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/events/events.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.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)
75 ban a node for a period of time
77 static void ctdb_ban_node(struct ctdb_recoverd *rec, uint32_t pnn, uint32_t ban_time)
80 struct ctdb_context *ctdb = rec->ctdb;
81 struct ctdb_ban_time bantime;
83 DEBUG(DEBUG_NOTICE,("Banning node %u for %u seconds\n", pnn, ban_time));
85 if (!ctdb_validate_pnn(ctdb, pnn)) {
86 DEBUG(DEBUG_ERR,("Bad pnn %u in ctdb_ban_node\n", pnn));
91 bantime.time = ban_time;
93 ret = ctdb_ctrl_set_ban(ctdb, CONTROL_TIMEOUT(), pnn, &bantime);
95 DEBUG(DEBUG_ERR,(__location__ " Failed to ban node %d\n", pnn));
101 enum monitor_result { MONITOR_OK, MONITOR_RECOVERY_NEEDED, MONITOR_ELECTION_NEEDED, MONITOR_FAILED};
105 run the "recovered" eventscript on all nodes
107 static int run_recovered_eventscript(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, const char *caller)
112 tmp_ctx = talloc_new(ctdb);
113 CTDB_NO_MEMORY(ctdb, tmp_ctx);
115 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
116 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_END_RECOVERY,
118 CONTROL_TIMEOUT(), false, tdb_null,
121 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event when called from %s\n", caller));
123 talloc_free(tmp_ctx);
127 talloc_free(tmp_ctx);
132 remember the trouble maker
134 static void ctdb_set_culprit_count(struct ctdb_recoverd *rec, uint32_t culprit, uint32_t count)
136 struct ctdb_context *ctdb = talloc_get_type(rec->ctdb, struct ctdb_context);
137 struct ctdb_banning_state *ban_state;
139 if (culprit > ctdb->num_nodes) {
140 DEBUG(DEBUG_ERR,("Trying to set culprit %d but num_nodes is %d\n", culprit, ctdb->num_nodes));
144 if (ctdb->nodes[culprit]->ban_state == NULL) {
145 ctdb->nodes[culprit]->ban_state = talloc_zero(ctdb->nodes[culprit], struct ctdb_banning_state);
146 CTDB_NO_MEMORY_VOID(ctdb, ctdb->nodes[culprit]->ban_state);
150 ban_state = ctdb->nodes[culprit]->ban_state;
151 if (timeval_elapsed(&ban_state->last_reported_time) > ctdb->tunable.recovery_grace_period) {
152 /* this was the first time in a long while this node
153 misbehaved so we will forgive any old transgressions.
155 ban_state->count = 0;
158 ban_state->count += count;
159 ban_state->last_reported_time = timeval_current();
160 rec->last_culprit_node = culprit;
164 remember the trouble maker
166 static void ctdb_set_culprit(struct ctdb_recoverd *rec, uint32_t culprit)
168 ctdb_set_culprit_count(rec, culprit, 1);
172 /* this callback is called for every node that failed to execute the
175 static void startrecovery_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
177 struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
179 DEBUG(DEBUG_ERR, (__location__ " Node %u failed the startrecovery event. Setting it as recovery fail culprit\n", node_pnn));
181 ctdb_set_culprit(rec, node_pnn);
185 run the "startrecovery" eventscript on all nodes
187 static int run_startrecovery_eventscript(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap)
191 struct ctdb_context *ctdb = rec->ctdb;
193 tmp_ctx = talloc_new(ctdb);
194 CTDB_NO_MEMORY(ctdb, tmp_ctx);
196 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
197 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_START_RECOVERY,
199 CONTROL_TIMEOUT(), false, tdb_null,
201 startrecovery_fail_callback,
203 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event. Recovery failed.\n"));
204 talloc_free(tmp_ctx);
208 talloc_free(tmp_ctx);
212 static void async_getcap_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
214 if ( (outdata.dsize != sizeof(uint32_t)) || (outdata.dptr == NULL) ) {
215 DEBUG(DEBUG_ERR, (__location__ " Invalid length/pointer for getcap callback : %u %p\n", (unsigned)outdata.dsize, outdata.dptr));
218 if (node_pnn < ctdb->num_nodes) {
219 ctdb->nodes[node_pnn]->capabilities = *((uint32_t *)outdata.dptr);
224 update the node capabilities for all connected nodes
226 static int update_capabilities(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
231 tmp_ctx = talloc_new(ctdb);
232 CTDB_NO_MEMORY(ctdb, tmp_ctx);
234 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
235 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_CAPABILITIES,
239 async_getcap_callback, NULL,
241 DEBUG(DEBUG_ERR, (__location__ " Failed to read node capabilities.\n"));
242 talloc_free(tmp_ctx);
246 talloc_free(tmp_ctx);
250 static void set_recmode_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
252 struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
254 DEBUG(DEBUG_ERR,("Failed to freeze node %u during recovery. Set it as ban culprit for %d credits\n", node_pnn, rec->nodemap->num));
255 ctdb_set_culprit_count(rec, node_pnn, rec->nodemap->num);
258 static void transaction_start_fail_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
260 struct ctdb_recoverd *rec = talloc_get_type(callback_data, struct ctdb_recoverd);
262 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));
263 ctdb_set_culprit_count(rec, node_pnn, rec->nodemap->num);
267 change recovery mode on all nodes
269 static int set_recovery_mode(struct ctdb_context *ctdb, struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, uint32_t rec_mode)
275 tmp_ctx = talloc_new(ctdb);
276 CTDB_NO_MEMORY(ctdb, tmp_ctx);
278 /* freeze all nodes */
279 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
280 if (rec_mode == CTDB_RECOVERY_ACTIVE) {
283 for (i=1; i<=NUM_DB_PRIORITIES; i++) {
284 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_FREEZE,
289 set_recmode_fail_callback,
291 DEBUG(DEBUG_ERR, (__location__ " Unable to freeze nodes. Recovery failed.\n"));
292 talloc_free(tmp_ctx);
299 data.dsize = sizeof(uint32_t);
300 data.dptr = (unsigned char *)&rec_mode;
302 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECMODE,
308 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode. Recovery failed.\n"));
309 talloc_free(tmp_ctx);
313 talloc_free(tmp_ctx);
318 change recovery master on all node
320 static int set_recovery_master(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, uint32_t pnn)
326 tmp_ctx = talloc_new(ctdb);
327 CTDB_NO_MEMORY(ctdb, tmp_ctx);
329 data.dsize = sizeof(uint32_t);
330 data.dptr = (unsigned char *)&pnn;
332 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
333 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECMASTER,
335 CONTROL_TIMEOUT(), false, data,
338 DEBUG(DEBUG_ERR, (__location__ " Unable to set recmaster. Recovery failed.\n"));
339 talloc_free(tmp_ctx);
343 talloc_free(tmp_ctx);
347 /* update all remote nodes to use the same db priority that we have
348 this can fail if the remove node has not yet been upgraded to
349 support this function, so we always return success and never fail
350 a recovery if this call fails.
352 static int update_db_priority_on_remote_nodes(struct ctdb_context *ctdb,
353 struct ctdb_node_map *nodemap,
354 uint32_t pnn, struct ctdb_dbid_map *dbmap, TALLOC_CTX *mem_ctx)
359 nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
361 /* step through all local databases */
362 for (db=0; db<dbmap->num;db++) {
364 struct ctdb_db_priority db_prio;
367 db_prio.db_id = dbmap->dbs[db].dbid;
368 ret = ctdb_ctrl_get_db_priority(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, dbmap->dbs[db].dbid, &db_prio.priority);
370 DEBUG(DEBUG_ERR,(__location__ " Failed to read database priority from local node for db 0x%08x\n", dbmap->dbs[db].dbid));
374 DEBUG(DEBUG_INFO,("Update DB priority for db 0x%08x to %u\n", dbmap->dbs[db].dbid, db_prio.priority));
376 data.dptr = (uint8_t *)&db_prio;
377 data.dsize = sizeof(db_prio);
379 if (ctdb_client_async_control(ctdb,
380 CTDB_CONTROL_SET_DB_PRIORITY,
382 CONTROL_TIMEOUT(), false, data,
385 DEBUG(DEBUG_ERR,(__location__ " Failed to set DB priority for 0x%08x\n", db_prio.db_id));
393 ensure all other nodes have attached to any databases that we have
395 static int create_missing_remote_databases(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
396 uint32_t pnn, struct ctdb_dbid_map *dbmap, TALLOC_CTX *mem_ctx)
399 struct ctdb_dbid_map *remote_dbmap;
401 /* verify that all other nodes have all our databases */
402 for (j=0; j<nodemap->num; j++) {
403 /* we dont need to ourself ourselves */
404 if (nodemap->nodes[j].pnn == pnn) {
407 /* dont check nodes that are unavailable */
408 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
412 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
413 mem_ctx, &remote_dbmap);
415 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node %u\n", pnn));
419 /* step through all local databases */
420 for (db=0; db<dbmap->num;db++) {
424 for (i=0;i<remote_dbmap->num;i++) {
425 if (dbmap->dbs[db].dbid == remote_dbmap->dbs[i].dbid) {
429 /* the remote node already have this database */
430 if (i!=remote_dbmap->num) {
433 /* ok so we need to create this database */
434 ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), pnn, dbmap->dbs[db].dbid,
437 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbname from node %u\n", pnn));
440 ctdb_ctrl_createdb(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
441 mem_ctx, name, dbmap->dbs[db].persistent);
443 DEBUG(DEBUG_ERR, (__location__ " Unable to create remote db:%s\n", name));
454 ensure we are attached to any databases that anyone else is attached to
456 static int create_missing_local_databases(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
457 uint32_t pnn, struct ctdb_dbid_map **dbmap, TALLOC_CTX *mem_ctx)
460 struct ctdb_dbid_map *remote_dbmap;
462 /* verify that we have all database any other node has */
463 for (j=0; j<nodemap->num; j++) {
464 /* we dont need to ourself ourselves */
465 if (nodemap->nodes[j].pnn == pnn) {
468 /* dont check nodes that are unavailable */
469 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
473 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
474 mem_ctx, &remote_dbmap);
476 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node %u\n", pnn));
480 /* step through all databases on the remote node */
481 for (db=0; db<remote_dbmap->num;db++) {
484 for (i=0;i<(*dbmap)->num;i++) {
485 if (remote_dbmap->dbs[db].dbid == (*dbmap)->dbs[i].dbid) {
489 /* we already have this db locally */
490 if (i!=(*dbmap)->num) {
493 /* ok so we need to create this database and
496 ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
497 remote_dbmap->dbs[db].dbid, mem_ctx, &name);
499 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbname from node %u\n",
500 nodemap->nodes[j].pnn));
503 ctdb_ctrl_createdb(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, name,
504 remote_dbmap->dbs[db].persistent);
506 DEBUG(DEBUG_ERR, (__location__ " Unable to create local db:%s\n", name));
509 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, dbmap);
511 DEBUG(DEBUG_ERR, (__location__ " Unable to reread dbmap on node %u\n", pnn));
522 pull the remote database contents from one node into the recdb
524 static int pull_one_remote_database(struct ctdb_context *ctdb, uint32_t srcnode,
525 struct tdb_wrap *recdb, uint32_t dbid,
530 struct ctdb_marshall_buffer *reply;
531 struct ctdb_rec_data *rec;
533 TALLOC_CTX *tmp_ctx = talloc_new(recdb);
535 ret = ctdb_ctrl_pulldb(ctdb, srcnode, dbid, CTDB_LMASTER_ANY, tmp_ctx,
536 CONTROL_TIMEOUT(), &outdata);
538 DEBUG(DEBUG_ERR,(__location__ " Unable to copy db from node %u\n", srcnode));
539 talloc_free(tmp_ctx);
543 reply = (struct ctdb_marshall_buffer *)outdata.dptr;
545 if (outdata.dsize < offsetof(struct ctdb_marshall_buffer, data)) {
546 DEBUG(DEBUG_ERR,(__location__ " invalid data in pulldb reply\n"));
547 talloc_free(tmp_ctx);
551 rec = (struct ctdb_rec_data *)&reply->data[0];
555 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec), i++) {
557 struct ctdb_ltdb_header *hdr;
560 key.dptr = &rec->data[0];
561 key.dsize = rec->keylen;
562 data.dptr = &rec->data[key.dsize];
563 data.dsize = rec->datalen;
565 hdr = (struct ctdb_ltdb_header *)data.dptr;
567 if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
568 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
569 talloc_free(tmp_ctx);
573 /* fetch the existing record, if any */
574 existing = tdb_fetch(recdb->tdb, key);
576 if (existing.dptr != NULL) {
577 struct ctdb_ltdb_header header;
578 if (existing.dsize < sizeof(struct ctdb_ltdb_header)) {
579 DEBUG(DEBUG_CRIT,(__location__ " Bad record size %u from node %u\n",
580 (unsigned)existing.dsize, srcnode));
582 talloc_free(tmp_ctx);
585 header = *(struct ctdb_ltdb_header *)existing.dptr;
587 if (!(header.rsn < hdr->rsn ||
588 (header.dmaster != ctdb->recovery_master && header.rsn == hdr->rsn))) {
593 if (tdb_store(recdb->tdb, key, data, TDB_REPLACE) != 0) {
594 DEBUG(DEBUG_CRIT,(__location__ " Failed to store record\n"));
595 talloc_free(tmp_ctx);
600 talloc_free(tmp_ctx);
606 pull all the remote database contents into the recdb
608 static int pull_remote_database(struct ctdb_context *ctdb,
609 struct ctdb_recoverd *rec,
610 struct ctdb_node_map *nodemap,
611 struct tdb_wrap *recdb, uint32_t dbid,
616 /* pull all records from all other nodes across onto this node
617 (this merges based on rsn)
619 for (j=0; j<nodemap->num; j++) {
620 /* dont merge from nodes that are unavailable */
621 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
624 if (pull_one_remote_database(ctdb, nodemap->nodes[j].pnn, recdb, dbid, persistent) != 0) {
625 DEBUG(DEBUG_ERR,(__location__ " Failed to pull remote database from node %u\n",
626 nodemap->nodes[j].pnn));
627 ctdb_set_culprit_count(rec, nodemap->nodes[j].pnn, nodemap->num);
637 update flags on all active nodes
639 static int update_flags_on_all_nodes(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap, uint32_t pnn, uint32_t flags)
643 ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), pnn, flags, ~flags);
645 DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
653 ensure all nodes have the same vnnmap we do
655 static int update_vnnmap_on_all_nodes(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap,
656 uint32_t pnn, struct ctdb_vnn_map *vnnmap, TALLOC_CTX *mem_ctx)
660 /* push the new vnn map out to all the nodes */
661 for (j=0; j<nodemap->num; j++) {
662 /* dont push to nodes that are unavailable */
663 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
667 ret = ctdb_ctrl_setvnnmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn, mem_ctx, vnnmap);
669 DEBUG(DEBUG_ERR, (__location__ " Unable to set vnnmap for node %u\n", pnn));
679 struct vacuum_info *next, *prev;
680 struct ctdb_recoverd *rec;
682 struct ctdb_db_context *ctdb_db;
683 struct ctdb_marshall_buffer *recs;
684 struct ctdb_rec_data *r;
687 static void vacuum_fetch_next(struct vacuum_info *v);
690 called when a vacuum fetch has completed - just free it and do the next one
692 static void vacuum_fetch_callback(struct ctdb_client_call_state *state)
694 struct vacuum_info *v = talloc_get_type(state->async.private_data, struct vacuum_info);
696 vacuum_fetch_next(v);
701 process the next element from the vacuum list
703 static void vacuum_fetch_next(struct vacuum_info *v)
705 struct ctdb_call call;
706 struct ctdb_rec_data *r;
708 while (v->recs->count) {
709 struct ctdb_client_call_state *state;
711 struct ctdb_ltdb_header *hdr;
714 call.call_id = CTDB_NULL_FUNC;
715 call.flags = CTDB_IMMEDIATE_MIGRATION;
718 v->r = (struct ctdb_rec_data *)(r->length + (uint8_t *)r);
721 call.key.dptr = &r->data[0];
722 call.key.dsize = r->keylen;
724 /* ensure we don't block this daemon - just skip a record if we can't get
726 if (tdb_chainlock_nonblock(v->ctdb_db->ltdb->tdb, call.key) != 0) {
730 data = tdb_fetch(v->ctdb_db->ltdb->tdb, call.key);
731 if (data.dptr == NULL) {
732 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
736 if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
738 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
742 hdr = (struct ctdb_ltdb_header *)data.dptr;
743 if (hdr->dmaster == v->rec->ctdb->pnn) {
744 /* its already local */
746 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
752 state = ctdb_call_send(v->ctdb_db, &call);
753 tdb_chainunlock(v->ctdb_db->ltdb->tdb, call.key);
755 DEBUG(DEBUG_ERR,(__location__ " Failed to setup vacuum fetch call\n"));
759 state->async.fn = vacuum_fetch_callback;
760 state->async.private_data = v;
769 destroy a vacuum info structure
771 static int vacuum_info_destructor(struct vacuum_info *v)
773 DLIST_REMOVE(v->rec->vacuum_info, v);
779 handler for vacuum fetch
781 static void vacuum_fetch_handler(struct ctdb_context *ctdb, uint64_t srvid,
782 TDB_DATA data, void *private_data)
784 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
785 struct ctdb_marshall_buffer *recs;
787 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
789 struct ctdb_dbid_map *dbmap=NULL;
790 bool persistent = false;
791 struct ctdb_db_context *ctdb_db;
792 struct ctdb_rec_data *r;
794 struct vacuum_info *v;
796 recs = (struct ctdb_marshall_buffer *)data.dptr;
797 r = (struct ctdb_rec_data *)&recs->data[0];
799 if (recs->count == 0) {
800 talloc_free(tmp_ctx);
806 for (v=rec->vacuum_info;v;v=v->next) {
807 if (srcnode == v->srcnode && recs->db_id == v->ctdb_db->db_id) {
808 /* we're already working on records from this node */
809 talloc_free(tmp_ctx);
814 /* work out if the database is persistent */
815 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &dbmap);
817 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from local node\n"));
818 talloc_free(tmp_ctx);
822 for (i=0;i<dbmap->num;i++) {
823 if (dbmap->dbs[i].dbid == recs->db_id) {
824 persistent = dbmap->dbs[i].persistent;
828 if (i == dbmap->num) {
829 DEBUG(DEBUG_ERR, (__location__ " Unable to find db_id 0x%x on local node\n", recs->db_id));
830 talloc_free(tmp_ctx);
834 /* find the name of this database */
835 if (ctdb_ctrl_getdbname(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, recs->db_id, tmp_ctx, &name) != 0) {
836 DEBUG(DEBUG_ERR,(__location__ " Failed to get name of db 0x%x\n", recs->db_id));
837 talloc_free(tmp_ctx);
842 ctdb_db = ctdb_attach(ctdb, name, persistent, 0);
843 if (ctdb_db == NULL) {
844 DEBUG(DEBUG_ERR,(__location__ " Failed to attach to database '%s'\n", name));
845 talloc_free(tmp_ctx);
849 v = talloc_zero(rec, struct vacuum_info);
851 DEBUG(DEBUG_CRIT,(__location__ " Out of memory\n"));
852 talloc_free(tmp_ctx);
857 v->srcnode = srcnode;
858 v->ctdb_db = ctdb_db;
859 v->recs = talloc_memdup(v, recs, data.dsize);
860 if (v->recs == NULL) {
861 DEBUG(DEBUG_CRIT,(__location__ " Out of memory\n"));
863 talloc_free(tmp_ctx);
866 v->r = (struct ctdb_rec_data *)&v->recs->data[0];
868 DLIST_ADD(rec->vacuum_info, v);
870 talloc_set_destructor(v, vacuum_info_destructor);
872 vacuum_fetch_next(v);
873 talloc_free(tmp_ctx);
878 called when ctdb_wait_timeout should finish
880 static void ctdb_wait_handler(struct event_context *ev, struct timed_event *te,
881 struct timeval yt, void *p)
883 uint32_t *timed_out = (uint32_t *)p;
888 wait for a given number of seconds
890 static void ctdb_wait_timeout(struct ctdb_context *ctdb, uint32_t secs)
892 uint32_t timed_out = 0;
893 event_add_timed(ctdb->ev, ctdb, timeval_current_ofs(secs, 0), ctdb_wait_handler, &timed_out);
895 event_loop_once(ctdb->ev);
900 called when an election times out (ends)
902 static void ctdb_election_timeout(struct event_context *ev, struct timed_event *te,
903 struct timeval t, void *p)
905 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
906 rec->election_timeout = NULL;
908 DEBUG(DEBUG_WARNING,(__location__ " Election timed out\n"));
913 wait for an election to finish. It finished election_timeout seconds after
914 the last election packet is received
916 static void ctdb_wait_election(struct ctdb_recoverd *rec)
918 struct ctdb_context *ctdb = rec->ctdb;
919 while (rec->election_timeout) {
920 event_loop_once(ctdb->ev);
925 Update our local flags from all remote connected nodes.
926 This is only run when we are or we belive we are the recovery master
928 static int update_local_flags(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap)
931 struct ctdb_context *ctdb = rec->ctdb;
932 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
934 /* get the nodemap for all active remote nodes and verify
935 they are the same as for this node
937 for (j=0; j<nodemap->num; j++) {
938 struct ctdb_node_map *remote_nodemap=NULL;
941 if (nodemap->nodes[j].flags & NODE_FLAGS_DISCONNECTED) {
944 if (nodemap->nodes[j].pnn == ctdb->pnn) {
948 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
949 mem_ctx, &remote_nodemap);
951 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from remote node %u\n",
952 nodemap->nodes[j].pnn));
953 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
954 talloc_free(mem_ctx);
955 return MONITOR_FAILED;
957 if (nodemap->nodes[j].flags != remote_nodemap->nodes[j].flags) {
958 /* We should tell our daemon about this so it
959 updates its flags or else we will log the same
960 message again in the next iteration of recovery.
961 Since we are the recovery master we can just as
962 well update the flags on all nodes.
964 ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn, nodemap->nodes[j].flags, ~nodemap->nodes[j].flags);
966 DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
970 /* Update our local copy of the flags in the recovery
973 DEBUG(DEBUG_NOTICE,("Remote node %u had flags 0x%x, local had 0x%x - updating local\n",
974 nodemap->nodes[j].pnn, remote_nodemap->nodes[j].flags,
975 nodemap->nodes[j].flags));
976 nodemap->nodes[j].flags = remote_nodemap->nodes[j].flags;
978 talloc_free(remote_nodemap);
980 talloc_free(mem_ctx);
985 /* Create a new random generation ip.
986 The generation id can not be the INVALID_GENERATION id
988 static uint32_t new_generation(void)
993 generation = random();
995 if (generation != INVALID_GENERATION) {
1005 create a temporary working database
1007 static struct tdb_wrap *create_recdb(struct ctdb_context *ctdb, TALLOC_CTX *mem_ctx)
1010 struct tdb_wrap *recdb;
1013 /* open up the temporary recovery database */
1014 name = talloc_asprintf(mem_ctx, "%s/recdb.tdb.%u",
1015 ctdb->db_directory_state,
1022 tdb_flags = TDB_NOLOCK;
1023 if (ctdb->valgrinding) {
1024 tdb_flags |= TDB_NOMMAP;
1026 tdb_flags |= TDB_DISALLOW_NESTING;
1028 recdb = tdb_wrap_open(mem_ctx, name, ctdb->tunable.database_hash_size,
1029 tdb_flags, O_RDWR|O_CREAT|O_EXCL, 0600);
1030 if (recdb == NULL) {
1031 DEBUG(DEBUG_CRIT,(__location__ " Failed to create temp recovery database '%s'\n", name));
1041 a traverse function for pulling all relevent records from recdb
1044 struct ctdb_context *ctdb;
1045 struct ctdb_marshall_buffer *recdata;
1051 static int traverse_recdb(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *p)
1053 struct recdb_data *params = (struct recdb_data *)p;
1054 struct ctdb_rec_data *rec;
1055 struct ctdb_ltdb_header *hdr;
1057 /* skip empty records */
1058 if (data.dsize <= sizeof(struct ctdb_ltdb_header)) {
1062 /* update the dmaster field to point to us */
1063 hdr = (struct ctdb_ltdb_header *)data.dptr;
1064 if (!params->persistent) {
1065 hdr->dmaster = params->ctdb->pnn;
1068 /* add the record to the blob ready to send to the nodes */
1069 rec = ctdb_marshall_record(params->recdata, 0, key, NULL, data);
1071 params->failed = true;
1074 params->recdata = talloc_realloc_size(NULL, params->recdata, rec->length + params->len);
1075 if (params->recdata == NULL) {
1076 DEBUG(DEBUG_CRIT,(__location__ " Failed to expand recdata to %u (%u records)\n",
1077 rec->length + params->len, params->recdata->count));
1078 params->failed = true;
1081 params->recdata->count++;
1082 memcpy(params->len+(uint8_t *)params->recdata, rec, rec->length);
1083 params->len += rec->length;
1090 push the recdb database out to all nodes
1092 static int push_recdb_database(struct ctdb_context *ctdb, uint32_t dbid,
1094 struct tdb_wrap *recdb, struct ctdb_node_map *nodemap)
1096 struct recdb_data params;
1097 struct ctdb_marshall_buffer *recdata;
1099 TALLOC_CTX *tmp_ctx;
1102 tmp_ctx = talloc_new(ctdb);
1103 CTDB_NO_MEMORY(ctdb, tmp_ctx);
1105 recdata = talloc_zero(recdb, struct ctdb_marshall_buffer);
1106 CTDB_NO_MEMORY(ctdb, recdata);
1108 recdata->db_id = dbid;
1111 params.recdata = recdata;
1112 params.len = offsetof(struct ctdb_marshall_buffer, data);
1113 params.failed = false;
1114 params.persistent = persistent;
1116 if (tdb_traverse_read(recdb->tdb, traverse_recdb, ¶ms) == -1) {
1117 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1118 talloc_free(params.recdata);
1119 talloc_free(tmp_ctx);
1123 if (params.failed) {
1124 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1125 talloc_free(params.recdata);
1126 talloc_free(tmp_ctx);
1130 recdata = params.recdata;
1132 outdata.dptr = (void *)recdata;
1133 outdata.dsize = params.len;
1135 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
1136 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_PUSH_DB,
1138 CONTROL_TIMEOUT(), false, outdata,
1141 DEBUG(DEBUG_ERR,(__location__ " Failed to push recdb records to nodes for db 0x%x\n", dbid));
1142 talloc_free(recdata);
1143 talloc_free(tmp_ctx);
1147 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pushed remote database 0x%x of size %u\n",
1148 dbid, recdata->count));
1150 talloc_free(recdata);
1151 talloc_free(tmp_ctx);
1158 go through a full recovery on one database
1160 static int recover_database(struct ctdb_recoverd *rec,
1161 TALLOC_CTX *mem_ctx,
1165 struct ctdb_node_map *nodemap,
1166 uint32_t transaction_id)
1168 struct tdb_wrap *recdb;
1170 struct ctdb_context *ctdb = rec->ctdb;
1172 struct ctdb_control_wipe_database w;
1175 recdb = create_recdb(ctdb, mem_ctx);
1176 if (recdb == NULL) {
1180 /* pull all remote databases onto the recdb */
1181 ret = pull_remote_database(ctdb, rec, nodemap, recdb, dbid, persistent);
1183 DEBUG(DEBUG_ERR, (__location__ " Unable to pull remote database 0x%x\n", dbid));
1187 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pulled remote database 0x%x\n", dbid));
1189 /* wipe all the remote databases. This is safe as we are in a transaction */
1191 w.transaction_id = transaction_id;
1193 data.dptr = (void *)&w;
1194 data.dsize = sizeof(w);
1196 nodes = list_of_active_nodes(ctdb, nodemap, recdb, true);
1197 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_WIPE_DATABASE,
1199 CONTROL_TIMEOUT(), false, data,
1202 DEBUG(DEBUG_ERR, (__location__ " Unable to wipe database. Recovery failed.\n"));
1207 /* push out the correct database. This sets the dmaster and skips
1208 the empty records */
1209 ret = push_recdb_database(ctdb, dbid, persistent, recdb, nodemap);
1215 /* all done with this database */
1222 reload the nodes file
1224 static void reload_nodes_file(struct ctdb_context *ctdb)
1227 ctdb_load_nodes_file(ctdb);
1230 static int ctdb_reload_remote_public_ips(struct ctdb_context *ctdb,
1231 struct ctdb_recoverd *rec,
1232 struct ctdb_node_map *nodemap,
1238 if (ctdb->num_nodes != nodemap->num) {
1239 DEBUG(DEBUG_ERR, (__location__ " ctdb->num_nodes (%d) != nodemap->num (%d) invalid param\n",
1240 ctdb->num_nodes, nodemap->num));
1242 *culprit = ctdb->pnn;
1247 for (j=0; j<nodemap->num; j++) {
1248 /* release any existing data */
1249 if (ctdb->nodes[j]->known_public_ips) {
1250 talloc_free(ctdb->nodes[j]->known_public_ips);
1251 ctdb->nodes[j]->known_public_ips = NULL;
1253 if (ctdb->nodes[j]->available_public_ips) {
1254 talloc_free(ctdb->nodes[j]->available_public_ips);
1255 ctdb->nodes[j]->available_public_ips = NULL;
1258 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
1262 /* grab a new shiny list of public ips from the node */
1263 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1265 ctdb->nodes[j]->pnn,
1268 &ctdb->nodes[j]->known_public_ips);
1270 DEBUG(DEBUG_ERR,("Failed to read known public ips from node : %u\n",
1271 ctdb->nodes[j]->pnn));
1273 *culprit = ctdb->nodes[j]->pnn;
1278 if (verify_remote_ip_allocation(ctdb, ctdb->nodes[j]->known_public_ips)) {
1279 DEBUG(DEBUG_ERR,("Node %d has inconsistent public ip allocation and needs update.\n", ctdb->nodes[j]->pnn));
1280 rec->need_takeover_run = true;
1283 /* grab a new shiny list of public ips from the node */
1284 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1286 ctdb->nodes[j]->pnn,
1288 CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE,
1289 &ctdb->nodes[j]->available_public_ips);
1291 DEBUG(DEBUG_ERR,("Failed to read available public ips from node : %u\n",
1292 ctdb->nodes[j]->pnn));
1294 *culprit = ctdb->nodes[j]->pnn;
1304 we are the recmaster, and recovery is needed - start a recovery run
1306 static int do_recovery(struct ctdb_recoverd *rec,
1307 TALLOC_CTX *mem_ctx, uint32_t pnn,
1308 struct ctdb_node_map *nodemap, struct ctdb_vnn_map *vnnmap)
1310 struct ctdb_context *ctdb = rec->ctdb;
1312 uint32_t generation;
1313 struct ctdb_dbid_map *dbmap;
1316 struct timeval start_time;
1317 uint32_t culprit = (uint32_t)-1;
1319 DEBUG(DEBUG_NOTICE, (__location__ " Starting do_recovery\n"));
1321 /* if recovery fails, force it again */
1322 rec->need_recovery = true;
1324 for (i=0; i<ctdb->num_nodes; i++) {
1325 struct ctdb_banning_state *ban_state;
1327 if (ctdb->nodes[i]->ban_state == NULL) {
1330 ban_state = (struct ctdb_banning_state *)ctdb->nodes[i]->ban_state;
1331 if (ban_state->count < 2*ctdb->num_nodes) {
1334 DEBUG(DEBUG_NOTICE,("Node %u has caused %u recoveries recently - banning it for %u seconds\n",
1335 ctdb->nodes[i]->pnn, ban_state->count,
1336 ctdb->tunable.recovery_ban_period));
1337 ctdb_ban_node(rec, ctdb->nodes[i]->pnn, ctdb->tunable.recovery_ban_period);
1338 ban_state->count = 0;
1342 if (ctdb->tunable.verify_recovery_lock != 0) {
1343 DEBUG(DEBUG_ERR,("Taking out recovery lock from recovery daemon\n"));
1344 start_time = timeval_current();
1345 if (!ctdb_recovery_lock(ctdb, true)) {
1346 ctdb_set_culprit(rec, pnn);
1347 DEBUG(DEBUG_ERR,("Unable to get recovery lock - aborting recovery\n"));
1350 ctdb_ctrl_report_recd_lock_latency(ctdb, CONTROL_TIMEOUT(), timeval_elapsed(&start_time));
1351 DEBUG(DEBUG_NOTICE,("Recovery lock taken successfully by recovery daemon\n"));
1354 DEBUG(DEBUG_NOTICE, (__location__ " Recovery initiated due to problem with node %u\n", rec->last_culprit_node));
1356 /* get a list of all databases */
1357 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, &dbmap);
1359 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node :%u\n", pnn));
1363 /* we do the db creation before we set the recovery mode, so the freeze happens
1364 on all databases we will be dealing with. */
1366 /* verify that we have all the databases any other node has */
1367 ret = create_missing_local_databases(ctdb, nodemap, pnn, &dbmap, mem_ctx);
1369 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing local databases\n"));
1373 /* verify that all other nodes have all our databases */
1374 ret = create_missing_remote_databases(ctdb, nodemap, pnn, dbmap, mem_ctx);
1376 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing remote databases\n"));
1379 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - created remote databases\n"));
1381 /* update the database priority for all remote databases */
1382 ret = update_db_priority_on_remote_nodes(ctdb, nodemap, pnn, dbmap, mem_ctx);
1384 DEBUG(DEBUG_ERR, (__location__ " Unable to set db priority on remote nodes\n"));
1386 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated db priority for all databases\n"));
1389 /* set recovery mode to active on all nodes */
1390 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
1392 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
1396 /* execute the "startrecovery" event script on all nodes */
1397 ret = run_startrecovery_eventscript(rec, nodemap);
1399 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
1404 update all nodes to have the same flags that we have
1406 for (i=0;i<nodemap->num;i++) {
1407 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1411 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1413 DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1418 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1420 /* pick a new generation number */
1421 generation = new_generation();
1423 /* change the vnnmap on this node to use the new generation
1424 number but not on any other nodes.
1425 this guarantees that if we abort the recovery prematurely
1426 for some reason (a node stops responding?)
1427 that we can just return immediately and we will reenter
1428 recovery shortly again.
1429 I.e. we deliberately leave the cluster with an inconsistent
1430 generation id to allow us to abort recovery at any stage and
1431 just restart it from scratch.
1433 vnnmap->generation = generation;
1434 ret = ctdb_ctrl_setvnnmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, vnnmap);
1436 DEBUG(DEBUG_ERR, (__location__ " Unable to set vnnmap for node %u\n", pnn));
1440 data.dptr = (void *)&generation;
1441 data.dsize = sizeof(uint32_t);
1443 nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
1444 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_START,
1446 CONTROL_TIMEOUT(), false, data,
1448 transaction_start_fail_callback,
1450 DEBUG(DEBUG_ERR, (__location__ " Unable to start transactions. Recovery failed.\n"));
1451 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_CANCEL,
1453 CONTROL_TIMEOUT(), false, tdb_null,
1457 DEBUG(DEBUG_ERR,("Failed to cancel recovery transaction\n"));
1462 DEBUG(DEBUG_NOTICE,(__location__ " started transactions on all nodes\n"));
1464 for (i=0;i<dbmap->num;i++) {
1465 ret = recover_database(rec, mem_ctx,
1467 dbmap->dbs[i].persistent,
1468 pnn, nodemap, generation);
1470 DEBUG(DEBUG_ERR, (__location__ " Failed to recover database 0x%x\n", dbmap->dbs[i].dbid));
1475 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - starting database commits\n"));
1477 /* commit all the changes */
1478 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_COMMIT,
1480 CONTROL_TIMEOUT(), false, data,
1483 DEBUG(DEBUG_ERR, (__location__ " Unable to commit recovery changes. Recovery failed.\n"));
1487 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - committed databases\n"));
1490 /* update the capabilities for all nodes */
1491 ret = update_capabilities(ctdb, nodemap);
1493 DEBUG(DEBUG_ERR, (__location__ " Unable to update node capabilities.\n"));
1497 /* build a new vnn map with all the currently active and
1499 generation = new_generation();
1500 vnnmap = talloc(mem_ctx, struct ctdb_vnn_map);
1501 CTDB_NO_MEMORY(ctdb, vnnmap);
1502 vnnmap->generation = generation;
1504 vnnmap->map = talloc_zero_array(vnnmap, uint32_t, vnnmap->size);
1505 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1506 for (i=j=0;i<nodemap->num;i++) {
1507 if (nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE) {
1510 if (!(ctdb->nodes[i]->capabilities & CTDB_CAP_LMASTER)) {
1511 /* this node can not be an lmaster */
1512 DEBUG(DEBUG_DEBUG, ("Node %d cant be a LMASTER, skipping it\n", i));
1517 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1518 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1519 vnnmap->map[j++] = nodemap->nodes[i].pnn;
1522 if (vnnmap->size == 0) {
1523 DEBUG(DEBUG_NOTICE, ("No suitable lmasters found. Adding local node (recmaster) anyway.\n"));
1525 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1526 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1527 vnnmap->map[0] = pnn;
1530 /* update to the new vnnmap on all nodes */
1531 ret = update_vnnmap_on_all_nodes(ctdb, nodemap, pnn, vnnmap, mem_ctx);
1533 DEBUG(DEBUG_ERR, (__location__ " Unable to update vnnmap on all nodes\n"));
1537 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated vnnmap\n"));
1539 /* update recmaster to point to us for all nodes */
1540 ret = set_recovery_master(ctdb, nodemap, pnn);
1542 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery master\n"));
1546 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated recmaster\n"));
1549 update all nodes to have the same flags that we have
1551 for (i=0;i<nodemap->num;i++) {
1552 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1556 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1558 DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1563 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1565 /* disable recovery mode */
1566 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_NORMAL);
1568 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to normal on cluster\n"));
1572 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - disabled recovery mode\n"));
1575 tell nodes to takeover their public IPs
1577 ret = ctdb_reload_remote_public_ips(ctdb, rec, nodemap, &culprit);
1579 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
1583 rec->need_takeover_run = false;
1584 ret = ctdb_takeover_run(ctdb, nodemap);
1586 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses\n"));
1589 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - takeip finished\n"));
1591 /* execute the "recovered" event script on all nodes */
1592 ret = run_recovered_eventscript(ctdb, nodemap, "do_recovery");
1594 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Recovery process failed.\n"));
1598 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - finished the recovered event\n"));
1600 /* send a message to all clients telling them that the cluster
1601 has been reconfigured */
1602 ctdb_send_message(ctdb, CTDB_BROADCAST_CONNECTED, CTDB_SRVID_RECONFIGURE, tdb_null);
1604 DEBUG(DEBUG_NOTICE, (__location__ " Recovery complete\n"));
1606 rec->need_recovery = false;
1608 /* we managed to complete a full recovery, make sure to forgive
1609 any past sins by the nodes that could now participate in the
1612 DEBUG(DEBUG_ERR,("Resetting ban count to 0 for all nodes\n"));
1613 for (i=0;i<nodemap->num;i++) {
1614 struct ctdb_banning_state *ban_state;
1616 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1620 ban_state = (struct ctdb_banning_state *)ctdb->nodes[nodemap->nodes[i].pnn]->ban_state;
1621 if (ban_state == NULL) {
1625 ban_state->count = 0;
1629 /* We just finished a recovery successfully.
1630 We now wait for rerecovery_timeout before we allow
1631 another recovery to take place.
1633 DEBUG(DEBUG_NOTICE, (__location__ " New recoveries supressed for the rerecovery timeout\n"));
1634 ctdb_wait_timeout(ctdb, ctdb->tunable.rerecovery_timeout);
1635 DEBUG(DEBUG_NOTICE, (__location__ " Rerecovery timeout elapsed. Recovery reactivated.\n"));
1642 elections are won by first checking the number of connected nodes, then
1643 the priority time, then the pnn
1645 struct election_message {
1646 uint32_t num_connected;
1647 struct timeval priority_time;
1649 uint32_t node_flags;
1653 form this nodes election data
1655 static void ctdb_election_data(struct ctdb_recoverd *rec, struct election_message *em)
1658 struct ctdb_node_map *nodemap;
1659 struct ctdb_context *ctdb = rec->ctdb;
1663 em->pnn = rec->ctdb->pnn;
1664 em->priority_time = rec->priority_time;
1666 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, rec, &nodemap);
1668 DEBUG(DEBUG_ERR,(__location__ " unable to get election data\n"));
1672 rec->node_flags = nodemap->nodes[ctdb->pnn].flags;
1673 em->node_flags = rec->node_flags;
1675 for (i=0;i<nodemap->num;i++) {
1676 if (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED)) {
1677 em->num_connected++;
1681 /* we shouldnt try to win this election if we cant be a recmaster */
1682 if ((ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1683 em->num_connected = 0;
1684 em->priority_time = timeval_current();
1687 talloc_free(nodemap);
1691 see if the given election data wins
1693 static bool ctdb_election_win(struct ctdb_recoverd *rec, struct election_message *em)
1695 struct election_message myem;
1698 ctdb_election_data(rec, &myem);
1700 /* we cant win if we dont have the recmaster capability */
1701 if ((rec->ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1705 /* we cant win if we are banned */
1706 if (rec->node_flags & NODE_FLAGS_BANNED) {
1710 /* we cant win if we are stopped */
1711 if (rec->node_flags & NODE_FLAGS_STOPPED) {
1715 /* we will automatically win if the other node is banned */
1716 if (em->node_flags & NODE_FLAGS_BANNED) {
1720 /* we will automatically win if the other node is banned */
1721 if (em->node_flags & NODE_FLAGS_STOPPED) {
1725 /* try to use the most connected node */
1727 cmp = (int)myem.num_connected - (int)em->num_connected;
1730 /* then the longest running node */
1732 cmp = timeval_compare(&em->priority_time, &myem.priority_time);
1736 cmp = (int)myem.pnn - (int)em->pnn;
1743 send out an election request
1745 static int send_election_request(struct ctdb_recoverd *rec, uint32_t pnn, bool update_recmaster)
1748 TDB_DATA election_data;
1749 struct election_message emsg;
1751 struct ctdb_context *ctdb = rec->ctdb;
1753 srvid = CTDB_SRVID_RECOVERY;
1755 ctdb_election_data(rec, &emsg);
1757 election_data.dsize = sizeof(struct election_message);
1758 election_data.dptr = (unsigned char *)&emsg;
1761 /* send an election message to all active nodes */
1762 DEBUG(DEBUG_INFO,(__location__ " Send election request to all active nodes\n"));
1763 ctdb_send_message(ctdb, CTDB_BROADCAST_ALL, srvid, election_data);
1766 /* A new node that is already frozen has entered the cluster.
1767 The existing nodes are not frozen and dont need to be frozen
1768 until the election has ended and we start the actual recovery
1770 if (update_recmaster == true) {
1771 /* first we assume we will win the election and set
1772 recoverymaster to be ourself on the current node
1774 ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), pnn, pnn);
1776 DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request\n"));
1786 this function will unban all nodes in the cluster
1788 static void unban_all_nodes(struct ctdb_context *ctdb)
1791 struct ctdb_node_map *nodemap;
1792 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
1794 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
1796 DEBUG(DEBUG_ERR,(__location__ " failed to get nodemap to unban all nodes\n"));
1800 for (i=0;i<nodemap->num;i++) {
1801 if ( (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED))
1802 && (nodemap->nodes[i].flags & NODE_FLAGS_BANNED) ) {
1803 ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[i].pnn, 0, NODE_FLAGS_BANNED);
1807 talloc_free(tmp_ctx);
1812 we think we are winning the election - send a broadcast election request
1814 static void election_send_request(struct event_context *ev, struct timed_event *te, struct timeval t, void *p)
1816 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
1819 ret = send_election_request(rec, ctdb_get_pnn(rec->ctdb), false);
1821 DEBUG(DEBUG_ERR,("Failed to send election request!\n"));
1824 talloc_free(rec->send_election_te);
1825 rec->send_election_te = NULL;
1829 handler for memory dumps
1831 static void mem_dump_handler(struct ctdb_context *ctdb, uint64_t srvid,
1832 TDB_DATA data, void *private_data)
1834 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
1837 struct rd_memdump_reply *rd;
1839 if (data.dsize != sizeof(struct rd_memdump_reply)) {
1840 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
1841 talloc_free(tmp_ctx);
1844 rd = (struct rd_memdump_reply *)data.dptr;
1846 dump = talloc_zero(tmp_ctx, TDB_DATA);
1848 DEBUG(DEBUG_ERR, (__location__ " Failed to allocate memory for memdump\n"));
1849 talloc_free(tmp_ctx);
1852 ret = ctdb_dump_memory(ctdb, dump);
1854 DEBUG(DEBUG_ERR, (__location__ " ctdb_dump_memory() failed\n"));
1855 talloc_free(tmp_ctx);
1859 DEBUG(DEBUG_ERR, ("recovery master memory dump\n"));
1861 ret = ctdb_send_message(ctdb, rd->pnn, rd->srvid, *dump);
1863 DEBUG(DEBUG_ERR,("Failed to send rd memdump reply message\n"));
1864 talloc_free(tmp_ctx);
1868 talloc_free(tmp_ctx);
1872 handler for reload_nodes
1874 static void reload_nodes_handler(struct ctdb_context *ctdb, uint64_t srvid,
1875 TDB_DATA data, void *private_data)
1877 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1879 DEBUG(DEBUG_ERR, (__location__ " Reload nodes file from recovery daemon\n"));
1881 reload_nodes_file(rec->ctdb);
1885 static void reenable_ip_check(struct event_context *ev, struct timed_event *te,
1886 struct timeval yt, void *p)
1888 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
1890 talloc_free(rec->ip_check_disable_ctx);
1891 rec->ip_check_disable_ctx = NULL;
1894 static void disable_ip_check_handler(struct ctdb_context *ctdb, uint64_t srvid,
1895 TDB_DATA data, void *private_data)
1897 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1900 if (rec->ip_check_disable_ctx != NULL) {
1901 talloc_free(rec->ip_check_disable_ctx);
1902 rec->ip_check_disable_ctx = NULL;
1905 if (data.dsize != sizeof(uint32_t)) {
1906 DEBUG(DEBUG_ERR,(__location__ " Wrong size for data :%lu "
1907 "expexting %lu\n", (long unsigned)data.dsize,
1908 (long unsigned)sizeof(uint32_t)));
1911 if (data.dptr == NULL) {
1912 DEBUG(DEBUG_ERR,(__location__ " No data recaived\n"));
1916 timeout = *((uint32_t *)data.dptr);
1917 DEBUG(DEBUG_NOTICE,("Disabling ip check for %u seconds\n", timeout));
1919 rec->ip_check_disable_ctx = talloc_new(rec);
1920 CTDB_NO_MEMORY_VOID(ctdb, rec->ip_check_disable_ctx);
1922 event_add_timed(ctdb->ev, rec->ip_check_disable_ctx, timeval_current_ofs(timeout, 0), reenable_ip_check, rec);
1927 handler for ip reallocate, just add it to the list of callers and
1928 handle this later in the monitor_cluster loop so we do not recurse
1929 with other callers to takeover_run()
1931 static void ip_reallocate_handler(struct ctdb_context *ctdb, uint64_t srvid,
1932 TDB_DATA data, void *private_data)
1934 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1935 struct ip_reallocate_list *caller;
1937 if (data.dsize != sizeof(struct rd_memdump_reply)) {
1938 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
1942 if (rec->ip_reallocate_ctx == NULL) {
1943 rec->ip_reallocate_ctx = talloc_new(rec);
1944 CTDB_NO_MEMORY_FATAL(ctdb, rec->ip_reallocate_ctx);
1947 caller = talloc(rec->ip_reallocate_ctx, struct ip_reallocate_list);
1948 CTDB_NO_MEMORY_FATAL(ctdb, caller);
1950 caller->rd = (struct rd_memdump_reply *)talloc_steal(caller, data.dptr);
1951 caller->next = rec->reallocate_callers;
1952 rec->reallocate_callers = caller;
1957 static void process_ipreallocate_requests(struct ctdb_context *ctdb, struct ctdb_recoverd *rec)
1959 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
1962 struct ip_reallocate_list *callers;
1965 DEBUG(DEBUG_INFO, ("recovery master forced ip reallocation\n"));
1967 /* update the list of public ips that a node can handle for
1970 ret = ctdb_reload_remote_public_ips(ctdb, rec, rec->nodemap, &culprit);
1972 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
1974 rec->need_takeover_run = true;
1977 ret = ctdb_takeover_run(ctdb, rec->nodemap);
1979 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
1981 rec->need_takeover_run = true;
1985 result.dsize = sizeof(int32_t);
1986 result.dptr = (uint8_t *)&ret;
1988 for (callers=rec->reallocate_callers; callers; callers=callers->next) {
1990 /* Someone that sent srvid==0 does not want a reply */
1991 if (callers->rd->srvid == 0) {
1994 DEBUG(DEBUG_INFO,("Sending ip reallocate reply message to "
1995 "%u:%llu\n", (unsigned)callers->rd->pnn,
1996 (unsigned long long)callers->rd->srvid));
1997 ret = ctdb_send_message(ctdb, callers->rd->pnn, callers->rd->srvid, result);
1999 DEBUG(DEBUG_ERR,("Failed to send ip reallocate reply "
2000 "message to %u:%llu\n",
2001 (unsigned)callers->rd->pnn,
2002 (unsigned long long)callers->rd->srvid));
2006 talloc_free(tmp_ctx);
2007 talloc_free(rec->ip_reallocate_ctx);
2008 rec->ip_reallocate_ctx = NULL;
2009 rec->reallocate_callers = NULL;
2015 handler for recovery master elections
2017 static void election_handler(struct ctdb_context *ctdb, uint64_t srvid,
2018 TDB_DATA data, void *private_data)
2020 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2022 struct election_message *em = (struct election_message *)data.dptr;
2023 TALLOC_CTX *mem_ctx;
2025 /* we got an election packet - update the timeout for the election */
2026 talloc_free(rec->election_timeout);
2027 rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2028 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2029 ctdb_election_timeout, rec);
2031 mem_ctx = talloc_new(ctdb);
2033 /* someone called an election. check their election data
2034 and if we disagree and we would rather be the elected node,
2035 send a new election message to all other nodes
2037 if (ctdb_election_win(rec, em)) {
2038 if (!rec->send_election_te) {
2039 rec->send_election_te = event_add_timed(ctdb->ev, rec,
2040 timeval_current_ofs(0, 500000),
2041 election_send_request, rec);
2043 talloc_free(mem_ctx);
2044 /*unban_all_nodes(ctdb);*/
2049 talloc_free(rec->send_election_te);
2050 rec->send_election_te = NULL;
2052 if (ctdb->tunable.verify_recovery_lock != 0) {
2053 /* release the recmaster lock */
2054 if (em->pnn != ctdb->pnn &&
2055 ctdb->recovery_lock_fd != -1) {
2056 close(ctdb->recovery_lock_fd);
2057 ctdb->recovery_lock_fd = -1;
2058 unban_all_nodes(ctdb);
2062 /* ok, let that guy become recmaster then */
2063 ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), ctdb_get_pnn(ctdb), em->pnn);
2065 DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request"));
2066 talloc_free(mem_ctx);
2070 talloc_free(mem_ctx);
2076 force the start of the election process
2078 static void force_election(struct ctdb_recoverd *rec, uint32_t pnn,
2079 struct ctdb_node_map *nodemap)
2082 struct ctdb_context *ctdb = rec->ctdb;
2084 DEBUG(DEBUG_INFO,(__location__ " Force an election\n"));
2086 /* set all nodes to recovery mode to stop all internode traffic */
2087 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
2089 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
2093 talloc_free(rec->election_timeout);
2094 rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2095 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2096 ctdb_election_timeout, rec);
2098 ret = send_election_request(rec, pnn, true);
2100 DEBUG(DEBUG_ERR, (__location__ " failed to initiate recmaster election"));
2104 /* wait for a few seconds to collect all responses */
2105 ctdb_wait_election(rec);
2111 handler for when a node changes its flags
2113 static void monitor_handler(struct ctdb_context *ctdb, uint64_t srvid,
2114 TDB_DATA data, void *private_data)
2117 struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
2118 struct ctdb_node_map *nodemap=NULL;
2119 TALLOC_CTX *tmp_ctx;
2120 uint32_t changed_flags;
2122 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2123 int disabled_flag_changed;
2125 if (data.dsize != sizeof(*c)) {
2126 DEBUG(DEBUG_ERR,(__location__ "Invalid data in ctdb_node_flag_change\n"));
2130 tmp_ctx = talloc_new(ctdb);
2131 CTDB_NO_MEMORY_VOID(ctdb, tmp_ctx);
2133 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
2135 DEBUG(DEBUG_ERR,(__location__ "ctdb_ctrl_getnodemap failed in monitor_handler\n"));
2136 talloc_free(tmp_ctx);
2141 for (i=0;i<nodemap->num;i++) {
2142 if (nodemap->nodes[i].pnn == c->pnn) break;
2145 if (i == nodemap->num) {
2146 DEBUG(DEBUG_CRIT,(__location__ "Flag change for non-existant node %u\n", c->pnn));
2147 talloc_free(tmp_ctx);
2151 changed_flags = c->old_flags ^ c->new_flags;
2153 if (nodemap->nodes[i].flags != c->new_flags) {
2154 DEBUG(DEBUG_NOTICE,("Node %u has changed flags - now 0x%x was 0x%x\n", c->pnn, c->new_flags, c->old_flags));
2157 disabled_flag_changed = (nodemap->nodes[i].flags ^ c->new_flags) & NODE_FLAGS_DISABLED;
2159 nodemap->nodes[i].flags = c->new_flags;
2161 ret = ctdb_ctrl_getrecmaster(ctdb, tmp_ctx, CONTROL_TIMEOUT(),
2162 CTDB_CURRENT_NODE, &ctdb->recovery_master);
2165 ret = ctdb_ctrl_getrecmode(ctdb, tmp_ctx, CONTROL_TIMEOUT(),
2166 CTDB_CURRENT_NODE, &ctdb->recovery_mode);
2170 ctdb->recovery_master == ctdb->pnn &&
2171 ctdb->recovery_mode == CTDB_RECOVERY_NORMAL) {
2172 /* Only do the takeover run if the perm disabled or unhealthy
2173 flags changed since these will cause an ip failover but not
2175 If the node became disconnected or banned this will also
2176 lead to an ip address failover but that is handled
2179 if (disabled_flag_changed) {
2180 rec->need_takeover_run = true;
2184 talloc_free(tmp_ctx);
2188 handler for when we need to push out flag changes ot all other nodes
2190 static void push_flags_handler(struct ctdb_context *ctdb, uint64_t srvid,
2191 TDB_DATA data, void *private_data)
2194 struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
2195 struct ctdb_node_map *nodemap=NULL;
2196 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2200 /* find the recovery master */
2201 ret = ctdb_ctrl_getrecmaster(ctdb, tmp_ctx, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &recmaster);
2203 DEBUG(DEBUG_ERR, (__location__ " Unable to get recmaster from local node\n"));
2204 talloc_free(tmp_ctx);
2208 /* read the node flags from the recmaster */
2209 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), recmaster, tmp_ctx, &nodemap);
2211 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from node %u\n", c->pnn));
2212 talloc_free(tmp_ctx);
2215 if (c->pnn >= nodemap->num) {
2216 DEBUG(DEBUG_ERR,(__location__ " Nodemap from recmaster does not contain node %d\n", c->pnn));
2217 talloc_free(tmp_ctx);
2221 /* send the flags update to all connected nodes */
2222 nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
2224 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_MODIFY_FLAGS,
2225 nodes, 0, CONTROL_TIMEOUT(),
2229 DEBUG(DEBUG_ERR, (__location__ " ctdb_control to modify node flags failed\n"));
2231 talloc_free(tmp_ctx);
2235 talloc_free(tmp_ctx);
2239 struct verify_recmode_normal_data {
2241 enum monitor_result status;
2244 static void verify_recmode_normal_callback(struct ctdb_client_control_state *state)
2246 struct verify_recmode_normal_data *rmdata = talloc_get_type(state->async.private_data, struct verify_recmode_normal_data);
2249 /* one more node has responded with recmode data*/
2252 /* if we failed to get the recmode, then return an error and let
2253 the main loop try again.
2255 if (state->state != CTDB_CONTROL_DONE) {
2256 if (rmdata->status == MONITOR_OK) {
2257 rmdata->status = MONITOR_FAILED;
2262 /* if we got a response, then the recmode will be stored in the
2265 if (state->status != CTDB_RECOVERY_NORMAL) {
2266 DEBUG(DEBUG_NOTICE, (__location__ " Node:%u was in recovery mode. Restart recovery process\n", state->c->hdr.destnode));
2267 rmdata->status = MONITOR_RECOVERY_NEEDED;
2274 /* verify that all nodes are in normal recovery mode */
2275 static enum monitor_result verify_recmode(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
2277 struct verify_recmode_normal_data *rmdata;
2278 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
2279 struct ctdb_client_control_state *state;
2280 enum monitor_result status;
2283 rmdata = talloc(mem_ctx, struct verify_recmode_normal_data);
2284 CTDB_NO_MEMORY_FATAL(ctdb, rmdata);
2286 rmdata->status = MONITOR_OK;
2288 /* loop over all active nodes and send an async getrecmode call to
2290 for (j=0; j<nodemap->num; j++) {
2291 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
2294 state = ctdb_ctrl_getrecmode_send(ctdb, mem_ctx,
2296 nodemap->nodes[j].pnn);
2297 if (state == NULL) {
2298 /* we failed to send the control, treat this as
2299 an error and try again next iteration
2301 DEBUG(DEBUG_ERR,("Failed to call ctdb_ctrl_getrecmode_send during monitoring\n"));
2302 talloc_free(mem_ctx);
2303 return MONITOR_FAILED;
2306 /* set up the callback functions */
2307 state->async.fn = verify_recmode_normal_callback;
2308 state->async.private_data = rmdata;
2310 /* one more control to wait for to complete */
2315 /* now wait for up to the maximum number of seconds allowed
2316 or until all nodes we expect a response from has replied
2318 while (rmdata->count > 0) {
2319 event_loop_once(ctdb->ev);
2322 status = rmdata->status;
2323 talloc_free(mem_ctx);
2328 struct verify_recmaster_data {
2329 struct ctdb_recoverd *rec;
2332 enum monitor_result status;
2335 static void verify_recmaster_callback(struct ctdb_client_control_state *state)
2337 struct verify_recmaster_data *rmdata = talloc_get_type(state->async.private_data, struct verify_recmaster_data);
2340 /* one more node has responded with recmaster data*/
2343 /* if we failed to get the recmaster, then return an error and let
2344 the main loop try again.
2346 if (state->state != CTDB_CONTROL_DONE) {
2347 if (rmdata->status == MONITOR_OK) {
2348 rmdata->status = MONITOR_FAILED;
2353 /* if we got a response, then the recmaster will be stored in the
2356 if (state->status != rmdata->pnn) {
2357 DEBUG(DEBUG_ERR,("Node %d does not agree we are the recmaster. Need a new recmaster election\n", state->c->hdr.destnode));
2358 ctdb_set_culprit(rmdata->rec, state->c->hdr.destnode);
2359 rmdata->status = MONITOR_ELECTION_NEEDED;
2366 /* verify that all nodes agree that we are the recmaster */
2367 static enum monitor_result verify_recmaster(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap, uint32_t pnn)
2369 struct ctdb_context *ctdb = rec->ctdb;
2370 struct verify_recmaster_data *rmdata;
2371 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
2372 struct ctdb_client_control_state *state;
2373 enum monitor_result status;
2376 rmdata = talloc(mem_ctx, struct verify_recmaster_data);
2377 CTDB_NO_MEMORY_FATAL(ctdb, rmdata);
2381 rmdata->status = MONITOR_OK;
2383 /* loop over all active nodes and send an async getrecmaster call to
2385 for (j=0; j<nodemap->num; j++) {
2386 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
2389 state = ctdb_ctrl_getrecmaster_send(ctdb, mem_ctx,
2391 nodemap->nodes[j].pnn);
2392 if (state == NULL) {
2393 /* we failed to send the control, treat this as
2394 an error and try again next iteration
2396 DEBUG(DEBUG_ERR,("Failed to call ctdb_ctrl_getrecmaster_send during monitoring\n"));
2397 talloc_free(mem_ctx);
2398 return MONITOR_FAILED;
2401 /* set up the callback functions */
2402 state->async.fn = verify_recmaster_callback;
2403 state->async.private_data = rmdata;
2405 /* one more control to wait for to complete */
2410 /* now wait for up to the maximum number of seconds allowed
2411 or until all nodes we expect a response from has replied
2413 while (rmdata->count > 0) {
2414 event_loop_once(ctdb->ev);
2417 status = rmdata->status;
2418 talloc_free(mem_ctx);
2423 /* called to check that the local allocation of public ip addresses is ok.
2425 static int verify_local_ip_allocation(struct ctdb_context *ctdb, struct ctdb_recoverd *rec, uint32_t pnn)
2427 TALLOC_CTX *mem_ctx = talloc_new(NULL);
2428 struct ctdb_control_get_ifaces *ifaces = NULL;
2429 struct ctdb_all_public_ips *ips = NULL;
2430 struct ctdb_uptime *uptime1 = NULL;
2431 struct ctdb_uptime *uptime2 = NULL;
2433 bool need_iface_check = false;
2434 bool need_takeover_run = false;
2436 ret = ctdb_ctrl_uptime(ctdb, mem_ctx, CONTROL_TIMEOUT(),
2437 CTDB_CURRENT_NODE, &uptime1);
2439 DEBUG(DEBUG_ERR, ("Unable to get uptime from local node %u\n", pnn));
2440 talloc_free(mem_ctx);
2445 /* read the interfaces from the local node */
2446 ret = ctdb_ctrl_get_ifaces(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, mem_ctx, &ifaces);
2448 DEBUG(DEBUG_ERR, ("Unable to get interfaces from local node %u\n", pnn));
2449 talloc_free(mem_ctx);
2454 need_iface_check = true;
2455 } else if (rec->ifaces->num != ifaces->num) {
2456 need_iface_check = true;
2457 } else if (memcmp(rec->ifaces, ifaces, talloc_get_size(ifaces)) != 0) {
2458 need_iface_check = true;
2461 if (need_iface_check) {
2462 DEBUG(DEBUG_NOTICE, ("The interfaces status has changed on "
2463 "local node %u - force takeover run\n",
2465 need_takeover_run = true;
2468 /* read the ip allocation from the local node */
2469 ret = ctdb_ctrl_get_public_ips(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, mem_ctx, &ips);
2471 DEBUG(DEBUG_ERR, ("Unable to get public ips from local node %u\n", pnn));
2472 talloc_free(mem_ctx);
2476 ret = ctdb_ctrl_uptime(ctdb, mem_ctx, CONTROL_TIMEOUT(),
2477 CTDB_CURRENT_NODE, &uptime2);
2479 DEBUG(DEBUG_ERR, ("Unable to get uptime from local node %u\n", pnn));
2480 talloc_free(mem_ctx);
2484 /* skip the check if the startrecovery time has changed */
2485 if (timeval_compare(&uptime1->last_recovery_started,
2486 &uptime2->last_recovery_started) != 0) {
2487 DEBUG(DEBUG_NOTICE, (__location__ " last recovery time changed while we read the public ip list. skipping public ip address check\n"));
2488 talloc_free(mem_ctx);
2492 /* skip the check if the endrecovery time has changed */
2493 if (timeval_compare(&uptime1->last_recovery_finished,
2494 &uptime2->last_recovery_finished) != 0) {
2495 DEBUG(DEBUG_NOTICE, (__location__ " last recovery time changed while we read the public ip list. skipping public ip address check\n"));
2496 talloc_free(mem_ctx);
2500 /* skip the check if we have started but not finished recovery */
2501 if (timeval_compare(&uptime1->last_recovery_finished,
2502 &uptime1->last_recovery_started) != 1) {
2503 DEBUG(DEBUG_NOTICE, (__location__ " in the middle of recovery or ip reallocation. skipping public ip address check\n"));
2504 talloc_free(mem_ctx);
2509 talloc_free(rec->ifaces);
2510 rec->ifaces = talloc_steal(rec, ifaces);
2512 /* verify that we have the ip addresses we should have
2513 and we dont have ones we shouldnt have.
2514 if we find an inconsistency we set recmode to
2515 active on the local node and wait for the recmaster
2516 to do a full blown recovery
2518 for (j=0; j<ips->num; j++) {
2519 if (ips->ips[j].pnn == pnn) {
2520 if (!ctdb_sys_have_ip(&ips->ips[j].addr)) {
2521 DEBUG(DEBUG_CRIT,("Public address '%s' is missing and we should serve this ip\n",
2522 ctdb_addr_to_str(&ips->ips[j].addr)));
2523 need_takeover_run = true;
2526 if (ctdb_sys_have_ip(&ips->ips[j].addr)) {
2527 DEBUG(DEBUG_CRIT,("We are still serving a public address '%s' that we should not be serving.\n",
2528 ctdb_addr_to_str(&ips->ips[j].addr)));
2529 need_takeover_run = true;
2534 if (need_takeover_run) {
2535 struct takeover_run_reply rd;
2538 DEBUG(DEBUG_CRIT,("Trigger takeoverrun\n"));
2542 data.dptr = (uint8_t *)&rd;
2543 data.dsize = sizeof(rd);
2545 ret = ctdb_send_message(ctdb, rec->recmaster, CTDB_SRVID_TAKEOVER_RUN, data);
2547 DEBUG(DEBUG_ERR,(__location__ " Failed to send ipreallocate to recmaster :%d\n", (int)rec->recmaster));
2550 talloc_free(mem_ctx);
2555 static void async_getnodemap_callback(struct ctdb_context *ctdb, uint32_t node_pnn, int32_t res, TDB_DATA outdata, void *callback_data)
2557 struct ctdb_node_map **remote_nodemaps = callback_data;
2559 if (node_pnn >= ctdb->num_nodes) {
2560 DEBUG(DEBUG_ERR,(__location__ " pnn from invalid node\n"));
2564 remote_nodemaps[node_pnn] = (struct ctdb_node_map *)talloc_steal(remote_nodemaps, outdata.dptr);
2568 static int get_remote_nodemaps(struct ctdb_context *ctdb, TALLOC_CTX *mem_ctx,
2569 struct ctdb_node_map *nodemap,
2570 struct ctdb_node_map **remote_nodemaps)
2574 nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
2575 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_NODEMAP,
2577 CONTROL_TIMEOUT(), false, tdb_null,
2578 async_getnodemap_callback,
2580 remote_nodemaps) != 0) {
2581 DEBUG(DEBUG_ERR, (__location__ " Unable to pull all remote nodemaps\n"));
2589 enum reclock_child_status { RECLOCK_CHECKING, RECLOCK_OK, RECLOCK_FAILED, RECLOCK_TIMEOUT};
2590 struct ctdb_check_reclock_state {
2591 struct ctdb_context *ctdb;
2592 struct timeval start_time;
2595 struct timed_event *te;
2596 struct fd_event *fde;
2597 enum reclock_child_status status;
2600 /* when we free the reclock state we must kill any child process.
2602 static int check_reclock_destructor(struct ctdb_check_reclock_state *state)
2604 struct ctdb_context *ctdb = state->ctdb;
2606 ctdb_ctrl_report_recd_lock_latency(ctdb, CONTROL_TIMEOUT(), timeval_elapsed(&state->start_time));
2608 if (state->fd[0] != -1) {
2609 close(state->fd[0]);
2612 if (state->fd[1] != -1) {
2613 close(state->fd[1]);
2616 kill(state->child, SIGKILL);
2621 called if our check_reclock child times out. this would happen if
2622 i/o to the reclock file blocks.
2624 static void ctdb_check_reclock_timeout(struct event_context *ev, struct timed_event *te,
2625 struct timeval t, void *private_data)
2627 struct ctdb_check_reclock_state *state = talloc_get_type(private_data,
2628 struct ctdb_check_reclock_state);
2630 DEBUG(DEBUG_ERR,(__location__ " check_reclock child process hung/timedout CFS slow to grant locks?\n"));
2631 state->status = RECLOCK_TIMEOUT;
2634 /* this is called when the child process has completed checking the reclock
2635 file and has written data back to us through the pipe.
2637 static void reclock_child_handler(struct event_context *ev, struct fd_event *fde,
2638 uint16_t flags, void *private_data)
2640 struct ctdb_check_reclock_state *state= talloc_get_type(private_data,
2641 struct ctdb_check_reclock_state);
2645 /* we got a response from our child process so we can abort the
2648 talloc_free(state->te);
2651 ret = read(state->fd[0], &c, 1);
2652 if (ret != 1 || c != RECLOCK_OK) {
2653 DEBUG(DEBUG_ERR,(__location__ " reclock child process returned error %d\n", c));
2654 state->status = RECLOCK_FAILED;
2659 state->status = RECLOCK_OK;
2663 static int check_recovery_lock(struct ctdb_context *ctdb)
2666 struct ctdb_check_reclock_state *state;
2667 pid_t parent = getpid();
2669 if (ctdb->recovery_lock_fd == -1) {
2670 DEBUG(DEBUG_CRIT,("recovery master doesn't have the recovery lock\n"));
2674 state = talloc(ctdb, struct ctdb_check_reclock_state);
2675 CTDB_NO_MEMORY(ctdb, state);
2678 state->start_time = timeval_current();
2679 state->status = RECLOCK_CHECKING;
2683 ret = pipe(state->fd);
2686 DEBUG(DEBUG_CRIT,(__location__ " Failed to open pipe for check_reclock child\n"));
2690 state->child = fork();
2691 if (state->child == (pid_t)-1) {
2692 DEBUG(DEBUG_CRIT,(__location__ " fork() failed in check_reclock child\n"));
2693 close(state->fd[0]);
2695 close(state->fd[1]);
2701 if (state->child == 0) {
2702 char cc = RECLOCK_OK;
2703 close(state->fd[0]);
2706 if (pread(ctdb->recovery_lock_fd, &cc, 1, 0) == -1) {
2707 DEBUG(DEBUG_CRIT,("failed read from recovery_lock_fd - %s\n", strerror(errno)));
2708 cc = RECLOCK_FAILED;
2711 write(state->fd[1], &cc, 1);
2712 /* make sure we die when our parent dies */
2713 while (kill(parent, 0) == 0 || errno != ESRCH) {
2715 write(state->fd[1], &cc, 1);
2719 close(state->fd[1]);
2721 set_close_on_exec(state->fd[0]);
2723 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d for check_recovery_lock\n", state->fd[0]));
2725 talloc_set_destructor(state, check_reclock_destructor);
2727 state->te = event_add_timed(ctdb->ev, state, timeval_current_ofs(15, 0),
2728 ctdb_check_reclock_timeout, state);
2729 if (state->te == NULL) {
2730 DEBUG(DEBUG_CRIT,(__location__ " Failed to create a timed event for reclock child\n"));
2735 state->fde = event_add_fd(ctdb->ev, state, state->fd[0],
2736 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
2737 reclock_child_handler,
2740 if (state->fde == NULL) {
2741 DEBUG(DEBUG_CRIT,(__location__ " Failed to create an fd event for reclock child\n"));
2746 while (state->status == RECLOCK_CHECKING) {
2747 event_loop_once(ctdb->ev);
2750 if (state->status == RECLOCK_FAILED) {
2751 DEBUG(DEBUG_ERR,(__location__ " reclock child failed when checking file\n"));
2752 close(ctdb->recovery_lock_fd);
2753 ctdb->recovery_lock_fd = -1;
2762 static int update_recovery_lock_file(struct ctdb_context *ctdb)
2764 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
2765 const char *reclockfile;
2767 if (ctdb_ctrl_getreclock(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &reclockfile) != 0) {
2768 DEBUG(DEBUG_ERR,("Failed to read reclock file from daemon\n"));
2769 talloc_free(tmp_ctx);
2773 if (reclockfile == NULL) {
2774 if (ctdb->recovery_lock_file != NULL) {
2775 DEBUG(DEBUG_ERR,("Reclock file disabled\n"));
2776 talloc_free(ctdb->recovery_lock_file);
2777 ctdb->recovery_lock_file = NULL;
2778 if (ctdb->recovery_lock_fd != -1) {
2779 close(ctdb->recovery_lock_fd);
2780 ctdb->recovery_lock_fd = -1;
2783 ctdb->tunable.verify_recovery_lock = 0;
2784 talloc_free(tmp_ctx);
2788 if (ctdb->recovery_lock_file == NULL) {
2789 ctdb->recovery_lock_file = talloc_strdup(ctdb, reclockfile);
2790 if (ctdb->recovery_lock_fd != -1) {
2791 close(ctdb->recovery_lock_fd);
2792 ctdb->recovery_lock_fd = -1;
2794 talloc_free(tmp_ctx);
2799 if (!strcmp(reclockfile, ctdb->recovery_lock_file)) {
2800 talloc_free(tmp_ctx);
2804 talloc_free(ctdb->recovery_lock_file);
2805 ctdb->recovery_lock_file = talloc_strdup(ctdb, reclockfile);
2806 ctdb->tunable.verify_recovery_lock = 0;
2807 if (ctdb->recovery_lock_fd != -1) {
2808 close(ctdb->recovery_lock_fd);
2809 ctdb->recovery_lock_fd = -1;
2812 talloc_free(tmp_ctx);
2817 the main monitoring loop
2819 static void monitor_cluster(struct ctdb_context *ctdb)
2822 TALLOC_CTX *mem_ctx=NULL;
2823 struct ctdb_node_map *nodemap=NULL;
2824 struct ctdb_node_map *recmaster_nodemap=NULL;
2825 struct ctdb_node_map **remote_nodemaps=NULL;
2826 struct ctdb_vnn_map *vnnmap=NULL;
2827 struct ctdb_vnn_map *remote_vnnmap=NULL;
2828 int32_t debug_level;
2830 struct ctdb_recoverd *rec;
2832 DEBUG(DEBUG_NOTICE,("monitor_cluster starting\n"));
2834 rec = talloc_zero(ctdb, struct ctdb_recoverd);
2835 CTDB_NO_MEMORY_FATAL(ctdb, rec);
2839 rec->priority_time = timeval_current();
2841 /* register a message port for sending memory dumps */
2842 ctdb_set_message_handler(ctdb, CTDB_SRVID_MEM_DUMP, mem_dump_handler, rec);
2844 /* register a message port for recovery elections */
2845 ctdb_set_message_handler(ctdb, CTDB_SRVID_RECOVERY, election_handler, rec);
2847 /* when nodes are disabled/enabled */
2848 ctdb_set_message_handler(ctdb, CTDB_SRVID_SET_NODE_FLAGS, monitor_handler, rec);
2850 /* when we are asked to puch out a flag change */
2851 ctdb_set_message_handler(ctdb, CTDB_SRVID_PUSH_NODE_FLAGS, push_flags_handler, rec);
2853 /* register a message port for vacuum fetch */
2854 ctdb_set_message_handler(ctdb, CTDB_SRVID_VACUUM_FETCH, vacuum_fetch_handler, rec);
2856 /* register a message port for reloadnodes */
2857 ctdb_set_message_handler(ctdb, CTDB_SRVID_RELOAD_NODES, reload_nodes_handler, rec);
2859 /* register a message port for performing a takeover run */
2860 ctdb_set_message_handler(ctdb, CTDB_SRVID_TAKEOVER_RUN, ip_reallocate_handler, rec);
2862 /* register a message port for disabling the ip check for a short while */
2863 ctdb_set_message_handler(ctdb, CTDB_SRVID_DISABLE_IP_CHECK, disable_ip_check_handler, rec);
2867 talloc_free(mem_ctx);
2870 mem_ctx = talloc_new(ctdb);
2872 DEBUG(DEBUG_CRIT,(__location__ " Failed to create temporary context\n"));
2876 /* we only check for recovery once every second */
2877 ctdb_wait_timeout(ctdb, ctdb->tunable.recover_interval);
2879 /* verify that the main daemon is still running */
2880 if (kill(ctdb->ctdbd_pid, 0) != 0) {
2881 DEBUG(DEBUG_CRIT,("CTDB daemon is no longer available. Shutting down recovery daemon\n"));
2885 /* ping the local daemon to tell it we are alive */
2886 ctdb_ctrl_recd_ping(ctdb);
2888 if (rec->election_timeout) {
2889 /* an election is in progress */
2893 /* read the debug level from the parent and update locally */
2894 ret = ctdb_ctrl_get_debuglevel(ctdb, CTDB_CURRENT_NODE, &debug_level);
2896 DEBUG(DEBUG_ERR, (__location__ " Failed to read debuglevel from parent\n"));
2899 LogLevel = debug_level;
2902 /* We must check if we need to ban a node here but we want to do this
2903 as early as possible so we dont wait until we have pulled the node
2904 map from the local node. thats why we have the hardcoded value 20
2906 for (i=0; i<ctdb->num_nodes; i++) {
2907 struct ctdb_banning_state *ban_state;
2909 if (ctdb->nodes[i]->ban_state == NULL) {
2912 ban_state = (struct ctdb_banning_state *)ctdb->nodes[i]->ban_state;
2913 if (ban_state->count < 20) {
2916 DEBUG(DEBUG_NOTICE,("Node %u has caused %u recoveries recently - banning it for %u seconds\n",
2917 ctdb->nodes[i]->pnn, ban_state->count,
2918 ctdb->tunable.recovery_ban_period));
2919 ctdb_ban_node(rec, ctdb->nodes[i]->pnn, ctdb->tunable.recovery_ban_period);
2920 ban_state->count = 0;
2923 /* get relevant tunables */
2924 ret = ctdb_ctrl_get_all_tunables(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &ctdb->tunable);
2926 DEBUG(DEBUG_ERR,("Failed to get tunables - retrying\n"));
2930 /* get the current recovery lock file from the server */
2931 if (update_recovery_lock_file(ctdb) != 0) {
2932 DEBUG(DEBUG_ERR,("Failed to update the recovery lock file\n"));
2936 /* Make sure that if recovery lock verification becomes disabled when
2939 if (ctdb->tunable.verify_recovery_lock == 0) {
2940 if (ctdb->recovery_lock_fd != -1) {
2941 close(ctdb->recovery_lock_fd);
2942 ctdb->recovery_lock_fd = -1;
2946 pnn = ctdb_ctrl_getpnn(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE);
2947 if (pnn == (uint32_t)-1) {
2948 DEBUG(DEBUG_ERR,("Failed to get local pnn - retrying\n"));
2952 /* get the vnnmap */
2953 ret = ctdb_ctrl_getvnnmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, &vnnmap);
2955 DEBUG(DEBUG_ERR, (__location__ " Unable to get vnnmap from node %u\n", pnn));
2960 /* get number of nodes */
2962 talloc_free(rec->nodemap);
2963 rec->nodemap = NULL;
2966 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), pnn, rec, &rec->nodemap);
2968 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from node %u\n", pnn));
2971 nodemap = rec->nodemap;
2973 /* check which node is the recovery master */
2974 ret = ctdb_ctrl_getrecmaster(ctdb, mem_ctx, CONTROL_TIMEOUT(), pnn, &rec->recmaster);
2976 DEBUG(DEBUG_ERR, (__location__ " Unable to get recmaster from node %u\n", pnn));
2980 /* if we are not the recmaster we can safely ignore any ip reallocate requests */
2981 if (rec->recmaster != pnn) {
2982 if (rec->ip_reallocate_ctx != NULL) {
2983 talloc_free(rec->ip_reallocate_ctx);
2984 rec->ip_reallocate_ctx = NULL;
2985 rec->reallocate_callers = NULL;
2988 /* if there are takeovers requested, perform it and notify the waiters */
2989 if (rec->reallocate_callers) {
2990 process_ipreallocate_requests(ctdb, rec);
2993 if (rec->recmaster == (uint32_t)-1) {
2994 DEBUG(DEBUG_NOTICE,(__location__ " Initial recovery master set - forcing election\n"));
2995 force_election(rec, pnn, nodemap);
3000 /* if the local daemon is STOPPED, we verify that the databases are
3001 also frozen and thet the recmode is set to active
3003 if (nodemap->nodes[pnn].flags & NODE_FLAGS_STOPPED) {
3004 ret = ctdb_ctrl_getrecmode(ctdb, mem_ctx, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, &ctdb->recovery_mode);
3006 DEBUG(DEBUG_ERR,(__location__ " Failed to read recmode from local node\n"));
3008 if (ctdb->recovery_mode == CTDB_RECOVERY_NORMAL) {
3009 DEBUG(DEBUG_ERR,("Node is stopped but recovery mode is not active. Activate recovery mode and lock databases\n"));
3011 ret = ctdb_ctrl_freeze_priority(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, 1);
3013 DEBUG(DEBUG_ERR,(__location__ " Failed to freeze node due to node being STOPPED\n"));
3016 ret = ctdb_ctrl_setrecmode(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, CTDB_RECOVERY_ACTIVE);
3018 DEBUG(DEBUG_ERR,(__location__ " Failed to activate recovery mode due to node being stopped\n"));
3025 /* If the local node is stopped, verify we are not the recmaster
3026 and yield this role if so
3028 if ((nodemap->nodes[pnn].flags & NODE_FLAGS_STOPPED) && (rec->recmaster == pnn)) {
3029 DEBUG(DEBUG_ERR,("Local node is STOPPED. Yielding recmaster role\n"));
3030 force_election(rec, pnn, nodemap);
3034 /* check that we (recovery daemon) and the local ctdb daemon
3035 agrees on whether we are banned or not
3039 /* remember our own node flags */
3040 rec->node_flags = nodemap->nodes[pnn].flags;
3042 /* count how many active nodes there are */
3043 rec->num_active = 0;
3044 rec->num_connected = 0;
3045 for (i=0; i<nodemap->num; i++) {
3046 if (!(nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE)) {
3049 if (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED)) {
3050 rec->num_connected++;
3055 /* verify that the recmaster node is still active */
3056 for (j=0; j<nodemap->num; j++) {
3057 if (nodemap->nodes[j].pnn==rec->recmaster) {
3062 if (j == nodemap->num) {
3063 DEBUG(DEBUG_ERR, ("Recmaster node %u not in list. Force reelection\n", rec->recmaster));
3064 force_election(rec, pnn, nodemap);
3068 /* if recovery master is disconnected we must elect a new recmaster */
3069 if (nodemap->nodes[j].flags & NODE_FLAGS_DISCONNECTED) {
3070 DEBUG(DEBUG_NOTICE, ("Recmaster node %u is disconnected. Force reelection\n", nodemap->nodes[j].pnn));
3071 force_election(rec, pnn, nodemap);
3075 /* grap the nodemap from the recovery master to check if it is banned */
3076 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
3077 mem_ctx, &recmaster_nodemap);
3079 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from recovery master %u\n",
3080 nodemap->nodes[j].pnn));
3085 if (recmaster_nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3086 DEBUG(DEBUG_NOTICE, ("Recmaster node %u no longer available. Force reelection\n", nodemap->nodes[j].pnn));
3087 force_election(rec, pnn, nodemap);
3092 /* verify that we have all ip addresses we should have and we dont
3093 * have addresses we shouldnt have.
3095 if (ctdb->do_checkpublicip) {
3096 if (rec->ip_check_disable_ctx == NULL) {
3097 if (verify_local_ip_allocation(ctdb, rec, pnn) != 0) {
3098 DEBUG(DEBUG_ERR, (__location__ " Public IPs were inconsistent.\n"));
3104 /* if we are not the recmaster then we do not need to check
3105 if recovery is needed
3107 if (pnn != rec->recmaster) {
3112 /* ensure our local copies of flags are right */
3113 ret = update_local_flags(rec, nodemap);
3114 if (ret == MONITOR_ELECTION_NEEDED) {
3115 DEBUG(DEBUG_NOTICE,("update_local_flags() called for a re-election.\n"));
3116 force_election(rec, pnn, nodemap);
3119 if (ret != MONITOR_OK) {
3120 DEBUG(DEBUG_ERR,("Unable to update local flags\n"));
3124 if (ctdb->num_nodes != nodemap->num) {
3125 DEBUG(DEBUG_ERR, (__location__ " ctdb->num_nodes (%d) != nodemap->num (%d) reloading nodes file\n", ctdb->num_nodes, nodemap->num));
3126 reload_nodes_file(ctdb);
3130 /* verify that all active nodes agree that we are the recmaster */
3131 switch (verify_recmaster(rec, nodemap, pnn)) {
3132 case MONITOR_RECOVERY_NEEDED:
3133 /* can not happen */
3135 case MONITOR_ELECTION_NEEDED:
3136 force_election(rec, pnn, nodemap);
3140 case MONITOR_FAILED:
3145 if (rec->need_recovery) {
3146 /* a previous recovery didn't finish */
3147 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3151 /* verify that all active nodes are in normal mode
3152 and not in recovery mode
3154 switch (verify_recmode(ctdb, nodemap)) {
3155 case MONITOR_RECOVERY_NEEDED:
3156 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3158 case MONITOR_FAILED:
3160 case MONITOR_ELECTION_NEEDED:
3161 /* can not happen */
3167 if (ctdb->tunable.verify_recovery_lock != 0) {
3168 /* we should have the reclock - check its not stale */
3169 ret = check_recovery_lock(ctdb);
3171 DEBUG(DEBUG_ERR,("Failed check_recovery_lock. Force a recovery\n"));
3172 ctdb_set_culprit(rec, ctdb->pnn);
3173 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3178 /* get the nodemap for all active remote nodes
3180 remote_nodemaps = talloc_array(mem_ctx, struct ctdb_node_map *, nodemap->num);
3181 if (remote_nodemaps == NULL) {
3182 DEBUG(DEBUG_ERR, (__location__ " failed to allocate remote nodemap array\n"));
3185 for(i=0; i<nodemap->num; i++) {
3186 remote_nodemaps[i] = NULL;
3188 if (get_remote_nodemaps(ctdb, mem_ctx, nodemap, remote_nodemaps) != 0) {
3189 DEBUG(DEBUG_ERR,(__location__ " Failed to read remote nodemaps\n"));
3193 /* verify that all other nodes have the same nodemap as we have
3195 for (j=0; j<nodemap->num; j++) {
3196 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3200 if (remote_nodemaps[j] == NULL) {
3201 DEBUG(DEBUG_ERR,(__location__ " Did not get a remote nodemap for node %d, restarting monitoring\n", j));
3202 ctdb_set_culprit(rec, j);
3207 /* if the nodes disagree on how many nodes there are
3208 then this is a good reason to try recovery
3210 if (remote_nodemaps[j]->num != nodemap->num) {
3211 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different node count. %u vs %u of the local node\n",
3212 nodemap->nodes[j].pnn, remote_nodemaps[j]->num, nodemap->num));
3213 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3214 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3218 /* if the nodes disagree on which nodes exist and are
3219 active, then that is also a good reason to do recovery
3221 for (i=0;i<nodemap->num;i++) {
3222 if (remote_nodemaps[j]->nodes[i].pnn != nodemap->nodes[i].pnn) {
3223 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different nodemap pnn for %d (%u vs %u).\n",
3224 nodemap->nodes[j].pnn, i,
3225 remote_nodemaps[j]->nodes[i].pnn, nodemap->nodes[i].pnn));
3226 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3227 do_recovery(rec, mem_ctx, pnn, nodemap,
3233 /* verify the flags are consistent
3235 for (i=0; i<nodemap->num; i++) {
3236 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
3240 if (nodemap->nodes[i].flags != remote_nodemaps[j]->nodes[i].flags) {
3241 DEBUG(DEBUG_ERR, (__location__ " Remote node:%u has different flags for node %u. It has 0x%02x vs our 0x%02x\n",
3242 nodemap->nodes[j].pnn,
3243 nodemap->nodes[i].pnn,
3244 remote_nodemaps[j]->nodes[i].flags,
3245 nodemap->nodes[j].flags));
3247 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));
3248 update_flags_on_all_nodes(ctdb, nodemap, nodemap->nodes[i].pnn, remote_nodemaps[j]->nodes[i].flags);
3249 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3250 do_recovery(rec, mem_ctx, pnn, nodemap,
3254 DEBUG(DEBUG_ERR,("Use flags 0x%02x from local recmaster node for cluster update of node %d flags\n", nodemap->nodes[i].flags, i));
3255 update_flags_on_all_nodes(ctdb, nodemap, nodemap->nodes[i].pnn, nodemap->nodes[i].flags);
3256 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3257 do_recovery(rec, mem_ctx, pnn, nodemap,
3266 /* there better be the same number of lmasters in the vnn map
3267 as there are active nodes or we will have to do a recovery
3269 if (vnnmap->size != rec->num_active) {
3270 DEBUG(DEBUG_ERR, (__location__ " The vnnmap count is different from the number of active nodes. %u vs %u\n",
3271 vnnmap->size, rec->num_active));
3272 ctdb_set_culprit(rec, ctdb->pnn);
3273 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3277 /* verify that all active nodes in the nodemap also exist in
3280 for (j=0; j<nodemap->num; j++) {
3281 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3284 if (nodemap->nodes[j].pnn == pnn) {
3288 for (i=0; i<vnnmap->size; i++) {
3289 if (vnnmap->map[i] == nodemap->nodes[j].pnn) {
3293 if (i == vnnmap->size) {
3294 DEBUG(DEBUG_ERR, (__location__ " Node %u is active in the nodemap but did not exist in the vnnmap\n",
3295 nodemap->nodes[j].pnn));
3296 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3297 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3303 /* verify that all other nodes have the same vnnmap
3304 and are from the same generation
3306 for (j=0; j<nodemap->num; j++) {
3307 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
3310 if (nodemap->nodes[j].pnn == pnn) {
3314 ret = ctdb_ctrl_getvnnmap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
3315 mem_ctx, &remote_vnnmap);
3317 DEBUG(DEBUG_ERR, (__location__ " Unable to get vnnmap from remote node %u\n",
3318 nodemap->nodes[j].pnn));
3322 /* verify the vnnmap generation is the same */
3323 if (vnnmap->generation != remote_vnnmap->generation) {
3324 DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different generation of vnnmap. %u vs %u (ours)\n",
3325 nodemap->nodes[j].pnn, remote_vnnmap->generation, vnnmap->generation));
3326 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3327 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3331 /* verify the vnnmap size is the same */
3332 if (vnnmap->size != remote_vnnmap->size) {
3333 DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different size of vnnmap. %u vs %u (ours)\n",
3334 nodemap->nodes[j].pnn, remote_vnnmap->size, vnnmap->size));
3335 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3336 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3340 /* verify the vnnmap is the same */
3341 for (i=0;i<vnnmap->size;i++) {
3342 if (remote_vnnmap->map[i] != vnnmap->map[i]) {
3343 DEBUG(DEBUG_ERR, (__location__ " Remote node %u has different vnnmap.\n",
3344 nodemap->nodes[j].pnn));
3345 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
3346 do_recovery(rec, mem_ctx, pnn, nodemap,
3353 /* we might need to change who has what IP assigned */
3354 if (rec->need_takeover_run) {
3355 uint32_t culprit = (uint32_t)-1;
3357 rec->need_takeover_run = false;
3359 /* update the list of public ips that a node can handle for
3362 ret = ctdb_reload_remote_public_ips(ctdb, rec, nodemap, &culprit);
3364 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
3366 ctdb_set_culprit(rec, culprit);
3367 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3371 /* execute the "startrecovery" event script on all nodes */
3372 ret = run_startrecovery_eventscript(rec, nodemap);
3374 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
3375 ctdb_set_culprit(rec, ctdb->pnn);
3376 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3380 ret = ctdb_takeover_run(ctdb, nodemap);
3382 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses - starting recovery\n"));
3383 ctdb_set_culprit(rec, ctdb->pnn);
3384 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3388 /* execute the "recovered" event script on all nodes */
3389 ret = run_recovered_eventscript(ctdb, nodemap, "monitor_cluster");
3391 // we cant check whether the event completed successfully
3392 // since this script WILL fail if the node is in recovery mode
3393 // and if that race happens, the code here would just cause a second
3394 // cascading recovery.
3396 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Update of public ips failed.\n"));
3397 ctdb_set_culprit(rec, ctdb->pnn);
3398 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3409 event handler for when the main ctdbd dies
3411 static void ctdb_recoverd_parent(struct event_context *ev, struct fd_event *fde,
3412 uint16_t flags, void *private_data)
3414 DEBUG(DEBUG_ALERT,("recovery daemon parent died - exiting\n"));
3419 called regularly to verify that the recovery daemon is still running
3421 static void ctdb_check_recd(struct event_context *ev, struct timed_event *te,
3422 struct timeval yt, void *p)
3424 struct ctdb_context *ctdb = talloc_get_type(p, struct ctdb_context);
3426 if (kill(ctdb->recoverd_pid, 0) != 0) {
3427 DEBUG(DEBUG_ERR,("Recovery daemon (pid:%d) is no longer running. Shutting down main daemon\n", (int)ctdb->recoverd_pid));
3429 ctdb_stop_recoverd(ctdb);
3430 ctdb_stop_keepalive(ctdb);
3431 ctdb_stop_monitoring(ctdb);
3432 ctdb_release_all_ips(ctdb);
3433 if (ctdb->methods != NULL) {
3434 ctdb->methods->shutdown(ctdb);
3436 ctdb_event_script(ctdb, CTDB_EVENT_SHUTDOWN);
3441 event_add_timed(ctdb->ev, ctdb,
3442 timeval_current_ofs(30, 0),
3443 ctdb_check_recd, ctdb);
3446 static void recd_sig_child_handler(struct event_context *ev,
3447 struct signal_event *se, int signum, int count,
3451 // struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
3456 pid = waitpid(-1, &status, WNOHANG);
3458 if (errno != ECHILD) {
3459 DEBUG(DEBUG_ERR, (__location__ " waitpid() returned error. errno:%s(%d)\n", strerror(errno),errno));
3464 DEBUG(DEBUG_DEBUG, ("RECD SIGCHLD from %d\n", (int)pid));
3470 startup the recovery daemon as a child of the main ctdb daemon
3472 int ctdb_start_recoverd(struct ctdb_context *ctdb)
3475 struct signal_event *se;
3477 if (pipe(fd) != 0) {
3481 ctdb->ctdbd_pid = getpid();
3483 ctdb->recoverd_pid = fork();
3484 if (ctdb->recoverd_pid == -1) {
3488 if (ctdb->recoverd_pid != 0) {
3490 event_add_timed(ctdb->ev, ctdb,
3491 timeval_current_ofs(30, 0),
3492 ctdb_check_recd, ctdb);
3498 srandom(getpid() ^ time(NULL));
3500 if (switch_from_server_to_client(ctdb) != 0) {
3501 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch recovery daemon into client mode. shutting down.\n"));
3505 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to recovery daemon\n", fd[0]));
3507 event_add_fd(ctdb->ev, ctdb, fd[0], EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
3508 ctdb_recoverd_parent, &fd[0]);
3510 /* set up a handler to pick up sigchld */
3511 se = event_add_signal(ctdb->ev, ctdb,
3513 recd_sig_child_handler,
3516 DEBUG(DEBUG_CRIT,("Failed to set up signal handler for SIGCHLD in recovery daemon\n"));
3520 monitor_cluster(ctdb);
3522 DEBUG(DEBUG_ALERT,("ERROR: ctdb_recoverd finished!?\n"));
3527 shutdown the recovery daemon
3529 void ctdb_stop_recoverd(struct ctdb_context *ctdb)
3531 if (ctdb->recoverd_pid == 0) {
3535 DEBUG(DEBUG_NOTICE,("Shutting down recovery daemon\n"));
3536 kill(ctdb->recoverd_pid, SIGTERM);