data, TDB_REPLACE) != 0) {
 623                         DEBUG(DEBUG_CRIT,(__location__ " Failed to store record\n"));
 624                         talloc_free(tmp_ctx);
 625                         return -1;
 626                 }
 627         }
 628
 629         talloc_free(tmp_ctx);
 630
 631         return 0;
 632 }
 633
 634
 635 struct pull_seqnum_cbdata {
 636         int failed;
 637         uint32_t pnn;
 638         uint64_t seqnum;
 639 };
 640
 641 static void pull_seqnum_cb(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
 642 {
 643         struct pull_seqnum_cbdata *cb_data = talloc_get_type(callback_data, struct pull_seqnum_cbdata);
 644         uint64_t seqnum;
 645
 646         if (cb_data->failed != 0) {
 647                 DEBUG(DEBUG_ERR, ("Got seqnum from node %d but we have already failed the entire operation\n", node_pnn));
 648                 return;
 649         }
 650
 651         if (res != 0) {
 652                 DEBUG(DEBUG_ERR, ("Error when pulling seqnum from node %d\n", node_pnn));
 653                 cb_data->failed = 1;
 654                 return;
 655         }
 656
 657         if (outdata.dsize != sizeof(uint64_t)) {
 658                 DEBUG(DEBUG_ERR, ("Error when reading pull seqnum from node %d, got %d bytes but expected %d\n", node_pnn, (int)outdata.dsize, (int)sizeof(uint64_t)));
 659                 cb_data->failed = -1;
 660                 return;
 661         }
 662
 663         seqnum = *((uint64_t *)outdata.dptr);
 664
 665         if (seqnum > cb_data->seqnum) {
 666                 cb_data->seqnum = seqnum;
 667                 cb_data->pnn = node_pnn;
 668         }
 669 }
 670
 671 static void pull_seqnum_fail_cb(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
 672 {
 673         struct pull_seqnum_cbdata *cb_data = talloc_get_type(callback_data, struct pull_seqnum_cbdata);
 674
 675         DEBUG(DEBUG_ERR, ("Failed to pull db seqnum from node %d\n", node_pnn));
 676         cb_data->failed = 1;
 677 }
 678
 679 static int pull_highest_seqnum_pdb(struct ctdb_context *ctdb,
 680                                 struct ctdb_recoverd *rec,
 681                                 struct ctdb_node_map *nodemap,
 682                                 struct tdb_wrap *recdb, uint32_t dbid)
 683 {
 684         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
 685         uint32_t *nodes;
 686         TDB_DATA data;
 687         uint32_t outdata[2];
 688         struct pull_seqnum_cbdata *cb_data;
 689
 690         DEBUG(DEBUG_NOTICE, ("Scan for highest seqnum pdb for db:0x%08x\n", dbid));
 691
 692         outdata[0] = dbid;
 693         outdata[1] = 0;
 694
 695         data.dsize = sizeof(outdata);
 696         data.dptr  = (uint8_t *)&outdata[0];
 697
 698         cb_data = talloc(tmp_ctx, struct pull_seqnum_cbdata);
 699         if (cb_data == NULL) {
 700                 DEBUG(DEBUG_ERR, ("Failed to allocate pull highest seqnum cb_data structure\n"));
 701                 talloc_free(tmp_ctx);
 702                 return -1;
 703         }
 704
 705         cb_data->failed = 0;
 706         cb_data->pnn    = -1;
 707         cb_data->seqnum = 0;
 708
 709         nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
 710         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_DB_SEQNUM,
 711                                         nodes, 0,
 712                                         CONTROL_TIMEOUT(), false, data,
 713                                         pull_seqnum_cb,
 714                                         pull_seqnum_fail_cb,
 715                                         cb_data) != 0) {
 716                 DEBUG(DEBUG_ERR, (__location__ " Failed to run async GET_DB_SEQNUM\n"));
 717
 718                 talloc_free(tmp_ctx);
 719                 return -1;
 720         }
 721
 722         if (cb_data->failed != 0) {
 723                 DEBUG(DEBUG_NOTICE, ("Failed to pull sequence numbers for DB 0x%08x\n", dbid));
 724                 talloc_free(tmp_ctx);
 725                 return -1;
 726         }
 727
 728         if (cb_data->seqnum == 0 || cb_data->pnn == -1) {
 729                 DEBUG(DEBUG_NOTICE, ("Failed to find a node with highest sequence numbers for DB 0x%08x\n", dbid));
 730                 talloc_free(tmp_ctx);
 731                 return -1;
 732         }
 733
 734         DEBUG(DEBUG_NOTICE, ("Pull persistent db:0x%08x from node %d with highest seqnum:%lld\n", dbid, cb_data->pnn, (long long)cb_data->seqnum));
 735
 736         if (pull_one_remote_database(ctdb, cb_data->pnn, recdb, dbid) != 0) {
 737                 DEBUG(DEBUG_ERR, ("Failed to pull higest seqnum database 0x%08x from node %d\n", dbid, cb_data->pnn));
 738                 talloc_free(tmp_ctx);
 739                 return -1;
 740         }
 741
 742         talloc_free(tmp_ctx);
 743         return 0;
 744 }
 745
 746
 747 /*
 748   pull all the remote database contents into the recdb
 749  */
 750 static int pull_remote_database(struct ctdb_context *ctdb,
 751                                 struct ctdb_recoverd *rec,
 752                                 struct ctdb_node_map *nodemap,
 753                                 struct tdb_wrap *recdb, uint32_t dbid,
 754                                 bool persistent)
 755 {
 756         int j;
 757
 758         if (persistent && ctdb->tunable.recover_pdb_by_seqnum != 0) {
 759                 int ret;
 760                 ret = pull_highest_seqnum_pdb(ctdb, rec, nodemap, recdb, dbid);
 761                 if (ret == 0) {
 762                         return 0;
 763                 }
 764         }
 765
 766         /* pull all records from all other nodes across onto this node
 767            (this merges based on rsn)
 768         */
 769         for (j=0; j<nodemap->num; j++) {
 770                 /* dont merge from nodes that are unavailable */
 771                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
 772                         continue;
 773                 }
 774                 if (pull_one_remote_database(ctdb, nodemap->nodes[j].pnn, recdb, dbid) != 0) {
 775                         DEBUG(DEBUG_ERR,(__location__ " Failed to pull remote database from node %u\n",
 776                                  nodemap->nodes[j].pnn));
 777                         ctdb_set_culprit_count(rec, nodemap->nodes[j].pnn, nodemap->num);
 778                         return -1;
 779                 }
 780         }
 781
 782         return 0;
 783 }
 784
 785
 786 /*
 787   update flags on all active nodes
 788  */
 789 static int update_flags_on_all_nodes(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, uint32_t pnn, uint32_t flags)
 790 {
 791         int ret;
 792
 793         ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), pnn, flags, ~flags);
 794                 if (ret != 0) {
 795                 DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
 796                 return -1;
 797         }
 798
 799         return 0;
 800 }
 801
 802 /*
 803   ensure all nodes have the same vnnmap we do
 804  */
 805 static int update_vnnmap_on_all_nodes(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
 806                                       uint32_t pnn, struct ctdb_vnn_map *vnnmap, TALLOC_CTX *mem_ctx)
 807 {
 808         int j, ret;
 809
 810         /* push the new vnn map out to all the nodes */
 811         for (j=0; j<nodemap->num; j++) {
 812                 /* dont push to nodes that are unavailable */
 813                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
 814                         continue;
 815                 }
 816
 817                 ret = ctdb_ctrl_setvnnmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn, mem_ctx, vnnmap);
 818                 if (ret != 0) {
 819                         DEBUG(DEBUG_ERR, (__location__ " Unable to set vnnmap for node %u\n", pnn));
 820                         return -1;
 821                 }
 822         }
 823
 824         return 0;
 825 }
 826
 827
 828 struct vacuum_info {
 829         struct vacuum_info *next, *prev;
 830         struct ctdb_recoverd *rec;
 831         uint32_t srcnode;
 832         struct ctdb_db_context *ctdb_db;
 833         struct ctdb_marshall_buffer *recs;
 834         struct ctdb_rec_data *r;
 835 };
 836
 837 static void vacuum_fetch_next(struct vacuum_info *v);
 838
 839 /*
 840   called when a vacuum fetch has completed - just free it and do the next one
 841  */
 842 static void vacuum_fetch_callback(struct ctdb_client_call_state *state)
 843 {
 844         struct vacuum_info *v = talloc_get_type(state->async.private_data, struct vacuum_info);
 845         talloc_free(state);
 846         vacuum_fetch_next(v);
 847 }
 848
 849
 850 /*
 851   process the next element from the vacuum list
 852 */
 853 static void vacuum_fetch_next(struct vacuum_info *v)
 854 {
 855         struct ctdb_call call;
 856         struct ctdb_rec_data *r;
 857
 858         while (v->recs->count) {
 859                 struct ctdb_client_call_state *state;
 860                 TDB_DATA data;
 861                 struct ctdb_ltdb_header *hdr;
 862
 863                 ZERO_STRUCT(call);
 864                 call.call_id = CTDB_NULL_FUNC;
 865                 call.flags = CTDB_IMMEDIATE_MIGRATION;
 866                 call.flags |= CTDB_CALL_FLAG_VACUUM_MIGRATION;
 867
 868                 r = v->r;
 869                 v->r = (struct ctdb_rec_data *)(r->length + (uint8_t *)r);
 870                 v->recs->count--;
 871
 872                 call.key.dptr = &r->data[0];
 873                 call.key.dsize = r->keylen;
 874
 875                 /* ensure we don't block this daemon - just skip a record if we can't get
 876                    the chainlock */
 877                 if (tdb_chainlock_nonblock(v->ctdb_db->ltdb->tdb, call.key) != 0) {
 878                         continue;
 879                 }
 880
 881                 data = tdb_fetch(v->ctdb_db->ltdb->tdb, call.key);
 882                 if (data.dptr == NULL) {
 883                         tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
 884                         continue;
 885                 }
 886
 887                 if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
 888                         free(data.dptr);
 889                         tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
 890                         continue;
 891                 }
 892
 893                 hdr = (struct ctdb_ltdb_header *)data.dptr;
 894                 if (hdr->dmaster == v->rec->ctdb->pnn) {
 895                         /* its already local */
 896                         free(data.dptr);
 897                         tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
 898                         continue;
 899                 }
 900
 901                 free(data.dptr);
 902
 903                 state = ctdb_call_send(v->ctdb_db, &call);
 904                 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
 905                 if (state == NULL) {
 906                         DEBUG(DEBUG_ERR,(__location__ " Failed to setup vacuum fetch call\n"));
 907                         talloc_free(v);
 908                         return;
 909                 }
 910                 state->async.fn = vacuum_fetch_callback;
 911                 state->async.private_data = v;
 912                 return;
 913         }
 914
 915         talloc_free(v);
 916 }
 917
 918
 919 /*
 920   destroy a vacuum info structure
 921  */
 922 static int vacuum_info_destructor(struct vacuum_info *v)
 923 {
 924         DLIST_REMOVE(v->rec->vacuum_info, v);
 925         return 0;
 926 }
 927
 928
 929 /*
 930   handler for vacuum fetch
 931 */
 932 static void vacuum_fetch_handler(struct ctdb_context *ctdb, uint64_t srvid,
 933                                  TDB_DATA data, void *private_data)
 934 {
 935         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
 936         struct ctdb_marshall_buffer *recs;
 937         int ret, i;
 938         TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
 939         const char *name;
 940         struct ctdb_dbid_map *dbmap=NULL;
 941         bool persistent = false;
 942         struct ctdb_db_context *ctdb_db;
 943         struct ctdb_rec_data *r;
 944         uint32_t srcnode;
 945         struct vacuum_info *v;
 946
 947         recs = (struct ctdb_marshall_buffer *)data.dptr;
 948         r = (struct ctdb_rec_data *)&recs->data[0];
 949
 950         if (recs->count == 0) {
 951                 talloc_free(tmp_ctx);
 952                 return;
 953         }
 954
 955         srcnode = r->reqid;
 956
 957         for (v=rec->vacuum_info;v;v=v->next) {
 958                 if (srcnode == v->srcnode && recs->db_id == v->ctdb_db->db_id) {
 959                         /* we're already working on records from this node */
 960                         talloc_free(tmp_ctx);
 961                         return;
 962                 }
 963         }
 964
 965         /* work out if the database is persistent */
 966         ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &dbmap);
 967         if (ret != 0) {
 968                 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from local node\n"));
 969                 talloc_free(tmp_ctx);
 970                 return;
 971         }
 972
 973         for (i=0;i<dbmap->num;i++) {
 974                 if (dbmap->dbs[i].dbid == recs->db_id) {
 975                         persistent = dbmap->dbs[i].flags & CTDB_DB_FLAGS_PERSISTENT;
 976                         break;
 977                 }
 978         }
 979         if (i == dbmap->num) {
 980                 DEBUG(DEBUG_ERR, (__location__ " Unable to find db_id 0x%x on local node\n", recs->db_id));
 981                 talloc_free(tmp_ctx);
 982                 return;
 983         }
 984
 985         /* find the name of this database */
 986         if (ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, recs->db_id, tmp_ctx, &name) != 0) {
 987                 DEBUG(DEBUG_ERR,(__location__ " Failed to get name of db 0x%x\n", recs->db_id));
 988                 talloc_free(tmp_ctx);
 989                 return;
 990         }
 991
 992         /* attach to it */
 993         ctdb_db = ctdb_attach(ctdb, CONTROL_TIMEOUT(), name, persistent, 0);
 994         if (ctdb_db == NULL) {
 995                 DEBUG(DEBUG_ERR,(__location__ " Failed to attach to database '%s'\n", name));
 996                 talloc_free(tmp_ctx);
 997                 return;
 998         }
 999
1000         v = talloc_zero(rec, struct vacuum_info);
1001         if (v == NULL) {
1002                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory\n"));
1003                 talloc_free(tmp_ctx);
1004                 return;
1005         }
1006
1007         v->rec = rec;
1008         v->srcnode = srcnode;
1009         v->ctdb_db = ctdb_db;
1010         v->recs = talloc_memdup(v, recs, data.dsize);
1011         if (v->recs == NULL) {
1012                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory\n"));
1013                 talloc_free(v);
1014                 talloc_free(tmp_ctx);
1015                 return;
1016         }
1017         v->r =  (struct ctdb_rec_data *)&v->recs->data[0];
1018
1019         DLIST_ADD(rec->vacuum_info, v);
1020
1021         talloc_set_destructor(v, vacuum_info_destructor);
1022
1023         vacuum_fetch_next(v);
1024         talloc_free(tmp_ctx);
1025 }
1026
1027
1028 /*
1029   called when ctdb_wait_timeout should finish
1030  */
1031 static void ctdb_wait_handler(struct event_context *ev, struct timed_event *te,
1032                               struct timeval yt, void *p)
1033 {
1034         uint32_t *timed_out = (uint32_t *)p;
1035         (*timed_out) = 1;
1036 }
1037
1038 /*
1039   wait for a given number of seconds
1040  */
1041 static void ctdb_wait_timeout(struct ctdb_context *ctdb, double secs)
1042 {
1043         uint32_t timed_out = 0;
1044         time_t usecs = (secs - (time_t)secs) * 1000000;
1045         event_add_timed(ctdb->ev, ctdb, timeval_current_ofs(secs, usecs), ctdb_wait_handler, &timed_out);
1046         while (!timed_out) {
1047                 event_loop_once(ctdb->ev);
1048         }
1049 }
1050
1051 /*
1052   called when an election times out (ends)
1053  */
1054 static void ctdb_election_timeout(struct event_context *ev, struct timed_event *te,
1055                                   struct timeval t, void *p)
1056 {
1057         struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
1058         rec->election_timeout = NULL;
1059         fast_start = false;
1060
1061         DEBUG(DEBUG_WARNING,(__location__ " Election timed out\n"));
1062 }
1063
1064
1065 /*
1066   wait for an election to finish. It finished election_timeout seconds after
1067   the last election packet is received
1068  */
1069 static void ctdb_wait_election(struct ctdb_recoverd *rec)
1070 {
1071         struct ctdb_context *ctdb = rec->ctdb;
1072         while (rec->election_timeout) {
1073                 event_loop_once(ctdb->ev);
1074         }
1075 }
1076
1077 /*
1078   Update our local flags from all remote connected nodes.
1079   This is only run when we are or we belive we are the recovery master
1080  */
1081 static int update_local_flags(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap)
1082 {
1083         int j;
1084         struct ctdb_context *ctdb = rec->ctdb;
1085         TALLOC_CTX *mem_ctx = talloc_new(ctdb);
1086
1087         /* get the nodemap for all active remote nodes and verify
1088            they are the same as for this node
1089          */
1090         for (j=0; j<nodemap->num; j++) {
1091                 struct ctdb_node_map *remote_nodemap=NULL;
1092                 int ret;
1093
1094                 if (nodemap->nodes[j].flags & NODE_FLAGS_DISCONNECTED) {
1095                         continue;
1096                 }
1097                 if (nodemap->nodes[j].pnn == ctdb->pnn) {
1098                         continue;
1099                 }
1100
1101                 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
1102                                            mem_ctx, &remote_nodemap);
1103                 if (ret != 0) {
1104                         DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from remote node %u\n",
1105                                   nodemap->nodes[j].pnn));
1106                         ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
1107                         talloc_free(mem_ctx);
1108                         return MONITOR_FAILED;
1109                 }
1110                 if (nodemap->nodes[j].flags != remote_nodemap->nodes[j].flags) {
1111                         /* We should tell our daemon about this so it
1112                            updates its flags or else we will log the same
1113                            message again in the next iteration of recovery.
1114                            Since we are the recovery master we can just as
1115                            well update the flags on all nodes.
1116                         */
1117                         ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn, remote_nodemap->nodes[j].flags, ~remote_nodemap->nodes[j].flags);
1118                         if (ret != 0) {
1119                                 DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
1120                                 return -1;
1121                         }
1122
1123                         /* Update our local copy of the flags in the recovery
1124                            daemon.
1125                         */
1126                         DEBUG(DEBUG_NOTICE,("Remote node %u had flags 0x%x, local had 0x%x - updating local\n",
1127                                  nodemap->nodes[j].pnn, remote_nodemap->nodes[j].flags,
1128                                  nodemap->nodes[j].flags));
1129                         nodemap->nodes[j].flags = remote_nodemap->nodes[j].flags;
1130                 }
1131                 talloc_free(remote_nodemap);
1132         }
1133         talloc_free(mem_ctx);
1134         return MONITOR_OK;
1135 }
1136
1137
1138 /* Create a new random generation ip.
1139    The generation id can not be the INVALID_GENERATION id
1140 */
1141 static uint32_t new_generation(void)
1142 {
1143         uint32_t generation;
1144
1145         while (1) {
1146                 generation = random();
1147
1148                 if (generation != INVALID_GENERATION) {
1149                         break;
1150                 }
1151         }
1152
1153         return generation;
1154 }
1155
1156
1157 /*
1158   create a temporary working database
1159  */
1160 static struct tdb_wrap *create_recdb(struct ctdb_context *ctdb, TALLOC_CTX *mem_ctx)
1161 {
1162         char *name;
1163         struct tdb_wrap *recdb;
1164         unsigned tdb_flags;
1165
1166         /* open up the temporary recovery database */
1167         name = talloc_asprintf(mem_ctx, "%s/recdb.tdb.%u",
1168                                ctdb->db_directory_state,
1169                                ctdb->pnn);
1170         if (name == NULL) {
1171                 return NULL;
1172         }
1173         unlink(name);
1174
1175         tdb_flags = TDB_NOLOCK;
1176         if (ctdb->valgrinding) {
1177                 tdb_flags |= TDB_NOMMAP;
1178         }
1179         tdb_flags |= TDB_DISALLOW_NESTING;
1180
1181         recdb = tdb_wrap_open(mem_ctx, name, ctdb->tunable.database_hash_size,
1182                               tdb_flags, O_RDWR|O_CREAT|O_EXCL, 0600);
1183         if (recdb == NULL) {
1184                 DEBUG(DEBUG_CRIT,(__location__ " Failed to create temp recovery database '%s'\n", name));
1185         }
1186
1187         talloc_free(name);
1188
1189         return recdb;
1190 }
1191
1192
1193 /*
1194    a traverse function for pulling all relevant records from recdb
1195  */
1196 struct recdb_data {
1197         struct ctdb_context *ctdb;
1198         struct ctdb_marshall_buffer *recdata;
1199         uint32_t len;
1200         uint32_t allocated_len;
1201         bool failed;
1202         bool persistent;
1203 };
1204
1205 static int traverse_recdb(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *p)
1206 {
1207         struct recdb_data *params = (struct recdb_data *)p;
1208         struct ctdb_rec_data *rec;
1209         struct ctdb_ltdb_header *hdr;
1210
1211         /*
1212          * skip empty records - but NOT for persistent databases:
1213          *
1214          * The record-by-record mode of recovery deletes empty records.
1215          * For persistent databases, this can lead to data corruption
1216          * by deleting records that should be there:
1217          *
1218          * - Assume the cluster has been running for a while.
1219          *
1220          * - A record R in a persistent database has been created and
1221          *   deleted a couple of times, the last operation being deletion,
1222          *   leaving an empty record with a high RSN, say 10.
1223          *
1224          * - Now a node N is turned off.
1225          *
1226          * - This leaves the local database copy of D on N with the empty
1227          *   copy of R and RSN 10. On all other nodes, the recovery has deleted
1228          *   the copy of record R.
1229          *
1230          * - Now the record is created again while node N is turned off.
1231          *   This creates R with RSN = 1 on all nodes except for N.
1232          *
1233          * - Now node N is turned on again. The following recovery will chose
1234          *   the older empty copy of R due to RSN 10 > RSN 1.
1235          *
1236          * ==> Hence the record is gone after the recovery.
1237          *
1238          * On databases like Samba's registry, this can damage the higher-level
1239          * data structures built from the various tdb-level records.
1240          */
1241         if (!params->persistent && data.dsize <= sizeof(struct ctdb_ltdb_header)) {
1242                 return 0;
1243         }
1244
1245         /* update the dmaster field to point to us */
1246         hdr = (struct ctdb_ltdb_header *)data.dptr;
1247         if (!params->persistent) {
1248                 hdr->dmaster = params->ctdb->pnn;
1249                 hdr->flags |= CTDB_REC_FLAG_MIGRATED_WITH_DATA;
1250         }
1251
1252         /* add the record to the blob ready to send to the nodes */
1253         rec = ctdb_marshall_record(params->recdata, 0, key, NULL, data);
1254         if (rec == NULL) {
1255                 params->failed = true;
1256                 return -1;
1257         }
1258         if (params->len + rec->length >= params->allocated_len) {
1259                 params->allocated_len = rec->length + params->len + params->ctdb->tunable.pulldb_preallocation_size;
1260                 params->recdata = talloc_realloc_size(NULL, params->recdata, params->allocated_len);
1261         }
1262         if (params->recdata == NULL) {
1263                 DEBUG(DEBUG_CRIT,(__location__ " Failed to expand recdata to %u (%u records)\n",
1264                          rec->length + params->len, params->recdata->count));
1265                 params->failed = true;
1266                 return -1;
1267         }
1268         params->recdata->count++;
1269         memcpy(params->len+(uint8_t *)params->recdata, rec, rec->length);
1270         params->len += rec->length;
1271         talloc_free(rec);
1272
1273         return 0;
1274 }
1275
1276 /*
1277   push the recdb database out to all nodes
1278  */
1279 static int push_recdb_database(struct ctdb_context *ctdb, uint32_t dbid,
1280                                bool persistent,
1281                                struct tdb_wrap *recdb, struct ctdb_node_map *nodemap)
1282 {
1283         struct recdb_data params;
1284         struct ctdb_marshall_buffer *recdata;
1285         TDB_DATA outdata;
1286         TALLOC_CTX *tmp_ctx;
1287         uint32_t *nodes;
1288
1289         tmp_ctx = talloc_new(ctdb);
1290         CTDB_NO_MEMORY(ctdb, tmp_ctx);
1291
1292         recdata = talloc_zero(recdb, struct ctdb_marshall_buffer);
1293         CTDB_NO_MEMORY(ctdb, recdata);
1294
1295         recdata->db_id = dbid;
1296
1297         params.ctdb = ctdb;
1298         params.recdata = recdata;
1299         params.len = offsetof(struct ctdb_marshall_buffer, data);
1300         params.allocated_len = params.len;
1301         params.failed = false;
1302         params.persistent = persistent;
1303
1304         if (tdb_traverse_read(recdb->tdb, traverse_recdb, &params) == -1) {
1305                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1306                 talloc_free(params.recdata);
1307                 talloc_free(tmp_ctx);
1308                 return -1;
1309         }
1310
1311         if (params.failed) {
1312                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1313                 talloc_free(params.recdata);
1314                 talloc_free(tmp_ctx);
1315                 return -1;
1316         }
1317
1318         recdata = params.recdata;
1319
1320         outdata.dptr = (void *)recdata;
1321         outdata.dsize = params.len;
1322
1323         nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
1324         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_PUSH_DB,
1325                                         nodes, 0,
1326                                         CONTROL_TIMEOUT(), false, outdata,
1327                                         NULL, NULL,
1328                                         NULL) != 0) {
1329                 DEBUG(DEBUG_ERR,(__location__ " Failed to push recdb records to nodes for db 0x%x\n", dbid));
1330                 talloc_free(recdata);
1331                 talloc_free(tmp_ctx);
1332                 return -1;
1333         }
1334
1335         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pushed remote database 0x%x of size %u\n",
1336                   dbid, recdata->count));
1337
1338         talloc_free(recdata);
1339         talloc_free(tmp_ctx);
1340
1341         return 0;
1342 }
1343
1344
1345 /*
1346   go through a full recovery on one database
1347  */
1348 static int recover_database(struct ctdb_recoverd *rec,
1349                             TALLOC_CTX *mem_ctx,
1350                             uint32_t dbid,
1351                             bool persistent,
1352                             uint32_t pnn,
1353                             struct ctdb_node_map *nodemap,
1354                             uint32_t transaction_id)
1355 {
1356         struct tdb_wrap *recdb;
1357         int ret;
1358         struct ctdb_context *ctdb = rec->ctdb;
1359         TDB_DATA data;
1360         struct ctdb_control_wipe_database w;
1361         uint32_t *nodes;
1362
1363         recdb = create_recdb(ctdb, mem_ctx);
1364         if (recdb == NULL) {
1365                 return -1;
1366         }
1367
1368         /* pull all remote databases onto the recdb */
1369         ret = pull_remote_database(ctdb, rec, nodemap, recdb, dbid, persistent);
1370         if (ret != 0) {
1371                 DEBUG(DEBUG_ERR, (__location__ " Unable to pull remote database 0x%x\n", dbid));
1372                 return -1;
1373         }
1374
1375         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pulled remote database 0x%x\n", dbid));
1376
1377         /* wipe all the remote databases. This is safe as we are in a transaction */
1378         w.db_id = dbid;
1379         w.transaction_id = transaction_id;
1380
1381         data.dptr = (void *)&w;
1382         data.dsize = sizeof(w);
1383
1384         nodes = list_of_active_nodes(ctdb, nodemap, recdb, true);
1385         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_WIPE_DATABASE,
1386                                         nodes, 0,
1387                                         CONTROL_TIMEOUT(), false, data,
1388                                         NULL, NULL,
1389                                         NULL) != 0) {
1390                 DEBUG(DEBUG_ERR, (__location__ " Unable to wipe database. Recovery failed.\n"));
1391                 talloc_free(recdb);
1392                 return -1;
1393         }
1394
1395         /* push out the correct database. This sets the dmaster and skips
1396            the empty records */
1397         ret = push_recdb_database(ctdb, dbid, persistent, recdb, nodemap);
1398         if (ret != 0) {
1399                 talloc_free(recdb);
1400                 return -1;
1401         }
1402
1403         /* all done with this database */
1404         talloc_free(recdb);
1405
1406         return 0;
1407 }
1408
1409 /*
1410   reload the nodes file
1411 */
1412 static void reload_nodes_file(struct ctdb_context *ctdb)
1413 {
1414         ctdb->nodes = NULL;
1415         ctdb_load_nodes_file(ctdb);
1416 }
1417
1418 static int ctdb_reload_remote_public_ips(struct ctdb_context *ctdb,
1419                                          struct ctdb_recoverd *rec,
1420                                          struct ctdb_node_map *nodemap,
1421                                          uint32_t *culprit)
1422 {
1423         int j;
1424         int ret;
1425
1426         if (ctdb->num_nodes != nodemap->num) {
1427                 DEBUG(DEBUG_ERR, (__location__ " ctdb->num_nodes (%d) != nodemap->num (%d) invalid param\n",
1428                                   ctdb->num_nodes, nodemap->num));
1429                 if (culprit) {
1430                         *culprit = ctdb->pnn;
1431                 }
1432                 return -1;
1433         }
1434
1435         for (j=0; j<nodemap->num; j++) {
1436                 /* For readability */
1437                 struct ctdb_node *node = ctdb->nodes[j];
1438
1439                 /* release any existing data */
1440                 if (node->known_public_ips) {
1441                         talloc_free(node->known_public_ips);
1442                         node->known_public_ips = NULL;
1443                 }
1444                 if (node->available_public_ips) {
1445                         talloc_free(node->available_public_ips);
1446                         node->available_public_ips = NULL;
1447                 }
1448
1449                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
1450                         continue;
1451                 }
1452
1453                 /* Retrieve the list of known public IPs from the node */
1454                 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1455                                         CONTROL_TIMEOUT(),
1456                                         node->pnn,
1457                                         ctdb->nodes,
1458                                         0,
1459                                         &node->known_public_ips);
1460                 if (ret != 0) {
1461                         DEBUG(DEBUG_ERR,
1462                               ("Failed to read known public IPs from node: %u\n",
1463                                node->pnn));
1464                         if (culprit) {
1465                                 *culprit = node->pnn;
1466                         }
1467                         return -1;
1468                 }
1469
1470                 if (ctdb->do_checkpublicip &&
1471                     (rec->ip_check_disable_ctx == NULL) &&
1472                     verify_remote_ip_allocation(ctdb,
1473                                                  node->known_public_ips,
1474                                                  node->pnn)) {
1475                         DEBUG(DEBUG_ERR,("Trigger IP reallocation\n"));
1476                         rec->need_takeover_run = true;
1477                 }
1478
1479                 /* Retrieve the list of available public IPs from the node */
1480                 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1481                                         CONTROL_TIMEOUT(),
1482                                         node->pnn,
1483                                         ctdb->nodes,
1484                                         CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE,
1485                                         &node->available_public_ips);
1486                 if (ret != 0) {
1487                         DEBUG(DEBUG_ERR,
1488                               ("Failed to read available public IPs from node: %u\n",
1489                                node->pnn));
1490                         if (culprit) {
1491                                 *culprit = node->pnn;
1492                         }
1493                         return -1;
1494                 }
1495         }
1496
1497         return 0;
1498 }
1499
1500 /* when we start a recovery, make sure all nodes use the same reclock file
1501    setting
1502 */
1503 static int sync_recovery_lock_file_across_cluster(struct ctdb_recoverd *rec)
1504 {
1505         struct ctdb_context *ctdb = rec->ctdb;
1506         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1507         TDB_DATA data;
1508         uint32_t *nodes;
1509
1510         if (ctdb->recovery_lock_file == NULL) {
1511                 data.dptr  = NULL;
1512                 data.dsize = 0;
1513         } else {
1514                 data.dsize = strlen(ctdb->recovery_lock_file) + 1;
1515                 data.dptr  = (uint8_t *)ctdb->recovery_lock_file;
1516         }
1517
1518         nodes = list_of_active_nodes(ctdb, rec->nodemap, tmp_ctx, true);
1519         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECLOCK_FILE,
1520                                         nodes, 0,
1521                                         CONTROL_TIMEOUT(),
1522                                         false, data,
1523                                         NULL, NULL,
1524                                         rec) != 0) {
1525                 DEBUG(DEBUG_ERR, (__location__ " Failed to sync reclock file settings\n"));
1526                 talloc_free(tmp_ctx);
1527                 return -1;
1528         }
1529
1530         talloc_free(tmp_ctx);
1531         return 0;
1532 }
1533
1534
1535 /*
1536  * this callback is called for every node that failed to execute ctdb_takeover_run()
1537  * and set flag to re-run takeover run.
1538  */
1539 static void takeover_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
1540 {
1541         DEBUG(DEBUG_ERR, ("Node %u failed the takeover run\n", node_pnn));
1542
1543         if (callback_data != NULL) {
1544                 struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
1545
1546                 DEBUG(DEBUG_ERR, ("Setting node %u as recovery fail culprit\n", node_pnn));
1547
1548                 ctdb_set_culprit(rec, node_pnn);
1549                 rec->need_takeover_run = true;
1550         }
1551 }
1552
1553
1554 static void ban_misbehaving_nodes(struct ctdb_recoverd *rec, bool *self_ban)
1555 {
1556         struct ctdb_context *ctdb = rec->ctdb;
1557         int i;
1558         struct ctdb_banning_state *ban_state;
1559
1560         *self_ban = false;
1561         for (i=0; i<ctdb->num_nodes; i++) {
1562                 if (ctdb->nodes[i]->ban_state == NULL) {
1563                         continue;
1564                 }
1565                 ban_state = (struct ctdb_banning_state *)ctdb->nodes[i]->ban_state;
1566                 if (ban_state->count < 2*ctdb->num_nodes) {
1567                         continue;
1568                 }
1569
1570                 DEBUG(DEBUG_NOTICE,("Node %u reached %u banning credits - banning it for %u seconds\n",
1571                         ctdb->nodes[i]->pnn, ban_state->count,
1572                         ctdb->tunable.recovery_ban_period));
1573                 ctdb_ban_node(rec, ctdb->nodes[i]->pnn, ctdb->tunable.recovery_ban_period);
1574                 ban_state->count = 0;
1575
1576                 /* Banning ourself? */
1577                 if (ctdb->nodes[i]->pnn == rec->ctdb->pnn) {
1578                         *self_ban = true;
1579                 }
1580         }
1581 }
1582
1583
1584 /*
1585   we are the recmaster, and recovery is needed - start a recovery run
1586  */
1587 static int do_recovery(struct ctdb_recoverd *rec,
1588                        TALLOC_CTX *mem_ctx, uint32_t pnn,
1589                        struct ctdb_node_map *nodemap, struct ctdb_vnn_map *vnnmap)
1590 {
1591         struct ctdb_context *ctdb = rec->ctdb;
1592         int i, j, ret;
1593         uint32_t generation;
1594         struct ctdb_dbid_map *dbmap;
1595         TDB_DATA data;
1596         uint32_t *nodes;
1597         struct timeval start_time;
1598         uint32_t culprit = (uint32_t)-1;
1599         bool self_ban;
1600
1601         DEBUG(DEBUG_NOTICE, (__location__ " Starting do_recovery\n"));
1602
1603         /* if recovery fails, force it again */
1604         rec->need_recovery = true;
1605
1606         ban_misbehaving_nodes(rec, &self_ban);
1607         if (self_ban) {
1608                 DEBUG(DEBUG_NOTICE, ("This node was banned, aborting recovery\n"));
1609                 return -1;
1610         }
1611
1612         if (ctdb->tunable.verify_recovery_lock != 0) {
1613                 DEBUG(DEBUG_ERR,("Taking out recovery lock from recovery daemon\n"));
1614                 start_time = timeval_current();
1615                 if (!ctdb_recovery_lock(ctdb, true)) {
1616                         DEBUG(DEBUG_ERR,("Unable to get recovery lock - aborting recovery "
1617                                          "and ban ourself for %u seconds\n",
1618                                          ctdb->tunable.recovery_ban_period));
1619                         ctdb_ban_node(rec, pnn, ctdb->tunable.recovery_ban_period);
1620                         return -1;
1621                 }
1622                 ctdb_ctrl_report_recd_lock_latency(ctdb, CONTROL_TIMEOUT(), timeval_elapsed(&start_time));
1623                 DEBUG(DEBUG_NOTICE,("Recovery lock taken successfully by recovery daemon\n"));
1624         }
1625
1626         DEBUG(DEBUG_NOTICE, (__location__ " Recovery initiated due to problem with node %u\n", rec->last_culprit_node));
1627
1628         /* get a list of all databases */
1629         ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, &dbmap);
1630         if (ret != 0) {
1631                 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node :%u\n", pnn));
1632                 return -1;
1633         }
1634
1635         /* we do the db creation before we set the recovery mode, so the freeze happens
1636            on all databases we will be dealing with. */
1637
1638         /* verify that we have all the databases any other node has */
1639         ret = create_missing_local_databases(ctdb, nodemap, pnn, &dbmap, mem_ctx);
1640         if (ret != 0) {
1641                 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing local databases\n"));
1642                 return -1;
1643         }
1644
1645         /* verify that all other nodes have all our databases */
1646         ret = create_missing_remote_databases(ctdb, nodemap, pnn, dbmap, mem_ctx);
1647         if (ret != 0) {
1648                 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing remote databases\n"));
1649                 return -1;
1650         }
1651         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - created remote databases\n"));
1652
1653         /* update the database priority for all remote databases */
1654         ret = update_db_priority_on_remote_nodes(ctdb, nodemap, pnn, dbmap, mem_ctx);
1655         if (ret != 0) {
1656                 DEBUG(DEBUG_ERR, (__location__ " Unable to set db priority on remote nodes\n"));
1657         }
1658         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated db priority for all databases\n"));
1659
1660
1661         /* update all other nodes to use the same setting for reclock files
1662            as the local recovery master.
1663         */
1664         sync_recovery_lock_file_across_cluster(rec);
1665
1666         /* set recovery mode to active on all nodes */
1667         ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
1668         if (ret != 0) {
1669                 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
1670                 return -1;
1671         }
1672
1673         /* execute the "startrecovery" event script on all nodes */
1674         ret = run_startrecovery_eventscript(rec, nodemap);
1675         if (ret!=0) {
1676                 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
1677                 return -1;
1678         }
1679
1680         /*
1681           update all nodes to have the same flags that we have
1682          */
1683         for (i=0;i<nodemap->num;i++) {
1684                 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1685                         continue;
1686                 }
1687
1688                 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1689                 if (ret != 0) {
1690                         DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1691                         return -1;
1692                 }
1693         }
1694
1695         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1696
1697         /* pick a new generation number */
1698         generation = new_generation();
1699
1700         /* change the vnnmap on this node to use the new generation
1701            number but not on any other nodes.
1702            this guarantees that if we abort the recovery prematurely
1703            for some reason (a node stops responding?)
1704            that we can just return immediately and we will reenter
1705            recovery shortly again.
1706            I.e. we deliberately leave the cluster with an inconsistent
1707            generation id to allow us to abort recovery at any stage and
1708            just restart it from scratch.
1709          */
1710         vnnmap->generation = generation;
1711         ret = ctdb_ctrl_setvnnmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, vnnmap);
1712         if (ret != 0) {
1713                 DEBUG(DEBUG_ERR, (__location__ " Unable to set vnnmap for node %u\n", pnn));
1714                 return -1;
1715         }
1716
1717         data.dptr = (void *)&generation;
1718         data.dsize = sizeof(uint32_t);
1719
1720         nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
1721         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_START,
1722                                         nodes, 0,
1723                                         CONTROL_TIMEOUT(), false, data,
1724                                         NULL,
1725                                         transaction_start_fail_callback,
1726                                         rec) != 0) {
1727                 DEBUG(DEBUG_ERR, (__location__ " Unable to start transactions. Recovery failed.\n"));
1728                 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_CANCEL,
1729                                         nodes, 0,
1730                                         CONTROL_TIMEOUT(), false, tdb_null,
1731                                         NULL,
1732                                         NULL,
1733                                         NULL) != 0) {
1734                         DEBUG(DEBUG_ERR,("Failed to cancel recovery transaction\n"));
1735                 }
1736                 return -1;
1737         }
1738
1739         DEBUG(DEBUG_NOTICE,(__location__ " started transactions on all nodes\n"));
1740
1741         for (i=0;i<dbmap->num;i++) {
1742                 ret = recover_database(rec, mem_ctx,
1743                                        dbmap->dbs[i].dbid,
1744                                        dbmap->dbs[i].flags & CTDB_DB_FLAGS_PERSISTENT,
1745                                        pnn, nodemap, generation);
1746                 if (ret != 0) {
1747                         DEBUG(DEBUG_ERR, (__location__ " Failed to recover database 0x%x\n", dbmap->dbs[i].dbid));
1748                         return -1;
1749                 }
1750         }
1751
1752         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - starting database commits\n"));
1753
1754         /* commit all the changes */
1755         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_COMMIT,
1756                                         nodes, 0,
1757                                         CONTROL_TIMEOUT(), false, data,
1758                                         NULL, NULL,
1759                                         NULL) != 0) {
1760                 DEBUG(DEBUG_ERR, (__location__ " Unable to commit recovery changes. Recovery failed.\n"));
1761                 return -1;
1762         }
1763
1764         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - committed databases\n"));
1765
1766
1767         /* update the capabilities for all nodes */
1768         ret = update_capabilities(ctdb, nodemap);
1769         if (ret!=0) {
1770                 DEBUG(DEBUG_ERR, (__location__ " Unable to update node capabilities.\n"));
1771                 return -1;
1772         }
1773
1774         /* build a new vnn map with all the currently active and
1775            unbanned nodes */
1776         generation = new_generation();
1777         vnnmap = talloc(mem_ctx, struct ctdb_vnn_map);
1778         CTDB_NO_MEMORY(ctdb, vnnmap);
1779         vnnmap->generation = generation;
1780         vnnmap->size = 0;
1781         vnnmap->map = talloc_zero_array(vnnmap, uint32_t, vnnmap->size);
1782         CTDB_NO_MEMORY(ctdb, vnnmap->map);
1783         for (i=j=0;i<nodemap->num;i++) {
1784                 if (nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE) {
1785                         continue;
1786                 }
1787                 if (!(ctdb->nodes[i]->capabilities & CTDB_CAP_LMASTER)) {
1788                         /* this node can not be an lmaster */
1789                         DEBUG(DEBUG_DEBUG, ("Node %d cant be a LMASTER, skipping it\n", i));
1790                         continue;
1791                 }
1792
1793                 vnnmap->size++;
1794                 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1795                 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1796                 vnnmap->map[j++] = nodemap->nodes[i].pnn;
1797
1798         }
1799         if (vnnmap->size == 0) {
1800                 DEBUG(DEBUG_NOTICE, ("No suitable lmasters found. Adding local node (recmaster) anyway.\n"));
1801                 vnnmap->size++;
1802                 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1803                 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1804                 vnnmap->map[0] = pnn;
1805         }
1806
1807         /* update to the new vnnmap on all nodes */
1808         ret = update_vnnmap_on_all_nodes(ctdb, nodemap, pnn, vnnmap, mem_ctx);
1809         if (ret != 0) {
1810                 DEBUG(DEBUG_ERR, (__location__ " Unable to update vnnmap on all nodes\n"));
1811                 return -1;
1812         }
1813
1814         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated vnnmap\n"));
1815
1816         /* update recmaster to point to us for all nodes */
1817         ret = set_recovery_master(ctdb, nodemap, pnn);
1818         if (ret!=0) {
1819                 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery master\n"));
1820                 return -1;
1821         }
1822
1823         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated recmaster\n"));
1824
1825         /*
1826           update all nodes to have the same flags that we have
1827          */
1828         for (i=0;i<nodemap->num;i++) {
1829                 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1830                         continue;
1831                 }
1832
1833                 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1834                 if (ret != 0) {
1835                         DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1836                         return -1;
1837                 }
1838         }
1839
1840         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1841
1842         /* disable recovery mode */
1843         ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_NORMAL);
1844         if (ret != 0) {
1845                 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to normal on cluster\n"));
1846                 return -1;
1847         }
1848
1849         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - disabled recovery mode\n"));
1850
1851         /* Fetch known/available public IPs from each active node */
1852         ret = ctdb_reload_remote_public_ips(ctdb, rec, nodemap, &culprit);
1853         if (ret != 0) {
1854                 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
1855                                  culprit));
1856                 rec->need_takeover_run = true;
1857                 return -1;
1858         }
1859         rec->need_takeover_run = false;
1860         ret = ctdb_takeover_run(ctdb, nodemap, takeover_fail_callback, NULL);
1861         if (ret != 0) {
1862                 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses. ctdb_takeover_run() failed.\n"));
1863                 rec->need_takeover_run = true;
1864         }
1865
1866         /* execute the "recovered" event script on all nodes */
1867         ret = run_recovered_eventscript(rec, nodemap, "do_recovery");
1868         if (ret!=0) {
1869                 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Recovery process failed.\n"));
1870                 return -1;
1871         }
1872
1873         DEBUG(DEBUG_NOTICE, (__location__ " Recovery - finished the recovered event\n"));
1874
1875         /* send a message to all clients telling them that the cluster
1876            has been reconfigured */
1877         ctdb_client_send_message(ctdb, CTDB_BROADCAST_CONNECTED, CTDB_SRVID_RECONFIGURE, tdb_null);
1878
1879         DEBUG(DEBUG_NOTICE, (__location__ " Recovery complete\n"));
1880
1881         rec->need_recovery = false;
1882
1883         /* we managed to complete a full recovery, make sure to forgive
1884            any past sins by the nodes that could now participate in the
1885            recovery.
1886         */
1887         DEBUG(DEBUG_ERR,("Resetting ban count to 0 for all nodes\n"));
1888         for (i=0;i<nodemap->num;i++) {
1889                 struct ctdb_banning_state *ban_state;
1890
1891                 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1892                         continue;
1893                 }
1894
1895                 ban_state = (struct ctdb_banning_state *)ctdb->nodes[nodemap->nodes[i].pnn]->ban_state;
1896                 if (ban_state == NULL) {
1897                         continue;
1898                 }
1899
1900                 ban_state->count = 0;
1901         }
1902
1903
1904         /* We just finished a recovery successfully.
1905            We now wait for rerecovery_timeout before we allow
1906            another recovery to take place.
1907         */
1908         DEBUG(DEBUG_NOTICE, ("Just finished a recovery. New recoveries will now be supressed for the rerecovery timeout (%d seconds)\n", ctdb->tunable.rerecovery_timeout));
1909         ctdb_wait_timeout(ctdb, ctdb->tunable.rerecovery_timeout);
1910         DEBUG(DEBUG_NOTICE, ("The rerecovery timeout has elapsed. We now allow recoveries to trigger again.\n"));
1911
1912         return 0;
1913 }
1914
1915
1916 /*
1917   elections are won by first checking the number of connected nodes, then
1918   the priority time, then the pnn
1919  */
1920 struct election_message {
1921         uint32_t num_connected;
1922         struct timeval priority_time;
1923         uint32_t pnn;
1924         uint32_t node_flags;
1925 };
1926
1927 /*
1928   form this nodes election data
1929  */
1930 static void ctdb_election_data(struct ctdb_recoverd *rec, struct election_message *em)
1931 {
1932         int ret, i;
1933         struct ctdb_node_map *nodemap;
1934         struct ctdb_context *ctdb = rec->ctdb;
1935
1936         ZERO_STRUCTP(em);
1937
1938         em->pnn = rec->ctdb->pnn;
1939         em->priority_time = rec->priority_time;
1940
1941         ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, rec, &nodemap);
1942         if (ret != 0) {
1943                 DEBUG(DEBUG_ERR,(__location__ " unable to get election data\n"));
1944                 return;
1945         }
1946
1947         rec->node_flags = nodemap->nodes[ctdb->pnn].flags;
1948         em->node_flags = rec->node_flags;
1949
1950         for (i=0;i<nodemap->num;i++) {
1951                 if (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED)) {
1952                         em->num_connected++;
1953                 }
1954         }
1955
1956         /* we shouldnt try to win this election if we cant be a recmaster */
1957         if ((ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1958                 em->num_connected = 0;
1959                 em->priority_time = timeval_current();
1960         }
1961
1962         talloc_free(nodemap);
1963 }
1964
1965 /*
1966   see if the given election data wins
1967  */
1968 static bool ctdb_election_win(struct ctdb_recoverd *rec, struct election_message *em)
1969 {
1970         struct election_message myem;
1971         int cmp = 0;
1972
1973         ctdb_election_data(rec, &myem);
1974
1975         /* we cant win if we dont have the recmaster capability */
1976         if ((rec->ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1977                 return false;
1978         }
1979
1980         /* we cant win if we are banned */
1981         if (rec->node_flags & NODE_FLAGS_BANNED) {
1982                 return false;
1983         }
1984
1985         /* we cant win if we are stopped */
1986         if (rec->node_flags & NODE_FLAGS_STOPPED) {
1987                 return false;
1988         }
1989
1990         /* we will automatically win if the other node is banned */
1991         if (em->node_flags & NODE_FLAGS_BANNED) {
1992                 return true;
1993         }
1994
1995         /* we will automatically win if the other node is banned */
1996         if (em->node_flags & NODE_FLAGS_STOPPED) {
1997                 return true;
1998         }
1999
2000         /* try to use the most connected node */
2001         if (cmp == 0) {
2002                 cmp = (int)myem.num_connected - (int)em->num_connected;
2003         }
2004
2005         /* then the longest running node */
2006         if (cmp == 0) {
2007                 cmp = timeval_compare(&em->priority_time, &myem.priority_time);
2008         }
2009
2010         if (cmp == 0) {
2011                 cmp = (int)myem.pnn - (int)em->pnn;
2012         }
2013
2014         return cmp > 0;
2015 }
2016
2017 /*
2018   send out an election request
2019  */
2020 static int send_election_request(struct ctdb_recoverd *rec, uint32_t pnn, bool update_recmaster)
2021 {
2022         int ret;
2023         TDB_DATA election_data;
2024         struct election_message emsg;
2025         uint64_t srvid;
2026         struct ctdb_context *ctdb = rec->ctdb;
2027
2028         srvid = CTDB_SRVID_RECOVERY;
2029
2030         ctdb_election_data(rec, &emsg);
2031
2032         election_data.dsize = sizeof(struct election_message);
2033         election_data.dptr  = (unsigned char *)&emsg;
2034
2035
2036         /* send an election message to all active nodes */
2037         DEBUG(DEBUG_INFO,(__location__ " Send election request to all active nodes\n"));
2038         ctdb_client_send_message(ctdb, CTDB_BROADCAST_ALL, srvid, election_data);
2039
2040
2041         /* A new node that is already frozen has entered the cluster.
2042            The existing nodes are not frozen and dont need to be frozen
2043            until the election has ended and we start the actual recovery
2044         */
2045         if (update_recmaster == true) {
2046                 /* first we assume we will win the election and set
2047                    recoverymaster to be ourself on the current node
2048                  */
2049                 ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), pnn, pnn);
2050                 if (ret != 0) {
2051                         DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request\n"));
2052                         return -1;
2053                 }
2054         }
2055
2056
2057         return 0;
2058 }
2059
2060 /*
2061   this function will unban all nodes in the cluster
2062 */
2063 static void unban_all_nodes(struct ctdb_context *ctdb)
2064 {
2065         int ret, i;
2066         struct ctdb_node_map *nodemap;
2067         TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2068
2069         ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
2070         if (ret != 0) {
2071                 DEBUG(DEBUG_ERR,(__location__ " failed to get nodemap to unban all nodes\n"));
2072                 return;
2073         }
2074
2075         for (i=0;i<nodemap->num;i++) {
2076                 if ( (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED))
2077                   && (nodemap->nodes[i].flags & NODE_FLAGS_BANNED) ) {
2078                         ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[i].pnn, 0, NODE_FLAGS_BANNED);
2079                 }
2080         }
2081
2082         talloc_free(tmp_ctx);
2083 }
2084
2085
2086 /*
2087   we think we are winning the election - send a broadcast election request
2088  */
2089 static void election_send_request(struct event_context *ev, struct timed_event *te, struct timeval t, void *p)
2090 {
2091         struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
2092         int ret;
2093
2094         ret = send_election_request(rec, ctdb_get_pnn(rec->ctdb), false);
2095         if (ret != 0) {
2096                 DEBUG(DEBUG_ERR,("Failed to send election request!\n"));
2097         }
2098
2099         talloc_free(rec->send_election_te);
2100         rec->send_election_te = NULL;
2101 }
2102
2103 /*
2104   handler for memory dumps
2105 */
2106 static void mem_dump_handler(struct ctdb_context *ctdb, uint64_t srvid,
2107                              TDB_DATA data, void *private_data)
2108 {
2109         TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2110         TDB_DATA *dump;
2111         int ret;
2112         struct rd_memdump_reply *rd;
2113
2114         if (data.dsize != sizeof(struct rd_memdump_reply)) {
2115                 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
2116                 talloc_free(tmp_ctx);
2117                 return;
2118         }
2119         rd = (struct rd_memdump_reply *)data.dptr;
2120
2121         dump = talloc_zero(tmp_ctx, TDB_DATA);
2122         if (dump == NULL) {
2123                 DEBUG(DEBUG_ERR, (__location__ " Failed to allocate memory for memdump\n"));
2124                 talloc_free(tmp_ctx);
2125                 return;
2126         }
2127         ret = ctdb_dump_memory(ctdb, dump);
2128         if (ret != 0) {
2129                 DEBUG(DEBUG_ERR, (__location__ " ctdb_dump_memory() failed\n"));
2130                 talloc_free(tmp_ctx);
2131                 return;
2132         }
2133
2134 DEBUG(DEBUG_ERR, ("recovery master memory dump\n"));
2135
2136         ret = ctdb_client_send_message(ctdb, rd->pnn, rd->srvid, *dump);
2137         if (ret != 0) {
2138                 DEBUG(DEBUG_ERR,("Failed to send rd memdump reply message\n"));
2139                 talloc_free(tmp_ctx);
2140                 return;
2141         }
2142
2143         talloc_free(tmp_ctx);
2144 }
2145
2146 /*
2147   handler for getlog
2148 */
2149 static void getlog_handler(struct ctdb_context *ctdb, uint64_t srvid,
2150                            TDB_DATA data, void *private_data)
2151 {
2152         struct ctdb_get_log_addr *log_addr;
2153         pid_t child;
2154
2155         if (data.dsize != sizeof(struct ctdb_get_log_addr)) {
2156                 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
2157                 return;
2158         }
2159         log_addr = (struct ctdb_get_log_addr *)data.dptr;
2160
2161         child = ctdb_fork_no_free_ringbuffer(ctdb);
2162         if (child == (pid_t)-1) {
2163                 DEBUG(DEBUG_ERR,("Failed to fork a log collector child\n"));
2164                 return;
2165         }
2166
2167         if (child == 0) {
2168                 ctdb_set_process_name("ctdb_rec_log_collector");
2169                 if (switch_from_server_to_client(ctdb, "recoverd-log-collector") != 0) {
2170                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch log collector child into client mode.\n"));
2171                         _exit(1);
2172                 }
2173                 ctdb_collect_log(ctdb, log_addr);
2174                 _exit(0);
2175         }
2176 }
2177
2178 /*
2179   handler for clearlog
2180 */
2181 static void clearlog_handler(struct ctdb_context *ctdb, uint64_t srvid,
2182                              TDB_DATA data, void *private_data)
2183 {
2184         ctdb_clear_log(ctdb);
2185 }
2186
2187 /*
2188   handler for reload_nodes
2189 */
2190 static void reload_nodes_handler(struct ctdb_context *ctdb, uint64_t srvid,
2191                              TDB_DATA data, void *private_data)
2192 {
2193         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2194
2195         DEBUG(DEBUG_ERR, (__location__ " Reload nodes file from recovery daemon\n"));
2196
2197         reload_nodes_file(rec->ctdb);
2198 }
2199
2200
2201 static void reenable_ip_check(struct event_context *ev, struct timed_event *te,
2202                               struct timeval yt, void *p)
2203 {
2204         struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
2205
2206         talloc_free(rec->ip_check_disable_ctx);
2207         rec->ip_check_disable_ctx = NULL;
2208 }
2209
2210
2211 static void ctdb_rebalance_timeout(struct event_context *ev, struct timed_event *te,
2212                                   struct timeval t, void *p)
2213 {
2214         struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
2215         struct ctdb_context *ctdb = rec->ctdb;
2216         int ret;
2217
2218         DEBUG(DEBUG_NOTICE,("Rebalance all nodes that have had ip assignment changes.\n"));
2219
2220         ret = ctdb_takeover_run(ctdb, rec->nodemap, takeover_fail_callback, NULL);
2221         if (ret != 0) {
2222                 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses. ctdb_takeover_run() failed.\n"));
2223                 rec->need_takeover_run = true;
2224         }
2225
2226         talloc_free(rec->deferred_rebalance_ctx);
2227         rec->deferred_rebalance_ctx = NULL;
2228 }
2229
2230
2231 static void recd_node_rebalance_handler(struct ctdb_context *ctdb, uint64_t srvid,
2232                              TDB_DATA data, void *private_data)
2233 {
2234         uint32_t pnn;
2235         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2236
2237         if (data.dsize != sizeof(uint32_t)) {
2238                 DEBUG(DEBUG_ERR,(__location__ " Incorrect size of node rebalance message. Was %zd but expected %zd bytes\n", data.dsize, sizeof(uint32_t)));
2239                 return;
2240         }
2241
2242         if (ctdb->tunable.deferred_rebalance_on_node_add == 0) {
2243                 return;
2244         }
2245
2246         pnn = *(uint32_t *)&data.dptr[0];
2247
2248         lcp2_forcerebalance(ctdb, pnn);
2249         DEBUG(DEBUG_NOTICE,("Received message to perform node rebalancing for node %d\n", pnn));
2250
2251         if (rec->deferred_rebalance_ctx != NULL) {
2252                 talloc_free(rec->deferred_rebalance_ctx);
2253         }
2254         rec->deferred_rebalance_ctx = talloc_new(rec);
2255         event_add_timed(ctdb->ev, rec->deferred_rebalance_ctx,
2256                         timeval_current_ofs(ctdb->tunable.deferred_rebalance_on_node_add, 0),
2257                         ctdb_rebalance_timeout, rec);
2258 }
2259
2260
2261
2262 static void recd_update_ip_handler(struct ctdb_context *ctdb, uint64_t srvid,
2263                              TDB_DATA data, void *private_data)
2264 {
2265         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2266         struct ctdb_public_ip *ip;
2267
2268         if (rec->recmaster != rec->ctdb->pnn) {
2269                 DEBUG(DEBUG_INFO,("Not recmaster, ignore update ip message\n"));
2270                 return;
2271         }
2272
2273         if (data.dsize != sizeof(struct ctdb_public_ip)) {
2274                 DEBUG(DEBUG_ERR,(__location__ " Incorrect size of recd update ip message. Was %zd but expected %zd bytes\n", data.dsize, sizeof(struct ctdb_public_ip)));
2275                 return;
2276         }
2277
2278         ip = (struct ctdb_public_ip *)data.dptr;
2279
2280         update_ip_assignment_tree(rec->ctdb, ip);
2281 }
2282
2283
2284 static void disable_ip_check_handler(struct ctdb_context *ctdb, uint64_t srvid,
2285                              TDB_DATA data, void *private_data)
2286 {
2287         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2288         uint32_t timeout;
2289
2290         if (rec->ip_check_disable_ctx != NULL) {
2291                 talloc_free(rec->ip_check_disable_ctx);
2292                 rec->ip_check_disable_ctx = NULL;
2293         }
2294
2295         if (data.dsize != sizeof(uint32_t)) {
2296                 DEBUG(DEBUG_ERR,(__location__ " Wrong size for data :%lu "
2297                                  "expexting %lu\n", (long unsigned)data.dsize,
2298                                  (long unsigned)sizeof(uint32_t)));
2299                 return;
2300         }
2301         if (data.dptr == NULL) {
2302                 DEBUG(DEBUG_ERR,(__location__ " No data recaived\n"));
2303                 return;
2304         }
2305
2306         timeout = *((uint32_t *)data.dptr);
2307
2308         if (timeout == 0) {
2309                 DEBUG(DEBUG_NOTICE,("Reenabling ip check\n"));
2310                 return;
2311         }
2312
2313         DEBUG(DEBUG_NOTICE,("Disabling ip check for %u seconds\n", timeout));
2314
2315         rec->ip_check_disable_ctx = talloc_new(rec);
2316         CTDB_NO_MEMORY_VOID(ctdb, rec->ip_check_disable_ctx);
2317
2318         event_add_timed(ctdb->ev, rec->ip_check_disable_ctx, timeval_current_ofs(timeout, 0), reenable_ip_check, rec);
2319 }
2320
2321
2322 /*
2323   handler for reload all ips.
2324 */
2325 static void ip_reloadall_handler(struct ctdb_context *ctdb, uint64_t srvid,
2326                              TDB_DATA data, void *private_data)
2327 {
2328         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2329
2330         if (data.dsize != sizeof(struct reloadips_all_reply)) {
2331                 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
2332                 return;
2333         }
2334
2335         reload_all_ips_request = (struct reloadips_all_reply *)talloc_steal(rec, data.dptr);
2336
2337         DEBUG(DEBUG_NOTICE,("RELOAD_ALL_IPS message received from node:%d srvid:%d\n", reload_all_ips_request->pnn, (int)reload_all_ips_request->srvid));
2338         return;
2339 }
2340
2341 static void async_reloadips_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
2342 {
2343         uint32_t *status = callback_data;
2344
2345         if (res != 0) {
2346                 DEBUG(DEBUG_ERR,("Reload ips all failed on node %d\n", node_pnn));
2347                 *status = 1;
2348         }
2349 }
2350
2351 static int
2352 reload_all_ips(struct ctdb_context *ctdb, struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, struct reloadips_all_reply *rips)
2353 {
2354         TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2355         uint32_t *nodes;
2356         uint32_t status;
2357         int i;
2358
2359         DEBUG(DEBUG_ERR,("RELOAD ALL IPS on all active nodes\n"));
2360         for (i = 0; i< nodemap->num; i++) {
2361                 if (nodemap->nodes[i].flags != 0) {
2362                         DEBUG(DEBUG_ERR, ("Can not reload ips on all nodes. Node %d is not up and healthy\n", i));
2363                         talloc_free(tmp_ctx);
2364                         return -1;
2365                 }
2366         }
2367
2368         /* send the flags update to all connected nodes */
2369         nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
2370         status = 0;
2371         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_RELOAD_PUBLIC_IPS,
2372                                         nodes, 0,
2373                                         CONTROL_TIMEOUT(),
2374                                         false, tdb_null,
2375                                         async_reloadips_callback, NULL,
2376                                         &status) != 0) {
2377                 DEBUG(DEBUG_ERR, (__location__ " Failed to reloadips on all nodes.\n"));
2378                 talloc_free(tmp_ctx);
2379                 return -1;
2380         }
2381
2382         if (status != 0) {
2383                 DEBUG(DEBUG_ERR, (__location__ " Failed to reloadips on all nodes.\n"));
2384                 talloc_free(tmp_ctx);
2385                 return -1;
2386         }
2387
2388         ctdb_client_send_message(ctdb, rips->pnn, rips->srvid, tdb_null);
2389
2390         talloc_free(tmp_ctx);
2391         return 0;
2392 }
2393
2394
2395 /*
2396   handler for ip reallocate, just add it to the list of callers and
2397   handle this later in the monitor_cluster loop so we do not recurse
2398   with other callers to takeover_run()
2399 */
2400 static void ip_reallocate_handler(struct ctdb_context *ctdb, uint64_t srvid,
2401                              TDB_DATA data, void *private_data)
2402 {
2403         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2404         struct ip_reallocate_list *caller;
2405
2406         if (data.dsize != sizeof(struct rd_memdump_reply)) {
2407                 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
2408                 return;
2409         }
2410
2411         if (rec->ip_reallocate_ctx == NULL) {
2412                 rec->ip_reallocate_ctx = talloc_new(rec);
2413                 CTDB_NO_MEMORY_FATAL(ctdb, rec->ip_reallocate_ctx);
2414         }
2415
2416         caller = talloc(rec->ip_reallocate_ctx, struct ip_reallocate_list);
2417         CTDB_NO_MEMORY_FATAL(ctdb, caller);
2418
2419         caller->rd   = (struct rd_memdump_reply *)talloc_steal(caller, data.dptr);
2420         caller->next = rec->reallocate_callers;
2421         rec->reallocate_callers = caller;
2422
2423         return;
2424 }
2425
2426 static void process_ipreallocate_requests(struct ctdb_context *ctdb, struct ctdb_recoverd *rec)
2427 {
2428         TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2429         TDB_DATA result;
2430         int32_t ret;
2431         struct ip_reallocate_list *callers;
2432         uint32_t culprit;
2433
2434         DEBUG(DEBUG_INFO, ("recovery master forced ip reallocation\n"));
2435
2436         /* update the list of public ips that a node can handle for
2437            all connected nodes
2438         */
2439         ret = ctdb_reload_remote_public_ips(ctdb, rec, rec->nodemap, &culprit);
2440         if (ret != 0) {
2441                 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
2442                                  culprit));
2443                 rec->need_takeover_run = true;
2444         }
2445         if (ret == 0) {
2446                 ret = ctdb_takeover_run(ctdb, rec->nodemap, takeover_fail_callback, NULL);
2447                 if (ret != 0) {
2448                         DEBUG(DEBUG_ERR,("Failed to reallocate addresses: ctdb_takeover_run() failed.\n"));
2449                         rec->need_takeover_run = true;
2450                 }
2451         }
2452
2453         result.dsize = sizeof(int32_t);
2454         result.dptr  = (uint8_t *)&ret;
2455
2456         for (callers=rec->reallocate_callers; callers; callers=callers->next) {
2457
2458                 /* Someone that sent srvid==0 does not want a reply */
2459                 if (callers->rd->srvid == 0) {
2460                         continue;
2461                 }
2462                 DEBUG(DEBUG_INFO,("Sending ip reallocate reply message to "
2463                                   "%u:%llu\n", (unsigned)callers->rd->pnn,
2464                                   (unsigned long long)callers->rd->srvid));
2465                 ret = ctdb_client_send_message(ctdb, callers->rd->pnn, callers->rd->srvid, result);
2466                 if (ret != 0) {
2467                         DEBUG(DEBUG_ERR,("Failed to send ip reallocate reply "
2468                                          "message to %u:%llu\n",
2469                                          (unsigned)callers->rd->pnn,
2470                                          (unsigned long long)callers->rd->srvid));
2471                 }
2472         }
2473
2474         talloc_free(tmp_ctx);
2475         talloc_free(rec->ip_reallocate_ctx);
2476         rec->ip_reallocate_ctx = NULL;
2477         rec->reallocate_callers = NULL;
2478
2479 }
2480
2481
2482 /*
2483   handler for recovery master elections
2484 */
2485 static void election_handler(struct ctdb_context *ctdb, uint64_t srvid,
2486                              TDB_DATA data, void *private_data)
2487 {
2488         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2489         int ret;
2490         struct election_message *em = (struct election_message *)data.dptr;
2491         TALLOC_CTX *mem_ctx;
2492
2493         /* we got an election packet - update the timeout for the election */
2494         talloc_free(rec->election_timeout);
2495         rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2496                                                 fast_start ?
2497                                                 timeval_current_ofs(0, 500000) :
2498                                                 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2499                                                 ctdb_election_timeout, rec);
2500
2501         mem_ctx = talloc_new(ctdb);
2502
2503         /* someone called an election. check their election data
2504            and if we disagree and we would rather be the elected node,
2505            send a new election message to all other nodes
2506          */
2507         if (ctdb_election_win(rec, em)) {
2508                 if (!rec->send_election_te) {
2509                         rec->send_election_te = event_add_timed(ctdb->ev, rec,
2510                                                                 timeval_current_ofs(0, 500000),
2511                                                                 election_send_request, rec);
2512                 }
2513                 talloc_free(mem_ctx);
2514                 /*unban_all_nodes(ctdb);*/
2515                 return;
2516         }
2517
2518         /* we didn't win */
2519         talloc_free(rec->send_election_te);
2520         rec->send_election_te = NULL;
2521
2522         if (ctdb->tunable.verify_recovery_lock != 0) {
2523                 /* release the recmaster lock */
2524                 if (em->pnn != ctdb->pnn &&
2525                     ctdb->recovery_lock_fd != -1) {
2526                         close(ctdb->recovery_lock_fd);
2527                         ctdb->recovery_lock_fd = -1;
2528                         unban_all_nodes(ctdb);
2529                 }
2530         }
2531
2532         /* ok, let that guy become recmaster then */
2533         ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), ctdb_get_pnn(ctdb), em->pnn);
2534         if (ret != 0) {
2535                 DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request"));
2536                 talloc_free(mem_ctx);
2537                 return;
2538         }
2539
2540         talloc_free(mem_ctx);
2541         return;
2542 }
2543
2544
2545 /*
2546   force the start of the election process
2547  */
2548 static void force_election(struct ctdb_recoverd *rec, uint32_t pnn,
2549                            struct ctdb_node_map *nodemap)
2550 {
2551         int ret;
2552         struct ctdb_context *ctdb = rec->ctdb;
2553
2554         DEBUG(DEBUG_INFO,(__location__ " Force an election\n"));
2555
2556         /* set all nodes to recovery mode to stop all internode traffic */
2557         ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
2558         if (ret != 0) {
2559                 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
2560                 return;
2561         }
2562
2563         talloc_free(rec->election_timeout);
2564         rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2565                                                 fast_start ?
2566                                                 timeval_current_ofs(0, 500000) :
2567                                                 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2568                                                 ctdb_election_timeout, rec);
2569
2570         ret = send_election_request(rec, pnn, true);
2571         if (ret!=0) {
2572                 DEBUG(DEBUG_ERR, (__location__ " failed to initiate recmaster election"));
2573                 return;
2574         }
2575
2576         /* wait for a few seconds to collect all responses */
2577         ctdb_wait_election(rec);
2578 }
2579
2580
2581
2582 /*
2583   handler for when a node changes its flags
2584 */
2585 static void monitor_handler(struct ctdb_context *ctdb, uint64_t srvid,
2586                             TDB_DATA data, void *private_data)
2587 {
2588         int ret;
2589         struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
2590         struct ctdb_node_map *nodemap=NULL;
2591         TALLOC_CTX *tmp_ctx;
2592         int i;
2593         struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2594         int disabled_flag_changed;
2595
2596         if (data.dsize != sizeof(*c)) {
2597                 DEBUG(DEBUG_ERR,(__location__ "Invalid data in ctdb_node_flag_change\n"));
2598                 return;
2599         }
2600
2601         tmp_ctx = talloc_new(ctdb);
2602         CTDB_NO_MEMORY_VOID(ctdb, tmp_ctx);
2603
2604         ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
2605         if (ret != 0) {
2606                 DEBUG(DEBUG_ERR,(__location__ "ctdb_ctrl_getnodemap failed in monitor_handler\n"));
2607                 talloc_free(tmp_ctx);
2608                 return;
2609         }
2610
2611
2612         for (i=0;i<nodemap->num;i++) {
2613                 if (nodemap->nodes[i].pnn == c->pnn) break;
2614         }
2615
2616         if (i == nodemap->num) {
2617                 DEBUG(DEBUG_CRIT,(__location__ "Flag change for non-existant node %u\n", c->pnn));
2618                 talloc_free(tmp_ctx);
2619                 return;
2620         }
2621
2622         if (nodemap->nodes[i].flags != c->new_flags) {
2623                 DEBUG(DEBUG_NOTICE,("Node %u has changed flags - now 0x%x  was 0x%x\n", c->pnn, c->new_flags, nodemap->nodes[i].flags));
2624         }
2625
2626         disabled_flag_changed =  (nodemap->nodes[i].flags ^ c->new_flags) & NODE_FLAGS_DISABLED;
2627
2628         nodemap->nodes[i].flags = c->new_flags;
2629
2630         ret = ctdb_ctrl_getrecmaster(ctdb, tmp_ctx, CONTROL_TIMEOUT(),
2631                                      CTDB_CURRENT_NODE, &ctdb->recovery_master);
2632
2633         if (ret == 0) {
2634                 ret = ctdb_ctrl_getrecmode(ctdb, tmp_ctx, CONTROL_TIMEOUT(),
2635                                            CTDB_CURRENT_NODE, &ctdb->recovery_mode);
2636         }
2637
2638         if (ret == 0 &&
2639             ctdb->recovery_master == ctdb->pnn &&
2640             ctdb->recovery_mode == CTDB_RECOVERY_NORMAL) {
2641                 /* Only do the takeover run if the perm disabled or unhealthy
2642                    flags changed since these will cause an ip failover but not
2643                    a recovery.
2644                    If the node became disconnected or banned this will also
2645                    lead to an ip address failover but that is handled
2646                    during recovery
2647                 */
2648                 if (disabled_flag_changed) {
2649                         rec->need_takeover_run = true;
2650                 }
2651         }
2652
2653         talloc_free(tmp_ctx);
2654 }
2655
2656 /*
2657   handler for when we need to push out flag changes ot all other nodes
2658 */
2659 static void push_flags_handler(struct ctdb_context *ctdb, uint64_t srvid,
2660                             TDB_DATA data, void *private_data)
2661 {
2662         int ret;
2663         struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
2664         struct ctdb_node_map *nodemap=NULL;
2665         TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2666         uint32_t recmaster;
2667         uint32_t *nodes;
2668
2669         /* find the recovery master */
2670         ret = ctdb_ctrl_getrecmaster(ctdb, tmp_ctx, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &recmaster);
2671         if (ret != 0) {
2672                 DEBUG(DEBUG_ERR, (__location__ " Unable to get recmaster from local node\n"));
2673                 talloc_free(tmp_ctx);
2674                 return;
2675         }
2676
2677         /* read the node flags from the recmaster */
2678         ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), recmaster, tmp_ctx, &nodemap);
2679         if (ret != 0) {
2680                 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from node %u\n", c->pnn));
2681                 talloc_free(tmp_ctx);
2682                 return;
2683         }
2684         if (c->pnn >= nodemap->num) {
2685                 DEBUG(DEBUG_ERR,(__location__ " Nodemap from recmaster does not contain node %d\n", c->pnn));
2686                 talloc_free(tmp_ctx);
2687                 return;
2688         }
2689
2690         /* send the flags update to all connected nodes */
2691         nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
2692
2693         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_MODIFY_FLAGS,
2694                                       nodes, 0, CONTROL_TIMEOUT(),
2695                                       false, data,
2696                                       NULL, NULL,
2697                                       NULL) != 0) {
2698                 DEBUG(DEBUG_ERR, (__location__ " ctdb_control to modify node flags failed\n"));
2699
2700                 talloc_free(tmp_ctx);
2701                 return;
2702         }
2703
2704         talloc_free(tmp_ctx);
2705 }
2706
2707
2708 struct verify_recmode_normal_data {
2709         uint32_t count;
2710         enum monitor_result status;
2711 };
2712
2713 static void verify_recmode_normal_callback(struct ctdb_client_control_state *state)
2714 {
2715         struct verify_recmode_normal_data *rmdata = talloc_get_type(state->async.private_data, struct verify_recmode_normal_data);
2716
2717
2718         /* one more node has responded with recmode data*/
2719         rmdata->count--;
2720
2721         /* if we failed to get the recmode, then return an error and let
2722            the main loop try again.
2723         */
2724         if (state->state != CTDB_CONTROL_DONE) {
2725                 if (rmdata->status == MONITOR_OK) {
2726                         rmdata->status = MONITOR_FAILED;
2727                 }
2728                 return;
2729         }
2730
2731         /* if we got a response, then the recmode will be stored in the
2732            status field
2733         */
2734         if (state->status != CTDB_RECOVERY_NORMAL) {
2735                 DEBUG(DEBUG_NOTICE, ("Node:%u was in recovery mode. Start recovery process\n", state->c->hdr.destnode));
2736                 rmdata->status = MONITOR_RECOVERY_NEEDED;
2737         }
2738
2739         return;
2740 }
2741
2742
2743 /* verify that all nodes are in normal recovery mode */
2744 static enum monitor_result verify_recmode(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
2745 {
2746         struct verify_recmode_normal_data *rmdata;
2747         TALLOC_CTX *mem_ctx = talloc_new(ctdb);
2748         struct ctdb_client_control_state *state;
2749         enum monitor_result status;
2750         int j;
2751
2752         rmdata = talloc(mem_ctx, struct verify_recmode_normal_data);
2753         CTDB_NO_MEMORY_FATAL(ctdb, rmdata);
2754         rmdata->count  = 0;
2755         rmdata->status = MONITOR_OK;
2756
2757         /* loop over all active nodes and send an async getrecmode call to
2758            them*/
2759         for (j=0; j<nodemap->num; j++) {
2760                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
2761                         continue;
2762                 }
2763                 state = ctdb_ctrl_getrecmode_send(ctdb, mem_ctx,
2764                                         CONTROL_TIMEOUT(),
2765                                         nodemap->nodes[j].pnn);
2766                 if (state == NULL) {
2767                         /* we failed to send the control, treat this as
2768                            an error and try again next iteration
2769                         */
2770                         DEBUG(DEBUG_ERR,("Failed to call ctdb_ctrl_getrecmode_send during monitoring\n"));
2771                         talloc_free(mem_ctx);
2772                         return MONITOR_FAILED;
2773                 }
2774
2775                 /* set up the callback functions */
2776                 state->async.fn = verify_recmode_normal_callback;
2777                 state->async.private_data = rmdata;
2778
2779                 /* one more control to wait for to complete */
2780                 rmdata->count++;
2781         }
2782
2783
2784         /* now wait for up to the maximum number of seconds allowed
2785            or until all nodes we expect a response from has replied
2786         */
2787         while (rmdata->count > 0) {
2788                 event_loop_once(ctdb->ev);
2789         }
2790
2791         status = rmdata->status;
2792         talloc_free(mem_ctx);
2793         return status;
2794 }
2795
2796
2797 struct verify_recmaster_data {
2798         struct ctdb_recoverd *rec;
2799         uint32_t count;
2800         uint32_t pnn;
2801         enum monitor_result status;
2802 };
2803
2804 static void verify_recmaster_callback(struct ctdb_client_control_state *state)
2805 {
2806         struct verify_recmaster_data *rmdata = talloc_get_type(state->async.private_data, struct verify_recmaster_data);
2807
2808
2809         /* one more node has responded with recmaster data*/
2810         rmdata->count--;
2811
2812         /* if we failed to get the recmaster, then return an error and let
2813            the main loop try again.
2814         */
2815         if (state->state != CTDB_CONTROL_DONE) {
2816                 if (rmdata->status == MONITOR_OK) {
2817                         rmdata->status = MONITOR_FAILED;
2818                 }
2819                 return;
2820         }
2821
2822         /* if we got a response, then the recmaster will be stored in the
2823            status field
2824         */
2825         if (state->status != rmdata->pnn) {
2826                 DEBUG(DEBUG_ERR,("Node %d does not agree we are the recmaster. Need a new recmaster election\n", state->c->hdr.destnode));
2827                 ctdb_set_culprit(rmdata->rec, state->c->hdr.destnode);
2828                 rmdata->status = MONITOR_ELECTION_NEEDED;
2829         }
2830
2831         return;
2832 }
2833
2834
2835 /* verify that all nodes agree that we are the recmaster */
2836 static enum monitor_result verify_recmaster(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, uint32_t pnn)
2837 {
2838         struct ctdb_context *ctdb = rec->ctdb;
2839         struct verify_recmaster_data *rmdata;
2840         TALLOC_CTX *mem_ctx = talloc_new(ctdb);
2841         struct ctdb_client_control_state *state;
2842         enum monitor_result status;
2843         int j;
2844
2845         rmdata = talloc(mem_ctx, struct verify_recmaster_data);
2846         CTDB_NO_MEMORY_FATAL(ctdb, rmdata);
2847         rmdata->rec    = rec;
2848         rmdata->count  = 0;
2849         rmdata->pnn    = pnn;
2850         rmdata->status = MONITOR_OK;
2851
2852         /* loop over all active nodes and send an async getrecmaster call to
2853            them*/
2854         for (j=0; j<nodemap->num; j++) {
2855                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
2856                         continue;
2857                 }
2858                 state = ctdb_ctrl_getrecmaster_send(ctdb, mem_ctx,
2859                                         CONTROL_TIMEOUT(),
2860                                         nodemap->nodes[j].pnn);
2861                 if (state == NULL) {
2862                         /* we failed to send the control, treat this as
2863                            an error and try again next iteration
2864                         */
2865                         DEBUG(DEBUG_ERR,("Failed to call ctdb_ctrl_getrecmaster_send during monitoring\n"));
2866                         talloc_free(mem_ctx);
2867                         return MONITOR_FAILED;
2868                 }
2869
2870                 /* set up the callback functions */
2871                 state->async.fn = verify_recmaster_callback;
2872                 state->async.private_data = rmdata;
2873
2874                 /* one more control to wait for to complete */
2875                 rmdata->count++;
2876         }
2877
2878
2879         /* now wait for up to the maximum number of seconds allowed
2880            or until all nodes we expect a response from has replied
2881         */
2882         while (rmdata->count > 0) {
2883                 event_loop_once(ctdb->ev);
2884         }
2885
2886         status = rmdata->status;
2887         talloc_free(mem_ctx);
2888         return status;
2889 }
2890
2891 static bool interfaces_have_changed(struct ctdb_context *ctdb,
2892                                     struct ctdb_recoverd *rec)
2893 {
2894         struct ctdb_control_get_ifaces *ifaces = NULL;
2895         TALLOC_CTX *mem_ctx;
2896         bool ret = false;
2897
2898         mem_ctx = talloc_new(NULL);
2899
2900         /* Read the interfaces from the local node */
2901         if (ctdb_ctrl_get_ifaces(ctdb, CONTROL_TIMEOUT(),
2902                                  CTDB_CURRENT_NODE, mem_ctx, &ifaces) != 0) {
2903                 DEBUG(DEBUG_ERR, ("Unable to get interfaces from local node %u\n", ctdb->pnn));
2904                 /* We could return an error.  However, this will be
2905                  * rare so we'll decide that the interfaces have
2906                  * actually changed, just in case.
2907                  */
2908                 talloc_free(mem_ctx);
2909                 return true;
2910         }
2911
2912         if (!rec->ifaces) {
2913                 /* We haven't been here before so things have changed */
2914                 ret = true;
2915         } else if (rec->ifaces->num != ifaces->num) {
2916                 /* Number of interfaces has changed */
2917                 ret = true;
2918         } else {
2919                 /* See if interface names or link states have changed */
2920                 int i;
2921                 for (i = 0; i < rec->ifaces->num; i++) {
2922                         struct ctdb_control_iface_info * iface = &rec->ifaces->ifaces[i];
2923                         if (strcmp(iface->name, ifaces->ifaces[i].name) != 0 ||
2924                             iface->link_state != ifaces->ifaces[i].link_state) {
2925                                 ret = true;
2926                                 break;
2927                         }
2928                 }
2929         }
2930
2931         talloc_free(rec->ifaces);
2932         rec->ifaces = talloc_steal(rec, ifaces);
2933
2934         talloc_free(mem_ctx);
2935         return ret;
2936 }
2937
2938 /* called to check that the local allocation of public ip addresses is ok.
2939 */
2940 static int verify_local_ip_allocation(struct ctdb_context *ctdb, struct ctdb_recoverd *rec, uint32_t pnn, struct ctdb_node_map *nodemap)
2941 {
2942         TALLOC_CTX *mem_ctx = talloc_new(NULL);
2943         struct ctdb_uptime *uptime1 = NULL;
2944         struct ctdb_uptime *uptime2 = NULL;
2945         int ret, j;
2946         bool need_takeover_run = false;
2947
2948         ret = ctdb_ctrl_uptime(ctdb, mem_ctx, CONTROL_TIMEOUT(),
2949                                 CTDB_CURRENT_NODE, &uptime1);
2950         if (ret != 0) {
2951                 DEBUG(DEBUG_ERR, ("Unable to get uptime from local node %u\n", pnn));
2952                 talloc_free(mem_ctx);
2953                 return -1;
2954         }
2955
2956         if (interfaces_have_changed(ctdb, rec)) {
2957                 DEBUG(DEBUG_NOTICE, ("The interfaces status has changed on "
2958                                      "local node %u - force takeover run\n",
2959                                      pnn));
2960                 need_takeover_run = true;
2961         }
2962
2963         ret = ctdb_ctrl_uptime(ctdb, mem_ctx, CONTROL_TIMEOUT(),
2964                                 CTDB_CURRENT_NODE, &uptime2);
2965         if (ret != 0) {
2966                 DEBUG(DEBUG_ERR, ("Unable to get uptime from local node %u\n", pnn));
2967                 talloc_free(mem_ctx);
2968                 return -1;
2969         }
2970
2971         /* skip the check if the startrecovery time has changed */
2972         if (timeval_compare(&uptime1->last_recovery_started,
2973                             &uptime2->last_recovery_started) != 0) {
2974                 DEBUG(DEBUG_NOTICE, (__location__ " last recovery time changed while we read the public ip list. skipping public ip address check\n"));
2975                 talloc_free(mem_ctx);
2976                 return 0;
2977         }
2978
2979         /* skip the check if the endrecovery time has changed */
2980         if (timeval_compare(&uptime1->last_recovery_finished,
2981                             &uptime2->last_recovery_finished) != 0) {
2982                 DEBUG(DEBUG_NOTICE, (__location__ " last recovery time changed while we read the public ip list. skipping public ip address check\n"));
2983                 talloc_free(mem_ctx);
2984                 return 0;
2985         }
2986
2987         /* skip the check if we have started but not finished recovery */
2988         if (timeval_compare(&uptime1->last_recovery_finished,
2989                             &uptime1->last_recovery_started) != 1) {
2990                 DEBUG(DEBUG_INFO, (__location__ " in the middle of recovery or ip reallocation. skipping public ip address check\n"));
2991                 talloc_free(mem_ctx);
2992
2993                 return 0;
2994         }
2995
2996         /* verify that we have the ip addresses we should have
2997            and we dont have ones we shouldnt have.
2998            if we find an inconsistency we set recmode to
2999            active on the local node and wait for the recmaster
3000            to do a full blown recovery.
3001            also if the pnn is -1 and we are healthy and can host the ip
3002            we also request a ip reallocation.
3003         */
3004         if (ctdb->tunable.disable_ip_failover == 0) {
3005                 struct ctdb_all_public_ips *ips = NULL;
3006
3007                 /* read the *available* IPs from the local node */
3008                 ret = ctdb_ctrl_get_public_ips_flags(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, mem_ctx, CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE, &ips);
3009                 if (ret != 0) {
3010                         DEBUG(DEBUG_ERR, ("Unable to get available public IPs from local node %u\n", pnn));
3011                         talloc_free(mem_ctx);
3012                         return -1;
3013                 }
3014
3015                 for (j=0; j<ips->num; j++) {
3016                         if (ips->ips[j].pnn == -1 &&
3017                             nodemap->nodes[pnn].flags == 0) {
3018                                 DEBUG(DEBUG_CRIT,("Public IP '%s' is not assigned and we could serve it\n",
3019                                                   ctdb_addr_to_str(&ips->ips[j].addr)));
3020                                 need_takeover_run = true;
3021                         }
3022                 }
3023
3024                 talloc_free(ips);
3025
3026                 /* read the *known* IPs from the local node */
3027                 ret = ctdb_ctrl_get_public_ips_flags(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, mem_ctx, 0, &ips);
3028                 if (ret != 0) {
3029                         DEBUG(DEBUG_ERR, ("Unable to get known public IPs from local node %u\n", pnn));
3030                         talloc_free(mem_ctx);
3031                         return -1;
3032                 }
3033
3034                 for (j=0; j<ips->num; j++) {
3035                         if (ips->ips[j].pnn == pnn) {
3036                                 if (ctdb->do_checkpublicip && !ctdb_sys_have_ip(&ips->ips[j].addr)) {
3037                                         DEBUG(DEBUG_CRIT,("Public IP '%s' is assigned to us but not on an interface\n",
3038                                                 ctdb_addr_to_str(&ips->ips[j].addr)));
3039                                         need_takeover_run = true;
3040                                 }
3041                         } else {
3042                                 if (ctdb->do_checkpublicip &&
3043                                     ctdb_sys_have_ip(&ips->ips[j].addr)) {
3044
3045                                         DEBUG(DEBUG_CRIT,("We are still serving a public IP '%s' that we should not be serving. Removing it\n",
3046                                                 ctdb_addr_to_str(&ips->ips[j].addr)));
3047
3048                                         if (ctdb_ctrl_release_ip(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &ips->ips[j]) != 0) {
3049                                                 DEBUG(DEBUG_ERR,("Failed to release local IP address\n"));
3050                                         }
3051                                 }
3052                         }
3053                 }
3054         }
3055
3056         if (need_takeover_run) {
3057                 struct takeover_run_reply rd;
3058                 TDB_DATA data;
3059
3060                 DEBUG(DEBUG_CRIT,("Trigger takeoverrun\n"));
3061
3062                 rd.pnn = ctdb->pnn;
3063                 rd.srvid = 0;
3064                 data.dptr = (uint8_t *)&rd;
3065                 data.dsize = sizeof(rd);
3066
3067                 ret = ctdb_client_send_message(ctdb, rec->recmaster, CTDB_SRVID_TAKEOVER_RUN, data);
3068                 if (ret != 0) {
3069                         DEBUG(DEBUG_ERR,(__location__ " Failed to send ipreallocate to recmaster :%d\n", (int)rec->recmaster));
3070                 }
3071         }
3072         talloc_free(mem_ctx);
3073         return 0;
3074 }
3075
3076
3077 static void async_getnodemap_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
3078 {
3079         struct ctdb_node_map **remote_nodemaps = callback_data;
3080
3081         if (node_pnn >= ctdb->num_nodes) {
3082                 DEBUG(DEBUG_ERR,(__location__ " pnn from invalid node\n"));
3083                 return;
3084         }
3085
3086         remote_nodemaps[node_pnn] = (struct ctdb_node_map *)talloc_steal(remote_nodemaps, outdata.dptr);
3087
3088 }
3089
3090 static int get_remote_nodemaps(struct ctdb_context *ctdb, TALLOC_CTX *mem_ctx,
3091         struct ctdb_node_map *nodemap,
3092         struct ctdb_node_map **remote_nodemaps)
3093 {
3094         uint32_t *nodes;
3095
3096         nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
3097         if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_NODEMAP,
3098                                         nodes, 0,
3099                                         CONTROL_TIMEOUT(), false, tdb_null,
3100                                         async_getnodemap_callback,
3101                                         NULL,
3102                                         remote_nodemaps) != 0) {
3103                 DEBUG(DEBUG_ERR, (__location__ " Unable to pull all remote nodemaps\n"));
3104
3105                 return -1;
3106         }
3107
3108         return 0;
3109 }
3110
3111 enum reclock_child_status { RECLOCK_CHECKING, RECLOCK_OK, RECLOCK_FAILED, RECLOCK_TIMEOUT};
3112 struct ctdb_check_reclock_state {
3113         struct ctdb_context *ctdb;
3114         struct timeval start_time;
3115         int fd[2];
3116         pid_t child;
3117         struct timed_event *te;
3118         struct fd_event *fde;
3119         enum reclock_child_status status;
3120 };
3121
3122 /* when we free the reclock state we must kill any child process.
3123 */
3124 static int check_reclock_destructor(struct ctdb_check_reclock_state *state)
3125 {
3126         struct ctdb_context *ctdb = state->ctdb;
3127
3128         ctdb_ctrl_report_recd_lock_latency(ctdb, CONTROL_TIMEOUT(), timeval_elapsed(&state->start_time));
3129
3130         if (state->fd[0] != -1) {
3131                 close(state->fd[0]);
3132                 state->fd[0] = -1;
3133         }
3134         if (state->fd[1] != -1) {
3135                 close(state->fd[1]);
3136                 state->fd[1] = -1;
3137         }
3138         ctdb_kill(ctdb, state->child, SIGKILL);
3139         return 0;
3140 }
3141
3142 /*
3143   called if our check_reclock child times out. this would happen if
3144   i/o to the reclock file blocks.
3145  */
3146 static void ctdb_check_reclock_timeout(struct event_context *ev, struct timed_event *te,
3147                                          struct timeval t, void *private_data)
3148 {
3149         struct ctdb_check_reclock_state *state = talloc_get_type(private_data,
3150                                            struct ctdb_check_reclock_state);
3151
3152         DEBUG(DEBUG_ERR,(__location__ " check_reclock child process hung/timedout CFS slow to grant locks?\n"));
3153         state->status = RECLOCK_TIMEOUT;
3154 }
3155
3156 /* this is called when the child process has completed checking the reclock
3157    file and has written data back to us through the pipe.
3158 */
3159 static void reclock_child_handler(struct event_context *ev, struct fd_event *fde,
3160                              uint16_t flags, void *private_data)
3161 {
3162         struct ctdb_check_reclock_state *state= talloc_get_type(private_data,
3163                                              struct ctdb_check_reclock_state);
3164         char c = 0;
3165         int ret;
3166
3167         /* we got a response from our child process so we can abort the
3168            timeout.
3169         */
3170         talloc_free(state->te);
3171         state->te = NULL;
3172
3173         ret = read(state->fd[0], &c, 1);
3174         if (ret != 1 || c != RECLOCK_OK) {
3175                 DEBUG(DEBUG_ERR,(__location__ " reclock child process returned error %d\n", c));
3176                 state->status = RECLOCK_FAILED;
3177
3178                 return;
3179         }
3180
3181         state->status = RECLOCK_OK;
3182         return;
3183 }
3184
3185 static int check_recovery_lock(struct ctdb_context *ctdb)
3186 {
3187         int ret;
3188         struct ctdb_check_reclock_state *state;
3189         pid_t parent = getpid();
3190
3191         if (ctdb->recovery_lock_fd == -1) {
3192                 DEBUG(DEBUG_CRIT,("recovery master doesn't have the recovery lock\n"));
3193                 return -1;
3194         }
3195
3196         state = talloc(ctdb, struct ctdb_check_reclock_state);
3197         CTDB_NO_MEMORY(ctdb, state);
3198
3199         state->ctdb = ctdb;
3200         state->start_time = timeval_current();
3201         state->status = RECLOCK_CHECKING;
3202         state->fd[0] = -1;
3203         state->fd[1] = -1;
3204
3205         ret = pipe(state->fd);
3206         if (ret != 0) {
3207                 talloc_free(state);
3208                 DEBUG(DEBUG_CRIT,(__location__ " Failed to open pipe for check_reclock child\n"));
3209                 return -1;
3210         }
3211
3212         state->child = ctdb_fork(ctdb);
3213         if (state->child == (pid_t)-1) {
3214                 DEBUG(DEBUG_CRIT,(__location__ " fork() failed in check_reclock child\n"));
3215                 close(state->fd[0]);
3216                 state->fd[0] = -1;
3217                 close(state->fd[1]);
3218                 state->fd[1] = -1;
3219                 talloc_free(state);
3220                 return -1;
3221         }
3222
3223         if (state->child == 0) {
3224                 char cc = RECLOCK_OK;
3225                 close(state->fd[0]);
3226                 state->fd[0] = -1;
3227
3228                 ctdb_set_process_name("ctdb_rec_reclock");
3229                 debug_extra = talloc_asprintf(NULL, "recovery-lock:");
3230                 if (pread(ctdb->recovery_lock_fd, &cc, 1, 0) == -1) {
3231                         DEBUG(DEBUG_CRIT,("failed read from recovery_lock_fd - %s\n", strerror(errno)));
3232                         cc = RECLOCK_FAILED;
3233                 }
3234
3235                 write(state->fd[1], &cc, 1);
3236                 /* make sure we die when our parent dies */
3237                 while (ctdb_kill(ctdb, parent, 0) == 0 || errno != ESRCH) {
3238                         sleep(5);
3239                 }
3240                 _exit(0);
3241         }
3242         close(state->fd[1]);
3243         state->fd[1] = -1;
3244         set_close_on_exec(state->fd[0]);
3245
3246         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d for check_recovery_lock\n", state->fd[0]));
3247
3248         talloc_set_destructor(state, check_reclock_destructor);
3249
3250         state->te = event_add_timed(ctdb->ev, state, timeval_current_ofs(15, 0),
3251                                     ctdb_check_reclock_timeout, state);
3252         if (state->te == NULL) {
3253                 DEBUG(DEBUG_CRIT,(__location__ " Failed to create a timed event for reclock child\n"));
3254                 talloc_free(state);
3255                 return -1;
3256         }
3257
3258         state->fde = event_add_fd(ctdb->ev, state, state->fd[0],
3259                                 EVENT_FD_READ,
3260                                 reclock_child_handler,
3261                                 (void *)state);
3262
3263         if (state->fde == NULL) {
3264                 DEBUG(DEBUG_CRIT,(__location__ " Failed to create an fd event for reclock child\n"));
3265                 talloc_free(state);
3266                 return -1;
3267         }
3268         tevent_fd_set_auto_close(state->fde);
3269
3270         while (state->status == RECLOCK_CHECKING) {
3271                 event_loop_once(ctdb->ev);
3272         }
3273
3274         if (state->status == RECLOCK_FAILED) {
3275                 DEBUG(DEBUG_ERR,(__location__ " reclock child failed when checking file\n"));
3276                 close(ctdb->recovery_lock_fd);
3277                 ctdb->recovery_lock_fd = -1;
3278                 talloc_free(state);
3279                 return -1;
3280         }
3281
3282         talloc_free(state);
3283         return 0;
3284 }
3285
3286 static int update_recovery_lock_file(struct ctdb_context *ctdb)
3287 {
3288         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
3289         const char *reclockfile;
3290
3291         if (ctdb_ctrl_getreclock(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &reclockfile) != 0) {
3292                 DEBUG(DEBUG_ERR,("Failed to read reclock file from daemon\n"));
3293                 talloc_free(tmp_ctx);
3294                 return -1;
3295         }
3296
3297         if (reclockfile == NULL) {
3298                 if (ctdb->recovery_lock_file != NULL) {
3299                         DEBUG(DEBUG_ERR,("Reclock file disabled\n"));
3300                         talloc_free(ctdb->recovery_lock_file);
3301                         ctdb->recovery_lock_file = NULL;
3302                         if (ctdb->recovery_lock_fd != -1) {
3303                                 close(ctdb->recovery_lock_fd);
3304                                 ctdb->recovery_lock_fd = -1;
3305                         }
3306                 }
3307                 ctdb->tunable.verify_recovery_lock = 0;
3308                 talloc_free(tmp_ctx);
3309                 return 0;
3310         }
3311
3312         if (ctdb->recovery_lock_file == NULL) {
3313                 ctdb->recovery_lock_file = talloc_strdup(ctdb, reclockfile);
3314                 if (ctdb->recovery_lock_fd != -1) {
3315                         close(ctdb->recovery_lock_fd);
3316                         ctdb->recovery_lock_fd = -1;
3317                 }
3318                 talloc_free(tmp_ctx);
3319                 return 0;
3320         }
3321
3322
3323         if (!strcmp(reclockfile, ctdb->recovery_lock_file)) {
3324                 talloc_free(tmp_ctx);
3325                 return 0;
3326         }
3327
3328         talloc_free(ctdb->recovery_lock_file);
3329         ctdb->recovery_lock_file = talloc_strdup(ctdb, reclockfile);
3330         ctdb->tunable.verify_recovery_lock = 0;
3331         if (ctdb->recovery_lock_fd != -1) {
3332                 close(ctdb->recovery_lock_fd);
3333                 ctdb->recovery_lock_fd = -1;
3334         }
3335
3336         talloc_free(tmp_ctx);
3337         return 0;
3338 }
3339
3340 static void main_loop(struct ctdb_context *ctdb, struct ctdb_recoverd *rec,
3341                       TALLOC_CTX *mem_ctx)
3342 {
3343         uint32_t pnn;
3344         struct ctdb_node_map *nodemap=NULL;
3345         struct ctdb_node_map *recmaster_nodemap=NULL;
3346         struct ctdb_node_map **remote_nodemaps=NULL;
3347         struct ctdb_vnn_map *vnnmap=NULL;
3348         struct ctdb_vnn_map *remote_vnnmap=NULL;
3349         int32_t debug_level;
3350         int i, j, ret;
3351         bool self_ban;
3352
3353
3354         /* verify that the main daemon is still running */
3355         if (ctdb_kill(ctdb, ctdb->ctdbd_pid, 0) != 0) {
3356                 DEBUG(DEBUG_CRIT,("CTDB daemon is no longer available. Shutting down recovery daemon\n"));
3357                 exit(-1);
3358         }
3359
3360         /* ping the local daemon to tell it we are alive */
3361         ctdb_ctrl_recd_ping(ctdb);
3362
3363         if (rec->election_timeout) {
3364                 /* an election is in progress */
3365                 return;
3366         }
3367
3368         /* read the debug level from the parent and update locally */
3369         ret = ctdb_ctrl_get_debuglevel(ctdb, CTDB_CURRENT_NODE, &debug_level);
3370         if (ret !=0) {
3371                 DEBUG(DEBUG_ERR, (__location__ " Failed to read debuglevel from parent\n"));
3372                 return;
3373         }
3374         LogLevel = debug_level;
3375
3376         /* get relevant tunables */
3377         ret = ctdb_ctrl_get_all_tunables(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &ctdb->tunable);
3378         if (ret != 0) {
3379                 DEBUG(DEBUG_ERR,("Failed to get tunables - retrying\n"));
3380                 return;
3381         }
3382
3383         /* get the current recovery lock file from the server */
3384         if (update_recovery_lock_file(ctdb) != 0) {
3385                 DEBUG(DEBUG_ERR,("Failed to update the recovery lock file\n"));
3386                 return;
3387         }
3388
3389         /* Make sure that if recovery lock verification becomes disabled when
3390            we close the file
3391         */
3392         if (ctdb->tunable.verify_recovery_lock == 0) {
3393                 if (ctdb->recovery_lock_fd != -1) {
3394                         close(ctdb->recovery_lock_fd);
3395                         ctdb->recovery_lock_fd = -1;
3396                 }
3397         }
3398
3399         pnn = ctdb_ctrl_getpnn(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE);
3400         if (pnn == (uint32_t)-1) {
3401                 DEBUG(DEBUG_ERR,("Failed to get local pnn - retrying\n"));
3402                 return;
3403         }
3404
3405         /* get the vnnmap */
3406         ret = ctdb_ctrl_getvnnmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, &vnnmap);
3407         if (ret != 0) {
3408                 DEBUG(DEBUG_ERR, (__location__ " Unable to get vnnmap from node %u\n", pnn));
3409                 return;
3410         }
3411
3412
3413         /* get number of nodes */
3414         if (rec->nodemap) {
3415                 talloc_free(rec->nodemap);
3416                 rec->nodemap = NULL;
3417                 nodemap=NULL;
3418         }
3419         ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), pnn, rec, &rec->nodemap);
3420         if (ret != 0) {
3421                 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from node %u\n", pnn));
3422                 return;
3423         }
3424         nodemap = rec->nodemap;
3425
3426         /* remember our own node flags */
3427         rec->node_flags = nodemap->nodes[pnn].flags;
3428
3429         ban_misbehaving_nodes(rec, &self_ban);
3430         if (self_ban) {
3431                 DEBUG(DEBUG_NOTICE, ("This node was banned, restart main_loop\n"));
3432                 return;
3433         }
3434
3435         /* if the local daemon is STOPPED or BANNED, we verify that the databases are
3436            also frozen and that the recmode is set to active.
3437         */
3438         if (rec->node_flags & (NODE_FLAGS_STOPPED | NODE_FLAGS_BANNED)) {
3439                 ret = ctdb_ctrl_getrecmode(ctdb, mem_ctx, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &ctdb->recovery_mode);
3440                 if (ret != 0) {
3441                         DEBUG(DEBUG_ERR,(__location__ " Failed to read recmode from local node\n"));
3442                 }
3443                 if (ctdb->recovery_mode == CTDB_RECOVERY_NORMAL) {
3444                         DEBUG(DEBUG_ERR,("Node is stopped or banned but recovery mode is not active. Activate recovery mode and lock databases\n"));
3445
3446                         ret = ctdb_ctrl_freeze_priority(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, 1);
3447                         if (ret != 0) {
3448                                 DEBUG(DEBUG_ERR,(__location__ " Failed to freeze node in STOPPED or BANNED state\n"));
3449                                 return;
3450                         }
3451                         ret = ctdb_ctrl_setrecmode(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, CTDB_RECOVERY_ACTIVE);
3452                         if (ret != 0) {
3453                                 DEBUG(DEBUG_ERR,(__location__ " Failed to activate recovery mode in STOPPED or BANNED state\n"));
3454
3455                                 return;
3456                         }
3457                 }
3458
3459                 /* If this node is stopped or banned then it is not the recovery
3460                  * master, so don't do anything. This prevents stopped or banned
3461                  * node from starting election and sending unnecessary controls.
3462                  */
3463                 return;
3464         }
3465
3466         /* check which node is the recovery master */
3467         ret = ctdb_ctrl_getrecmaster(ctdb, mem_ctx, CONTROL_TIMEOUT(), pnn, &rec->recmaster);
3468         if (ret != 0) {
3469                 DEBUG(DEBUG_ERR, (__location__ " Unable to get recmaster from node %u\n", pnn));
3470                 return;
3471         }
3472
3473         /* if we are not the recmaster we can safely ignore any ip reallocate requests */
3474         if (rec->recmaster != pnn) {
3475                 if (rec->ip_reallocate_ctx != NULL) {
3476                         talloc_free(rec->ip_reallocate_ctx);
3477                         rec->ip_reallocate_ctx = NULL;
3478                         rec->reallocate_callers = NULL;
3479                 }
3480         }
3481
3482         /* This is a special case.  When recovery daemon is started, recmaster
3483          * is set to -1.  If a node is not started in stopped state, then
3484          * start election to decide recovery master
3485          */
3486         if (rec->recmaster == (uint32_t)-1) {
3487                 DEBUG(DEBUG_NOTICE,(__location__ " Initial recovery master set - forcing election\n"));
3488                 force_election(rec, pnn, nodemap);
3489                 return;
3490         }
3491
3492         /* update the capabilities for all nodes */
3493         ret = update_capabilities(ctdb, nodemap);
3494         if (ret != 0) {
3495                 DEBUG(DEBUG_ERR, (__location__ " Unable to update node capabilities.\n"));
3496                 return;
3497         }
3498
3499         /*
3500          * If the current recmaster does not have CTDB_CAP_RECMASTER,
3501          * but we have, then force an election and try to become the new
3502          * recmaster.
3503          */
3504         if ((rec->ctdb->nodes[rec->recmaster]->capabilities & CTDB_CAP_RECMASTER) == 0 &&
3505             (rec->ctdb->capabilities & CTDB_CAP_RECMASTER) &&
3506              !(nodemap->nodes[pnn].flags & NODE_FLAGS_INACTIVE)) {
3507                 DEBUG(DEBUG_ERR, (__location__ " Current recmaster node %u does not have CAP_RECMASTER,"
3508                                   " but we (node %u) have - force an election\n",
3509                                   rec->recmaster, pnn));
3510                 force_election(rec, pnn, nodemap);
3511                 return;
3512         }
3513
3514         /* count how many active nodes there are */
3515         rec->num_active    = 0;
3516         rec->num_connected = 0;
3517         for (i=0; i<nodemap->num; i++) {
3518                 if (!(nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE)) {
3519                         rec->num_active++;
3520                 }
3521                 if (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED)) {
3522                         rec->num_connected++;
3523                 }
3524         }
3525
3526
3527         /* verify that the recmaster node is still active */
3528         for (j=0; j<nodemap->num; j++) {
3529                 if (nodemap->nodes[j].pnn==rec->recmaster) {
3530                         break;
3531                 }
3532         }
3533
3534         if (j == nodemap->num) {
3535                 DEBUG(DEBUG_ERR, ("Recmaster node %u not in list. Force reelection\n", rec->recmaster));
3536                 force_election(rec, pnn, nodemap);
3537                 return;
3538         }
3539
3540         /* if recovery master is disconnected we must elect a new recmaster */
3541         if (nodemap->nodes[j].flags & NODE_FLAGS_DISCONNECTED) {
3542                 DEBUG(DEBUG_NOTICE, ("Recmaster node %u is disconnected. Force reelection\n", nodemap->nodes[j].pnn));
3543                 force_election(rec, pnn, nodemap);
3544                 return;
3545         }
3546
3547         /* get nodemap from the recovery master to check if it is inactive */
3548         ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
3549                                    mem_ctx, &recmaster_nodemap);
3550         if (ret != 0) {
3551                 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from recovery master %u\n",
3552                           nodemap->nodes[j].pnn));
3553                 return;
3554         }
3555
3556
3557         if ((recmaster_nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) &&
3558             (rec->node_flags & NODE_FLAGS_INACTIVE) == 0) {
3559                 DEBUG(DEBUG_NOTICE, ("Recmaster node %u no longer available. Force reelection\n", nodemap->nodes[j].pnn));
3560                 /*
3561                  * update our nodemap to carry the recmaster's notion of
3562                  * its own flags, so that we don't keep freezing the
3563                  * inactive recmaster node...
3564                  */
3565                 nodemap->nodes[j].flags = recmaster_nodemap->nodes[j].flags;
3566                 force_election(rec, pnn, nodemap);
3567                 return;
3568         }
3569
3570         /* verify that we have all ip addresses we should have and we dont
3571          * have addresses we shouldnt have.
3572          */
3573         if (ctdb->tunable.disable_ip_failover == 0) {
3574                 if (rec->ip_check_disable_ctx == NULL) {
3575                         if (verify_local_ip_allocation(ctdb, rec, pnn, nodemap) != 0) {
3576                                 DEBUG(DEBUG_ERR, (__location__ " Public IPs were inconsistent.\n"));
3577                         }
3578                 }
3579         }
3580
3581
3582         /* if we are not the recmaster then we do not need to check
3583            if recovery is needed
3584          */
3585         if (pnn != rec->recmaster) {
3586                 return;
3587         }
3588
3589
3590         /* ensure our local copies of flags are right */
3591         ret = update_local_flags(rec, nodemap);
3592         if (ret == MONITOR_ELECTION_NEEDED) {
3593                 DEBUG(DEBUG_NOTICE,("update_local_flags() called for a re-election.\n"));
3594                 force_election(rec, pnn, nodemap);
3595                 return;
3596         }
3597         if (ret != MONITOR_OK) {
3598                 DEBUG(DEBUG_ERR,("Unable to update local flags\n"));
3599                 return;
3600         }
3601
3602         if (ctdb->num_nodes != nodemap->num) {
3603                 DEBUG(DEBUG_ERR, (__location__ " ctdb->num_nodes (%d) != nodemap->num (%d) reloading nodes file\n", ctdb->num_nodes, nodemap->num));
3604                 reload_nodes_file(ctdb);
3605                 return;
3606         }
3607
3608         /* verify that all active nodes agree that we are the recmaster */
3609         switch (verify_recmaster(rec, nodemap, pnn)) {
3610         case MONITOR_RECOVERY_NEEDED:
3611                 /* can not happen */
3612                 return;
3613         case MONITOR_ELECTION_NEEDED:
3614                 force_election(rec, pnn, nodemap);
3615                 return;
3616         case MONITOR_OK:
3617                 break;
3618         case MONITOR_FAILED:
3619                 return;
3620         }
3621
3622
3623         if (rec->need_recovery) {
3624                 /* a previous recovery didn't finish */
3625                 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3626                 return;
3627         }
3628
3629         /* verify that all active nodes are in normal mode
3630            and not in recovery mode
3631         */
3632         switch (verify_recmode(ctdb, nodemap)) {
3633         case MONITOR_RECOVERY_NEEDED:
3634                 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3635                 return;
3636         case MONITOR_FAILED:
3637                 return;
3638         case MONITOR_ELECTION_NEEDED:
3639                 /* can not happen */
3640         case MONITOR_OK:
3641                 break;
3642         }
3643
3644
3645         if (ctdb->tunable.verify_recovery_lock != 0) {
3646                 /* we should have the reclock - check its not stale */
3647                 ret = check_recovery_lock(ctdb);
3648                 if (ret != 0) {
3649                         DEBUG(DEBUG_ERR,("Failed check_recovery_lock. Force a recovery\n"));
3650                         ctdb_set_culprit(rec, ctdb->pnn);
3651                         do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3652                         return;
3653                 }
3654         }
3655
3656
3657         /* is there a pending reload all ips ? */
3658         if (reload_all_ips_request != NULL) {
3659                 reload_all_ips(ctdb, rec, nodemap, reload_all_ips_request);
3660                 talloc_free(reload_all_ips_request);
3661                 reload_all_ips_request = NULL;
3662         }
3663
3664         /* if there are takeovers requested, perform it and notify the waiters */
3665         if (rec->reallocate_callers) {
3666                 process_ipreallocate_requests(ctdb, rec);
3667         }
3668
3669         /* get the nodemap for all active remote nodes
3670          */
3671         remote_nodemaps = talloc_array(mem_ctx, struct ctdb_node_map *, nodemap->num);
3672         if (remote_nodemaps == NULL) {
3673                 DEBUG(DEBUG_ERR, (__location__ " failed to allocate remote nodemap array\n"));
3674                 return;
3675         }
3676         for(i=0; i<nodemap->num; i++) {
3677                 remote_nodemaps[i] = NULL;
3678         }
3679         if (get_remote_nodemaps(ctdb, mem_ctx, nodemap, remote_nodemaps) != 0) {
3680                 DEBUG(DEBUG_ERR,(__location__ " Failed to read remote nodemaps\n"));
3681                 return;
3682         }
3683
3684         /* verify that all other nodes have the same nodemap as we have
3685         */
3686         for (j=0; j<nodemap->num; j++) {
3687                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3688                         continue;
3689                 }
3690
3691                 if (remote_nodemaps[j] == NULL) {
3692                         DEBUG(DEBUG_ERR,(__location__ " Did not get a remote nodemap for node %d, restarting monitoring\n", j));
3693                         ctdb_set_culprit(rec, j);
3694
3695                         return;
3696                 }
3697
3698                 /* if the nodes disagree on how many nodes there are
3699                    then this is a good reason to try recovery
3700                  */
3701                 if (remote_nodemaps[j]->num != nodemap->num) {
3702                         DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different node count. %u vs %u of the local node\n",
3703                                   nodemap->nodes[j].pnn, remote_nodemaps[j]->num, nodemap->num));
3704                         ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3705                         do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3706                         return;
3707                 }
3708
3709                 /* if the nodes disagree on which nodes exist and are
3710                    active, then that is also a good reason to do recovery
3711                  */
3712                 for (i=0;i<nodemap->num;i++) {
3713                         if (remote_nodemaps[j]->nodes[i].pnn != nodemap->nodes[i].pnn) {
3714                                 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different nodemap pnn for %d (%u vs %u).\n",
3715                                           nodemap->nodes[j].pnn, i,
3716                                           remote_nodemaps[j]->nodes[i].pnn, nodemap->nodes[i].pnn));
3717                                 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3718                                 do_recovery(rec, mem_ctx, pnn, nodemap,
3719                                             vnnmap);
3720                                 return;
3721                         }
3722                 }
3723
3724                 /* verify the flags are consistent
3725                 */
3726                 for (i=0; i<nodemap->num; i++) {
3727                         if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
3728                                 continue;
3729                         }
3730
3731                         if (nodemap->nodes[i].flags != remote_nodemaps[j]->nodes[i].flags) {
3732                                 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different flags for node %u. It has 0x%02x vs our 0x%02x\n",
3733                                   nodemap->nodes[j].pnn,
3734                                   nodemap->nodes[i].pnn,
3735                                   remote_nodemaps[j]->nodes[i].flags,
3736                                   nodemap->nodes[i].flags));
3737                                 if (i == j) {
3738                                         DEBUG(DEBUG_ERR,("Use flags 0x%02x from remote node %d for cluster update of its own flags\n", remote_nodemaps[j]->nodes[i].flags, j));
3739                                         update_flags_on_all_nodes(ctdb, nodemap, nodemap->nodes[i].pnn, remote_nodemaps[j]->nodes[i].flags);
3740                                         ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3741                                         do_recovery(rec, mem_ctx, pnn, nodemap,
3742                                                     vnnmap);
3743                                         return;
3744                                 } else {
3745                                         DEBUG(DEBUG_ERR,("Use flags 0x%02x from local recmaster node for cluster update of node %d flags\n", nodemap->nodes[i].flags, i));
3746                                         update_flags_on_all_nodes(ctdb, nodemap, nodemap->nodes[i].pnn, nodemap->nodes[i].flags);
3747                                         ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3748                                         do_recovery(rec, mem_ctx, pnn, nodemap,
3749                                                     vnnmap);
3750                                         return;
3751                                 }
3752                         }
3753                 }
3754         }
3755
3756
3757         /* there better be the same number of lmasters in the vnn map
3758            as there are active nodes or we will have to do a recovery
3759          */
3760         if (vnnmap->size != rec->num_active) {
3761                 DEBUG(DEBUG_ERR, (__location__ " The vnnmap count is different from the number of active nodes. %u vs %u\n",
3762                           vnnmap->size, rec->num_active));
3763                 ctdb_set_culprit(rec, ctdb->pnn);
3764                 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3765                 return;
3766         }
3767
3768         /* verify that all active nodes in the nodemap also exist in
3769            the vnnmap.
3770          */
3771         for (j=0; j<nodemap->num; j++) {
3772                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3773                         continue;
3774                 }
3775                 if (nodemap->nodes[j].pnn == pnn) {
3776                         continue;
3777                 }
3778
3779                 for (i=0; i<vnnmap->size; i++) {
3780                         if (vnnmap->map[i] == nodemap->nodes[j].pnn) {
3781                                 break;
3782                         }
3783                 }
3784                 if (i == vnnmap->size) {
3785                         DEBUG(DEBUG_ERR, (__location__ " Node %u is active in the nodemap but did not exist in the vnnmap\n",
3786                                   nodemap->nodes[j].pnn));
3787                         ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3788                         do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3789                         return;
3790                 }
3791         }
3792
3793
3794         /* verify that all other nodes have the same vnnmap
3795            and are from the same generation
3796          */
3797         for (j=0; j<nodemap->num; j++) {
3798                 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3799                         continue;
3800                 }
3801                 if (nodemap->nodes[j].pnn == pnn) {
3802                         continue;
3803                 }
3804
3805                 ret = ctdb_ctrl_getvnnmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
3806                                           mem_ctx, &remote_vnnmap);
3807                 if (ret != 0) {
3808                         DEBUG(DEBUG_ERR, (__location__ " Unable to get vnnmap from remote node %u\n",
3809                                   nodemap->nodes[j].pnn));
3810                         return;
3811                 }
3812
3813                 /* verify the vnnmap generation is the same */
3814                 if (vnnmap->generation != remote_vnnmap->generation) {
3815                         DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different generation of vnnmap. %u vs %u (ours)\n",
3816                                   nodemap->nodes[j].pnn, remote_vnnmap->generation, vnnmap->generation));
3817                         ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3818                         do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3819                         return;
3820                 }
3821
3822                 /* verify the vnnmap size is the same */
3823                 if (vnnmap->size != remote_vnnmap->size) {
3824                         DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different size of vnnmap. %u vs %u (ours)\n",
3825                                   nodemap->nodes[j].pnn, remote_vnnmap->size, vnnmap->size));
3826                         ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3827                         do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3828                         return;
3829                 }
3830
3831                 /* verify the vnnmap is the same */
3832                 for (i=0;i<vnnmap->size;i++) {
3833                         if (remote_vnnmap->map[i] != vnnmap->map[i]) {
3834                                 DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different vnnmap.\n",
3835                                           nodemap->nodes[j].pnn));
3836                                 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3837                                 do_recovery(rec, mem_ctx, pnn, nodemap,
3838                                             vnnmap);
3839                                 return;
3840                         }
3841                 }
3842         }
3843
3844         /* we might need to change who has what IP assigned */
3845         if (rec->need_takeover_run) {
3846                 uint32_t culprit = (uint32_t)-1;
3847
3848                 rec->need_takeover_run = false;
3849
3850                 /* update the list of public ips that a node can handle for
3851                    all connected nodes
3852                 */
3853                 ret = ctdb_reload_remote_public_ips(ctdb, rec, nodemap, &culprit);
3854                 if (ret != 0) {
3855                         DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
3856                                          culprit));
3857                         rec->need_takeover_run = true;
3858                         return;
3859                 }
3860
3861                 /* execute the "startrecovery" event script on all nodes */
3862                 ret = run_startrecovery_eventscript(rec, nodemap);
3863                 if (ret!=0) {
3864                         DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
3865                         ctdb_set_culprit(rec, ctdb->pnn);
3866                         do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3867                         return;
3868                 }
3869
3870                 /* If takeover run fails, then the offending nodes are
3871                  * assigned ban culprit counts. And we re-try takeover.
3872                  * If takeover run fails repeatedly, the node would get
3873                  * banned.
3874                  *
3875                  * If rec->need_takeover_run is not set to true at this
3876                  * failure, monitoring is disabled cluster-wide (via
3877                  * startrecovery eventscript) and will not get enabled.
3878                  */
3879                 ret = ctdb_takeover_run(ctdb, nodemap, takeover_fail_callback, rec);
3880                 if (ret != 0) {
3881                         DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses. Trying again\n"));
3882                         return;
3883                 }
3884
3885                 /* execute the "recovered" event script on all nodes */
3886                 ret = run_recovered_eventscript(rec, nodemap, "monitor_cluster");
3887 #if 0
3888 // we cant check whether the event completed successfully
3889 // since this script WILL fail if the node is in recovery mode
3890 // and if that race happens, the code here would just cause a second
3891 // cascading recovery.
3892                 if (ret!=0) {
3893                         DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Update of public ips failed.\n"));
3894                         ctdb_set_culprit(rec, ctdb->pnn);
3895                         do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3896                 }
3897 #endif
3898         }
3899 }
3900
3901 /*
3902   the main monitoring loop
3903  */
3904 static void monitor_cluster(struct ctdb_context *ctdb)
3905 {
3906         struct ctdb_recoverd *rec;
3907
3908         DEBUG(DEBUG_NOTICE,("monitor_cluster starting\n"));
3909
3910         rec = talloc_zero(ctdb, struct ctdb_recoverd);
3911         CTDB_NO_MEMORY_FATAL(ctdb, rec);
3912
3913         rec->ctdb = ctdb;
3914
3915         rec->priority_time = timeval_current();
3916
3917         /* register a message port for sending memory dumps */
3918         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_MEM_DUMP, mem_dump_handler, rec);
3919
3920         /* register a message port for requesting logs */
3921         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_GETLOG, getlog_handler, rec);
3922
3923         /* register a message port for clearing logs */
3924         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_CLEARLOG, clearlog_handler, rec);
3925
3926         /* register a message port for recovery elections */
3927         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RECOVERY, election_handler, rec);
3928
3929         /* when nodes are disabled/enabled */
3930         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_SET_NODE_FLAGS, monitor_handler, rec);
3931
3932         /* when we are asked to puch out a flag change */
3933         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_PUSH_NODE_FLAGS, push_flags_handler, rec);
3934
3935         /* register a message port for vacuum fetch */
3936         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_VACUUM_FETCH, vacuum_fetch_handler, rec);
3937
3938         /* register a message port for reloadnodes  */
3939         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RELOAD_NODES, reload_nodes_handler, rec);
3940
3941         /* register a message port for performing a takeover run */
3942         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_TAKEOVER_RUN, ip_reallocate_handler, rec);
3943
3944         /* register a message port for performing a reload all ips */
3945         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RELOAD_ALL_IPS, ip_reloadall_handler, rec);
3946
3947         /* register a message port for disabling the ip check for a short while */
3948         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_DISABLE_IP_CHECK, disable_ip_check_handler, rec);
3949
3950         /* register a message port for updating the recovery daemons node assignment for an ip */
3951         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RECD_UPDATE_IP, recd_update_ip_handler, rec);
3952
3953         /* register a message port for forcing a rebalance of a node next
3954            reallocation */
3955         ctdb_client_set_message_handler(ctdb, CTDB_SRVID_REBALANCE_NODE, recd_node_rebalance_handler, rec);
3956
3957         for (;;) {
3958                 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
3959                 struct timeval start;
3960                 double elapsed;
3961
3962                 if (!mem_ctx) {
3963                         DEBUG(DEBUG_CRIT,(__location__
3964                                           " Failed to create temp context\n"));
3965                         exit(-1);
3966                 }
3967
3968                 start = timeval_current();
3969                 main_loop(ctdb, rec, mem_ctx);
3970                 talloc_free(mem_ctx);
3971
3972                 /* we only check for recovery once every second */
3973                 elapsed = timeval_elapsed(&start);
3974                 if (elapsed < ctdb->tunable.recover_interval) {
3975                         ctdb_wait_timeout(ctdb, ctdb->tunable.recover_interval
3976                                           - elapsed);
3977                 }
3978         }
3979 }
3980
3981 /*
3982   event handler for when the main ctdbd dies
3983  */
3984 static void ctdb_recoverd_parent(struct event_context *ev, struct fd_event *fde,
3985                                  uint16_t flags, void *private_data)
3986 {
3987         DEBUG(DEBUG_ALERT,("recovery daemon parent died - exiting\n"));
3988         _exit(1);
3989 }
3990
3991 /*
3992   called regularly to verify that the recovery daemon is still running
3993  */
3994 static void ctdb_check_recd(struct event_context *ev, struct timed_event *te,
3995                               struct timeval yt, void *p)
3996 {
3997         struct ctdb_context *ctdb = talloc_get_type(p, struct ctdb_context);
3998
3999         if (ctdb_kill(ctdb, ctdb->recoverd_pid, 0) != 0) {
4000                 DEBUG(DEBUG_ERR,("Recovery daemon (pid:%d) is no longer running. Trying to restart recovery daemon.\n", (int)ctdb->recoverd_pid));
4001
4002                 event_add_timed(ctdb->ev, ctdb, timeval_zero(),
4003                                 ctdb_restart_recd, ctdb);
4004
4005                 return;
4006         }
4007
4008         event_add_timed(ctdb->ev, ctdb->recd_ctx,
4009                         timeval_current_ofs(30, 0),
4010                         ctdb_check_recd, ctdb);
4011 }
4012
4013 static void recd_sig_child_handler(struct event_context *ev,
4014         struct signal_event *se, int signum, int count,
4015         void *dont_care,
4016         void *private_data)
4017 {
4018 //      struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
4019         int status;
4020         pid_t pid = -1;
4021
4022         while (pid != 0) {
4023                 pid = waitpid(-1, &status, WNOHANG);
4024                 if (pid == -1) {
4025                         if (errno != ECHILD) {
4026                                 DEBUG(DEBUG_ERR, (__location__ " waitpid() returned error. errno:%s(%d)\n", strerror(errno),errno));
4027                         }
4028                         return;
4029                 }
4030                 if (pid > 0) {
4031                         DEBUG(DEBUG_DEBUG, ("RECD SIGCHLD from %d\n", (int)pid));
4032                 }
4033         }
4034 }
4035
4036 /*
4037   startup the recovery daemon as a child of the main ctdb daemon
4038  */
4039 int ctdb_start_recoverd(struct ctdb_context *ctdb)
4040 {
4041         int fd[2];
4042         struct signal_event *se;
4043         struct tevent_fd *fde;
4044
4045         if (pipe(fd) != 0) {
4046                 return -1;
4047         }
4048
4049         ctdb->ctdbd_pid = getpid();
4050
4051         ctdb->recoverd_pid = ctdb_fork_no_free_ringbuffer(ctdb);
4052         if (ctdb->recoverd_pid == -1) {
4053                 return -1;
4054         }
4055
4056         if (ctdb->recoverd_pid != 0) {
4057                 talloc_free(ctdb->recd_ctx);
4058                 ctdb->recd_ctx = talloc_new(ctdb);
4059                 CTDB_NO_MEMORY(ctdb, ctdb->recd_ctx);
4060
4061                 close(fd[0]);
4062                 event_add_timed(ctdb->ev, ctdb->recd_ctx,
4063                                 timeval_current_ofs(30, 0),
4064                                 ctdb_check_recd, ctdb);
4065                 return 0;
4066         }
4067
4068         close(fd[1]);
4069
4070         srandom(getpid() ^ time(NULL));
4071
4072         /* Clear the log ringbuffer */
4073         ctdb_clear_log(ctdb);
4074
4075         ctdb_set_process_name("ctdb_recovered");
4076         if (switch_from_server_to_client(ctdb, "recoverd") != 0) {
4077                 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch recovery daemon into client mode. shutting down.\n"));
4078                 exit(1);
4079         }
4080
4081         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to recovery daemon\n", fd[0]));
4082
4083         fde = event_add_fd(ctdb->ev, ctdb, fd[0], EVENT_FD_READ,
4084                      ctdb_recoverd_parent, &fd[0]);
4085         tevent_fd_set_auto_close(fde);
4086
4087         /* set up a handler to pick up sigchld */
4088         se = event_add_signal(ctdb->ev, ctdb,
4089                                      SIGCHLD, 0,
4090                                      recd_sig_child_handler,
4091                                      ctdb);
4092         if (se == NULL) {
4093                 DEBUG(DEBUG_CRIT,("Failed to set up signal handler for SIGCHLD in recovery daemon\n"));
4094                 exit(1);
4095         }
4096
4097         monitor_cluster(ctdb);
4098
4099         DEBUG(DEBUG_ALERT,("ERROR: ctdb_recoverd finished!?\n"));
4100         return -1;
4101 }
4102
4103 /*
4104   shutdown the recovery daemon
4105  */
4106 void ctdb_stop_recoverd(struct ctdb_context *ctdb)
4107 {
4108         if (ctdb->recoverd_pid == 0) {
4109                 return;
4110         }
4111
4112         DEBUG(DEBUG_NOTICE,("Shutting down recovery daemon\n"));
4113         ctdb_kill(ctdb, ctdb->recoverd_pid, SIGTERM);
4114
4115         TALLOC_FREE(ctdb->recd_ctx);
4116         TALLOC_FREE(ctdb->recd_ping_count);
4117 }
4118
4119 static void ctdb_restart_recd(struct event_context *ev, struct timed_event *te,
4120                        struct timeval t, void *private_data)
4121 {
4122         struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
4123
4124         DEBUG(DEBUG_ERR,("Restarting recovery daemon\n"));
4125         ctdb_stop_recoverd(ctdb);
4126         ctdb_start_recoverd(ctdb);
4127 }