4 Copyright (C) Ronnie Sahlberg 2007
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.
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.
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/>.
21 #include "lib/tevent/tevent.h"
22 #include "system/filesys.h"
23 #include "system/time.h"
24 #include "system/network.h"
25 #include "system/wait.h"
28 #include "../include/ctdb_client.h"
29 #include "../include/ctdb_private.h"
31 #include "dlinklist.h"
34 /* list of "ctdb ipreallocate" processes to call back when we have
35 finished the takeover run.
37 struct ip_reallocate_list {
38 struct ip_reallocate_list *next;
39 struct rd_memdump_reply *rd;
42 struct ctdb_banning_state {
44 struct timeval last_reported_time;
48 private state of recovery daemon
50 struct ctdb_recoverd {
51 struct ctdb_context *ctdb;
54 uint32_t num_connected;
55 uint32_t last_culprit_node;
56 struct ctdb_node_map *nodemap;
57 struct timeval priority_time;
58 bool need_takeover_run;
61 struct timed_event *send_election_te;
62 struct timed_event *election_timeout;
63 struct vacuum_info *vacuum_info;
64 TALLOC_CTX *ip_reallocate_ctx;
65 struct ip_reallocate_list *reallocate_callers;
66 TALLOC_CTX *ip_check_disable_ctx;
67 struct ctdb_control_get_ifaces *ifaces;
70 #define CONTROL_TIMEOUT() timeval_current_ofs(ctdb->tunable.recover_timeout, 0)
71 #define MONITOR_TIMEOUT() timeval_current_ofs(ctdb->tunable.recover_interval, 0)
73 static void ctdb_restart_recd(struct event_context *ev, struct timed_event *te, struct timeval t, void *private_data);
76 ban a node for a period of time
78 static void ctdb_ban_node(struct ctdb_recoverd *rec, uint32_t pnn, uint32_t ban_time)
81 struct ctdb_context *ctdb = rec->ctdb;
82 struct ctdb_ban_time bantime;
84 DEBUG(DEBUG_NOTICE,("Banning node %u for %u seconds\n", pnn, ban_time));
86 if (!ctdb_validate_pnn(ctdb, pnn)) {
87 DEBUG(DEBUG_ERR,("Bad pnn %u in ctdb_ban_node\n", pnn));
92 bantime.time = ban_time;
94 ret = ctdb_ctrl_set_ban(ctdb, CONTROL_TIMEOUT(), pnn, &bantime);
96 DEBUG(DEBUG_ERR,(__location__ " Failed to ban node %d\n", pnn));
102 enum monitor_result { MONITOR_OK, MONITOR_RECOVERY_NEEDED, MONITOR_ELECTION_NEEDED, MONITOR_FAILED};
106 run the "recovered" eventscript on all nodes
108 static int run_recovered_eventscript(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, const char *caller)
113 tmp_ctx = talloc_new(ctdb);
114 CTDB_NO_MEMORY(ctdb, tmp_ctx);
116 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
117 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_END_RECOVERY,
119 CONTROL_TIMEOUT(), false, tdb_null,
122 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event when called from %s\n", caller));
124 talloc_free(tmp_ctx);
128 talloc_free(tmp_ctx);
133 remember the trouble maker
135 static void ctdb_set_culprit_count(struct ctdb_recoverd *rec, uint32_t culprit, uint32_t count)
137 struct ctdb_context *ctdb = talloc_get_type(rec->ctdb, struct ctdb_context);
138 struct ctdb_banning_state *ban_state;
140 if (culprit > ctdb->num_nodes) {
141 DEBUG(DEBUG_ERR,("Trying to set culprit %d but num_nodes is %d\n", culprit, ctdb->num_nodes));
145 if (ctdb->nodes[culprit]->ban_state == NULL) {
146 ctdb->nodes[culprit]->ban_state = talloc_zero(ctdb->nodes[culprit], struct ctdb_banning_state);
147 CTDB_NO_MEMORY_VOID(ctdb, ctdb->nodes[culprit]->ban_state);
151 ban_state = ctdb->nodes[culprit]->ban_state;
152 if (timeval_elapsed(&ban_state->last_reported_time) > ctdb->tunable.recovery_grace_period) {
153 /* this was the first time in a long while this node
154 misbehaved so we will forgive any old transgressions.
156 ban_state->count = 0;
159 ban_state->count += count;
160 ban_state->last_reported_time = timeval_current();
161 rec->last_culprit_node = culprit;
165 remember the trouble maker
167 static void ctdb_set_culprit(struct ctdb_recoverd *rec, uint32_t culprit)
169 ctdb_set_culprit_count(rec, culprit, 1);
173 /* this callback is called for every node that failed to execute the
176 static void startrecovery_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
178 struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
180 DEBUG(DEBUG_ERR, (__location__ " Node %u failed the startrecovery event. Setting it as recovery fail culprit\n", node_pnn));
182 ctdb_set_culprit(rec, node_pnn);
186 run the "startrecovery" eventscript on all nodes
188 static int run_startrecovery_eventscript(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap)
192 struct ctdb_context *ctdb = rec->ctdb;
194 tmp_ctx = talloc_new(ctdb);
195 CTDB_NO_MEMORY(ctdb, tmp_ctx);
197 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
198 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_START_RECOVERY,
200 CONTROL_TIMEOUT(), false, tdb_null,
202 startrecovery_fail_callback,
204 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event. Recovery failed.\n"));
205 talloc_free(tmp_ctx);
209 talloc_free(tmp_ctx);
213 static void async_getcap_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
215 if ( (outdata.dsize != sizeof(uint32_t)) || (outdata.dptr == NULL) ) {
216 DEBUG(DEBUG_ERR, (__location__ " Invalid length/pointer for getcap callback : %u %p\n", (unsigned)outdata.dsize, outdata.dptr));
219 if (node_pnn < ctdb->num_nodes) {
220 ctdb->nodes[node_pnn]->capabilities = *((uint32_t *)outdata.dptr);
225 update the node capabilities for all connected nodes
227 static int update_capabilities(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
232 tmp_ctx = talloc_new(ctdb);
233 CTDB_NO_MEMORY(ctdb, tmp_ctx);
235 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
236 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_CAPABILITIES,
240 async_getcap_callback, NULL,
242 DEBUG(DEBUG_ERR, (__location__ " Failed to read node capabilities.\n"));
243 talloc_free(tmp_ctx);
247 talloc_free(tmp_ctx);
251 static void set_recmode_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
253 struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
255 DEBUG(DEBUG_ERR,("Failed to freeze node %u during recovery. Set it as ban culprit for %d credits\n", node_pnn, rec->nodemap->num));
256 ctdb_set_culprit_count(rec, node_pnn, rec->nodemap->num);
259 static void transaction_start_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
261 struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
263 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));
264 ctdb_set_culprit_count(rec, node_pnn, rec->nodemap->num);
268 change recovery mode on all nodes
270 static int set_recovery_mode(struct ctdb_context *ctdb, struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, uint32_t rec_mode)
276 tmp_ctx = talloc_new(ctdb);
277 CTDB_NO_MEMORY(ctdb, tmp_ctx);
279 /* freeze all nodes */
280 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
281 if (rec_mode == CTDB_RECOVERY_ACTIVE) {
284 for (i=1; i<=NUM_DB_PRIORITIES; i++) {
285 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_FREEZE,
290 set_recmode_fail_callback,
292 DEBUG(DEBUG_ERR, (__location__ " Unable to freeze nodes. Recovery failed.\n"));
293 talloc_free(tmp_ctx);
300 data.dsize = sizeof(uint32_t);
301 data.dptr = (unsigned char *)&rec_mode;
303 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECMODE,
309 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode. Recovery failed.\n"));
310 talloc_free(tmp_ctx);
314 talloc_free(tmp_ctx);
319 change recovery master on all node
321 static int set_recovery_master(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, uint32_t pnn)
327 tmp_ctx = talloc_new(ctdb);
328 CTDB_NO_MEMORY(ctdb, tmp_ctx);
330 data.dsize = sizeof(uint32_t);
331 data.dptr = (unsigned char *)&pnn;
333 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
334 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECMASTER,
336 CONTROL_TIMEOUT(), false, data,
339 DEBUG(DEBUG_ERR, (__location__ " Unable to set recmaster. Recovery failed.\n"));
340 talloc_free(tmp_ctx);
344 talloc_free(tmp_ctx);
348 /* update all remote nodes to use the same db priority that we have
349 this can fail if the remove node has not yet been upgraded to
350 support this function, so we always return success and never fail
351 a recovery if this call fails.
353 static int update_db_priority_on_remote_nodes(struct ctdb_context *ctdb,
354 struct ctdb_node_map *nodemap,
355 uint32_t pnn, struct ctdb_dbid_map *dbmap, TALLOC_CTX *mem_ctx)
360 nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
362 /* step through all local databases */
363 for (db=0; db<dbmap->num;db++) {
365 struct ctdb_db_priority db_prio;
368 db_prio.db_id = dbmap->dbs[db].dbid;
369 ret = ctdb_ctrl_get_db_priority(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, dbmap->dbs[db].dbid, &db_prio.priority);
371 DEBUG(DEBUG_ERR,(__location__ " Failed to read database priority from local node for db 0x%08x\n", dbmap->dbs[db].dbid));
375 DEBUG(DEBUG_INFO,("Update DB priority for db 0x%08x to %u\n", dbmap->dbs[db].dbid, db_prio.priority));
377 data.dptr = (uint8_t *)&db_prio;
378 data.dsize = sizeof(db_prio);
380 if (ctdb_client_async_control(ctdb,
381 CTDB_CONTROL_SET_DB_PRIORITY,
383 CONTROL_TIMEOUT(), false, data,
386 DEBUG(DEBUG_ERR,(__location__ " Failed to set DB priority for 0x%08x\n", db_prio.db_id));
394 ensure all other nodes have attached to any databases that we have
396 static int create_missing_remote_databases(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
397 uint32_t pnn, struct ctdb_dbid_map *dbmap, TALLOC_CTX *mem_ctx)
400 struct ctdb_dbid_map *remote_dbmap;
402 /* verify that all other nodes have all our databases */
403 for (j=0; j<nodemap->num; j++) {
404 /* we dont need to ourself ourselves */
405 if (nodemap->nodes[j].pnn == pnn) {
408 /* dont check nodes that are unavailable */
409 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
413 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
414 mem_ctx, &remote_dbmap);
416 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node %u\n", pnn));
420 /* step through all local databases */
421 for (db=0; db<dbmap->num;db++) {
425 for (i=0;i<remote_dbmap->num;i++) {
426 if (dbmap->dbs[db].dbid == remote_dbmap->dbs[i].dbid) {
430 /* the remote node already have this database */
431 if (i!=remote_dbmap->num) {
434 /* ok so we need to create this database */
435 ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), pnn, dbmap->dbs[db].dbid,
438 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbname from node %u\n", pnn));
441 ctdb_ctrl_createdb(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
443 dbmap->dbs[db].flags & CTDB_DB_FLAGS_PERSISTENT);
445 DEBUG(DEBUG_ERR, (__location__ " Unable to create remote db:%s\n", name));
456 ensure we are attached to any databases that anyone else is attached to
458 static int create_missing_local_databases(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
459 uint32_t pnn, struct ctdb_dbid_map **dbmap, TALLOC_CTX *mem_ctx)
462 struct ctdb_dbid_map *remote_dbmap;
464 /* verify that we have all database any other node has */
465 for (j=0; j<nodemap->num; j++) {
466 /* we dont need to ourself ourselves */
467 if (nodemap->nodes[j].pnn == pnn) {
470 /* dont check nodes that are unavailable */
471 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
475 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
476 mem_ctx, &remote_dbmap);
478 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node %u\n", pnn));
482 /* step through all databases on the remote node */
483 for (db=0; db<remote_dbmap->num;db++) {
486 for (i=0;i<(*dbmap)->num;i++) {
487 if (remote_dbmap->dbs[db].dbid == (*dbmap)->dbs[i].dbid) {
491 /* we already have this db locally */
492 if (i!=(*dbmap)->num) {
495 /* ok so we need to create this database and
498 ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
499 remote_dbmap->dbs[db].dbid, mem_ctx, &name);
501 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbname from node %u\n",
502 nodemap->nodes[j].pnn));
505 ctdb_ctrl_createdb(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, name,
506 remote_dbmap->dbs[db].flags & CTDB_DB_FLAGS_PERSISTENT);
508 DEBUG(DEBUG_ERR, (__location__ " Unable to create local db:%s\n", name));
511 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, dbmap);
513 DEBUG(DEBUG_ERR, (__location__ " Unable to reread dbmap on node %u\n", pnn));
524 pull the remote database contents from one node into the recdb
526 static int pull_one_remote_database(struct ctdb_context *ctdb, uint32_t srcnode,
527 struct tdb_wrap *recdb, uint32_t dbid,
532 struct ctdb_marshall_buffer *reply;
533 struct ctdb_rec_data *rec;
535 TALLOC_CTX *tmp_ctx = talloc_new(recdb);
537 ret = ctdb_ctrl_pulldb(ctdb, srcnode, dbid, CTDB_LMASTER_ANY, tmp_ctx,
538 CONTROL_TIMEOUT(), &outdata);
540 DEBUG(DEBUG_ERR,(__location__ " Unable to copy db from node %u\n", srcnode));
541 talloc_free(tmp_ctx);
545 reply = (struct ctdb_marshall_buffer *)outdata.dptr;
547 if (outdata.dsize < offsetof(struct ctdb_marshall_buffer, data)) {
548 DEBUG(DEBUG_ERR,(__location__ " invalid data in pulldb reply\n"));
549 talloc_free(tmp_ctx);
553 rec = (struct ctdb_rec_data *)&reply->data[0];
557 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec), i++) {
559 struct ctdb_ltdb_header *hdr;
562 key.dptr = &rec->data[0];
563 key.dsize = rec->keylen;
564 data.dptr = &rec->data[key.dsize];
565 data.dsize = rec->datalen;
567 hdr = (struct ctdb_ltdb_header *)data.dptr;
569 if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
570 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
571 talloc_free(tmp_ctx);
575 /* fetch the existing record, if any */
576 existing = tdb_fetch(recdb->tdb, key);
578 if (existing.dptr != NULL) {
579 struct ctdb_ltdb_header header;
580 if (existing.dsize < sizeof(struct ctdb_ltdb_header)) {
581 DEBUG(DEBUG_CRIT,(__location__ " Bad record size %u from node %u\n",
582 (unsigned)existing.dsize, srcnode));
584 talloc_free(tmp_ctx);
587 header = *(struct ctdb_ltdb_header *)existing.dptr;
589 if (!(header.rsn < hdr->rsn ||
590 (header.dmaster != ctdb->recovery_master && header.rsn == hdr->rsn))) {
595 if (tdb_store(recdb->tdb, key, data, TDB_REPLACE) != 0) {
596 DEBUG(DEBUG_CRIT,(__location__ " Failed to store record\n"));
597 talloc_free(tmp_ctx);
602 talloc_free(tmp_ctx);
608 pull all the remote database contents into the recdb
610 static int pull_remote_database(struct ctdb_context *ctdb,
611 struct ctdb_recoverd *rec,
612 struct ctdb_node_map *nodemap,
613 struct tdb_wrap *recdb, uint32_t dbid,
618 /* pull all records from all other nodes across onto this node
619 (this merges based on rsn)
621 for (j=0; j<nodemap->num; j++) {
622 /* dont merge from nodes that are unavailable */
623 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
626 if (pull_one_remote_database(ctdb, nodemap->nodes[j].pnn, recdb, dbid, persistent) != 0) {
627 DEBUG(DEBUG_ERR,(__location__ " Failed to pull remote database from node %u\n",
628 nodemap->nodes[j].pnn));
629 ctdb_set_culprit_count(rec, nodemap->nodes[j].pnn, nodemap->num);
639 update flags on all active nodes
641 static int update_flags_on_all_nodes(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, uint32_t pnn, uint32_t flags)
645 ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), pnn, flags, ~flags);
647 DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
655 ensure all nodes have the same vnnmap we do
657 static int update_vnnmap_on_all_nodes(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
658 uint32_t pnn, struct ctdb_vnn_map *vnnmap, TALLOC_CTX *mem_ctx)
662 /* push the new vnn map out to all the nodes */
663 for (j=0; j<nodemap->num; j++) {
664 /* dont push to nodes that are unavailable */
665 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
669 ret = ctdb_ctrl_setvnnmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn, mem_ctx, vnnmap);
671 DEBUG(DEBUG_ERR, (__location__ " Unable to set vnnmap for node %u\n", pnn));
681 struct vacuum_info *next, *prev;
682 struct ctdb_recoverd *rec;
684 struct ctdb_db_context *ctdb_db;
685 struct ctdb_marshall_buffer *recs;
686 struct ctdb_rec_data *r;
689 static void vacuum_fetch_next(struct vacuum_info *v);
692 called when a vacuum fetch has completed - just free it and do the next one
694 static void vacuum_fetch_callback(struct ctdb_client_call_state *state)
696 struct vacuum_info *v = talloc_get_type(state->async.private_data, struct vacuum_info);
698 vacuum_fetch_next(v);
703 process the next element from the vacuum list
705 static void vacuum_fetch_next(struct vacuum_info *v)
707 struct ctdb_call call;
708 struct ctdb_rec_data *r;
710 while (v->recs->count) {
711 struct ctdb_client_call_state *state;
713 struct ctdb_ltdb_header *hdr;
716 call.call_id = CTDB_NULL_FUNC;
717 call.flags = CTDB_IMMEDIATE_MIGRATION;
718 call.flags |= CTDB_CALL_FLAG_VACUUM_MIGRATION;
721 v->r = (struct ctdb_rec_data *)(r->length + (uint8_t *)r);
724 call.key.dptr = &r->data[0];
725 call.key.dsize = r->keylen;
727 /* ensure we don't block this daemon - just skip a record if we can't get
729 if (tdb_chainlock_nonblock(v->ctdb_db->ltdb->tdb, call.key) != 0) {
733 data = tdb_fetch(v->ctdb_db->ltdb->tdb, call.key);
734 if (data.dptr == NULL) {
735 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
739 if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
741 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
745 hdr = (struct ctdb_ltdb_header *)data.dptr;
746 if (hdr->dmaster == v->rec->ctdb->pnn) {
747 /* its already local */
749 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
755 state = ctdb_call_send(v->ctdb_db, &call);
756 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
758 DEBUG(DEBUG_ERR,(__location__ " Failed to setup vacuum fetch call\n"));
762 state->async.fn = vacuum_fetch_callback;
763 state->async.private_data = v;
772 destroy a vacuum info structure
774 static int vacuum_info_destructor(struct vacuum_info *v)
776 DLIST_REMOVE(v->rec->vacuum_info, v);
782 handler for vacuum fetch
784 static void vacuum_fetch_handler(struct ctdb_context *ctdb, uint64_t srvid,
785 TDB_DATA data, void *private_data)
787 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
788 struct ctdb_marshall_buffer *recs;
790 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
792 struct ctdb_dbid_map *dbmap=NULL;
793 bool persistent = false;
794 struct ctdb_db_context *ctdb_db;
795 struct ctdb_rec_data *r;
797 struct vacuum_info *v;
799 recs = (struct ctdb_marshall_buffer *)data.dptr;
800 r = (struct ctdb_rec_data *)&recs->data[0];
802 if (recs->count == 0) {
803 talloc_free(tmp_ctx);
809 for (v=rec->vacuum_info;v;v=v->next) {
810 if (srcnode == v->srcnode && recs->db_id == v->ctdb_db->db_id) {
811 /* we're already working on records from this node */
812 talloc_free(tmp_ctx);
817 /* work out if the database is persistent */
818 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &dbmap);
820 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from local node\n"));
821 talloc_free(tmp_ctx);
825 for (i=0;i<dbmap->num;i++) {
826 if (dbmap->dbs[i].dbid == recs->db_id) {
827 persistent = dbmap->dbs[i].flags & CTDB_DB_FLAGS_PERSISTENT;
831 if (i == dbmap->num) {
832 DEBUG(DEBUG_ERR, (__location__ " Unable to find db_id 0x%x on local node\n", recs->db_id));
833 talloc_free(tmp_ctx);
837 /* find the name of this database */
838 if (ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, recs->db_id, tmp_ctx, &name) != 0) {
839 DEBUG(DEBUG_ERR,(__location__ " Failed to get name of db 0x%x\n", recs->db_id));
840 talloc_free(tmp_ctx);
845 ctdb_db = ctdb_attach(ctdb, CONTROL_TIMEOUT(), name, persistent, 0);
846 if (ctdb_db == NULL) {
847 DEBUG(DEBUG_ERR,(__location__ " Failed to attach to database '%s'\n", name));
848 talloc_free(tmp_ctx);
852 v = talloc_zero(rec, struct vacuum_info);
854 DEBUG(DEBUG_CRIT,(__location__ " Out of memory\n"));
855 talloc_free(tmp_ctx);
860 v->srcnode = srcnode;
861 v->ctdb_db = ctdb_db;
862 v->recs = talloc_memdup(v, recs, data.dsize);
863 if (v->recs == NULL) {
864 DEBUG(DEBUG_CRIT,(__location__ " Out of memory\n"));
866 talloc_free(tmp_ctx);
869 v->r = (struct ctdb_rec_data *)&v->recs->data[0];
871 DLIST_ADD(rec->vacuum_info, v);
873 talloc_set_destructor(v, vacuum_info_destructor);
875 vacuum_fetch_next(v);
876 talloc_free(tmp_ctx);
881 called when ctdb_wait_timeout should finish
883 static void ctdb_wait_handler(struct event_context *ev, struct timed_event *te,
884 struct timeval yt, void *p)
886 uint32_t *timed_out = (uint32_t *)p;
891 wait for a given number of seconds
893 static void ctdb_wait_timeout(struct ctdb_context *ctdb, double secs)
895 uint32_t timed_out = 0;
896 time_t usecs = (secs - (time_t)secs) * 1000000;
897 event_add_timed(ctdb->ev, ctdb, timeval_current_ofs(secs, usecs), ctdb_wait_handler, &timed_out);
899 event_loop_once(ctdb->ev);
904 called when an election times out (ends)
906 static void ctdb_election_timeout(struct event_context *ev, struct timed_event *te,
907 struct timeval t, void *p)
909 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
910 rec->election_timeout = NULL;
913 DEBUG(DEBUG_WARNING,(__location__ " Election timed out\n"));
918 wait for an election to finish. It finished election_timeout seconds after
919 the last election packet is received
921 static void ctdb_wait_election(struct ctdb_recoverd *rec)
923 struct ctdb_context *ctdb = rec->ctdb;
924 while (rec->election_timeout) {
925 event_loop_once(ctdb->ev);
930 Update our local flags from all remote connected nodes.
931 This is only run when we are or we belive we are the recovery master
933 static int update_local_flags(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap)
936 struct ctdb_context *ctdb = rec->ctdb;
937 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
939 /* get the nodemap for all active remote nodes and verify
940 they are the same as for this node
942 for (j=0; j<nodemap->num; j++) {
943 struct ctdb_node_map *remote_nodemap=NULL;
946 if (nodemap->nodes[j].flags & NODE_FLAGS_DISCONNECTED) {
949 if (nodemap->nodes[j].pnn == ctdb->pnn) {
953 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
954 mem_ctx, &remote_nodemap);
956 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from remote node %u\n",
957 nodemap->nodes[j].pnn));
958 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
959 talloc_free(mem_ctx);
960 return MONITOR_FAILED;
962 if (nodemap->nodes[j].flags != remote_nodemap->nodes[j].flags) {
963 /* We should tell our daemon about this so it
964 updates its flags or else we will log the same
965 message again in the next iteration of recovery.
966 Since we are the recovery master we can just as
967 well update the flags on all nodes.
969 ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn, nodemap->nodes[j].flags, ~nodemap->nodes[j].flags);
971 DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
975 /* Update our local copy of the flags in the recovery
978 DEBUG(DEBUG_NOTICE,("Remote node %u had flags 0x%x, local had 0x%x - updating local\n",
979 nodemap->nodes[j].pnn, remote_nodemap->nodes[j].flags,
980 nodemap->nodes[j].flags));
981 nodemap->nodes[j].flags = remote_nodemap->nodes[j].flags;
983 talloc_free(remote_nodemap);
985 talloc_free(mem_ctx);
990 /* Create a new random generation ip.
991 The generation id can not be the INVALID_GENERATION id
993 static uint32_t new_generation(void)
998 generation = random();
1000 if (generation != INVALID_GENERATION) {
1010 create a temporary working database
1012 static struct tdb_wrap *create_recdb(struct ctdb_context *ctdb, TALLOC_CTX *mem_ctx)
1015 struct tdb_wrap *recdb;
1018 /* open up the temporary recovery database */
1019 name = talloc_asprintf(mem_ctx, "%s/recdb.tdb.%u",
1020 ctdb->db_directory_state,
1027 tdb_flags = TDB_NOLOCK;
1028 if (ctdb->valgrinding) {
1029 tdb_flags |= TDB_NOMMAP;
1031 tdb_flags |= TDB_DISALLOW_NESTING;
1033 recdb = tdb_wrap_open(mem_ctx, name, ctdb->tunable.database_hash_size,
1034 tdb_flags, O_RDWR|O_CREAT|O_EXCL, 0600);
1035 if (recdb == NULL) {
1036 DEBUG(DEBUG_CRIT,(__location__ " Failed to create temp recovery database '%s'\n", name));
1046 a traverse function for pulling all relevent records from recdb
1049 struct ctdb_context *ctdb;
1050 struct ctdb_marshall_buffer *recdata;
1056 static int traverse_recdb(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *p)
1058 struct recdb_data *params = (struct recdb_data *)p;
1059 struct ctdb_rec_data *rec;
1060 struct ctdb_ltdb_header *hdr;
1062 /* skip empty records */
1063 if (data.dsize <= sizeof(struct ctdb_ltdb_header)) {
1067 /* update the dmaster field to point to us */
1068 hdr = (struct ctdb_ltdb_header *)data.dptr;
1069 if (!params->persistent) {
1070 hdr->dmaster = params->ctdb->pnn;
1071 hdr->flags |= CTDB_REC_FLAG_MIGRATED_WITH_DATA;
1074 /* add the record to the blob ready to send to the nodes */
1075 rec = ctdb_marshall_record(params->recdata, 0, key, NULL, data);
1077 params->failed = true;
1080 params->recdata = talloc_realloc_size(NULL, params->recdata, rec->length + params->len);
1081 if (params->recdata == NULL) {
1082 DEBUG(DEBUG_CRIT,(__location__ " Failed to expand recdata to %u (%u records)\n",
1083 rec->length + params->len, params->recdata->count));
1084 params->failed = true;
1087 params->recdata->count++;
1088 memcpy(params->len+(uint8_t *)params->recdata, rec, rec->length);
1089 params->len += rec->length;
1096 push the recdb database out to all nodes
1098 static int push_recdb_database(struct ctdb_context *ctdb, uint32_t dbid,
1100 struct tdb_wrap *recdb, struct ctdb_node_map *nodemap)
1102 struct recdb_data params;
1103 struct ctdb_marshall_buffer *recdata;
1105 TALLOC_CTX *tmp_ctx;
1108 tmp_ctx = talloc_new(ctdb);
1109 CTDB_NO_MEMORY(ctdb, tmp_ctx);
1111 recdata = talloc_zero(recdb, struct ctdb_marshall_buffer);
1112 CTDB_NO_MEMORY(ctdb, recdata);
1114 recdata->db_id = dbid;
1117 params.recdata = recdata;
1118 params.len = offsetof(struct ctdb_marshall_buffer, data);
1119 params.failed = false;
1120 params.persistent = persistent;
1122 if (tdb_traverse_read(recdb->tdb, traverse_recdb, ¶ms) == -1) {
1123 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1124 talloc_free(params.recdata);
1125 talloc_free(tmp_ctx);
1129 if (params.failed) {
1130 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1131 talloc_free(params.recdata);
1132 talloc_free(tmp_ctx);
1136 recdata = params.recdata;
1138 outdata.dptr = (void *)recdata;
1139 outdata.dsize = params.len;
1141 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
1142 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_PUSH_DB,
1144 CONTROL_TIMEOUT(), false, outdata,
1147 DEBUG(DEBUG_ERR,(__location__ " Failed to push recdb records to nodes for db 0x%x\n", dbid));
1148 talloc_free(recdata);
1149 talloc_free(tmp_ctx);
1153 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pushed remote database 0x%x of size %u\n",
1154 dbid, recdata->count));
1156 talloc_free(recdata);
1157 talloc_free(tmp_ctx);
1164 go through a full recovery on one database
1166 static int recover_database(struct ctdb_recoverd *rec,
1167 TALLOC_CTX *mem_ctx,
1171 struct ctdb_node_map *nodemap,
1172 uint32_t transaction_id)
1174 struct tdb_wrap *recdb;
1176 struct ctdb_context *ctdb = rec->ctdb;
1178 struct ctdb_control_wipe_database w;
1181 recdb = create_recdb(ctdb, mem_ctx);
1182 if (recdb == NULL) {
1186 /* pull all remote databases onto the recdb */
1187 ret = pull_remote_database(ctdb, rec, nodemap, recdb, dbid, persistent);
1189 DEBUG(DEBUG_ERR, (__location__ " Unable to pull remote database 0x%x\n", dbid));
1193 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pulled remote database 0x%x\n", dbid));
1195 /* wipe all the remote databases. This is safe as we are in a transaction */
1197 w.transaction_id = transaction_id;
1199 data.dptr = (void *)&w;
1200 data.dsize = sizeof(w);
1202 nodes = list_of_active_nodes(ctdb, nodemap, recdb, true);
1203 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_WIPE_DATABASE,
1205 CONTROL_TIMEOUT(), false, data,
1208 DEBUG(DEBUG_ERR, (__location__ " Unable to wipe database. Recovery failed.\n"));
1213 /* push out the correct database. This sets the dmaster and skips
1214 the empty records */
1215 ret = push_recdb_database(ctdb, dbid, persistent, recdb, nodemap);
1221 /* all done with this database */
1228 reload the nodes file
1230 static void reload_nodes_file(struct ctdb_context *ctdb)
1233 ctdb_load_nodes_file(ctdb);
1236 static int ctdb_reload_remote_public_ips(struct ctdb_context *ctdb,
1237 struct ctdb_recoverd *rec,
1238 struct ctdb_node_map *nodemap,
1244 if (ctdb->num_nodes != nodemap->num) {
1245 DEBUG(DEBUG_ERR, (__location__ " ctdb->num_nodes (%d) != nodemap->num (%d) invalid param\n",
1246 ctdb->num_nodes, nodemap->num));
1248 *culprit = ctdb->pnn;
1253 for (j=0; j<nodemap->num; j++) {
1254 /* release any existing data */
1255 if (ctdb->nodes[j]->known_public_ips) {
1256 talloc_free(ctdb->nodes[j]->known_public_ips);
1257 ctdb->nodes[j]->known_public_ips = NULL;
1259 if (ctdb->nodes[j]->available_public_ips) {
1260 talloc_free(ctdb->nodes[j]->available_public_ips);
1261 ctdb->nodes[j]->available_public_ips = NULL;
1264 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
1268 /* grab a new shiny list of public ips from the node */
1269 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1271 ctdb->nodes[j]->pnn,
1274 &ctdb->nodes[j]->known_public_ips);
1276 DEBUG(DEBUG_ERR,("Failed to read known public ips from node : %u\n",
1277 ctdb->nodes[j]->pnn));
1279 *culprit = ctdb->nodes[j]->pnn;
1284 if (ctdb->tunable.disable_ip_failover == 0) {
1285 if (rec->ip_check_disable_ctx == NULL) {
1286 if (verify_remote_ip_allocation(ctdb, ctdb->nodes[j]->known_public_ips)) {
1287 DEBUG(DEBUG_ERR,("Node %d has inconsistent public ip allocation and needs update.\n", ctdb->nodes[j]->pnn));
1288 rec->need_takeover_run = true;
1293 /* grab a new shiny list of public ips from the node */
1294 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1296 ctdb->nodes[j]->pnn,
1298 CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE,
1299 &ctdb->nodes[j]->available_public_ips);
1301 DEBUG(DEBUG_ERR,("Failed to read available public ips from node : %u\n",
1302 ctdb->nodes[j]->pnn));
1304 *culprit = ctdb->nodes[j]->pnn;
1313 /* when we start a recovery, make sure all nodes use the same reclock file
1316 static int sync_recovery_lock_file_across_cluster(struct ctdb_recoverd *rec)
1318 struct ctdb_context *ctdb = rec->ctdb;
1319 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1323 if (ctdb->recovery_lock_file == NULL) {
1327 data.dsize = strlen(ctdb->recovery_lock_file) + 1;
1328 data.dptr = (uint8_t *)ctdb->recovery_lock_file;
1331 nodes = list_of_active_nodes(ctdb, rec->nodemap, tmp_ctx, true);
1332 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECLOCK_FILE,
1338 DEBUG(DEBUG_ERR, (__location__ " Failed to sync reclock file settings\n"));
1339 talloc_free(tmp_ctx);
1343 talloc_free(tmp_ctx);
1349 we are the recmaster, and recovery is needed - start a recovery run
1351 static int do_recovery(struct ctdb_recoverd *rec,
1352 TALLOC_CTX *mem_ctx, uint32_t pnn,
1353 struct ctdb_node_map *nodemap, struct ctdb_vnn_map *vnnmap)
1355 struct ctdb_context *ctdb = rec->ctdb;
1357 uint32_t generation;
1358 struct ctdb_dbid_map *dbmap;
1361 struct timeval start_time;
1362 uint32_t culprit = (uint32_t)-1;
1364 DEBUG(DEBUG_NOTICE, (__location__ " Starting do_recovery\n"));
1366 /* if recovery fails, force it again */
1367 rec->need_recovery = true;
1369 for (i=0; i<ctdb->num_nodes; i++) {
1370 struct ctdb_banning_state *ban_state;
1372 if (ctdb->nodes[i]->ban_state == NULL) {
1375 ban_state = (struct ctdb_banning_state *)ctdb->nodes[i]->ban_state;
1376 if (ban_state->count < 2*ctdb->num_nodes) {
1379 DEBUG(DEBUG_NOTICE,("Node %u has caused %u recoveries recently - banning it for %u seconds\n",
1380 ctdb->nodes[i]->pnn, ban_state->count,
1381 ctdb->tunable.recovery_ban_period));
1382 ctdb_ban_node(rec, ctdb->nodes[i]->pnn, ctdb->tunable.recovery_ban_period);
1383 ban_state->count = 0;
1387 if (ctdb->tunable.verify_recovery_lock != 0) {
1388 DEBUG(DEBUG_ERR,("Taking out recovery lock from recovery daemon\n"));
1389 start_time = timeval_current();
1390 if (!ctdb_recovery_lock(ctdb, true)) {
1391 DEBUG(DEBUG_ERR,("Unable to get recovery lock - aborting recovery "
1392 "and ban ourself for %u seconds\n",
1393 ctdb->tunable.recovery_ban_period));
1394 ctdb_ban_node(rec, pnn, ctdb->tunable.recovery_ban_period);
1397 ctdb_ctrl_report_recd_lock_latency(ctdb, CONTROL_TIMEOUT(), timeval_elapsed(&start_time));
1398 DEBUG(DEBUG_NOTICE,("Recovery lock taken successfully by recovery daemon\n"));
1401 DEBUG(DEBUG_NOTICE, (__location__ " Recovery initiated due to problem with node %u\n", rec->last_culprit_node));
1403 /* get a list of all databases */
1404 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, &dbmap);
1406 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node :%u\n", pnn));
1410 /* we do the db creation before we set the recovery mode, so the freeze happens
1411 on all databases we will be dealing with. */
1413 /* verify that we have all the databases any other node has */
1414 ret = create_missing_local_databases(ctdb, nodemap, pnn, &dbmap, mem_ctx);
1416 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing local databases\n"));
1420 /* verify that all other nodes have all our databases */
1421 ret = create_missing_remote_databases(ctdb, nodemap, pnn, dbmap, mem_ctx);
1423 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing remote databases\n"));
1426 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - created remote databases\n"));
1428 /* update the database priority for all remote databases */
1429 ret = update_db_priority_on_remote_nodes(ctdb, nodemap, pnn, dbmap, mem_ctx);
1431 DEBUG(DEBUG_ERR, (__location__ " Unable to set db priority on remote nodes\n"));
1433 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated db priority for all databases\n"));
1436 /* update all other nodes to use the same setting for reclock files
1437 as the local recovery master.
1439 sync_recovery_lock_file_across_cluster(rec);
1441 /* set recovery mode to active on all nodes */
1442 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
1444 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
1448 /* execute the "startrecovery" event script on all nodes */
1449 ret = run_startrecovery_eventscript(rec, nodemap);
1451 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
1456 update all nodes to have the same flags that we have
1458 for (i=0;i<nodemap->num;i++) {
1459 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1463 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1465 DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1470 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1472 /* pick a new generation number */
1473 generation = new_generation();
1475 /* change the vnnmap on this node to use the new generation
1476 number but not on any other nodes.
1477 this guarantees that if we abort the recovery prematurely
1478 for some reason (a node stops responding?)
1479 that we can just return immediately and we will reenter
1480 recovery shortly again.
1481 I.e. we deliberately leave the cluster with an inconsistent
1482 generation id to allow us to abort recovery at any stage and
1483 just restart it from scratch.
1485 vnnmap->generation = generation;
1486 ret = ctdb_ctrl_setvnnmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, vnnmap);
1488 DEBUG(DEBUG_ERR, (__location__ " Unable to set vnnmap for node %u\n", pnn));
1492 data.dptr = (void *)&generation;
1493 data.dsize = sizeof(uint32_t);
1495 nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
1496 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_START,
1498 CONTROL_TIMEOUT(), false, data,
1500 transaction_start_fail_callback,
1502 DEBUG(DEBUG_ERR, (__location__ " Unable to start transactions. Recovery failed.\n"));
1503 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_CANCEL,
1505 CONTROL_TIMEOUT(), false, tdb_null,
1509 DEBUG(DEBUG_ERR,("Failed to cancel recovery transaction\n"));
1514 DEBUG(DEBUG_NOTICE,(__location__ " started transactions on all nodes\n"));
1516 for (i=0;i<dbmap->num;i++) {
1517 ret = recover_database(rec, mem_ctx,
1519 dbmap->dbs[i].flags & CTDB_DB_FLAGS_PERSISTENT,
1520 pnn, nodemap, generation);
1522 DEBUG(DEBUG_ERR, (__location__ " Failed to recover database 0x%x\n", dbmap->dbs[i].dbid));
1527 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - starting database commits\n"));
1529 /* commit all the changes */
1530 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_COMMIT,
1532 CONTROL_TIMEOUT(), false, data,
1535 DEBUG(DEBUG_ERR, (__location__ " Unable to commit recovery changes. Recovery failed.\n"));
1539 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - committed databases\n"));
1542 /* update the capabilities for all nodes */
1543 ret = update_capabilities(ctdb, nodemap);
1545 DEBUG(DEBUG_ERR, (__location__ " Unable to update node capabilities.\n"));
1549 /* build a new vnn map with all the currently active and
1551 generation = new_generation();
1552 vnnmap = talloc(mem_ctx, struct ctdb_vnn_map);
1553 CTDB_NO_MEMORY(ctdb, vnnmap);
1554 vnnmap->generation = generation;
1556 vnnmap->map = talloc_zero_array(vnnmap, uint32_t, vnnmap->size);
1557 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1558 for (i=j=0;i<nodemap->num;i++) {
1559 if (nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE) {
1562 if (!(ctdb->nodes[i]->capabilities & CTDB_CAP_LMASTER)) {
1563 /* this node can not be an lmaster */
1564 DEBUG(DEBUG_DEBUG, ("Node %d cant be a LMASTER, skipping it\n", i));
1569 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1570 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1571 vnnmap->map[j++] = nodemap->nodes[i].pnn;
1574 if (vnnmap->size == 0) {
1575 DEBUG(DEBUG_NOTICE, ("No suitable lmasters found. Adding local node (recmaster) anyway.\n"));
1577 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1578 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1579 vnnmap->map[0] = pnn;
1582 /* update to the new vnnmap on all nodes */
1583 ret = update_vnnmap_on_all_nodes(ctdb, nodemap, pnn, vnnmap, mem_ctx);
1585 DEBUG(DEBUG_ERR, (__location__ " Unable to update vnnmap on all nodes\n"));
1589 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated vnnmap\n"));
1591 /* update recmaster to point to us for all nodes */
1592 ret = set_recovery_master(ctdb, nodemap, pnn);
1594 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery master\n"));
1598 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated recmaster\n"));
1601 update all nodes to have the same flags that we have
1603 for (i=0;i<nodemap->num;i++) {
1604 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1608 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1610 DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1615 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1617 /* disable recovery mode */
1618 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_NORMAL);
1620 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to normal on cluster\n"));
1624 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - disabled recovery mode\n"));
1627 tell nodes to takeover their public IPs
1629 ret = ctdb_reload_remote_public_ips(ctdb, rec, nodemap, &culprit);
1631 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
1633 rec->need_takeover_run = true;
1636 rec->need_takeover_run = false;
1637 ret = ctdb_takeover_run(ctdb, nodemap);
1639 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses. ctdb_takeover_run() failed.\n"));
1640 rec->need_takeover_run = true;
1643 /* execute the "recovered" event script on all nodes */
1644 ret = run_recovered_eventscript(ctdb, nodemap, "do_recovery");
1646 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Recovery process failed.\n"));
1650 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - finished the recovered event\n"));
1652 /* send a message to all clients telling them that the cluster
1653 has been reconfigured */
1654 ctdb_client_send_message(ctdb, CTDB_BROADCAST_CONNECTED, CTDB_SRVID_RECONFIGURE, tdb_null);
1656 DEBUG(DEBUG_NOTICE, (__location__ " Recovery complete\n"));
1658 rec->need_recovery = false;
1660 /* we managed to complete a full recovery, make sure to forgive
1661 any past sins by the nodes that could now participate in the
1664 DEBUG(DEBUG_ERR,("Resetting ban count to 0 for all nodes\n"));
1665 for (i=0;i<nodemap->num;i++) {
1666 struct ctdb_banning_state *ban_state;
1668 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1672 ban_state = (struct ctdb_banning_state *)ctdb->nodes[nodemap->nodes[i].pnn]->ban_state;
1673 if (ban_state == NULL) {
1677 ban_state->count = 0;
1681 /* We just finished a recovery successfully.
1682 We now wait for rerecovery_timeout before we allow
1683 another recovery to take place.
1685 DEBUG(DEBUG_NOTICE, ("Just finished a recovery. New recoveries will now be supressed for the rerecovery timeout (%d seconds)\n", ctdb->tunable.rerecovery_timeout));
1686 ctdb_wait_timeout(ctdb, ctdb->tunable.rerecovery_timeout);
1687 DEBUG(DEBUG_NOTICE, ("The rerecovery timeout has elapsed. We now allow recoveries to trigger again.\n"));
1694 elections are won by first checking the number of connected nodes, then
1695 the priority time, then the pnn
1697 struct election_message {
1698 uint32_t num_connected;
1699 struct timeval priority_time;
1701 uint32_t node_flags;
1705 form this nodes election data
1707 static void ctdb_election_data(struct ctdb_recoverd *rec, struct election_message *em)
1710 struct ctdb_node_map *nodemap;
1711 struct ctdb_context *ctdb = rec->ctdb;
1715 em->pnn = rec->ctdb->pnn;
1716 em->priority_time = rec->priority_time;
1718 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, rec, &nodemap);
1720 DEBUG(DEBUG_ERR,(__location__ " unable to get election data\n"));
1724 rec->node_flags = nodemap->nodes[ctdb->pnn].flags;
1725 em->node_flags = rec->node_flags;
1727 for (i=0;i<nodemap->num;i++) {
1728 if (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED)) {
1729 em->num_connected++;
1733 /* we shouldnt try to win this election if we cant be a recmaster */
1734 if ((ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1735 em->num_connected = 0;
1736 em->priority_time = timeval_current();
1739 talloc_free(nodemap);
1743 see if the given election data wins
1745 static bool ctdb_election_win(struct ctdb_recoverd *rec, struct election_message *em)
1747 struct election_message myem;
1750 ctdb_election_data(rec, &myem);
1752 /* we cant win if we dont have the recmaster capability */
1753 if ((rec->ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1757 /* we cant win if we are banned */
1758 if (rec->node_flags & NODE_FLAGS_BANNED) {
1762 /* we cant win if we are stopped */
1763 if (rec->node_flags & NODE_FLAGS_STOPPED) {
1767 /* we will automatically win if the other node is banned */
1768 if (em->node_flags & NODE_FLAGS_BANNED) {
1772 /* we will automatically win if the other node is banned */
1773 if (em->node_flags & NODE_FLAGS_STOPPED) {
1777 /* try to use the most connected node */
1779 cmp = (int)myem.num_connected - (int)em->num_connected;
1782 /* then the longest running node */
1784 cmp = timeval_compare(&em->priority_time, &myem.priority_time);
1788 cmp = (int)myem.pnn - (int)em->pnn;
1795 send out an election request
1797 static int send_election_request(struct ctdb_recoverd *rec, uint32_t pnn, bool update_recmaster)
1800 TDB_DATA election_data;
1801 struct election_message emsg;
1803 struct ctdb_context *ctdb = rec->ctdb;
1805 srvid = CTDB_SRVID_RECOVERY;
1807 ctdb_election_data(rec, &emsg);
1809 election_data.dsize = sizeof(struct election_message);
1810 election_data.dptr = (unsigned char *)&emsg;
1813 /* send an election message to all active nodes */
1814 DEBUG(DEBUG_INFO,(__location__ " Send election request to all active nodes\n"));
1815 ctdb_client_send_message(ctdb, CTDB_BROADCAST_ALL, srvid, election_data);
1818 /* A new node that is already frozen has entered the cluster.
1819 The existing nodes are not frozen and dont need to be frozen
1820 until the election has ended and we start the actual recovery
1822 if (update_recmaster == true) {
1823 /* first we assume we will win the election and set
1824 recoverymaster to be ourself on the current node
1826 ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), pnn, pnn);
1828 DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request\n"));
1838 this function will unban all nodes in the cluster
1840 static void unban_all_nodes(struct ctdb_context *ctdb)
1843 struct ctdb_node_map *nodemap;
1844 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
1846 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
1848 DEBUG(DEBUG_ERR,(__location__ " failed to get nodemap to unban all nodes\n"));
1852 for (i=0;i<nodemap->num;i++) {
1853 if ( (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED))
1854 && (nodemap->nodes[i].flags & NODE_FLAGS_BANNED) ) {
1855 ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[i].pnn, 0, NODE_FLAGS_BANNED);
1859 talloc_free(tmp_ctx);
1864 we think we are winning the election - send a broadcast election request
1866 static void election_send_request(struct event_context *ev, struct timed_event *te, struct timeval t, void *p)
1868 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
1871 ret = send_election_request(rec, ctdb_get_pnn(rec->ctdb), false);
1873 DEBUG(DEBUG_ERR,("Failed to send election request!\n"));
1876 talloc_free(rec->send_election_te);
1877 rec->send_election_te = NULL;
1881 handler for memory dumps
1883 static void mem_dump_handler(struct ctdb_context *ctdb, uint64_t srvid,
1884 TDB_DATA data, void *private_data)
1886 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
1889 struct rd_memdump_reply *rd;
1891 if (data.dsize != sizeof(struct rd_memdump_reply)) {
1892 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
1893 talloc_free(tmp_ctx);
1896 rd = (struct rd_memdump_reply *)data.dptr;
1898 dump = talloc_zero(tmp_ctx, TDB_DATA);
1900 DEBUG(DEBUG_ERR, (__location__ " Failed to allocate memory for memdump\n"));
1901 talloc_free(tmp_ctx);
1904 ret = ctdb_dump_memory(ctdb, dump);
1906 DEBUG(DEBUG_ERR, (__location__ " ctdb_dump_memory() failed\n"));
1907 talloc_free(tmp_ctx);
1911 DEBUG(DEBUG_ERR, ("recovery master memory dump\n"));
1913 ret = ctdb_client_send_message(ctdb, rd->pnn, rd->srvid, *dump);
1915 DEBUG(DEBUG_ERR,("Failed to send rd memdump reply message\n"));
1916 talloc_free(tmp_ctx);
1920 talloc_free(tmp_ctx);
1924 handler for reload_nodes
1926 static void reload_nodes_handler(struct ctdb_context *ctdb, uint64_t srvid,
1927 TDB_DATA data, void *private_data)
1929 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1931 DEBUG(DEBUG_ERR, (__location__ " Reload nodes file from recovery daemon\n"));
1933 reload_nodes_file(rec->ctdb);
1937 static void reenable_ip_check(struct event_context *ev, struct timed_event *te,
1938 struct timeval yt, void *p)
1940 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
1942 talloc_free(rec->ip_check_disable_ctx);
1943 rec->ip_check_disable_ctx = NULL;
1947 static void recd_update_ip_handler(struct ctdb_context *ctdb, uint64_t srvid,
1948 TDB_DATA data, void *private_data)
1950 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1951 struct ctdb_public_ip *ip;
1953 if (rec->recmaster != rec->ctdb->pnn) {
1954 DEBUG(DEBUG_INFO,("Not recmaster, ignore update ip message\n"));
1958 if (data.dsize != sizeof(struct ctdb_public_ip)) {
1959 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)));
1963 ip = (struct ctdb_public_ip *)data.dptr;
1965 update_ip_assignment_tree(rec->ctdb, ip);
1969 static void disable_ip_check_handler(struct ctdb_context *ctdb, uint64_t srvid,
1970 TDB_DATA data, void *private_data)
1972 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1975 if (rec->ip_check_disable_ctx != NULL) {
1976 talloc_free(rec->ip_check_disable_ctx);
1977 rec->ip_check_disable_ctx = NULL;
1980 if (data.dsize != sizeof(uint32_t)) {
1981 DEBUG(DEBUG_ERR,(__location__ " Wrong size for data :%lu "
1982 "expexting %lu\n", (long unsigned)data.dsize,
1983 (long unsigned)sizeof(uint32_t)));
1986 if (data.dptr == NULL) {
1987 DEBUG(DEBUG_ERR,(__location__ " No data recaived\n"));
1991 timeout = *((uint32_t *)data.dptr);
1992 DEBUG(DEBUG_NOTICE,("Disabling ip check for %u seconds\n", timeout));
1994 rec->ip_check_disable_ctx = talloc_new(rec);
1995 CTDB_NO_MEMORY_VOID(ctdb, rec->ip_check_disable_ctx);
1997 event_add_timed(ctdb->ev, rec->ip_check_disable_ctx, timeval_current_ofs(timeout, 0), reenable_ip_check, rec);
2002 handler for ip reallocate, just add it to the list of callers and
2003 handle this later in the monitor_cluster loop so we do not recurse
2004 with other callers to takeover_run()
2006 static void ip_reallocate_handler(struct ctdb_context *ctdb, uint64_t srvid,
2007 TDB_DATA data, void *private_data)
2009 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2010 struct ip_reallocate_list *caller;
2012 if (data.dsize != sizeof(struct rd_memdump_reply)) {
2013 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
2017 if (rec->ip_reallocate_ctx == NULL) {
2018 rec->ip_reallocate_ctx = talloc_new(rec);
2019 CTDB_NO_MEMORY_FATAL(ctdb, rec->ip_reallocate_ctx);
2022 caller = talloc(rec->ip_reallocate_ctx, struct ip_reallocate_list);
2023 CTDB_NO_MEMORY_FATAL(ctdb, caller);
2025 caller->rd = (struct rd_memdump_reply *)talloc_steal(caller, data.dptr);
2026 caller->next = rec->reallocate_callers;
2027 rec->reallocate_callers = caller;
2032 static void process_ipreallocate_requests(struct ctdb_context *ctdb, struct ctdb_recoverd *rec)
2034 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2037 struct ip_reallocate_list *callers;
2040 DEBUG(DEBUG_INFO, ("recovery master forced ip reallocation\n"));
2042 /* update the list of public ips that a node can handle for
2045 ret = ctdb_reload_remote_public_ips(ctdb, rec, rec->nodemap, &culprit);
2047 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
2049 rec->need_takeover_run = true;
2052 ret = ctdb_takeover_run(ctdb, rec->nodemap);
2054 DEBUG(DEBUG_ERR,("Failed to reallocate addresses: ctdb_takeover_run() failed.\n"));
2055 rec->need_takeover_run = true;
2059 result.dsize = sizeof(int32_t);
2060 result.dptr = (uint8_t *)&ret;
2062 for (callers=rec->reallocate_callers; callers; callers=callers->next) {
2064 /* Someone that sent srvid==0 does not want a reply */
2065 if (callers->rd->srvid == 0) {
2068 DEBUG(DEBUG_INFO,("Sending ip reallocate reply message to "
2069 "%u:%llu\n", (unsigned)callers->rd->pnn,
2070 (unsigned long long)callers->rd->srvid));
2071 ret = ctdb_client_send_message(ctdb, callers->rd->pnn, callers->rd->srvid, result);
2073 DEBUG(DEBUG_ERR,("Failed to send ip reallocate reply "
2074 "message to %u:%llu\n",
2075 (unsigned)callers->rd->pnn,
2076 (unsigned long long)callers->rd->srvid));
2080 talloc_free(tmp_ctx);
2081 talloc_free(rec->ip_reallocate_ctx);
2082 rec->ip_reallocate_ctx = NULL;
2083 rec->reallocate_callers = NULL;
2089 handler for recovery master elections
2091 static void election_handler(struct ctdb_context *ctdb, uint64_t srvid,
2092 TDB_DATA data, void *private_data)
2094 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2096 struct election_message *em = (struct election_message *)data.dptr;
2097 TALLOC_CTX *mem_ctx;
2099 /* we got an election packet - update the timeout for the election */
2100 talloc_free(rec->election_timeout);
2101 rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2103 timeval_current_ofs(0, 500000) :
2104 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2105 ctdb_election_timeout, rec);
2107 mem_ctx = talloc_new(ctdb);
2109 /* someone called an election. check their election data
2110 and if we disagree and we would rather be the elected node,
2111 send a new election message to all other nodes
2113 if (ctdb_election_win(rec, em)) {
2114 if (!rec->send_election_te) {
2115 rec->send_election_te = event_add_timed(ctdb->ev, rec,
2116 timeval_current_ofs(0, 500000),
2117 election_send_request, rec);
2119 talloc_free(mem_ctx);
2120 /*unban_all_nodes(ctdb);*/
2125 talloc_free(rec->send_election_te);
2126 rec->send_election_te = NULL;
2128 if (ctdb->tunable.verify_recovery_lock != 0) {
2129 /* release the recmaster lock */
2130 if (em->pnn != ctdb->pnn &&
2131 ctdb->recovery_lock_fd != -1) {
2132 close(ctdb->recovery_lock_fd);
2133 ctdb->recovery_lock_fd = -1;
2134 unban_all_nodes(ctdb);
2138 /* ok, let that guy become recmaster then */
2139 ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), ctdb_get_pnn(ctdb), em->pnn);
2141 DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request"));
2142 talloc_free(mem_ctx);
2146 talloc_free(mem_ctx);
2152 force the start of the election process
2154 static void force_election(struct ctdb_recoverd *rec, uint32_t pnn,
2155 struct ctdb_node_map *nodemap)
2158 struct ctdb_context *ctdb = rec->ctdb;
2160 DEBUG(DEBUG_INFO,(__location__ " Force an election\n"));
2162 /* set all nodes to recovery mode to stop all internode traffic */
2163 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
2165 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
2169 talloc_free(rec->election_timeout);
2170 rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2172 timeval_current_ofs(0, 500000) :
2173 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2174 ctdb_election_timeout, rec);
2176 ret = send_election_request(rec, pnn, true);
2178 DEBUG(DEBUG_ERR, (__location__ " failed to initiate recmaster election"));
2182 /* wait for a few seconds to collect all responses */
2183 ctdb_wait_election(rec);
2189 handler for when a node changes its flags
2191 static void monitor_handler(struct ctdb_context *ctdb, uint64_t srvid,
2192 TDB_DATA data, void *private_data)
2195 struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
2196 struct ctdb_node_map *nodemap=NULL;
2197 TALLOC_CTX *tmp_ctx;
2199 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2200 int disabled_flag_changed;
2202 if (data.dsize != sizeof(*c)) {
2203 DEBUG(DEBUG_ERR,(__location__ "Invalid data in ctdb_node_flag_change\n"));
2207 tmp_ctx = talloc_new(ctdb);
2208 CTDB_NO_MEMORY_VOID(ctdb, tmp_ctx);
2210 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
2212 DEBUG(DEBUG_ERR,(__location__ "ctdb_ctrl_getnodemap failed in monitor_handler\n"));
2213 talloc_free(tmp_ctx);
2218 for (i=0;i<nodemap->num;i++) {
2219 if (nodemap->nodes[i].pnn == c->pnn) break;
2222 if (i == nodemap->num) {
2223 DEBUG(DEBUG_CRIT,(__location__ "Flag change for non-existant node %u\n", c->pnn));
2224 talloc_free(tmp_ctx);
2228 if (nodemap->nodes[i].flags != c->new_flags) {
2229 DEBUG(DEBUG_NOTICE,("Node %u has changed flags - now 0x%x was 0x%x\n", c->pnn, c->new_flags, c->old_flags));
2232 disabled_flag_changed = (nodemap->nodes[i].flags ^ c->new_flags) & NODE_FLAGS_DISABLED;
2234 nodemap->nodes[i].flags = c->new_flags;
2236 ret = ctdb_ctrl_getrecmaster(ctdb, tmp_ctx, CONTROL_TIMEOUT(),
2237 CTDB_CURRENT_NODE, &ctdb->recovery_master);
2240 ret = ctdb_ctrl_getrecmode(ctdb, tmp_ctx, CONTROL_TIMEOUT(),
2241 CTDB_CURRENT_NODE, &ctdb->recovery_mode);
2245 ctdb->recovery_master == ctdb->pnn &&
2246 ctdb->recovery_mode == CTDB_RECOVERY_NORMAL) {
2247 /* Only do the takeover run if the perm disabled or unhealthy
2248 flags changed since these will cause an ip failover but not
2250 If the node became disconnected or banned this will also
2251 lead to an ip address failover but that is handled
2254 if (disabled_flag_changed) {
2255 rec->need_takeover_run = true;
2259 talloc_free(tmp_ctx);
2263 handler for when we need to push out flag changes ot all other nodes
2265 static void push_flags_handler(struct ctdb_context *ctdb, uint64_t srvid,
2266 TDB_DATA data, void *private_data)
2269 struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
2270 struct ctdb_node_map *nodemap=NULL;
2271 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2275 /* find the recovery master */
2276 ret = ctdb_ctrl_getrecmaster(ctdb, tmp_ctx, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &recmaster);
2278 DEBUG(DEBUG_ERR, (__location__ " Unable to get recmaster from local node\n"));
2279 talloc_free(tmp_ctx);
2283 /* read the node flags from the recmaster */
2284 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), recmaster, tmp_ctx, &nodemap);
2286 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from node %u\n", c->pnn));
2287 talloc_free(tmp_ctx);
2290 if (c->pnn >= nodemap->num) {
2291 DEBUG(DEBUG_ERR,(__location__ " Nodemap from recmaster does not contain node %d\n", c->pnn));
2292 talloc_free(tmp_ctx);
2296 /* send the flags update to all connected nodes */
2297 nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
2299 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_MODIFY_FLAGS,
2300 nodes, 0, CONTROL_TIMEOUT(),
2304 DEBUG(DEBUG_ERR, (__location__ " ctdb_control to modify node flags failed\n"));
2306 talloc_free(tmp_ctx);
2310 talloc_free(tmp_ctx);
2314 struct verify_recmode_normal_data {
2316 enum monitor_result status;
2319 static void verify_recmode_normal_callback(struct ctdb_client_control_state *state)
2321 struct verify_recmode_normal_data *rmdata = talloc_get_type(state->async.private_data, struct verify_recmode_normal_data);
2324 /* one more node has responded with recmode data*/
2327 /* if we failed to get the recmode, then return an error and let
2328 the main loop try again.
2330 if (state->state != CTDB_CONTROL_DONE) {
2331 if (rmdata->status == MONITOR_OK) {
2332 rmdata->status = MONITOR_FAILED;
2337 /* if we got a response, then the recmode will be stored in the
2340 if (state->status != CTDB_RECOVERY_NORMAL) {
2341 DEBUG(DEBUG_NOTICE, (__location__ " Node:%u was in recovery mode. Restart recovery process\n", state->c->hdr.destnode));
2342 rmdata->status = MONITOR_RECOVERY_NEEDED;
2349 /* verify that all nodes are in normal recovery mode */
2350 static enum monitor_result verify_recmode(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
2352 struct verify_recmode_normal_data *rmdata;
2353 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
2354 struct ctdb_client_control_state *state;
2355 enum monitor_result status;
2358 rmdata = talloc(mem_ctx, struct verify_recmode_normal_data);
2359 CTDB_NO_MEMORY_FATAL(ctdb, rmdata);
2361 rmdata->status = MONITOR_OK;
2363 /* loop over all active nodes and send an async getrecmode call to
2365 for (j=0; j<nodemap->num; j++) {
2366 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
2369 state = ctdb_ctrl_getrecmode_send(ctdb, mem_ctx,
2371 nodemap->nodes[j].pnn);
2372 if (state == NULL) {
2373 /* we failed to send the control, treat this as
2374 an error and try again next iteration
2376 DEBUG(DEBUG_ERR,("Failed to call ctdb_ctrl_getrecmode_send during monitoring\n"));
2377 talloc_free(mem_ctx);
2378 return MONITOR_FAILED;
2381 /* set up the callback functions */
2382 state->async.fn = verify_recmode_normal_callback;
2383 state->async.private_data = rmdata;
2385 /* one more control to wait for to complete */
2390 /* now wait for up to the maximum number of seconds allowed
2391 or until all nodes we expect a response from has replied
2393 while (rmdata->count > 0) {
2394 event_loop_once(ctdb->ev);
2397 status = rmdata->status;
2398 talloc_free(mem_ctx);
2403 struct verify_recmaster_data {
2404 struct ctdb_recoverd *rec;
2407 enum monitor_result status;
2410 static void verify_recmaster_callback(struct ctdb_client_control_state *state)
2412 struct verify_recmaster_data *rmdata = talloc_get_type(state->async.private_data, struct verify_recmaster_data);
2415 /* one more node has responded with recmaster data*/
2418 /* if we failed to get the recmaster, then return an error and let
2419 the main loop try again.
2421 if (state->state != CTDB_CONTROL_DONE) {
2422 if (rmdata->status == MONITOR_OK) {
2423 rmdata->status = MONITOR_FAILED;
2428 /* if we got a response, then the recmaster will be stored in the
2431 if (state->status != rmdata->pnn) {
2432 DEBUG(DEBUG_ERR,("Node %d does not agree we are the recmaster. Need a new recmaster election\n", state->c->hdr.destnode));
2433 ctdb_set_culprit(rmdata->rec, state->c->hdr.destnode);
2434 rmdata->status = MONITOR_ELECTION_NEEDED;
2441 /* verify that all nodes agree that we are the recmaster */
2442 static enum monitor_result verify_recmaster(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, uint32_t pnn)
2444 struct ctdb_context *ctdb = rec->ctdb;
2445 struct verify_recmaster_data *rmdata;
2446 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
2447 struct ctdb_client_control_state *state;
2448 enum monitor_result status;
2451 rmdata = talloc(mem_ctx, struct verify_recmaster_data);
2452 CTDB_NO_MEMORY_FATAL(ctdb, rmdata);
2456 rmdata->status = MONITOR_OK;
2458 /* loop over all active nodes and send an async getrecmaster call to
2460 for (j=0; j<nodemap->num; j++) {
2461 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
2464 state = ctdb_ctrl_getrecmaster_send(ctdb, mem_ctx,
2466 nodemap->nodes[j].pnn);
2467 if (state == NULL) {
2468 /* we failed to send the control, treat this as
2469 an error and try again next iteration
2471 DEBUG(DEBUG_ERR,("Failed to call ctdb_ctrl_getrecmaster_send during monitoring\n"));
2472 talloc_free(mem_ctx);
2473 return MONITOR_FAILED;
2476 /* set up the callback functions */
2477 state->async.fn = verify_recmaster_callback;
2478 state->async.private_data = rmdata;
2480 /* one more control to wait for to complete */
2485 /* now wait for up to the maximum number of seconds allowed
2486 or until all nodes we expect a response from has replied
2488 while (rmdata->count > 0) {
2489 event_loop_once(ctdb->ev);
2492 status = rmdata->status;
2493 talloc_free(mem_ctx);
2498 /* called to check that the local allocation of public ip addresses is ok.
2500 static int verify_local_ip_allocation(struct ctdb_context *ctdb, struct ctdb_recoverd *rec, uint32_t pnn, struct ctdb_node_map *nodemap)
2502 TALLOC_CTX *mem_ctx = talloc_new(NULL);
2503 struct ctdb_control_get_ifaces *ifaces = NULL;
2504 struct ctdb_all_public_ips *ips = NULL;
2505 struct ctdb_uptime *uptime1 = NULL;
2506 struct ctdb_uptime *uptime2 = NULL;
2508 bool need_iface_check = false;
2509 bool need_takeover_run = false;
2511 ret = ctdb_ctrl_uptime(ctdb, mem_ctx, CONTROL_TIMEOUT(),
2512 CTDB_CURRENT_NODE, &uptime1);
2514 DEBUG(DEBUG_ERR, ("Unable to get uptime from local node %u\n", pnn));
2515 talloc_free(mem_ctx);
2520 /* read the interfaces from the local node */
2521 ret = ctdb_ctrl_get_ifaces(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, mem_ctx, &ifaces);
2523 DEBUG(DEBUG_ERR, ("Unable to get interfaces from local node %u\n", pnn));
2524 talloc_free(mem_ctx);
2529 need_iface_check = true;
2530 } else if (rec->ifaces->num != ifaces->num) {
2531 need_iface_check = true;
2532 } else if (memcmp(rec->ifaces, ifaces, talloc_get_size(ifaces)) != 0) {
2533 need_iface_check = true;
2536 if (need_iface_check) {
2537 DEBUG(DEBUG_NOTICE, ("The interfaces status has changed on "
2538 "local node %u - force takeover run\n",
2540 need_takeover_run = true;
2543 /* read the ip allocation from the local node */
2544 ret = ctdb_ctrl_get_public_ips(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, mem_ctx, &ips);
2546 DEBUG(DEBUG_ERR, ("Unable to get public ips from local node %u\n", pnn));
2547 talloc_free(mem_ctx);
2551 ret = ctdb_ctrl_uptime(ctdb, mem_ctx, CONTROL_TIMEOUT(),
2552 CTDB_CURRENT_NODE, &uptime2);
2554 DEBUG(DEBUG_ERR, ("Unable to get uptime from local node %u\n", pnn));
2555 talloc_free(mem_ctx);
2559 /* skip the check if the startrecovery time has changed */
2560 if (timeval_compare(&uptime1->last_recovery_started,
2561 &uptime2->last_recovery_started) != 0) {
2562 DEBUG(DEBUG_NOTICE, (__location__ " last recovery time changed while we read the public ip list. skipping public ip address check\n"));
2563 talloc_free(mem_ctx);
2567 /* skip the check if the endrecovery time has changed */
2568 if (timeval_compare(&uptime1->last_recovery_finished,
2569 &uptime2->last_recovery_finished) != 0) {
2570 DEBUG(DEBUG_NOTICE, (__location__ " last recovery time changed while we read the public ip list. skipping public ip address check\n"));
2571 talloc_free(mem_ctx);
2575 /* skip the check if we have started but not finished recovery */
2576 if (timeval_compare(&uptime1->last_recovery_finished,
2577 &uptime1->last_recovery_started) != 1) {
2578 DEBUG(DEBUG_INFO, (__location__ " in the middle of recovery or ip reallocation. skipping public ip address check\n"));
2579 talloc_free(mem_ctx);
2584 talloc_free(rec->ifaces);
2585 rec->ifaces = talloc_steal(rec, ifaces);
2587 /* verify that we have the ip addresses we should have
2588 and we dont have ones we shouldnt have.
2589 if we find an inconsistency we set recmode to
2590 active on the local node and wait for the recmaster
2591 to do a full blown recovery.
2592 also if the pnn is -1 and we are healthy and can host the ip
2593 we also request a ip reallocation.
2595 if (ctdb->tunable.disable_ip_failover == 0) {
2596 for (j=0; j<ips->num; j++) {
2597 if (ips->ips[j].pnn == -1 && nodemap->nodes[pnn].flags == 0) {
2598 DEBUG(DEBUG_CRIT,("Public address '%s' is not assigned and we could serve this ip\n",
2599 ctdb_addr_to_str(&ips->ips[j].addr)));
2600 need_takeover_run = true;
2601 } else if (ips->ips[j].pnn == pnn) {
2602 if (!ctdb_sys_have_ip(&ips->ips[j].addr)) {
2603 DEBUG(DEBUG_CRIT,("Public address '%s' is missing and we should serve this ip\n",
2604 ctdb_addr_to_str(&ips->ips[j].addr)));
2605 need_takeover_run = true;
2608 if (ctdb_sys_have_ip(&ips->ips[j].addr)) {
2609 DEBUG(DEBUG_CRIT,("We are still serving a public address '%s' that we should not be serving.\n",
2610 ctdb_addr_to_str(&ips->ips[j].addr)));
2611 need_takeover_run = true;
2617 if (need_takeover_run) {
2618 struct takeover_run_reply rd;
2621 DEBUG(DEBUG_CRIT,("Trigger takeoverrun\n"));
2625 data.dptr = (uint8_t *)&rd;
2626 data.dsize = sizeof(rd);
2628 ret = ctdb_client_send_message(ctdb, rec->recmaster, CTDB_SRVID_TAKEOVER_RUN, data);
2630 DEBUG(DEBUG_ERR,(__location__ " Failed to send ipreallocate to recmaster :%d\n", (int)rec->recmaster));
2633 talloc_free(mem_ctx);
2638 static void async_getnodemap_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
2640 struct ctdb_node_map **remote_nodemaps = callback_data;
2642 if (node_pnn >= ctdb->num_nodes) {
2643 DEBUG(DEBUG_ERR,(__location__ " pnn from invalid node\n"));
2647 remote_nodemaps[node_pnn] = (struct ctdb_node_map *)talloc_steal(remote_nodemaps, outdata.dptr);
2651 static int get_remote_nodemaps(struct ctdb_context *ctdb, TALLOC_CTX *mem_ctx,
2652 struct ctdb_node_map *nodemap,
2653 struct ctdb_node_map **remote_nodemaps)
2657 nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
2658 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_NODEMAP,
2660 CONTROL_TIMEOUT(), false, tdb_null,
2661 async_getnodemap_callback,
2663 remote_nodemaps) != 0) {
2664 DEBUG(DEBUG_ERR, (__location__ " Unable to pull all remote nodemaps\n"));
2672 enum reclock_child_status { RECLOCK_CHECKING, RECLOCK_OK, RECLOCK_FAILED, RECLOCK_TIMEOUT};
2673 struct ctdb_check_reclock_state {
2674 struct ctdb_context *ctdb;
2675 struct timeval start_time;
2678 struct timed_event *te;
2679 struct fd_event *fde;
2680 enum reclock_child_status status;
2683 /* when we free the reclock state we must kill any child process.
2685 static int check_reclock_destructor(struct ctdb_check_reclock_state *state)
2687 struct ctdb_context *ctdb = state->ctdb;
2689 ctdb_ctrl_report_recd_lock_latency(ctdb, CONTROL_TIMEOUT(), timeval_elapsed(&state->start_time));
2691 if (state->fd[0] != -1) {
2692 close(state->fd[0]);
2695 if (state->fd[1] != -1) {
2696 close(state->fd[1]);
2699 kill(state->child, SIGKILL);
2704 called if our check_reclock child times out. this would happen if
2705 i/o to the reclock file blocks.
2707 static void ctdb_check_reclock_timeout(struct event_context *ev, struct timed_event *te,
2708 struct timeval t, void *private_data)
2710 struct ctdb_check_reclock_state *state = talloc_get_type(private_data,
2711 struct ctdb_check_reclock_state);
2713 DEBUG(DEBUG_ERR,(__location__ " check_reclock child process hung/timedout CFS slow to grant locks?\n"));
2714 state->status = RECLOCK_TIMEOUT;
2717 /* this is called when the child process has completed checking the reclock
2718 file and has written data back to us through the pipe.
2720 static void reclock_child_handler(struct event_context *ev, struct fd_event *fde,
2721 uint16_t flags, void *private_data)
2723 struct ctdb_check_reclock_state *state= talloc_get_type(private_data,
2724 struct ctdb_check_reclock_state);
2728 /* we got a response from our child process so we can abort the
2731 talloc_free(state->te);
2734 ret = read(state->fd[0], &c, 1);
2735 if (ret != 1 || c != RECLOCK_OK) {
2736 DEBUG(DEBUG_ERR,(__location__ " reclock child process returned error %d\n", c));
2737 state->status = RECLOCK_FAILED;
2742 state->status = RECLOCK_OK;
2746 static int check_recovery_lock(struct ctdb_context *ctdb)
2749 struct ctdb_check_reclock_state *state;
2750 pid_t parent = getpid();
2752 if (ctdb->recovery_lock_fd == -1) {
2753 DEBUG(DEBUG_CRIT,("recovery master doesn't have the recovery lock\n"));
2757 state = talloc(ctdb, struct ctdb_check_reclock_state);
2758 CTDB_NO_MEMORY(ctdb, state);
2761 state->start_time = timeval_current();
2762 state->status = RECLOCK_CHECKING;
2766 ret = pipe(state->fd);
2769 DEBUG(DEBUG_CRIT,(__location__ " Failed to open pipe for check_reclock child\n"));
2773 state->child = ctdb_fork(ctdb);
2774 if (state->child == (pid_t)-1) {
2775 DEBUG(DEBUG_CRIT,(__location__ " fork() failed in check_reclock child\n"));
2776 close(state->fd[0]);
2778 close(state->fd[1]);
2784 if (state->child == 0) {
2785 char cc = RECLOCK_OK;
2786 close(state->fd[0]);
2789 debug_extra = talloc_asprintf(NULL, "recovery-lock:");
2790 if (pread(ctdb->recovery_lock_fd, &cc, 1, 0) == -1) {
2791 DEBUG(DEBUG_CRIT,("failed read from recovery_lock_fd - %s\n", strerror(errno)));
2792 cc = RECLOCK_FAILED;
2795 write(state->fd[1], &cc, 1);
2796 /* make sure we die when our parent dies */
2797 while (kill(parent, 0) == 0 || errno != ESRCH) {
2799 write(state->fd[1], &cc, 1);
2803 close(state->fd[1]);
2805 set_close_on_exec(state->fd[0]);
2807 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d for check_recovery_lock\n", state->fd[0]));
2809 talloc_set_destructor(state, check_reclock_destructor);
2811 state->te = event_add_timed(ctdb->ev, state, timeval_current_ofs(15, 0),
2812 ctdb_check_reclock_timeout, state);
2813 if (state->te == NULL) {
2814 DEBUG(DEBUG_CRIT,(__location__ " Failed to create a timed event for reclock child\n"));
2819 state->fde = event_add_fd(ctdb->ev, state, state->fd[0],
2821 reclock_child_handler,
2824 if (state->fde == NULL) {
2825 DEBUG(DEBUG_CRIT,(__location__ " Failed to create an fd event for reclock child\n"));
2829 tevent_fd_set_auto_close(state->fde);
2831 while (state->status == RECLOCK_CHECKING) {
2832 event_loop_once(ctdb->ev);
2835 if (state->status == RECLOCK_FAILED) {
2836 DEBUG(DEBUG_ERR,(__location__ " reclock child failed when checking file\n"));
2837 close(ctdb->recovery_lock_fd);
2838 ctdb->recovery_lock_fd = -1;
2847 static int update_recovery_lock_file(struct ctdb_context *ctdb)
2849 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
2850 const char *reclockfile;
2852 if (ctdb_ctrl_getreclock(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &reclockfile) != 0) {
2853 DEBUG(DEBUG_ERR,("Failed to read reclock file from daemon\n"));
2854 talloc_free(tmp_ctx);
2858 if (reclockfile == NULL) {
2859 if (ctdb->recovery_lock_file != NULL) {
2860 DEBUG(DEBUG_ERR,("Reclock file disabled\n"));
2861 talloc_free(ctdb->recovery_lock_file);
2862 ctdb->recovery_lock_file = NULL;
2863 if (ctdb->recovery_lock_fd != -1) {
2864 close(ctdb->recovery_lock_fd);
2865 ctdb->recovery_lock_fd = -1;
2868 ctdb->tunable.verify_recovery_lock = 0;
2869 talloc_free(tmp_ctx);
2873 if (ctdb->recovery_lock_file == NULL) {
2874 ctdb->recovery_lock_file = talloc_strdup(ctdb, reclockfile);
2875 if (ctdb->recovery_lock_fd != -1) {
2876 close(ctdb->recovery_lock_fd);
2877 ctdb->recovery_lock_fd = -1;
2879 talloc_free(tmp_ctx);
2884 if (!strcmp(reclockfile, ctdb->recovery_lock_file)) {
2885 talloc_free(tmp_ctx);
2889 talloc_free(ctdb->recovery_lock_file);
2890 ctdb->recovery_lock_file = talloc_strdup(ctdb, reclockfile);
2891 ctdb->tunable.verify_recovery_lock = 0;
2892 if (ctdb->recovery_lock_fd != -1) {
2893 close(ctdb->recovery_lock_fd);
2894 ctdb->recovery_lock_fd = -1;
2897 talloc_free(tmp_ctx);
2901 static void main_loop(struct ctdb_context *ctdb, struct ctdb_recoverd *rec,
2902 TALLOC_CTX *mem_ctx)
2905 struct ctdb_node_map *nodemap=NULL;
2906 struct ctdb_node_map *recmaster_nodemap=NULL;
2907 struct ctdb_node_map **remote_nodemaps=NULL;
2908 struct ctdb_vnn_map *vnnmap=NULL;
2909 struct ctdb_vnn_map *remote_vnnmap=NULL;
2910 int32_t debug_level;
2915 /* verify that the main daemon is still running */
2916 if (kill(ctdb->ctdbd_pid, 0) != 0) {
2917 DEBUG(DEBUG_CRIT,("CTDB daemon is no longer available. Shutting down recovery daemon\n"));
2921 /* ping the local daemon to tell it we are alive */
2922 ctdb_ctrl_recd_ping(ctdb);
2924 if (rec->election_timeout) {
2925 /* an election is in progress */
2929 /* read the debug level from the parent and update locally */
2930 ret = ctdb_ctrl_get_debuglevel(ctdb, CTDB_CURRENT_NODE, &debug_level);
2932 DEBUG(DEBUG_ERR, (__location__ " Failed to read debuglevel from parent\n"));
2935 LogLevel = debug_level;
2938 /* We must check if we need to ban a node here but we want to do this
2939 as early as possible so we dont wait until we have pulled the node
2940 map from the local node. thats why we have the hardcoded value 20
2942 for (i=0; i<ctdb->num_nodes; i++) {
2943 struct ctdb_banning_state *ban_state;
2945 if (ctdb->nodes[i]->ban_state == NULL) {
2948 ban_state = (struct ctdb_banning_state *)ctdb->nodes[i]->ban_state;
2949 if (ban_state->count < 20) {
2952 DEBUG(DEBUG_NOTICE,("Node %u has caused %u recoveries recently - banning it for %u seconds\n",
2953 ctdb->nodes[i]->pnn, ban_state->count,
2954 ctdb->tunable.recovery_ban_period));
2955 ctdb_ban_node(rec, ctdb->nodes[i]->pnn, ctdb->tunable.recovery_ban_period);
2956 ban_state->count = 0;
2959 /* get relevant tunables */
2960 ret = ctdb_ctrl_get_all_tunables(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &ctdb->tunable);
2962 DEBUG(DEBUG_ERR,("Failed to get tunables - retrying\n"));
2966 /* get the current recovery lock file from the server */
2967 if (update_recovery_lock_file(ctdb) != 0) {
2968 DEBUG(DEBUG_ERR,("Failed to update the recovery lock file\n"));
2972 /* Make sure that if recovery lock verification becomes disabled when
2975 if (ctdb->tunable.verify_recovery_lock == 0) {
2976 if (ctdb->recovery_lock_fd != -1) {
2977 close(ctdb->recovery_lock_fd);
2978 ctdb->recovery_lock_fd = -1;
2982 pnn = ctdb_ctrl_getpnn(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE);
2983 if (pnn == (uint32_t)-1) {
2984 DEBUG(DEBUG_ERR,("Failed to get local pnn - retrying\n"));
2988 /* get the vnnmap */
2989 ret = ctdb_ctrl_getvnnmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, &vnnmap);
2991 DEBUG(DEBUG_ERR, (__location__ " Unable to get vnnmap from node %u\n", pnn));
2996 /* get number of nodes */
2998 talloc_free(rec->nodemap);
2999 rec->nodemap = NULL;
3002 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), pnn, rec, &rec->nodemap);
3004 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from node %u\n", pnn));
3007 nodemap = rec->nodemap;
3009 /* check which node is the recovery master */
3010 ret = ctdb_ctrl_getrecmaster(ctdb, mem_ctx, CONTROL_TIMEOUT(), pnn, &rec->recmaster);
3012 DEBUG(DEBUG_ERR, (__location__ " Unable to get recmaster from node %u\n", pnn));
3016 /* if we are not the recmaster we can safely ignore any ip reallocate requests */
3017 if (rec->recmaster != pnn) {
3018 if (rec->ip_reallocate_ctx != NULL) {
3019 talloc_free(rec->ip_reallocate_ctx);
3020 rec->ip_reallocate_ctx = NULL;
3021 rec->reallocate_callers = NULL;
3025 if (rec->recmaster == (uint32_t)-1) {
3026 DEBUG(DEBUG_NOTICE,(__location__ " Initial recovery master set - forcing election\n"));
3027 force_election(rec, pnn, nodemap);
3032 /* if the local daemon is STOPPED, we verify that the databases are
3033 also frozen and thet the recmode is set to active
3035 if (nodemap->nodes[pnn].flags & NODE_FLAGS_STOPPED) {
3036 ret = ctdb_ctrl_getrecmode(ctdb, mem_ctx, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &ctdb->recovery_mode);
3038 DEBUG(DEBUG_ERR,(__location__ " Failed to read recmode from local node\n"));
3040 if (ctdb->recovery_mode == CTDB_RECOVERY_NORMAL) {
3041 DEBUG(DEBUG_ERR,("Node is stopped but recovery mode is not active. Activate recovery mode and lock databases\n"));
3043 ret = ctdb_ctrl_freeze_priority(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, 1);
3045 DEBUG(DEBUG_ERR,(__location__ " Failed to freeze node due to node being STOPPED\n"));
3048 ret = ctdb_ctrl_setrecmode(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, CTDB_RECOVERY_ACTIVE);
3050 DEBUG(DEBUG_ERR,(__location__ " Failed to activate recovery mode due to node being stopped\n"));
3057 /* If the local node is stopped, verify we are not the recmaster
3058 and yield this role if so
3060 if ((nodemap->nodes[pnn].flags & NODE_FLAGS_STOPPED) && (rec->recmaster == pnn)) {
3061 DEBUG(DEBUG_ERR,("Local node is STOPPED. Yielding recmaster role\n"));
3062 force_election(rec, pnn, nodemap);
3066 /* check that we (recovery daemon) and the local ctdb daemon
3067 agrees on whether we are banned or not
3071 /* remember our own node flags */
3072 rec->node_flags = nodemap->nodes[pnn].flags;
3074 /* count how many active nodes there are */
3075 rec->num_active = 0;
3076 rec->num_connected = 0;
3077 for (i=0; i<nodemap->num; i++) {
3078 if (!(nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE)) {
3081 if (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED)) {
3082 rec->num_connected++;
3087 /* verify that the recmaster node is still active */
3088 for (j=0; j<nodemap->num; j++) {
3089 if (nodemap->nodes[j].pnn==rec->recmaster) {
3094 if (j == nodemap->num) {
3095 DEBUG(DEBUG_ERR, ("Recmaster node %u not in list. Force reelection\n", rec->recmaster));
3096 force_election(rec, pnn, nodemap);
3100 /* if recovery master is disconnected we must elect a new recmaster */
3101 if (nodemap->nodes[j].flags & NODE_FLAGS_DISCONNECTED) {
3102 DEBUG(DEBUG_NOTICE, ("Recmaster node %u is disconnected. Force reelection\n", nodemap->nodes[j].pnn));
3103 force_election(rec, pnn, nodemap);
3107 /* grap the nodemap from the recovery master to check if it is banned */
3108 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
3109 mem_ctx, &recmaster_nodemap);
3111 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from recovery master %u\n",
3112 nodemap->nodes[j].pnn));
3117 if (recmaster_nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3118 DEBUG(DEBUG_NOTICE, ("Recmaster node %u no longer available. Force reelection\n", nodemap->nodes[j].pnn));
3119 force_election(rec, pnn, nodemap);
3124 /* verify that we have all ip addresses we should have and we dont
3125 * have addresses we shouldnt have.
3127 if (ctdb->tunable.disable_ip_failover == 0) {
3128 if (rec->ip_check_disable_ctx == NULL) {
3129 if (verify_local_ip_allocation(ctdb, rec, pnn, nodemap) != 0) {
3130 DEBUG(DEBUG_ERR, (__location__ " Public IPs were inconsistent.\n"));
3136 /* if we are not the recmaster then we do not need to check
3137 if recovery is needed
3139 if (pnn != rec->recmaster) {
3144 /* ensure our local copies of flags are right */
3145 ret = update_local_flags(rec, nodemap);
3146 if (ret == MONITOR_ELECTION_NEEDED) {
3147 DEBUG(DEBUG_NOTICE,("update_local_flags() called for a re-election.\n"));
3148 force_election(rec, pnn, nodemap);
3151 if (ret != MONITOR_OK) {
3152 DEBUG(DEBUG_ERR,("Unable to update local flags\n"));
3156 if (ctdb->num_nodes != nodemap->num) {
3157 DEBUG(DEBUG_ERR, (__location__ " ctdb->num_nodes (%d) != nodemap->num (%d) reloading nodes file\n", ctdb->num_nodes, nodemap->num));
3158 reload_nodes_file(ctdb);
3162 /* verify that all active nodes agree that we are the recmaster */
3163 switch (verify_recmaster(rec, nodemap, pnn)) {
3164 case MONITOR_RECOVERY_NEEDED:
3165 /* can not happen */
3167 case MONITOR_ELECTION_NEEDED:
3168 force_election(rec, pnn, nodemap);
3172 case MONITOR_FAILED:
3177 if (rec->need_recovery) {
3178 /* a previous recovery didn't finish */
3179 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3183 /* verify that all active nodes are in normal mode
3184 and not in recovery mode
3186 switch (verify_recmode(ctdb, nodemap)) {
3187 case MONITOR_RECOVERY_NEEDED:
3188 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3190 case MONITOR_FAILED:
3192 case MONITOR_ELECTION_NEEDED:
3193 /* can not happen */
3199 if (ctdb->tunable.verify_recovery_lock != 0) {
3200 /* we should have the reclock - check its not stale */
3201 ret = check_recovery_lock(ctdb);
3203 DEBUG(DEBUG_ERR,("Failed check_recovery_lock. Force a recovery\n"));
3204 ctdb_set_culprit(rec, ctdb->pnn);
3205 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3210 /* if there are takeovers requested, perform it and notify the waiters */
3211 if (rec->reallocate_callers) {
3212 process_ipreallocate_requests(ctdb, rec);
3215 /* get the nodemap for all active remote nodes
3217 remote_nodemaps = talloc_array(mem_ctx, struct ctdb_node_map *, nodemap->num);
3218 if (remote_nodemaps == NULL) {
3219 DEBUG(DEBUG_ERR, (__location__ " failed to allocate remote nodemap array\n"));
3222 for(i=0; i<nodemap->num; i++) {
3223 remote_nodemaps[i] = NULL;
3225 if (get_remote_nodemaps(ctdb, mem_ctx, nodemap, remote_nodemaps) != 0) {
3226 DEBUG(DEBUG_ERR,(__location__ " Failed to read remote nodemaps\n"));
3230 /* verify that all other nodes have the same nodemap as we have
3232 for (j=0; j<nodemap->num; j++) {
3233 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3237 if (remote_nodemaps[j] == NULL) {
3238 DEBUG(DEBUG_ERR,(__location__ " Did not get a remote nodemap for node %d, restarting monitoring\n", j));
3239 ctdb_set_culprit(rec, j);
3244 /* if the nodes disagree on how many nodes there are
3245 then this is a good reason to try recovery
3247 if (remote_nodemaps[j]->num != nodemap->num) {
3248 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different node count. %u vs %u of the local node\n",
3249 nodemap->nodes[j].pnn, remote_nodemaps[j]->num, nodemap->num));
3250 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3251 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3255 /* if the nodes disagree on which nodes exist and are
3256 active, then that is also a good reason to do recovery
3258 for (i=0;i<nodemap->num;i++) {
3259 if (remote_nodemaps[j]->nodes[i].pnn != nodemap->nodes[i].pnn) {
3260 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different nodemap pnn for %d (%u vs %u).\n",
3261 nodemap->nodes[j].pnn, i,
3262 remote_nodemaps[j]->nodes[i].pnn, nodemap->nodes[i].pnn));
3263 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3264 do_recovery(rec, mem_ctx, pnn, nodemap,
3270 /* verify the flags are consistent
3272 for (i=0; i<nodemap->num; i++) {
3273 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
3277 if (nodemap->nodes[i].flags != remote_nodemaps[j]->nodes[i].flags) {
3278 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different flags for node %u. It has 0x%02x vs our 0x%02x\n",
3279 nodemap->nodes[j].pnn,
3280 nodemap->nodes[i].pnn,
3281 remote_nodemaps[j]->nodes[i].flags,
3282 nodemap->nodes[j].flags));
3284 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));
3285 update_flags_on_all_nodes(ctdb, nodemap, nodemap->nodes[i].pnn, remote_nodemaps[j]->nodes[i].flags);
3286 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3287 do_recovery(rec, mem_ctx, pnn, nodemap,
3291 DEBUG(DEBUG_ERR,("Use flags 0x%02x from local recmaster node for cluster update of node %d flags\n", nodemap->nodes[i].flags, i));
3292 update_flags_on_all_nodes(ctdb, nodemap, nodemap->nodes[i].pnn, nodemap->nodes[i].flags);
3293 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3294 do_recovery(rec, mem_ctx, pnn, nodemap,
3303 /* there better be the same number of lmasters in the vnn map
3304 as there are active nodes or we will have to do a recovery
3306 if (vnnmap->size != rec->num_active) {
3307 DEBUG(DEBUG_ERR, (__location__ " The vnnmap count is different from the number of active nodes. %u vs %u\n",
3308 vnnmap->size, rec->num_active));
3309 ctdb_set_culprit(rec, ctdb->pnn);
3310 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3314 /* verify that all active nodes in the nodemap also exist in
3317 for (j=0; j<nodemap->num; j++) {
3318 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3321 if (nodemap->nodes[j].pnn == pnn) {
3325 for (i=0; i<vnnmap->size; i++) {
3326 if (vnnmap->map[i] == nodemap->nodes[j].pnn) {
3330 if (i == vnnmap->size) {
3331 DEBUG(DEBUG_ERR, (__location__ " Node %u is active in the nodemap but did not exist in the vnnmap\n",
3332 nodemap->nodes[j].pnn));
3333 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3334 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3340 /* verify that all other nodes have the same vnnmap
3341 and are from the same generation
3343 for (j=0; j<nodemap->num; j++) {
3344 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3347 if (nodemap->nodes[j].pnn == pnn) {
3351 ret = ctdb_ctrl_getvnnmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
3352 mem_ctx, &remote_vnnmap);
3354 DEBUG(DEBUG_ERR, (__location__ " Unable to get vnnmap from remote node %u\n",
3355 nodemap->nodes[j].pnn));
3359 /* verify the vnnmap generation is the same */
3360 if (vnnmap->generation != remote_vnnmap->generation) {
3361 DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different generation of vnnmap. %u vs %u (ours)\n",
3362 nodemap->nodes[j].pnn, remote_vnnmap->generation, vnnmap->generation));
3363 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3364 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3368 /* verify the vnnmap size is the same */
3369 if (vnnmap->size != remote_vnnmap->size) {
3370 DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different size of vnnmap. %u vs %u (ours)\n",
3371 nodemap->nodes[j].pnn, remote_vnnmap->size, vnnmap->size));
3372 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3373 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3377 /* verify the vnnmap is the same */
3378 for (i=0;i<vnnmap->size;i++) {
3379 if (remote_vnnmap->map[i] != vnnmap->map[i]) {
3380 DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different vnnmap.\n",
3381 nodemap->nodes[j].pnn));
3382 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3383 do_recovery(rec, mem_ctx, pnn, nodemap,
3390 /* we might need to change who has what IP assigned */
3391 if (rec->need_takeover_run) {
3392 uint32_t culprit = (uint32_t)-1;
3394 rec->need_takeover_run = false;
3396 /* update the list of public ips that a node can handle for
3399 ret = ctdb_reload_remote_public_ips(ctdb, rec, nodemap, &culprit);
3401 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
3403 rec->need_takeover_run = true;
3407 /* execute the "startrecovery" event script on all nodes */
3408 ret = run_startrecovery_eventscript(rec, nodemap);
3410 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
3411 ctdb_set_culprit(rec, ctdb->pnn);
3412 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3416 ret = ctdb_takeover_run(ctdb, nodemap);
3418 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses. Try again later\n"));
3422 /* execute the "recovered" event script on all nodes */
3423 ret = run_recovered_eventscript(ctdb, nodemap, "monitor_cluster");
3425 // we cant check whether the event completed successfully
3426 // since this script WILL fail if the node is in recovery mode
3427 // and if that race happens, the code here would just cause a second
3428 // cascading recovery.
3430 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Update of public ips failed.\n"));
3431 ctdb_set_culprit(rec, ctdb->pnn);
3432 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3439 the main monitoring loop
3441 static void monitor_cluster(struct ctdb_context *ctdb)
3443 struct ctdb_recoverd *rec;
3445 DEBUG(DEBUG_NOTICE,("monitor_cluster starting\n"));
3447 rec = talloc_zero(ctdb, struct ctdb_recoverd);
3448 CTDB_NO_MEMORY_FATAL(ctdb, rec);
3452 rec->priority_time = timeval_current();
3454 /* register a message port for sending memory dumps */
3455 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_MEM_DUMP, mem_dump_handler, rec);
3457 /* register a message port for recovery elections */
3458 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RECOVERY, election_handler, rec);
3460 /* when nodes are disabled/enabled */
3461 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_SET_NODE_FLAGS, monitor_handler, rec);
3463 /* when we are asked to puch out a flag change */
3464 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_PUSH_NODE_FLAGS, push_flags_handler, rec);
3466 /* register a message port for vacuum fetch */
3467 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_VACUUM_FETCH, vacuum_fetch_handler, rec);
3469 /* register a message port for reloadnodes */
3470 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RELOAD_NODES, reload_nodes_handler, rec);
3472 /* register a message port for performing a takeover run */
3473 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_TAKEOVER_RUN, ip_reallocate_handler, rec);
3475 /* register a message port for disabling the ip check for a short while */
3476 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_DISABLE_IP_CHECK, disable_ip_check_handler, rec);
3478 /* register a message port for updating the recovery daemons node assignment for an ip */
3479 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RECD_UPDATE_IP, recd_update_ip_handler, rec);
3482 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
3483 struct timeval start;
3487 DEBUG(DEBUG_CRIT,(__location__
3488 " Failed to create temp context\n"));
3492 start = timeval_current();
3493 main_loop(ctdb, rec, mem_ctx);
3494 talloc_free(mem_ctx);
3496 /* we only check for recovery once every second */
3497 elapsed = timeval_elapsed(&start);
3498 if (elapsed < ctdb->tunable.recover_interval) {
3499 ctdb_wait_timeout(ctdb, ctdb->tunable.recover_interval
3506 event handler for when the main ctdbd dies
3508 static void ctdb_recoverd_parent(struct event_context *ev, struct fd_event *fde,
3509 uint16_t flags, void *private_data)
3511 DEBUG(DEBUG_ALERT,("recovery daemon parent died - exiting\n"));
3516 called regularly to verify that the recovery daemon is still running
3518 static void ctdb_check_recd(struct event_context *ev, struct timed_event *te,
3519 struct timeval yt, void *p)
3521 struct ctdb_context *ctdb = talloc_get_type(p, struct ctdb_context);
3523 if (kill(ctdb->recoverd_pid, 0) != 0) {
3524 DEBUG(DEBUG_ERR,("Recovery daemon (pid:%d) is no longer running. Trying to restart recovery daemon.\n", (int)ctdb->recoverd_pid));
3526 event_add_timed(ctdb->ev, ctdb, timeval_zero(),
3527 ctdb_restart_recd, ctdb);
3532 event_add_timed(ctdb->ev, ctdb,
3533 timeval_current_ofs(30, 0),
3534 ctdb_check_recd, ctdb);
3537 static void recd_sig_child_handler(struct event_context *ev,
3538 struct signal_event *se, int signum, int count,
3542 // struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
3547 pid = waitpid(-1, &status, WNOHANG);
3549 if (errno != ECHILD) {
3550 DEBUG(DEBUG_ERR, (__location__ " waitpid() returned error. errno:%s(%d)\n", strerror(errno),errno));
3555 DEBUG(DEBUG_DEBUG, ("RECD SIGCHLD from %d\n", (int)pid));
3561 startup the recovery daemon as a child of the main ctdb daemon
3563 int ctdb_start_recoverd(struct ctdb_context *ctdb)
3566 struct signal_event *se;
3567 struct tevent_fd *fde;
3569 if (pipe(fd) != 0) {
3573 ctdb->ctdbd_pid = getpid();
3575 ctdb->recoverd_pid = fork();
3576 if (ctdb->recoverd_pid == -1) {
3580 if (ctdb->recoverd_pid != 0) {
3582 event_add_timed(ctdb->ev, ctdb,
3583 timeval_current_ofs(30, 0),
3584 ctdb_check_recd, ctdb);
3590 srandom(getpid() ^ time(NULL));
3592 if (switch_from_server_to_client(ctdb, "recoverd") != 0) {
3593 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch recovery daemon into client mode. shutting down.\n"));
3597 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to recovery daemon\n", fd[0]));
3599 fde = event_add_fd(ctdb->ev, ctdb, fd[0], EVENT_FD_READ,
3600 ctdb_recoverd_parent, &fd[0]);
3601 tevent_fd_set_auto_close(fde);
3603 /* set up a handler to pick up sigchld */
3604 se = event_add_signal(ctdb->ev, ctdb,
3606 recd_sig_child_handler,
3609 DEBUG(DEBUG_CRIT,("Failed to set up signal handler for SIGCHLD in recovery daemon\n"));
3613 monitor_cluster(ctdb);
3615 DEBUG(DEBUG_ALERT,("ERROR: ctdb_recoverd finished!?\n"));
3620 shutdown the recovery daemon
3622 void ctdb_stop_recoverd(struct ctdb_context *ctdb)
3624 if (ctdb->recoverd_pid == 0) {
3628 DEBUG(DEBUG_NOTICE,("Shutting down recovery daemon\n"));
3629 kill(ctdb->recoverd_pid, SIGTERM);
3632 static void ctdb_restart_recd(struct event_context *ev, struct timed_event *te,
3633 struct timeval t, void *private_data)
3635 struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
3637 DEBUG(DEBUG_ERR,("Restarting recovery daemon\n"));
3638 ctdb_stop_recoverd(ctdb);
3639 ctdb_start_recoverd(ctdb);