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)
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, double secs)
892 uint32_t timed_out = 0;
893 time_t usecs = (secs - (time_t)secs) * 1000000;
894 event_add_timed(ctdb->ev, ctdb, timeval_current_ofs(secs, usecs), ctdb_wait_handler, &timed_out);
896 event_loop_once(ctdb->ev);
901 called when an election times out (ends)
903 static void ctdb_election_timeout(struct event_context *ev, struct timed_event *te,
904 struct timeval t, void *p)
906 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
907 rec->election_timeout = NULL;
910 DEBUG(DEBUG_WARNING,(__location__ " Election timed out\n"));
915 wait for an election to finish. It finished election_timeout seconds after
916 the last election packet is received
918 static void ctdb_wait_election(struct ctdb_recoverd *rec)
920 struct ctdb_context *ctdb = rec->ctdb;
921 while (rec->election_timeout) {
922 event_loop_once(ctdb->ev);
927 Update our local flags from all remote connected nodes.
928 This is only run when we are or we belive we are the recovery master
930 static int update_local_flags(struct ctdb_recoverd *rec, struct ctdb_node_map *nodemap)
933 struct ctdb_context *ctdb = rec->ctdb;
934 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
936 /* get the nodemap for all active remote nodes and verify
937 they are the same as for this node
939 for (j=0; j<nodemap->num; j++) {
940 struct ctdb_node_map *remote_nodemap=NULL;
943 if (nodemap->nodes[j].flags & NODE_FLAGS_DISCONNECTED) {
946 if (nodemap->nodes[j].pnn == ctdb->pnn) {
950 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn,
951 mem_ctx, &remote_nodemap);
953 DEBUG(DEBUG_ERR, (__location__ " Unable to get nodemap from remote node %u\n",
954 nodemap->nodes[j].pnn));
955 ctdb_set_culprit(rec, nodemap->nodes[j].pnn);
956 talloc_free(mem_ctx);
957 return MONITOR_FAILED;
959 if (nodemap->nodes[j].flags != remote_nodemap->nodes[j].flags) {
960 /* We should tell our daemon about this so it
961 updates its flags or else we will log the same
962 message again in the next iteration of recovery.
963 Since we are the recovery master we can just as
964 well update the flags on all nodes.
966 ret = ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[j].pnn, nodemap->nodes[j].flags, ~nodemap->nodes[j].flags);
968 DEBUG(DEBUG_ERR, (__location__ " Unable to update nodeflags on remote nodes\n"));
972 /* Update our local copy of the flags in the recovery
975 DEBUG(DEBUG_NOTICE,("Remote node %u had flags 0x%x, local had 0x%x - updating local\n",
976 nodemap->nodes[j].pnn, remote_nodemap->nodes[j].flags,
977 nodemap->nodes[j].flags));
978 nodemap->nodes[j].flags = remote_nodemap->nodes[j].flags;
980 talloc_free(remote_nodemap);
982 talloc_free(mem_ctx);
987 /* Create a new random generation ip.
988 The generation id can not be the INVALID_GENERATION id
990 static uint32_t new_generation(void)
995 generation = random();
997 if (generation != INVALID_GENERATION) {
1007 create a temporary working database
1009 static struct tdb_wrap *create_recdb(struct ctdb_context *ctdb, TALLOC_CTX *mem_ctx)
1012 struct tdb_wrap *recdb;
1015 /* open up the temporary recovery database */
1016 name = talloc_asprintf(mem_ctx, "%s/recdb.tdb.%u",
1017 ctdb->db_directory_state,
1024 tdb_flags = TDB_NOLOCK;
1025 if (ctdb->valgrinding) {
1026 tdb_flags |= TDB_NOMMAP;
1028 tdb_flags |= TDB_DISALLOW_NESTING;
1030 recdb = tdb_wrap_open(mem_ctx, name, ctdb->tunable.database_hash_size,
1031 tdb_flags, O_RDWR|O_CREAT|O_EXCL, 0600);
1032 if (recdb == NULL) {
1033 DEBUG(DEBUG_CRIT,(__location__ " Failed to create temp recovery database '%s'\n", name));
1043 a traverse function for pulling all relevent records from recdb
1046 struct ctdb_context *ctdb;
1047 struct ctdb_marshall_buffer *recdata;
1053 static int traverse_recdb(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *p)
1055 struct recdb_data *params = (struct recdb_data *)p;
1056 struct ctdb_rec_data *rec;
1057 struct ctdb_ltdb_header *hdr;
1059 /* skip empty records */
1060 if (data.dsize <= sizeof(struct ctdb_ltdb_header)) {
1064 /* update the dmaster field to point to us */
1065 hdr = (struct ctdb_ltdb_header *)data.dptr;
1066 if (!params->persistent) {
1067 hdr->dmaster = params->ctdb->pnn;
1070 /* add the record to the blob ready to send to the nodes */
1071 rec = ctdb_marshall_record(params->recdata, 0, key, NULL, data);
1073 params->failed = true;
1076 params->recdata = talloc_realloc_size(NULL, params->recdata, rec->length + params->len);
1077 if (params->recdata == NULL) {
1078 DEBUG(DEBUG_CRIT,(__location__ " Failed to expand recdata to %u (%u records)\n",
1079 rec->length + params->len, params->recdata->count));
1080 params->failed = true;
1083 params->recdata->count++;
1084 memcpy(params->len+(uint8_t *)params->recdata, rec, rec->length);
1085 params->len += rec->length;
1092 push the recdb database out to all nodes
1094 static int push_recdb_database(struct ctdb_context *ctdb, uint32_t dbid,
1096 struct tdb_wrap *recdb, struct ctdb_node_map *nodemap)
1098 struct recdb_data params;
1099 struct ctdb_marshall_buffer *recdata;
1101 TALLOC_CTX *tmp_ctx;
1104 tmp_ctx = talloc_new(ctdb);
1105 CTDB_NO_MEMORY(ctdb, tmp_ctx);
1107 recdata = talloc_zero(recdb, struct ctdb_marshall_buffer);
1108 CTDB_NO_MEMORY(ctdb, recdata);
1110 recdata->db_id = dbid;
1113 params.recdata = recdata;
1114 params.len = offsetof(struct ctdb_marshall_buffer, data);
1115 params.failed = false;
1116 params.persistent = persistent;
1118 if (tdb_traverse_read(recdb->tdb, traverse_recdb, ¶ms) == -1) {
1119 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1120 talloc_free(params.recdata);
1121 talloc_free(tmp_ctx);
1125 if (params.failed) {
1126 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse recdb database\n"));
1127 talloc_free(params.recdata);
1128 talloc_free(tmp_ctx);
1132 recdata = params.recdata;
1134 outdata.dptr = (void *)recdata;
1135 outdata.dsize = params.len;
1137 nodes = list_of_active_nodes(ctdb, nodemap, tmp_ctx, true);
1138 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_PUSH_DB,
1140 CONTROL_TIMEOUT(), false, outdata,
1143 DEBUG(DEBUG_ERR,(__location__ " Failed to push recdb records to nodes for db 0x%x\n", dbid));
1144 talloc_free(recdata);
1145 talloc_free(tmp_ctx);
1149 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pushed remote database 0x%x of size %u\n",
1150 dbid, recdata->count));
1152 talloc_free(recdata);
1153 talloc_free(tmp_ctx);
1160 go through a full recovery on one database
1162 static int recover_database(struct ctdb_recoverd *rec,
1163 TALLOC_CTX *mem_ctx,
1167 struct ctdb_node_map *nodemap,
1168 uint32_t transaction_id)
1170 struct tdb_wrap *recdb;
1172 struct ctdb_context *ctdb = rec->ctdb;
1174 struct ctdb_control_wipe_database w;
1177 recdb = create_recdb(ctdb, mem_ctx);
1178 if (recdb == NULL) {
1182 /* pull all remote databases onto the recdb */
1183 ret = pull_remote_database(ctdb, rec, nodemap, recdb, dbid, persistent);
1185 DEBUG(DEBUG_ERR, (__location__ " Unable to pull remote database 0x%x\n", dbid));
1189 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - pulled remote database 0x%x\n", dbid));
1191 /* wipe all the remote databases. This is safe as we are in a transaction */
1193 w.transaction_id = transaction_id;
1195 data.dptr = (void *)&w;
1196 data.dsize = sizeof(w);
1198 nodes = list_of_active_nodes(ctdb, nodemap, recdb, true);
1199 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_WIPE_DATABASE,
1201 CONTROL_TIMEOUT(), false, data,
1204 DEBUG(DEBUG_ERR, (__location__ " Unable to wipe database. Recovery failed.\n"));
1209 /* push out the correct database. This sets the dmaster and skips
1210 the empty records */
1211 ret = push_recdb_database(ctdb, dbid, persistent, recdb, nodemap);
1217 /* all done with this database */
1224 reload the nodes file
1226 static void reload_nodes_file(struct ctdb_context *ctdb)
1229 ctdb_load_nodes_file(ctdb);
1232 static int ctdb_reload_remote_public_ips(struct ctdb_context *ctdb,
1233 struct ctdb_recoverd *rec,
1234 struct ctdb_node_map *nodemap,
1240 if (ctdb->num_nodes != nodemap->num) {
1241 DEBUG(DEBUG_ERR, (__location__ " ctdb->num_nodes (%d) != nodemap->num (%d) invalid param\n",
1242 ctdb->num_nodes, nodemap->num));
1244 *culprit = ctdb->pnn;
1249 for (j=0; j<nodemap->num; j++) {
1250 /* release any existing data */
1251 if (ctdb->nodes[j]->known_public_ips) {
1252 talloc_free(ctdb->nodes[j]->known_public_ips);
1253 ctdb->nodes[j]->known_public_ips = NULL;
1255 if (ctdb->nodes[j]->available_public_ips) {
1256 talloc_free(ctdb->nodes[j]->available_public_ips);
1257 ctdb->nodes[j]->available_public_ips = NULL;
1260 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
1264 /* grab a new shiny list of public ips from the node */
1265 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1267 ctdb->nodes[j]->pnn,
1270 &ctdb->nodes[j]->known_public_ips);
1272 DEBUG(DEBUG_ERR,("Failed to read known public ips from node : %u\n",
1273 ctdb->nodes[j]->pnn));
1275 *culprit = ctdb->nodes[j]->pnn;
1280 if (ctdb->tunable.disable_ip_failover == 0) {
1281 if (rec->ip_check_disable_ctx == NULL) {
1282 if (verify_remote_ip_allocation(ctdb, ctdb->nodes[j]->known_public_ips)) {
1283 DEBUG(DEBUG_ERR,("Node %d has inconsistent public ip allocation and needs update.\n", ctdb->nodes[j]->pnn));
1284 rec->need_takeover_run = true;
1289 /* grab a new shiny list of public ips from the node */
1290 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1292 ctdb->nodes[j]->pnn,
1294 CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE,
1295 &ctdb->nodes[j]->available_public_ips);
1297 DEBUG(DEBUG_ERR,("Failed to read available public ips from node : %u\n",
1298 ctdb->nodes[j]->pnn));
1300 *culprit = ctdb->nodes[j]->pnn;
1309 /* when we start a recovery, make sure all nodes use the same reclock file
1312 static int sync_recovery_lock_file_across_cluster(struct ctdb_recoverd *rec)
1314 struct ctdb_context *ctdb = rec->ctdb;
1315 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1319 if (ctdb->recovery_lock_file == NULL) {
1323 data.dsize = strlen(ctdb->recovery_lock_file) + 1;
1324 data.dptr = (uint8_t *)ctdb->recovery_lock_file;
1327 nodes = list_of_active_nodes(ctdb, rec->nodemap, tmp_ctx, true);
1328 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_SET_RECLOCK_FILE,
1334 DEBUG(DEBUG_ERR, (__location__ " Failed to sync reclock file settings\n"));
1335 talloc_free(tmp_ctx);
1339 talloc_free(tmp_ctx);
1345 we are the recmaster, and recovery is needed - start a recovery run
1347 static int do_recovery(struct ctdb_recoverd *rec,
1348 TALLOC_CTX *mem_ctx, uint32_t pnn,
1349 struct ctdb_node_map *nodemap, struct ctdb_vnn_map *vnnmap)
1351 struct ctdb_context *ctdb = rec->ctdb;
1353 uint32_t generation;
1354 struct ctdb_dbid_map *dbmap;
1357 struct timeval start_time;
1358 uint32_t culprit = (uint32_t)-1;
1360 DEBUG(DEBUG_NOTICE, (__location__ " Starting do_recovery\n"));
1362 /* if recovery fails, force it again */
1363 rec->need_recovery = true;
1365 for (i=0; i<ctdb->num_nodes; i++) {
1366 struct ctdb_banning_state *ban_state;
1368 if (ctdb->nodes[i]->ban_state == NULL) {
1371 ban_state = (struct ctdb_banning_state *)ctdb->nodes[i]->ban_state;
1372 if (ban_state->count < 2*ctdb->num_nodes) {
1375 DEBUG(DEBUG_NOTICE,("Node %u has caused %u recoveries recently - banning it for %u seconds\n",
1376 ctdb->nodes[i]->pnn, ban_state->count,
1377 ctdb->tunable.recovery_ban_period));
1378 ctdb_ban_node(rec, ctdb->nodes[i]->pnn, ctdb->tunable.recovery_ban_period);
1379 ban_state->count = 0;
1383 if (ctdb->tunable.verify_recovery_lock != 0) {
1384 DEBUG(DEBUG_ERR,("Taking out recovery lock from recovery daemon\n"));
1385 start_time = timeval_current();
1386 if (!ctdb_recovery_lock(ctdb, true)) {
1387 DEBUG(DEBUG_ERR,("Unable to get recovery lock - aborting recovery "
1388 "and ban ourself for %u seconds\n",
1389 ctdb->tunable.recovery_ban_period));
1390 ctdb_ban_node(rec, pnn, ctdb->tunable.recovery_ban_period);
1393 ctdb_ctrl_report_recd_lock_latency(ctdb, CONTROL_TIMEOUT(), timeval_elapsed(&start_time));
1394 DEBUG(DEBUG_NOTICE,("Recovery lock taken successfully by recovery daemon\n"));
1397 DEBUG(DEBUG_NOTICE, (__location__ " Recovery initiated due to problem with node %u\n", rec->last_culprit_node));
1399 /* get a list of all databases */
1400 ret = ctdb_ctrl_getdbmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, &dbmap);
1402 DEBUG(DEBUG_ERR, (__location__ " Unable to get dbids from node :%u\n", pnn));
1406 /* we do the db creation before we set the recovery mode, so the freeze happens
1407 on all databases we will be dealing with. */
1409 /* verify that we have all the databases any other node has */
1410 ret = create_missing_local_databases(ctdb, nodemap, pnn, &dbmap, mem_ctx);
1412 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing local databases\n"));
1416 /* verify that all other nodes have all our databases */
1417 ret = create_missing_remote_databases(ctdb, nodemap, pnn, dbmap, mem_ctx);
1419 DEBUG(DEBUG_ERR, (__location__ " Unable to create missing remote databases\n"));
1422 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - created remote databases\n"));
1424 /* update the database priority for all remote databases */
1425 ret = update_db_priority_on_remote_nodes(ctdb, nodemap, pnn, dbmap, mem_ctx);
1427 DEBUG(DEBUG_ERR, (__location__ " Unable to set db priority on remote nodes\n"));
1429 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated db priority for all databases\n"));
1432 /* update all other nodes to use the same setting for reclock files
1433 as the local recovery master.
1435 sync_recovery_lock_file_across_cluster(rec);
1437 /* set recovery mode to active on all nodes */
1438 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
1440 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
1444 /* execute the "startrecovery" event script on all nodes */
1445 ret = run_startrecovery_eventscript(rec, nodemap);
1447 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
1452 update all nodes to have the same flags that we have
1454 for (i=0;i<nodemap->num;i++) {
1455 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1459 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1461 DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1466 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1468 /* pick a new generation number */
1469 generation = new_generation();
1471 /* change the vnnmap on this node to use the new generation
1472 number but not on any other nodes.
1473 this guarantees that if we abort the recovery prematurely
1474 for some reason (a node stops responding?)
1475 that we can just return immediately and we will reenter
1476 recovery shortly again.
1477 I.e. we deliberately leave the cluster with an inconsistent
1478 generation id to allow us to abort recovery at any stage and
1479 just restart it from scratch.
1481 vnnmap->generation = generation;
1482 ret = ctdb_ctrl_setvnnmap(ctdb, CONTROL_TIMEOUT(), pnn, mem_ctx, vnnmap);
1484 DEBUG(DEBUG_ERR, (__location__ " Unable to set vnnmap for node %u\n", pnn));
1488 data.dptr = (void *)&generation;
1489 data.dsize = sizeof(uint32_t);
1491 nodes = list_of_active_nodes(ctdb, nodemap, mem_ctx, true);
1492 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_START,
1494 CONTROL_TIMEOUT(), false, data,
1496 transaction_start_fail_callback,
1498 DEBUG(DEBUG_ERR, (__location__ " Unable to start transactions. Recovery failed.\n"));
1499 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_CANCEL,
1501 CONTROL_TIMEOUT(), false, tdb_null,
1505 DEBUG(DEBUG_ERR,("Failed to cancel recovery transaction\n"));
1510 DEBUG(DEBUG_NOTICE,(__location__ " started transactions on all nodes\n"));
1512 for (i=0;i<dbmap->num;i++) {
1513 ret = recover_database(rec, mem_ctx,
1515 dbmap->dbs[i].persistent,
1516 pnn, nodemap, generation);
1518 DEBUG(DEBUG_ERR, (__location__ " Failed to recover database 0x%x\n", dbmap->dbs[i].dbid));
1523 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - starting database commits\n"));
1525 /* commit all the changes */
1526 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_TRANSACTION_COMMIT,
1528 CONTROL_TIMEOUT(), false, data,
1531 DEBUG(DEBUG_ERR, (__location__ " Unable to commit recovery changes. Recovery failed.\n"));
1535 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - committed databases\n"));
1538 /* update the capabilities for all nodes */
1539 ret = update_capabilities(ctdb, nodemap);
1541 DEBUG(DEBUG_ERR, (__location__ " Unable to update node capabilities.\n"));
1545 /* build a new vnn map with all the currently active and
1547 generation = new_generation();
1548 vnnmap = talloc(mem_ctx, struct ctdb_vnn_map);
1549 CTDB_NO_MEMORY(ctdb, vnnmap);
1550 vnnmap->generation = generation;
1552 vnnmap->map = talloc_zero_array(vnnmap, uint32_t, vnnmap->size);
1553 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1554 for (i=j=0;i<nodemap->num;i++) {
1555 if (nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE) {
1558 if (!(ctdb->nodes[i]->capabilities & CTDB_CAP_LMASTER)) {
1559 /* this node can not be an lmaster */
1560 DEBUG(DEBUG_DEBUG, ("Node %d cant be a LMASTER, skipping it\n", i));
1565 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1566 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1567 vnnmap->map[j++] = nodemap->nodes[i].pnn;
1570 if (vnnmap->size == 0) {
1571 DEBUG(DEBUG_NOTICE, ("No suitable lmasters found. Adding local node (recmaster) anyway.\n"));
1573 vnnmap->map = talloc_realloc(vnnmap, vnnmap->map, uint32_t, vnnmap->size);
1574 CTDB_NO_MEMORY(ctdb, vnnmap->map);
1575 vnnmap->map[0] = pnn;
1578 /* update to the new vnnmap on all nodes */
1579 ret = update_vnnmap_on_all_nodes(ctdb, nodemap, pnn, vnnmap, mem_ctx);
1581 DEBUG(DEBUG_ERR, (__location__ " Unable to update vnnmap on all nodes\n"));
1585 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated vnnmap\n"));
1587 /* update recmaster to point to us for all nodes */
1588 ret = set_recovery_master(ctdb, nodemap, pnn);
1590 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery master\n"));
1594 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated recmaster\n"));
1597 update all nodes to have the same flags that we have
1599 for (i=0;i<nodemap->num;i++) {
1600 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1604 ret = update_flags_on_all_nodes(ctdb, nodemap, i, nodemap->nodes[i].flags);
1606 DEBUG(DEBUG_ERR, (__location__ " Unable to update flags on all nodes for node %d\n", i));
1611 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - updated flags\n"));
1613 /* disable recovery mode */
1614 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_NORMAL);
1616 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to normal on cluster\n"));
1620 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - disabled recovery mode\n"));
1623 tell nodes to takeover their public IPs
1625 ret = ctdb_reload_remote_public_ips(ctdb, rec, nodemap, &culprit);
1627 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
1631 rec->need_takeover_run = false;
1632 ret = ctdb_takeover_run(ctdb, nodemap);
1634 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses\n"));
1637 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - takeip finished\n"));
1639 /* execute the "recovered" event script on all nodes */
1640 ret = run_recovered_eventscript(ctdb, nodemap, "do_recovery");
1642 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Recovery process failed.\n"));
1646 DEBUG(DEBUG_NOTICE, (__location__ " Recovery - finished the recovered event\n"));
1648 /* send a message to all clients telling them that the cluster
1649 has been reconfigured */
1650 ctdb_client_send_message(ctdb, CTDB_BROADCAST_CONNECTED, CTDB_SRVID_RECONFIGURE, tdb_null);
1652 DEBUG(DEBUG_NOTICE, (__location__ " Recovery complete\n"));
1654 rec->need_recovery = false;
1656 /* we managed to complete a full recovery, make sure to forgive
1657 any past sins by the nodes that could now participate in the
1660 DEBUG(DEBUG_ERR,("Resetting ban count to 0 for all nodes\n"));
1661 for (i=0;i<nodemap->num;i++) {
1662 struct ctdb_banning_state *ban_state;
1664 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
1668 ban_state = (struct ctdb_banning_state *)ctdb->nodes[nodemap->nodes[i].pnn]->ban_state;
1669 if (ban_state == NULL) {
1673 ban_state->count = 0;
1677 /* We just finished a recovery successfully.
1678 We now wait for rerecovery_timeout before we allow
1679 another recovery to take place.
1681 DEBUG(DEBUG_NOTICE, ("Just finished a recovery. New recoveries will now be supressed for the rerecovery timeout (%d seconds)\n", ctdb->tunable.rerecovery_timeout));
1682 ctdb_wait_timeout(ctdb, ctdb->tunable.rerecovery_timeout);
1683 DEBUG(DEBUG_NOTICE, ("The rerecovery timeout has elapsed. We now allow recoveries to trigger again.\n"));
1690 elections are won by first checking the number of connected nodes, then
1691 the priority time, then the pnn
1693 struct election_message {
1694 uint32_t num_connected;
1695 struct timeval priority_time;
1697 uint32_t node_flags;
1701 form this nodes election data
1703 static void ctdb_election_data(struct ctdb_recoverd *rec, struct election_message *em)
1706 struct ctdb_node_map *nodemap;
1707 struct ctdb_context *ctdb = rec->ctdb;
1711 em->pnn = rec->ctdb->pnn;
1712 em->priority_time = rec->priority_time;
1714 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, rec, &nodemap);
1716 DEBUG(DEBUG_ERR,(__location__ " unable to get election data\n"));
1720 rec->node_flags = nodemap->nodes[ctdb->pnn].flags;
1721 em->node_flags = rec->node_flags;
1723 for (i=0;i<nodemap->num;i++) {
1724 if (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED)) {
1725 em->num_connected++;
1729 /* we shouldnt try to win this election if we cant be a recmaster */
1730 if ((ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1731 em->num_connected = 0;
1732 em->priority_time = timeval_current();
1735 talloc_free(nodemap);
1739 see if the given election data wins
1741 static bool ctdb_election_win(struct ctdb_recoverd *rec, struct election_message *em)
1743 struct election_message myem;
1746 ctdb_election_data(rec, &myem);
1748 /* we cant win if we dont have the recmaster capability */
1749 if ((rec->ctdb->capabilities & CTDB_CAP_RECMASTER) == 0) {
1753 /* we cant win if we are banned */
1754 if (rec->node_flags & NODE_FLAGS_BANNED) {
1758 /* we cant win if we are stopped */
1759 if (rec->node_flags & NODE_FLAGS_STOPPED) {
1763 /* we will automatically win if the other node is banned */
1764 if (em->node_flags & NODE_FLAGS_BANNED) {
1768 /* we will automatically win if the other node is banned */
1769 if (em->node_flags & NODE_FLAGS_STOPPED) {
1773 /* try to use the most connected node */
1775 cmp = (int)myem.num_connected - (int)em->num_connected;
1778 /* then the longest running node */
1780 cmp = timeval_compare(&em->priority_time, &myem.priority_time);
1784 cmp = (int)myem.pnn - (int)em->pnn;
1791 send out an election request
1793 static int send_election_request(struct ctdb_recoverd *rec, uint32_t pnn, bool update_recmaster)
1796 TDB_DATA election_data;
1797 struct election_message emsg;
1799 struct ctdb_context *ctdb = rec->ctdb;
1801 srvid = CTDB_SRVID_RECOVERY;
1803 ctdb_election_data(rec, &emsg);
1805 election_data.dsize = sizeof(struct election_message);
1806 election_data.dptr = (unsigned char *)&emsg;
1809 /* send an election message to all active nodes */
1810 DEBUG(DEBUG_INFO,(__location__ " Send election request to all active nodes\n"));
1811 ctdb_client_send_message(ctdb, CTDB_BROADCAST_ALL, srvid, election_data);
1814 /* A new node that is already frozen has entered the cluster.
1815 The existing nodes are not frozen and dont need to be frozen
1816 until the election has ended and we start the actual recovery
1818 if (update_recmaster == true) {
1819 /* first we assume we will win the election and set
1820 recoverymaster to be ourself on the current node
1822 ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), pnn, pnn);
1824 DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request\n"));
1834 this function will unban all nodes in the cluster
1836 static void unban_all_nodes(struct ctdb_context *ctdb)
1839 struct ctdb_node_map *nodemap;
1840 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
1842 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
1844 DEBUG(DEBUG_ERR,(__location__ " failed to get nodemap to unban all nodes\n"));
1848 for (i=0;i<nodemap->num;i++) {
1849 if ( (!(nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED))
1850 && (nodemap->nodes[i].flags & NODE_FLAGS_BANNED) ) {
1851 ctdb_ctrl_modflags(ctdb, CONTROL_TIMEOUT(), nodemap->nodes[i].pnn, 0, NODE_FLAGS_BANNED);
1855 talloc_free(tmp_ctx);
1860 we think we are winning the election - send a broadcast election request
1862 static void election_send_request(struct event_context *ev, struct timed_event *te, struct timeval t, void *p)
1864 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
1867 ret = send_election_request(rec, ctdb_get_pnn(rec->ctdb), false);
1869 DEBUG(DEBUG_ERR,("Failed to send election request!\n"));
1872 talloc_free(rec->send_election_te);
1873 rec->send_election_te = NULL;
1877 handler for memory dumps
1879 static void mem_dump_handler(struct ctdb_context *ctdb, uint64_t srvid,
1880 TDB_DATA data, void *private_data)
1882 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
1885 struct rd_memdump_reply *rd;
1887 if (data.dsize != sizeof(struct rd_memdump_reply)) {
1888 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
1889 talloc_free(tmp_ctx);
1892 rd = (struct rd_memdump_reply *)data.dptr;
1894 dump = talloc_zero(tmp_ctx, TDB_DATA);
1896 DEBUG(DEBUG_ERR, (__location__ " Failed to allocate memory for memdump\n"));
1897 talloc_free(tmp_ctx);
1900 ret = ctdb_dump_memory(ctdb, dump);
1902 DEBUG(DEBUG_ERR, (__location__ " ctdb_dump_memory() failed\n"));
1903 talloc_free(tmp_ctx);
1907 DEBUG(DEBUG_ERR, ("recovery master memory dump\n"));
1909 ret = ctdb_client_send_message(ctdb, rd->pnn, rd->srvid, *dump);
1911 DEBUG(DEBUG_ERR,("Failed to send rd memdump reply message\n"));
1912 talloc_free(tmp_ctx);
1916 talloc_free(tmp_ctx);
1920 handler for reload_nodes
1922 static void reload_nodes_handler(struct ctdb_context *ctdb, uint64_t srvid,
1923 TDB_DATA data, void *private_data)
1925 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1927 DEBUG(DEBUG_ERR, (__location__ " Reload nodes file from recovery daemon\n"));
1929 reload_nodes_file(rec->ctdb);
1933 static void reenable_ip_check(struct event_context *ev, struct timed_event *te,
1934 struct timeval yt, void *p)
1936 struct ctdb_recoverd *rec = talloc_get_type(p, struct ctdb_recoverd);
1938 talloc_free(rec->ip_check_disable_ctx);
1939 rec->ip_check_disable_ctx = NULL;
1943 static void recd_update_ip_handler(struct ctdb_context *ctdb, uint64_t srvid,
1944 TDB_DATA data, void *private_data)
1946 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1947 struct ctdb_public_ip *ip;
1949 if (rec->recmaster != rec->ctdb->pnn) {
1950 DEBUG(DEBUG_INFO,("Not recmaster, ignore update ip message\n"));
1954 if (data.dsize != sizeof(struct ctdb_public_ip)) {
1955 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)));
1959 ip = (struct ctdb_public_ip *)data.dptr;
1961 update_ip_assignment_tree(rec->ctdb, ip);
1965 static void disable_ip_check_handler(struct ctdb_context *ctdb, uint64_t srvid,
1966 TDB_DATA data, void *private_data)
1968 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
1971 if (rec->ip_check_disable_ctx != NULL) {
1972 talloc_free(rec->ip_check_disable_ctx);
1973 rec->ip_check_disable_ctx = NULL;
1976 if (data.dsize != sizeof(uint32_t)) {
1977 DEBUG(DEBUG_ERR,(__location__ " Wrong size for data :%lu "
1978 "expexting %lu\n", (long unsigned)data.dsize,
1979 (long unsigned)sizeof(uint32_t)));
1982 if (data.dptr == NULL) {
1983 DEBUG(DEBUG_ERR,(__location__ " No data recaived\n"));
1987 timeout = *((uint32_t *)data.dptr);
1988 DEBUG(DEBUG_NOTICE,("Disabling ip check for %u seconds\n", timeout));
1990 rec->ip_check_disable_ctx = talloc_new(rec);
1991 CTDB_NO_MEMORY_VOID(ctdb, rec->ip_check_disable_ctx);
1993 event_add_timed(ctdb->ev, rec->ip_check_disable_ctx, timeval_current_ofs(timeout, 0), reenable_ip_check, rec);
1998 handler for ip reallocate, just add it to the list of callers and
1999 handle this later in the monitor_cluster loop so we do not recurse
2000 with other callers to takeover_run()
2002 static void ip_reallocate_handler(struct ctdb_context *ctdb, uint64_t srvid,
2003 TDB_DATA data, void *private_data)
2005 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2006 struct ip_reallocate_list *caller;
2008 if (data.dsize != sizeof(struct rd_memdump_reply)) {
2009 DEBUG(DEBUG_ERR, (__location__ " Wrong size of return address.\n"));
2013 if (rec->ip_reallocate_ctx == NULL) {
2014 rec->ip_reallocate_ctx = talloc_new(rec);
2015 CTDB_NO_MEMORY_FATAL(ctdb, rec->ip_reallocate_ctx);
2018 caller = talloc(rec->ip_reallocate_ctx, struct ip_reallocate_list);
2019 CTDB_NO_MEMORY_FATAL(ctdb, caller);
2021 caller->rd = (struct rd_memdump_reply *)talloc_steal(caller, data.dptr);
2022 caller->next = rec->reallocate_callers;
2023 rec->reallocate_callers = caller;
2028 static void process_ipreallocate_requests(struct ctdb_context *ctdb, struct ctdb_recoverd *rec)
2030 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2033 struct ip_reallocate_list *callers;
2036 DEBUG(DEBUG_INFO, ("recovery master forced ip reallocation\n"));
2038 /* update the list of public ips that a node can handle for
2041 ret = ctdb_reload_remote_public_ips(ctdb, rec, rec->nodemap, &culprit);
2043 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
2045 rec->need_takeover_run = true;
2048 ret = ctdb_takeover_run(ctdb, rec->nodemap);
2050 DEBUG(DEBUG_ERR,("Failed to read public ips from remote node %d\n",
2052 rec->need_takeover_run = true;
2056 result.dsize = sizeof(int32_t);
2057 result.dptr = (uint8_t *)&ret;
2059 for (callers=rec->reallocate_callers; callers; callers=callers->next) {
2061 /* Someone that sent srvid==0 does not want a reply */
2062 if (callers->rd->srvid == 0) {
2065 DEBUG(DEBUG_INFO,("Sending ip reallocate reply message to "
2066 "%u:%llu\n", (unsigned)callers->rd->pnn,
2067 (unsigned long long)callers->rd->srvid));
2068 ret = ctdb_client_send_message(ctdb, callers->rd->pnn, callers->rd->srvid, result);
2070 DEBUG(DEBUG_ERR,("Failed to send ip reallocate reply "
2071 "message to %u:%llu\n",
2072 (unsigned)callers->rd->pnn,
2073 (unsigned long long)callers->rd->srvid));
2077 talloc_free(tmp_ctx);
2078 talloc_free(rec->ip_reallocate_ctx);
2079 rec->ip_reallocate_ctx = NULL;
2080 rec->reallocate_callers = NULL;
2086 handler for recovery master elections
2088 static void election_handler(struct ctdb_context *ctdb, uint64_t srvid,
2089 TDB_DATA data, void *private_data)
2091 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2093 struct election_message *em = (struct election_message *)data.dptr;
2094 TALLOC_CTX *mem_ctx;
2096 /* we got an election packet - update the timeout for the election */
2097 talloc_free(rec->election_timeout);
2098 rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2100 timeval_current_ofs(0, 500000) :
2101 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2102 ctdb_election_timeout, rec);
2104 mem_ctx = talloc_new(ctdb);
2106 /* someone called an election. check their election data
2107 and if we disagree and we would rather be the elected node,
2108 send a new election message to all other nodes
2110 if (ctdb_election_win(rec, em)) {
2111 if (!rec->send_election_te) {
2112 rec->send_election_te = event_add_timed(ctdb->ev, rec,
2113 timeval_current_ofs(0, 500000),
2114 election_send_request, rec);
2116 talloc_free(mem_ctx);
2117 /*unban_all_nodes(ctdb);*/
2122 talloc_free(rec->send_election_te);
2123 rec->send_election_te = NULL;
2125 if (ctdb->tunable.verify_recovery_lock != 0) {
2126 /* release the recmaster lock */
2127 if (em->pnn != ctdb->pnn &&
2128 ctdb->recovery_lock_fd != -1) {
2129 close(ctdb->recovery_lock_fd);
2130 ctdb->recovery_lock_fd = -1;
2131 unban_all_nodes(ctdb);
2135 /* ok, let that guy become recmaster then */
2136 ret = ctdb_ctrl_setrecmaster(ctdb, CONTROL_TIMEOUT(), ctdb_get_pnn(ctdb), em->pnn);
2138 DEBUG(DEBUG_ERR, (__location__ " failed to send recmaster election request"));
2139 talloc_free(mem_ctx);
2143 talloc_free(mem_ctx);
2149 force the start of the election process
2151 static void force_election(struct ctdb_recoverd *rec, uint32_t pnn,
2152 struct ctdb_node_map *nodemap)
2155 struct ctdb_context *ctdb = rec->ctdb;
2157 DEBUG(DEBUG_INFO,(__location__ " Force an election\n"));
2159 /* set all nodes to recovery mode to stop all internode traffic */
2160 ret = set_recovery_mode(ctdb, rec, nodemap, CTDB_RECOVERY_ACTIVE);
2162 DEBUG(DEBUG_ERR, (__location__ " Unable to set recovery mode to active on cluster\n"));
2166 talloc_free(rec->election_timeout);
2167 rec->election_timeout = event_add_timed(ctdb->ev, ctdb,
2169 timeval_current_ofs(0, 500000) :
2170 timeval_current_ofs(ctdb->tunable.election_timeout, 0),
2171 ctdb_election_timeout, rec);
2173 ret = send_election_request(rec, pnn, true);
2175 DEBUG(DEBUG_ERR, (__location__ " failed to initiate recmaster election"));
2179 /* wait for a few seconds to collect all responses */
2180 ctdb_wait_election(rec);
2186 handler for when a node changes its flags
2188 static void monitor_handler(struct ctdb_context *ctdb, uint64_t srvid,
2189 TDB_DATA data, void *private_data)
2192 struct ctdb_node_flag_change *c = (struct ctdb_node_flag_change *)data.dptr;
2193 struct ctdb_node_map *nodemap=NULL;
2194 TALLOC_CTX *tmp_ctx;
2195 uint32_t changed_flags;
2197 struct ctdb_recoverd *rec = talloc_get_type(private_data, struct ctdb_recoverd);
2198 int disabled_flag_changed;
2200 if (data.dsize != sizeof(*c)) {
2201 DEBUG(DEBUG_ERR,(__location__ "Invalid data in ctdb_node_flag_change\n"));
2205 tmp_ctx = talloc_new(ctdb);
2206 CTDB_NO_MEMORY_VOID(ctdb, tmp_ctx);
2208 ret = ctdb_ctrl_getnodemap(ctdb, CONTROL_TIMEOUT(), CTDB_CURRENT_NODE, tmp_ctx, &nodemap);
2210 DEBUG(DEBUG_ERR,(__location__ "ctdb_ctrl_getnodemap failed in monitor_handler\n"));
2211 talloc_free(tmp_ctx);
2216 for (i=0;i<nodemap->num;i++) {
2217 if (nodemap->nodes[i].pnn == c->pnn) break;
2220 if (i == nodemap->num) {
2221 DEBUG(DEBUG_CRIT,(__location__ "Flag change for non-existant node %u\n", c->pnn));
2222 talloc_free(tmp_ctx);
2226 changed_flags = c->old_flags ^ c->new_flags;
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_NOTICE, (__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 = fork();
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 ctdb_set_culprit(rec, culprit);
3404 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3408 /* execute the "startrecovery" event script on all nodes */
3409 ret = run_startrecovery_eventscript(rec, nodemap);
3411 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'startrecovery' event on cluster\n"));
3412 ctdb_set_culprit(rec, ctdb->pnn);
3413 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3417 ret = ctdb_takeover_run(ctdb, nodemap);
3419 DEBUG(DEBUG_ERR, (__location__ " Unable to setup public takeover addresses - starting recovery\n"));
3420 ctdb_set_culprit(rec, ctdb->pnn);
3421 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3425 /* execute the "recovered" event script on all nodes */
3426 ret = run_recovered_eventscript(ctdb, nodemap, "monitor_cluster");
3428 // we cant check whether the event completed successfully
3429 // since this script WILL fail if the node is in recovery mode
3430 // and if that race happens, the code here would just cause a second
3431 // cascading recovery.
3433 DEBUG(DEBUG_ERR, (__location__ " Unable to run the 'recovered' event on cluster. Update of public ips failed.\n"));
3434 ctdb_set_culprit(rec, ctdb->pnn);
3435 do_recovery(rec, mem_ctx, pnn, nodemap, vnnmap);
3442 the main monitoring loop
3444 static void monitor_cluster(struct ctdb_context *ctdb)
3446 struct ctdb_recoverd *rec;
3448 DEBUG(DEBUG_NOTICE,("monitor_cluster starting\n"));
3450 rec = talloc_zero(ctdb, struct ctdb_recoverd);
3451 CTDB_NO_MEMORY_FATAL(ctdb, rec);
3455 rec->priority_time = timeval_current();
3457 /* register a message port for sending memory dumps */
3458 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_MEM_DUMP, mem_dump_handler, rec);
3460 /* register a message port for recovery elections */
3461 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RECOVERY, election_handler, rec);
3463 /* when nodes are disabled/enabled */
3464 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_SET_NODE_FLAGS, monitor_handler, rec);
3466 /* when we are asked to puch out a flag change */
3467 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_PUSH_NODE_FLAGS, push_flags_handler, rec);
3469 /* register a message port for vacuum fetch */
3470 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_VACUUM_FETCH, vacuum_fetch_handler, rec);
3472 /* register a message port for reloadnodes */
3473 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RELOAD_NODES, reload_nodes_handler, rec);
3475 /* register a message port for performing a takeover run */
3476 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_TAKEOVER_RUN, ip_reallocate_handler, rec);
3478 /* register a message port for disabling the ip check for a short while */
3479 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_DISABLE_IP_CHECK, disable_ip_check_handler, rec);
3481 /* register a message port for updating the recovery daemons node assignment for an ip */
3482 ctdb_client_set_message_handler(ctdb, CTDB_SRVID_RECD_UPDATE_IP, recd_update_ip_handler, rec);
3485 TALLOC_CTX *mem_ctx = talloc_new(ctdb);
3486 struct timeval start;
3490 DEBUG(DEBUG_CRIT,(__location__
3491 " Failed to create temp context\n"));
3495 start = timeval_current();
3496 main_loop(ctdb, rec, mem_ctx);
3497 talloc_free(mem_ctx);
3499 /* we only check for recovery once every second */
3500 elapsed = timeval_elapsed(&start);
3501 if (elapsed < ctdb->tunable.recover_interval) {
3502 ctdb_wait_timeout(ctdb, ctdb->tunable.recover_interval
3509 event handler for when the main ctdbd dies
3511 static void ctdb_recoverd_parent(struct event_context *ev, struct fd_event *fde,
3512 uint16_t flags, void *private_data)
3514 DEBUG(DEBUG_ALERT,("recovery daemon parent died - exiting\n"));
3519 called regularly to verify that the recovery daemon is still running
3521 static void ctdb_check_recd(struct event_context *ev, struct timed_event *te,
3522 struct timeval yt, void *p)
3524 struct ctdb_context *ctdb = talloc_get_type(p, struct ctdb_context);
3526 if (kill(ctdb->recoverd_pid, 0) != 0) {
3527 DEBUG(DEBUG_ERR,("Recovery daemon (pid:%d) is no longer running. Shutting down main daemon\n", (int)ctdb->recoverd_pid));
3529 ctdb_stop_recoverd(ctdb);
3530 ctdb_stop_keepalive(ctdb);
3531 ctdb_stop_monitoring(ctdb);
3532 ctdb_release_all_ips(ctdb);
3533 if (ctdb->methods != NULL) {
3534 ctdb->methods->shutdown(ctdb);
3536 ctdb_event_script(ctdb, CTDB_EVENT_SHUTDOWN);
3541 event_add_timed(ctdb->ev, ctdb,
3542 timeval_current_ofs(30, 0),
3543 ctdb_check_recd, ctdb);
3546 static void recd_sig_child_handler(struct event_context *ev,
3547 struct signal_event *se, int signum, int count,
3551 // struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
3556 pid = waitpid(-1, &status, WNOHANG);
3558 if (errno != ECHILD) {
3559 DEBUG(DEBUG_ERR, (__location__ " waitpid() returned error. errno:%s(%d)\n", strerror(errno),errno));
3564 DEBUG(DEBUG_DEBUG, ("RECD SIGCHLD from %d\n", (int)pid));
3570 startup the recovery daemon as a child of the main ctdb daemon
3572 int ctdb_start_recoverd(struct ctdb_context *ctdb)
3575 struct signal_event *se;
3576 struct tevent_fd *fde;
3578 if (pipe(fd) != 0) {
3582 ctdb->ctdbd_pid = getpid();
3584 ctdb->recoverd_pid = fork();
3585 if (ctdb->recoverd_pid == -1) {
3589 if (ctdb->recoverd_pid != 0) {
3591 event_add_timed(ctdb->ev, ctdb,
3592 timeval_current_ofs(30, 0),
3593 ctdb_check_recd, ctdb);
3599 srandom(getpid() ^ time(NULL));
3601 if (switch_from_server_to_client(ctdb, "recoverd") != 0) {
3602 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch recovery daemon into client mode. shutting down.\n"));
3606 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to recovery daemon\n", fd[0]));
3608 fde = event_add_fd(ctdb->ev, ctdb, fd[0], EVENT_FD_READ,
3609 ctdb_recoverd_parent, &fd[0]);
3610 tevent_fd_set_auto_close(fde);
3612 /* set up a handler to pick up sigchld */
3613 se = event_add_signal(ctdb->ev, ctdb,
3615 recd_sig_child_handler,
3618 DEBUG(DEBUG_CRIT,("Failed to set up signal handler for SIGCHLD in recovery daemon\n"));
3622 monitor_cluster(ctdb);
3624 DEBUG(DEBUG_ALERT,("ERROR: ctdb_recoverd finished!?\n"));
3629 shutdown the recovery daemon
3631 void ctdb_stop_recoverd(struct ctdb_context *ctdb)
3633 if (ctdb->recoverd_pid == 0) {
3637 DEBUG(DEBUG_NOTICE,("Shutting down recovery daemon\n"));
3638 kill(ctdb->recoverd_pid, SIGTERM);