4 Copyright (C) Ronnie Sahlberg 2009
5 Copyright (C) Michael Adam 2010-2013
6 Copyright (C) Stefan Metzmacher 2010-2011
8 This program is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3 of the License, or
11 (at your option) any later version.
13 This program is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with this program; if not, see <http://www.gnu.org/licenses/>.
23 #include "system/network.h"
24 #include "system/filesys.h"
25 #include "system/time.h"
30 #include "lib/tdb_wrap/tdb_wrap.h"
31 #include "lib/util/dlinklist.h"
32 #include "lib/util/debug.h"
33 #include "lib/util/samba_util.h"
34 #include "lib/util/sys_rw.h"
35 #include "lib/util/util_process.h"
37 #include "ctdb_private.h"
38 #include "ctdb_client.h"
40 #include "common/rb_tree.h"
41 #include "common/common.h"
42 #include "common/logging.h"
44 #define TIMELIMIT() timeval_current_ofs(10, 0)
46 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
48 struct ctdb_vacuum_child_context {
49 struct ctdb_vacuum_child_context *next, *prev;
50 struct ctdb_vacuum_handle *vacuum_handle;
51 /* fd child writes status to */
54 enum vacuum_child_status status;
55 struct timeval start_time;
58 struct ctdb_vacuum_handle {
59 struct ctdb_db_context *ctdb_db;
60 struct ctdb_vacuum_child_context *child_ctx;
61 uint32_t fast_path_count;
65 /* a list of records to possibly delete */
67 struct ctdb_context *ctdb;
68 struct ctdb_db_context *ctdb_db;
69 struct tdb_context *dest_db;
70 trbt_tree_t *delete_list;
71 struct ctdb_marshall_buffer **vacuum_fetch_list;
77 uint32_t added_to_vacuum_fetch_list;
78 uint32_t added_to_delete_list;
92 uint32_t remote_error;
105 /* this structure contains the information for one record to be deleted */
106 struct delete_record_data {
107 struct ctdb_context *ctdb;
108 struct ctdb_db_context *ctdb_db;
109 struct ctdb_ltdb_header hdr;
114 struct delete_records_list {
115 struct ctdb_marshall_buffer *records;
116 struct vacuum_data *vdata;
119 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
120 const struct ctdb_ltdb_header *hdr,
124 * Store key and header in a tree, indexed by the key hash.
126 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
127 struct ctdb_db_context *ctdb_db,
129 const struct ctdb_ltdb_header *hdr,
132 struct delete_record_data *dd;
136 len = offsetof(struct delete_record_data, keydata) + key.dsize;
138 dd = (struct delete_record_data *)talloc_size(tree, len);
140 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
143 talloc_set_name_const(dd, "struct delete_record_data");
146 dd->ctdb_db = ctdb_db;
147 dd->key.dsize = key.dsize;
148 dd->key.dptr = dd->keydata;
149 memcpy(dd->keydata, key.dptr, key.dsize);
153 hash = ctdb_hash(&key);
155 trbt_insert32(tree, hash, dd);
160 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
161 struct ctdb_ltdb_header *hdr)
163 struct ctdb_context *ctdb = vdata->ctdb;
164 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
168 hash = ctdb_hash(&key);
170 if (trbt_lookup32(vdata->delete_list, hash)) {
171 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
175 ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
182 vdata->count.delete_list.total++;
188 * Add a record to the list of records to be sent
189 * to their lmaster with VACUUM_FETCH.
191 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
194 struct ctdb_context *ctdb = vdata->ctdb;
196 struct ctdb_marshall_buffer *vfl;
198 lmaster = ctdb_lmaster(ctdb, &key);
200 vfl = vdata->vacuum_fetch_list[lmaster];
202 vfl = ctdb_marshall_add(ctdb, vfl, vfl->db_id, ctdb->pnn,
203 key, NULL, tdb_null);
205 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
206 vdata->traverse_error = true;
210 vdata->vacuum_fetch_list[lmaster] = vfl;
216 static void ctdb_vacuum_event(struct tevent_context *ev,
217 struct tevent_timer *te,
218 struct timeval t, void *private_data);
220 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
222 struct ctdb_ltdb_header *header =
223 (struct ctdb_ltdb_header *)private_data;
225 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
229 *header = *(struct ctdb_ltdb_header *)data.dptr;
235 * traverse function for gathering the records that can be deleted
237 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
240 struct vacuum_data *vdata = talloc_get_type(private_data,
242 struct ctdb_context *ctdb = vdata->ctdb;
243 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
245 struct ctdb_ltdb_header *hdr;
248 vdata->count.db_traverse.total++;
250 lmaster = ctdb_lmaster(ctdb, &key);
251 if (lmaster >= ctdb->num_nodes) {
252 vdata->count.db_traverse.error++;
253 DEBUG(DEBUG_CRIT, (__location__
254 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
257 (unsigned)ctdb->num_nodes,
258 (unsigned)ctdb_hash(&key)));
262 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
263 /* it is not a deleted record */
264 vdata->count.db_traverse.skipped++;
268 hdr = (struct ctdb_ltdb_header *)data.dptr;
270 if (hdr->dmaster != ctdb->pnn) {
271 vdata->count.db_traverse.skipped++;
276 * Add the record to this process's delete_queue for processing
277 * in the subsequent traverse in the fast vacuum run.
279 res = insert_record_into_delete_queue(ctdb_db, hdr, key);
281 vdata->count.db_traverse.error++;
283 vdata->count.db_traverse.scheduled++;
290 * traverse the tree of records to delete and marshall them into
293 static int delete_marshall_traverse(void *param, void *data)
295 struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
296 struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
297 struct ctdb_marshall_buffer *m;
299 m = ctdb_marshall_add(recs, recs->records, recs->records->db_id,
300 recs->records->db_id,
301 dd->key, &dd->hdr, tdb_null);
303 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
312 * traverse function for the traversal of the delete_queue,
313 * the fast-path vacuuming list.
315 * - If the record has been migrated off the node
316 * or has been revived (filled with data) on the node,
317 * then skip the record.
319 * - If the current node is the record's lmaster and it is
320 * a record that has never been migrated with data, then
321 * delete the record from the local tdb.
323 * - If the current node is the record's lmaster and it has
324 * been migrated with data, then schedule it for the normal
325 * vacuuming procedure (i.e. add it to the delete_list).
327 * - If the current node is NOT the record's lmaster then
328 * add it to the list of records that are to be sent to
329 * the lmaster with the VACUUM_FETCH message.
331 static int delete_queue_traverse(void *param, void *data)
333 struct delete_record_data *dd =
334 talloc_get_type(data, struct delete_record_data);
335 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
336 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
337 struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
339 struct ctdb_ltdb_header header;
341 uint32_t hash = ctdb_hash(&(dd->key));
343 vdata->count.delete_queue.total++;
345 res = tdb_chainlock_nonblock(ctdb_db->ltdb->tdb, dd->key);
347 vdata->count.delete_queue.error++;
351 res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
352 vacuum_record_parser, &header);
357 if (header.dmaster != ctdb->pnn) {
358 /* The record has been migrated off the node. Skip. */
362 if (header.rsn != dd->hdr.rsn) {
364 * The record has been migrated off the node and back again.
365 * But not requeued for deletion. Skip it.
371 * We are dmaster, and the record has no data, and it has
372 * not been migrated after it has been queued for deletion.
374 * At this stage, the record could still have been revived locally
375 * and last been written with empty data. This can only be
376 * fixed with the addition of an active or delete flag. (TODO)
379 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
381 if (lmaster != ctdb->pnn) {
382 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
386 (__location__ " Error adding record to list "
387 "of records to send to lmaster.\n"));
388 vdata->count.delete_queue.error++;
390 vdata->count.delete_queue.added_to_vacuum_fetch_list++;
395 /* use header->flags or dd->hdr.flags ?? */
396 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
397 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
401 (__location__ " Error adding record to list "
402 "of records for deletion on lmaster.\n"));
403 vdata->count.delete_queue.error++;
405 vdata->count.delete_queue.added_to_delete_list++;
408 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
412 (__location__ " Error deleting record with key "
413 "hash [0x%08x] from local data base db[%s].\n",
414 hash, ctdb_db->db_name));
415 vdata->count.delete_queue.error++;
420 (__location__ " Deleted record with key hash "
421 "[0x%08x] from local data base db[%s].\n",
422 hash, ctdb_db->db_name));
423 vdata->count.delete_queue.deleted++;
429 vdata->count.delete_queue.skipped++;
432 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
438 * Delete the records that we are lmaster and dmaster for and
439 * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
442 static int delete_record_traverse(void *param, void *data)
444 struct delete_record_data *dd =
445 talloc_get_type(data, struct delete_record_data);
446 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
447 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
448 struct ctdb_context *ctdb = ctdb_db->ctdb;
450 struct ctdb_ltdb_header header;
452 uint32_t hash = ctdb_hash(&(dd->key));
454 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
457 (__location__ " Error getting chainlock on record with "
458 "key hash [0x%08x] on database db[%s].\n",
459 hash, ctdb_db->db_name));
460 vdata->count.delete_list.local_error++;
461 vdata->count.delete_list.left--;
467 * Verify that the record is still empty, its RSN has not
468 * changed and that we are still its lmaster and dmaster.
471 res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
472 vacuum_record_parser, &header);
477 if (header.flags & CTDB_REC_RO_FLAGS) {
478 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
479 "on database db[%s] has read-only flags. "
481 hash, ctdb_db->db_name));
485 if (header.dmaster != ctdb->pnn) {
486 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
487 "on database db[%s] has been migrated away. "
489 hash, ctdb_db->db_name));
493 if (header.rsn != dd->hdr.rsn) {
495 * The record has been migrated off the node and back again.
496 * But not requeued for deletion. Skip it.
498 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
499 "on database db[%s] seems to have been "
500 "migrated away and back again (with empty "
501 "data). skipping.\n",
502 hash, ctdb_db->db_name));
506 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
508 if (lmaster != ctdb->pnn) {
509 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
510 "delete list (key hash [0x%08x], db[%s]). "
511 "Strange! skipping.\n",
512 hash, ctdb_db->db_name));
516 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
520 (__location__ " Error deleting record with key hash "
521 "[0x%08x] from local data base db[%s].\n",
522 hash, ctdb_db->db_name));
523 vdata->count.delete_list.local_error++;
528 (__location__ " Deleted record with key hash [0x%08x] from "
529 "local data base db[%s].\n", hash, ctdb_db->db_name));
531 vdata->count.delete_list.deleted++;
535 vdata->count.delete_list.skipped++;
538 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
541 vdata->count.delete_list.left--;
547 * Traverse the delete_queue.
548 * Records are either deleted directly or filled
549 * into the delete list or the vacuum fetch lists
550 * for further processing.
552 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
553 struct vacuum_data *vdata)
558 ret = trbt_traversearray32(ctdb_db->delete_queue, 1,
559 delete_queue_traverse, vdata);
562 DEBUG(DEBUG_ERR, (__location__ " Error traversing "
563 "the delete queue.\n"));
566 sum = vdata->count.delete_queue.deleted
567 + vdata->count.delete_queue.skipped
568 + vdata->count.delete_queue.error
569 + vdata->count.delete_queue.added_to_delete_list
570 + vdata->count.delete_queue.added_to_vacuum_fetch_list;
572 if (vdata->count.delete_queue.total != sum) {
573 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
574 "counts for db[%s]: total[%u] != sum[%u]\n",
576 (unsigned)vdata->count.delete_queue.total,
580 if (vdata->count.delete_queue.total > 0) {
583 " fast vacuuming delete_queue traverse statistics: "
592 (unsigned)vdata->count.delete_queue.total,
593 (unsigned)vdata->count.delete_queue.deleted,
594 (unsigned)vdata->count.delete_queue.skipped,
595 (unsigned)vdata->count.delete_queue.error,
596 (unsigned)vdata->count.delete_queue.added_to_delete_list,
597 (unsigned)vdata->count.delete_queue.added_to_vacuum_fetch_list));
604 * read-only traverse of the database, looking for records that
605 * might be able to be vacuumed.
607 * This is not done each time but only every tunable
608 * VacuumFastPathCount times.
610 static void ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
611 struct vacuum_data *vdata)
615 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
616 if (ret == -1 || vdata->traverse_error) {
617 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
618 "'%s'\n", ctdb_db->db_name));
622 if (vdata->count.db_traverse.total > 0) {
625 " full vacuuming db traverse statistics: "
632 (unsigned)vdata->count.db_traverse.total,
633 (unsigned)vdata->count.db_traverse.skipped,
634 (unsigned)vdata->count.db_traverse.error,
635 (unsigned)vdata->count.db_traverse.scheduled));
642 * Process the vacuum fetch lists:
643 * For records for which we are not the lmaster, tell the lmaster to
646 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
647 struct vacuum_data *vdata)
650 struct ctdb_context *ctdb = ctdb_db->ctdb;
652 for (i = 0; i < ctdb->num_nodes; i++) {
654 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
656 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
660 if (vfl->count == 0) {
664 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
665 vfl->count, ctdb->nodes[i]->pnn,
668 data = ctdb_marshall_finish(vfl);
669 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
670 CTDB_SRVID_VACUUM_FETCH,
673 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
674 "fetch message to %u\n",
675 ctdb->nodes[i]->pnn));
683 * Process the delete list:
685 * This is the last step of vacuuming that consistently deletes
686 * those records that have been migrated with data and can hence
687 * not be deleted when leaving a node.
689 * In this step, the lmaster does the final deletion of those empty
690 * records that it is also dmaster for. It has ususally received
691 * at least some of these records previously from the former dmasters
692 * with the vacuum fetch message.
694 * 1) Send the records to all active nodes with the TRY_DELETE_RECORDS
695 * control. The remote notes delete their local copy.
696 * 2) The lmaster locally deletes its copies of all records that
697 * could successfully be deleted remotely in step #2.
699 static void ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
700 struct vacuum_data *vdata)
703 struct ctdb_context *ctdb = ctdb_db->ctdb;
704 struct delete_records_list *recs;
706 struct ctdb_node_map_old *nodemap;
707 uint32_t *active_nodes;
708 int num_active_nodes;
712 if (vdata->count.delete_list.total == 0) {
716 tmp_ctx = talloc_new(vdata);
717 if (tmp_ctx == NULL) {
718 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
722 vdata->count.delete_list.left = vdata->count.delete_list.total;
725 * get the list of currently active nodes
728 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
733 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
737 active_nodes = list_of_active_nodes(ctdb, nodemap,
738 nodemap, /* talloc context */
739 false /* include self */);
741 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
744 * Now delete the records all active nodes in a two-phase process:
745 * 1) tell all active remote nodes to delete all their copy
746 * 2) if all remote nodes deleted their record copy, delete it locally
749 recs = talloc_zero(tmp_ctx, struct delete_records_list);
751 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
757 * Send all records to all active nodes for deletion.
761 * Create a marshall blob from the remaining list of records to delete.
764 recs->records = (struct ctdb_marshall_buffer *)
765 talloc_zero_size(recs,
766 offsetof(struct ctdb_marshall_buffer, data));
767 if (recs->records == NULL) {
768 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
771 recs->records->db_id = ctdb_db->db_id;
773 ret = trbt_traversearray32(vdata->delete_list, 1,
774 delete_marshall_traverse, recs);
776 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
777 "delete list for second marshalling.\n"));
781 indata = ctdb_marshall_finish(recs->records);
783 for (i = 0; i < num_active_nodes; i++) {
784 struct ctdb_marshall_buffer *records;
785 struct ctdb_rec_data_old *rec;
789 ret = ctdb_control(ctdb, active_nodes[i], 0,
790 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
791 indata, recs, &outdata, &res,
793 if (ret != 0 || res != 0) {
794 DEBUG(DEBUG_ERR, ("Failed to delete records on "
795 "node %u: ret[%d] res[%d]\n",
796 active_nodes[i], ret, res));
801 * outdata contains the list of records coming back
802 * from the node: These are the records that the
803 * remote node could not delete. We remove these from
804 * the list to delete locally.
806 records = (struct ctdb_marshall_buffer *)outdata.dptr;
807 rec = (struct ctdb_rec_data_old *)&records->data[0];
808 while (records->count-- > 1) {
809 TDB_DATA reckey, recdata;
810 struct ctdb_ltdb_header *rechdr;
811 struct delete_record_data *dd;
813 reckey.dptr = &rec->data[0];
814 reckey.dsize = rec->keylen;
815 recdata.dptr = &rec->data[reckey.dsize];
816 recdata.dsize = rec->datalen;
818 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
819 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
822 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
823 recdata.dptr += sizeof(*rechdr);
824 recdata.dsize -= sizeof(*rechdr);
826 dd = (struct delete_record_data *)trbt_lookup32(
831 * The other node could not delete the
832 * record and it is the first node that
833 * failed. So we should remove it from
834 * the tree and update statistics.
837 vdata->count.delete_list.remote_error++;
838 vdata->count.delete_list.left--;
840 DEBUG(DEBUG_ERR, (__location__ " Failed to "
841 "find record with hash 0x%08x coming "
842 "back from TRY_DELETE_RECORDS "
843 "control in delete list.\n",
844 ctdb_hash(&reckey)));
845 vdata->count.delete_list.local_error++;
846 vdata->count.delete_list.left--;
849 rec = (struct ctdb_rec_data_old *)(rec->length + (uint8_t *)rec);
853 if (vdata->count.delete_list.left == 0) {
859 * Delete the remaining records locally.
861 * These records have successfully been deleted on all
862 * active remote nodes.
865 ret = trbt_traversearray32(vdata->delete_list, 1,
866 delete_record_traverse, vdata);
868 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
869 "delete list for deletion.\n"));
874 if (vdata->count.delete_list.left != 0) {
875 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
876 "there are %u records left for deletion after "
877 "processing delete list\n",
879 (unsigned)vdata->count.delete_list.left));
882 sum = vdata->count.delete_list.deleted
883 + vdata->count.delete_list.skipped
884 + vdata->count.delete_list.remote_error
885 + vdata->count.delete_list.local_error
886 + vdata->count.delete_list.left;
888 if (vdata->count.delete_list.total != sum) {
889 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
890 "delete list counts for db[%s]: total[%u] != sum[%u]\n",
892 (unsigned)vdata->count.delete_list.total,
896 if (vdata->count.delete_list.total > 0) {
899 " vacuum delete list statistics: "
908 (unsigned)vdata->count.delete_list.total,
909 (unsigned)vdata->count.delete_list.deleted,
910 (unsigned)vdata->count.delete_list.skipped,
911 (unsigned)vdata->count.delete_list.remote_error,
912 (unsigned)vdata->count.delete_list.local_error,
913 (unsigned)vdata->count.delete_list.left));
917 talloc_free(tmp_ctx);
923 * initialize the vacuum_data
925 static struct vacuum_data *ctdb_vacuum_init_vacuum_data(
926 struct ctdb_db_context *ctdb_db,
930 struct ctdb_context *ctdb = ctdb_db->ctdb;
931 struct vacuum_data *vdata;
933 vdata = talloc_zero(mem_ctx, struct vacuum_data);
935 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
939 vdata->ctdb = ctdb_db->ctdb;
940 vdata->ctdb_db = ctdb_db;
941 vdata->delete_list = trbt_create(vdata, 0);
942 if (vdata->delete_list == NULL) {
943 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
947 vdata->start = timeval_current();
949 vdata->count.delete_queue.added_to_delete_list = 0;
950 vdata->count.delete_queue.added_to_vacuum_fetch_list = 0;
951 vdata->count.delete_queue.deleted = 0;
952 vdata->count.delete_queue.skipped = 0;
953 vdata->count.delete_queue.error = 0;
954 vdata->count.delete_queue.total = 0;
955 vdata->count.db_traverse.scheduled = 0;
956 vdata->count.db_traverse.skipped = 0;
957 vdata->count.db_traverse.error = 0;
958 vdata->count.db_traverse.total = 0;
959 vdata->count.delete_list.total = 0;
960 vdata->count.delete_list.left = 0;
961 vdata->count.delete_list.remote_error = 0;
962 vdata->count.delete_list.local_error = 0;
963 vdata->count.delete_list.skipped = 0;
964 vdata->count.delete_list.deleted = 0;
966 /* the list needs to be of length num_nodes */
967 vdata->vacuum_fetch_list = talloc_zero_array(vdata,
968 struct ctdb_marshall_buffer *,
970 if (vdata->vacuum_fetch_list == NULL) {
971 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
974 for (i = 0; i < ctdb->num_nodes; i++) {
975 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
976 talloc_zero_size(vdata->vacuum_fetch_list,
977 offsetof(struct ctdb_marshall_buffer, data));
978 if (vdata->vacuum_fetch_list[i] == NULL) {
979 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
983 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
995 * - Always do the fast vacuuming run, which traverses
996 * the in-memory delete queue: these records have been
997 * scheduled for deletion.
998 * - Only if explicitly requested, the database is traversed
999 * in order to use the traditional heuristics on empty records
1000 * to trigger deletion.
1001 * This is done only every VacuumFastPathCount'th vacuuming run.
1003 * The traverse runs fill two lists:
1005 * - The delete_list:
1006 * This is the list of empty records the current
1007 * node is lmaster and dmaster for. These records are later
1008 * deleted first on other nodes and then locally.
1010 * The fast vacuuming run has a short cut for those records
1011 * that have never been migrated with data: these records
1012 * are immediately deleted locally, since they have left
1013 * no trace on other nodes.
1015 * - The vacuum_fetch lists
1016 * (one for each other lmaster node):
1017 * The records in this list are sent for deletion to
1018 * their lmaster in a bulk VACUUM_FETCH message.
1020 * The lmaster then migrates all these records to itelf
1021 * so that they can be vacuumed there.
1023 * This executes in the child context.
1025 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1026 bool full_vacuum_run)
1028 struct ctdb_context *ctdb = ctdb_db->ctdb;
1030 struct vacuum_data *vdata;
1031 TALLOC_CTX *tmp_ctx;
1033 DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1034 "%s db_id[0x%08x]\n",
1035 full_vacuum_run ? "full" : "fast",
1036 ctdb_db->db_name, ctdb_db->db_id));
1038 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1040 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1044 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1046 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1052 tmp_ctx = talloc_new(ctdb_db);
1053 if (tmp_ctx == NULL) {
1054 DEBUG(DEBUG_ERR, ("Out of memory!\n"));
1058 vdata = ctdb_vacuum_init_vacuum_data(ctdb_db, tmp_ctx);
1059 if (vdata == NULL) {
1060 talloc_free(tmp_ctx);
1064 if (full_vacuum_run) {
1065 ctdb_vacuum_traverse_db(ctdb_db, vdata);
1068 ctdb_process_delete_queue(ctdb_db, vdata);
1070 ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1072 ctdb_process_delete_list(ctdb_db, vdata);
1074 talloc_free(tmp_ctx);
1076 /* this ensures we run our event queue */
1077 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1083 * repack and vaccum a db
1084 * called from the child context
1086 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1087 bool full_vacuum_run)
1089 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1090 const char *name = ctdb_db->db_name;
1091 int freelist_size = 0;
1094 if (ctdb_vacuum_db(ctdb_db, full_vacuum_run) != 0) {
1095 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1098 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1099 if (freelist_size == -1) {
1100 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1105 * decide if a repack is necessary
1107 if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1112 DEBUG(DEBUG_INFO, ("Repacking %s with %u freelist entries\n",
1113 name, freelist_size));
1115 ret = tdb_repack(ctdb_db->ltdb->tdb);
1117 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1124 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1126 uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1131 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1133 double l = timeval_elapsed(&child_ctx->start_time);
1134 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1135 struct ctdb_context *ctdb = ctdb_db->ctdb;
1137 CTDB_UPDATE_DB_LATENCY(ctdb_db, "vacuum", vacuum.latency, l);
1138 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1140 if (child_ctx->child_pid != -1) {
1141 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1143 /* Bump the number of successful fast-path runs. */
1144 child_ctx->vacuum_handle->fast_path_count++;
1147 DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1149 tevent_add_timer(ctdb->ev, child_ctx->vacuum_handle,
1150 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1151 ctdb_vacuum_event, child_ctx->vacuum_handle);
1157 * this event is generated when a vacuum child process times out
1159 static void vacuum_child_timeout(struct tevent_context *ev,
1160 struct tevent_timer *te,
1161 struct timeval t, void *private_data)
1163 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1165 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1167 child_ctx->status = VACUUM_TIMEOUT;
1169 talloc_free(child_ctx);
1174 * this event is generated when a vacuum child process has completed
1176 static void vacuum_child_handler(struct tevent_context *ev,
1177 struct tevent_fd *fde,
1178 uint16_t flags, void *private_data)
1180 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1184 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1185 child_ctx->child_pid = -1;
1187 ret = sys_read(child_ctx->fd[0], &c, 1);
1188 if (ret != 1 || c != 0) {
1189 child_ctx->status = VACUUM_ERROR;
1190 DEBUG(DEBUG_ERR, ("A vacuum child process failed with an error for database %s. ret=%d c=%d\n", child_ctx->vacuum_handle->ctdb_db->db_name, ret, c));
1192 child_ctx->status = VACUUM_OK;
1195 talloc_free(child_ctx);
1199 * this event is called every time we need to start a new vacuum process
1201 static void ctdb_vacuum_event(struct tevent_context *ev,
1202 struct tevent_timer *te,
1203 struct timeval t, void *private_data)
1205 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1206 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1207 struct ctdb_context *ctdb = ctdb_db->ctdb;
1208 struct ctdb_vacuum_child_context *child_ctx;
1209 struct tevent_fd *fde;
1212 /* we don't vacuum if we are in recovery mode, or db frozen */
1213 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1214 ctdb_db_frozen(ctdb_db)) {
1215 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1216 ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ?
1217 "in recovery" : "frozen"));
1218 tevent_add_timer(ctdb->ev, vacuum_handle,
1219 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1220 ctdb_vacuum_event, vacuum_handle);
1224 /* Do not allow multiple vacuuming child processes to be active at the
1225 * same time. If there is vacuuming child process active, delay
1226 * new vacuuming event to stagger vacuuming events.
1228 if (ctdb->vacuumers != NULL) {
1229 tevent_add_timer(ctdb->ev, vacuum_handle,
1230 timeval_current_ofs(0, 500*1000),
1231 ctdb_vacuum_event, vacuum_handle);
1235 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1236 if (child_ctx == NULL) {
1237 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1238 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1242 ret = pipe(child_ctx->fd);
1244 talloc_free(child_ctx);
1245 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1246 tevent_add_timer(ctdb->ev, vacuum_handle,
1247 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1248 ctdb_vacuum_event, vacuum_handle);
1252 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1253 vacuum_handle->fast_path_count = 0;
1256 child_ctx->child_pid = ctdb_fork(ctdb);
1257 if (child_ctx->child_pid == (pid_t)-1) {
1258 close(child_ctx->fd[0]);
1259 close(child_ctx->fd[1]);
1260 talloc_free(child_ctx);
1261 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1262 tevent_add_timer(ctdb->ev, vacuum_handle,
1263 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1264 ctdb_vacuum_event, vacuum_handle);
1269 if (child_ctx->child_pid == 0) {
1271 bool full_vacuum_run = false;
1272 close(child_ctx->fd[0]);
1274 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1275 prctl_set_comment("ctdb_vacuum");
1276 if (switch_from_server_to_client(ctdb) != 0) {
1277 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1281 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1282 (vacuum_handle->fast_path_count == 0))
1284 full_vacuum_run = true;
1286 cc = ctdb_vacuum_and_repack_db(ctdb_db, full_vacuum_run);
1288 sys_write(child_ctx->fd[1], &cc, 1);
1292 set_close_on_exec(child_ctx->fd[0]);
1293 close(child_ctx->fd[1]);
1295 child_ctx->status = VACUUM_RUNNING;
1296 child_ctx->start_time = timeval_current();
1298 DLIST_ADD(ctdb->vacuumers, child_ctx);
1299 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1302 * Clear the fastpath vacuuming list in the parent.
1304 talloc_free(ctdb_db->delete_queue);
1305 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1306 if (ctdb_db->delete_queue == NULL) {
1307 /* fatal here? ... */
1308 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1309 "in parent context. Shutting down\n");
1312 tevent_add_timer(ctdb->ev, child_ctx,
1313 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1314 vacuum_child_timeout, child_ctx);
1316 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1318 fde = tevent_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1319 TEVENT_FD_READ, vacuum_child_handler, child_ctx);
1320 tevent_fd_set_auto_close(fde);
1322 vacuum_handle->child_ctx = child_ctx;
1323 child_ctx->vacuum_handle = vacuum_handle;
1326 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1328 /* Simply free them all. */
1329 while (ctdb->vacuumers) {
1330 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1331 ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1332 (int)ctdb->vacuumers->child_pid));
1333 /* vacuum_child_destructor kills it, removes from list */
1334 talloc_free(ctdb->vacuumers);
1338 /* this function initializes the vacuuming context for a database
1339 * starts the vacuuming events
1341 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1343 if (! ctdb_db_volatile(ctdb_db)) {
1345 ("Vacuuming is disabled for non-volatile database %s\n",
1350 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1351 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1353 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1354 ctdb_db->vacuum_handle->fast_path_count = 0;
1356 tevent_add_timer(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1357 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1358 ctdb_vacuum_event, ctdb_db->vacuum_handle);
1363 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1364 const struct ctdb_ltdb_header *hdr,
1367 struct delete_record_data *kd;
1370 hash = (uint32_t)ctdb_hash(&key);
1372 DEBUG(DEBUG_DEBUG, (__location__
1373 " remove_record_from_delete_queue: "
1378 "migrated_with_data[%s]\n",
1379 ctdb_db->db_name, ctdb_db->db_id,
1381 ctdb_lmaster(ctdb_db->ctdb, &key),
1382 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1384 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1386 DEBUG(DEBUG_DEBUG, (__location__
1387 " remove_record_from_delete_queue: "
1388 "record not in queue (hash[0x%08x])\n.",
1393 if ((kd->key.dsize != key.dsize) ||
1394 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1396 DEBUG(DEBUG_DEBUG, (__location__
1397 " remove_record_from_delete_queue: "
1398 "hash collision for key with hash[0x%08x] "
1399 "in db[%s] - skipping\n",
1400 hash, ctdb_db->db_name));
1404 DEBUG(DEBUG_DEBUG, (__location__
1405 " remove_record_from_delete_queue: "
1406 "removing key with hash[0x%08x]\n",
1415 * Insert a record into the ctdb_db context's delete queue,
1416 * handling hash collisions.
1418 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1419 const struct ctdb_ltdb_header *hdr,
1422 struct delete_record_data *kd;
1426 hash = (uint32_t)ctdb_hash(&key);
1428 DEBUG(DEBUG_DEBUG, (__location__ " schedule for deletion: db[%s] "
1432 "migrated_with_data[%s]\n",
1433 ctdb_db->db_name, ctdb_db->db_id,
1435 ctdb_lmaster(ctdb_db->ctdb, &key),
1436 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1438 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1440 if ((kd->key.dsize != key.dsize) ||
1441 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1444 (__location__ " schedule for deletion: "
1445 "hash collision for key hash [0x%08x]. "
1446 "Skipping the record.\n", hash));
1450 (__location__ " schedule for deletion: "
1451 "updating entry for key with hash [0x%08x].\n",
1456 ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1457 ctdb_db->delete_queue,
1461 (__location__ " schedule for deletion: error "
1462 "inserting key with hash [0x%08x] into delete queue\n",
1471 * Schedule a record for deletetion.
1472 * Called from the parent context.
1474 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1477 struct ctdb_control_schedule_for_deletion *dd;
1478 struct ctdb_db_context *ctdb_db;
1482 dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1484 ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1485 if (ctdb_db == NULL) {
1486 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1491 key.dsize = dd->keylen;
1494 ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1499 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1500 const struct ctdb_ltdb_header *hdr,
1504 struct ctdb_control_schedule_for_deletion *dd;
1508 if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1509 /* main daemon - directly queue */
1510 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1515 /* if we don't have a connection to the daemon we can not send
1516 a control. For example sometimes from update_record control child
1519 if (!ctdb_db->ctdb->can_send_controls) {
1524 /* child process: send the main daemon a control */
1525 indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1526 indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1527 if (indata.dptr == NULL) {
1528 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1531 dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1532 dd->db_id = ctdb_db->db_id;
1534 dd->keylen = key.dsize;
1535 memcpy(dd->key, key.dptr, key.dsize);
1537 ret = ctdb_control(ctdb_db->ctdb,
1540 CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1541 CTDB_CTRL_FLAG_NOREPLY, /* flags */
1546 NULL, /* timeout : NULL == wait forever */
1547 NULL); /* error message */
1549 talloc_free(indata.dptr);
1551 if (ret != 0 || status != 0) {
1552 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1553 "SCHEDULE_FOR_DELETION "
1563 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1564 const struct ctdb_ltdb_header *hdr,
1567 if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1569 * Only remove the record from the delete queue if called
1570 * in the main daemon.
1575 remove_record_from_delete_queue(ctdb_db, hdr, key);