4 Copyright (C) Ronnie Sahlberg 2009
5 Copyright (C) Michael Adam 2010-2011
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 "lib/events/events.h"
24 #include "lib/tdb/include/tdb.h"
25 #include "system/network.h"
26 #include "system/filesys.h"
27 #include "system/dir.h"
28 #include "../include/ctdb_private.h"
30 #include "lib/util/dlinklist.h"
31 #include "lib/events/events.h"
32 #include "../include/ctdb_private.h"
33 #include "../common/rb_tree.h"
35 #define TIMELIMIT() timeval_current_ofs(10, 0)
37 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
39 struct ctdb_vacuum_child_context {
40 struct ctdb_vacuum_child_context *next, *prev;
41 struct ctdb_vacuum_handle *vacuum_handle;
42 /* fd child writes status to */
45 enum vacuum_child_status status;
46 struct timeval start_time;
49 struct ctdb_vacuum_handle {
50 struct ctdb_db_context *ctdb_db;
51 struct ctdb_vacuum_child_context *child_ctx;
52 uint32_t fast_path_count;
56 /* a list of records to possibly delete */
58 uint32_t vacuum_limit;
59 uint32_t repack_limit;
60 struct ctdb_context *ctdb;
61 struct ctdb_db_context *ctdb_db;
62 struct tdb_context *dest_db;
63 trbt_tree_t *delete_list;
64 uint32_t delete_count;
65 struct ctdb_marshall_buffer **vacuum_fetch_list;
72 uint32_t fast_added_to_vacuum_fetch_list;
73 uint32_t fast_added_to_delete_list;
74 uint32_t fast_deleted;
75 uint32_t fast_skipped;
78 uint32_t full_added_to_vacuum_fetch_list;
79 uint32_t full_added_to_delete_list;
80 uint32_t full_skipped;
84 uint32_t delete_remote_error;
85 uint32_t delete_local_error;
86 uint32_t delete_deleted;
87 uint32_t delete_skipped;
90 /* this structure contains the information for one record to be deleted */
91 struct delete_record_data {
92 struct ctdb_context *ctdb;
93 struct ctdb_db_context *ctdb_db;
94 struct ctdb_ltdb_header hdr;
98 struct delete_records_list {
99 struct ctdb_marshall_buffer *records;
103 * Store key and header in a tree, indexed by the key hash.
105 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
106 struct ctdb_db_context *ctdb_db,
108 const struct ctdb_ltdb_header *hdr,
111 struct delete_record_data *dd;
114 dd = talloc_zero(tree, struct delete_record_data);
116 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
121 dd->ctdb_db = ctdb_db;
122 dd->key.dsize = key.dsize;
123 dd->key.dptr = talloc_memdup(dd, key.dptr, key.dsize);
124 if (dd->key.dptr == NULL) {
125 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
131 hash = (uint32_t)tdb_jenkins_hash(&key);
133 trbt_insert32(tree, hash, dd);
138 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
139 struct ctdb_ltdb_header *hdr)
141 struct ctdb_context *ctdb = vdata->ctdb;
142 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
146 hash = (uint32_t)tdb_jenkins_hash(&key);
148 if (trbt_lookup32(vdata->delete_list, hash)) {
149 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
153 ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
160 vdata->delete_count++;
166 * Add a record to the list of records to be sent
167 * to their lmaster with VACUUM_FETCH.
169 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
172 struct ctdb_context *ctdb = vdata->ctdb;
173 struct ctdb_rec_data *rec;
176 struct ctdb_marshall_buffer *vfl;
178 lmaster = ctdb_lmaster(ctdb, &key);
180 vfl = vdata->vacuum_fetch_list[lmaster];
182 rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
184 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
185 vdata->traverse_error = true;
189 old_size = talloc_get_size(vfl);
190 vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
192 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
193 vdata->traverse_error = true;
196 vdata->vacuum_fetch_list[lmaster] = vfl;
199 memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
208 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
209 struct timeval t, void *private_data);
213 * traverse function for gathering the records that can be deleted
215 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
217 struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
218 struct ctdb_context *ctdb = vdata->ctdb;
220 struct ctdb_ltdb_header *hdr;
225 lmaster = ctdb_lmaster(ctdb, &key);
226 if (lmaster >= ctdb->num_nodes) {
228 DEBUG(DEBUG_CRIT, (__location__
229 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
232 (unsigned)ctdb->num_nodes,
233 (unsigned)ctdb_hash(&key)));
237 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
238 /* it is not a deleted record */
239 vdata->full_skipped++;
243 hdr = (struct ctdb_ltdb_header *)data.dptr;
245 if (hdr->dmaster != ctdb->pnn) {
246 vdata->full_skipped++;
250 if (lmaster == ctdb->pnn) {
252 * We are both lmaster and dmaster, and the record is empty.
253 * So we should be able to delete it.
255 res = add_record_to_delete_list(vdata, key, hdr);
259 vdata->full_added_to_delete_list++;
263 * We are not lmaster.
264 * Add the record to the blob ready to send to the nodes.
266 res = add_record_to_vacuum_fetch_list(vdata, key);
270 vdata->full_added_to_vacuum_fetch_list++;
278 * traverse the tree of records to delete and marshall them into
281 static int delete_marshall_traverse(void *param, void *data)
283 struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
284 struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
285 struct ctdb_rec_data *rec;
288 rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
290 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
294 old_size = talloc_get_size(recs->records);
295 recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
296 if (recs->records == NULL) {
297 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
300 recs->records->count++;
301 memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
306 * traverse function for the traversal of the delete_queue,
307 * the fast-path vacuuming list.
309 * - If the record has been migrated off the node
310 * or has been revived (filled with data) on the node,
311 * then skip the record.
313 * - If the current node is the record's lmaster and it is
314 * a record that has never been migrated with data, then
315 * delete the record from the local tdb.
317 * - If the current node is the record's lmaster and it has
318 * been migrated with data, then schedule it for the normal
319 * vacuuming procedure (i.e. add it to the delete_list).
321 * - If the current node is NOT the record's lmaster then
322 * add it to the list of records that are to be sent to
323 * the lmaster with the VACUUM_FETCH message.
325 static int delete_queue_traverse(void *param, void *data)
327 struct delete_record_data *dd =
328 talloc_get_type(data, struct delete_record_data);
329 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
330 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
331 struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
333 struct ctdb_ltdb_header *header;
339 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
341 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
346 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
347 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
348 /* Does not exist or not a ctdb record. Skip. */
352 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
353 /* The record has been recycled (filled with data). Skip. */
357 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
359 if (header->dmaster != ctdb->pnn) {
360 /* The record has been migrated off the node. Skip. */
365 if (header->rsn != dd->hdr.rsn) {
367 * The record has been migrated off the node and back again.
368 * But not requeued for deletion. Skip it.
374 * We are dmaster, and the record has no data, and it has
375 * not been migrated after it has been queued for deletion.
377 * At this stage, the record could still have been revived locally
378 * and last been written with empty data. This can only be
379 * fixed with the addition of an active or delete flag. (TODO)
382 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
384 if (lmaster != ctdb->pnn) {
385 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
389 (__location__ " Error adding record to list "
390 "of records to send to lmaster.\n"));
393 vdata->fast_added_to_vacuum_fetch_list++;
398 /* use header->flags or dd->hdr.flags ?? */
399 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
400 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
404 (__location__ " Error adding record to list "
405 "of records for deletion on lmaster.\n"));
408 vdata->fast_added_to_delete_list++;
411 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
415 (__location__ " Error deleting record from local "
419 vdata->fast_deleted++;
426 vdata->fast_skipped++;
429 if (tdb_data.dptr != NULL) {
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;
453 bool deleted = false;
455 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
457 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
458 vdata->delete_local_error++;
463 * Verify that the record is still empty, its RSN has not
464 * changed and that we are still its lmaster and dmaster.
467 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
468 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
469 /* Does not exist or not a ctdb record. Skip. */
470 vdata->delete_skipped++;
474 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
475 /* The record has been recycled (filled with data). Skip. */
476 vdata->delete_skipped++;
480 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
482 if (header->dmaster != ctdb->pnn) {
483 /* The record has been migrated off the node. Skip. */
484 vdata->delete_skipped++;
489 if (header->rsn != dd->hdr.rsn) {
491 * The record has been migrated off the node and back again.
492 * But not requeued for deletion. Skip it.
494 vdata->delete_skipped++;
498 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
500 if (lmaster != ctdb->pnn) {
501 /* we are not lmaster - strange */
502 vdata->delete_skipped++;
506 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
510 (__location__ " Error deleting record from local "
512 vdata->delete_local_error++;
519 if (tdb_data.dptr != NULL) {
523 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
527 * successfully deleted the record locally.
528 * remove it from the list and update statistics.
531 vdata->delete_deleted++;
532 vdata->delete_left--;
539 * Fast vacuuming run:
540 * Traverse the delete_queue.
541 * This fills the same lists as the database traverse.
543 static void ctdb_vacuum_db_fast(struct ctdb_db_context *ctdb_db,
544 struct vacuum_data *vdata)
546 trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
548 if (vdata->fast_total > 0) {
551 " fast vacuuming delete_queue traverse statistics: "
560 (unsigned)vdata->fast_total,
561 (unsigned)vdata->fast_deleted,
562 (unsigned)vdata->fast_skipped,
563 (unsigned)vdata->fast_error,
564 (unsigned)vdata->fast_added_to_delete_list,
565 (unsigned)vdata->fast_added_to_vacuum_fetch_list));
573 * read-only traverse of the database, looking for records that
574 * might be able to be vacuumed.
576 * This is not done each time but only every tunable
577 * VacuumFastPathCount times.
579 static int ctdb_vacuum_db_full(struct ctdb_db_context *ctdb_db,
580 struct vacuum_data *vdata,
581 bool full_vacuum_run)
585 if (!full_vacuum_run) {
589 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
590 if (ret == -1 || vdata->traverse_error) {
591 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
592 "'%s'\n", ctdb_db->db_name));
596 if (vdata->full_total > 0) {
599 " full vacuuming db traverse statistics: "
607 (unsigned)vdata->full_total,
608 (unsigned)vdata->full_skipped,
609 (unsigned)vdata->full_error,
610 (unsigned)vdata->full_added_to_delete_list,
611 (unsigned)vdata->full_added_to_vacuum_fetch_list));
618 * Process the vacuum fetch lists:
619 * For records for which we are not the lmaster, tell the lmaster to
622 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
623 struct vacuum_data *vdata)
626 struct ctdb_context *ctdb = ctdb_db->ctdb;
628 for (i = 0; i < ctdb->num_nodes; i++) {
630 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
632 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
636 if (vfl->count == 0) {
640 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
641 vfl->count, ctdb->nodes[i]->pnn,
644 data.dsize = talloc_get_size(vfl);
645 data.dptr = (void *)vfl;
646 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn,
647 CTDB_SRVID_VACUUM_FETCH,
650 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
651 "fetch message to %u\n",
652 ctdb->nodes[i]->pnn));
661 * Proces the delete list:
662 * Send the records to delete to all other nodes with the
663 * try_delete_records control.
665 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
666 struct vacuum_data *vdata)
669 struct ctdb_context *ctdb = ctdb_db->ctdb;
671 vdata->delete_left = vdata->delete_count;
673 if (vdata->delete_count > 0) {
674 struct delete_records_list *recs;
675 TDB_DATA indata, outdata;
677 struct ctdb_node_map *nodemap;
678 uint32_t *active_nodes;
679 int num_active_nodes;
681 recs = talloc_zero(vdata, struct delete_records_list);
683 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
686 recs->records = (struct ctdb_marshall_buffer *)
687 talloc_zero_size(vdata,
688 offsetof(struct ctdb_marshall_buffer, data));
689 if (recs->records == NULL) {
690 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
693 recs->records->db_id = ctdb_db->db_id;
696 * traverse the tree of all records we want to delete and
697 * create a blob we can send to the other nodes.
699 trbt_traversearray32(vdata->delete_list, 1,
700 delete_marshall_traverse, recs);
702 indata.dsize = talloc_get_size(recs->records);
703 indata.dptr = (void *)recs->records;
706 * now tell all the active nodes to delete all these records
710 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
712 recs, /* talloc context */
715 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
719 active_nodes = list_of_active_nodes(ctdb, nodemap,
720 nodemap, /* talloc context */
721 false /* include self */);
723 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
725 for (i = 0; i < num_active_nodes; i++) {
726 struct ctdb_marshall_buffer *records;
727 struct ctdb_rec_data *rec;
729 ret = ctdb_control(ctdb, active_nodes[i], 0,
730 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
731 indata, recs, &outdata, &res,
733 if (ret != 0 || res != 0) {
734 DEBUG(DEBUG_ERR, ("Failed to delete records on "
735 "node %u: ret[%d] res[%d]\n",
736 active_nodes[i], ret, res));
741 * outdata contains the list of records coming back
742 * from the node: These are the records that the
743 * remote node could not delete.
745 * NOTE: There is a problem here:
747 * When a node failed to delete the record, but
748 * others succeeded, we may have created gaps in the
749 * history of the record. Hence when a node dies, an
750 * closed file handle might be resurrected or an open
751 * file handle might be lost, leading to blocked access
752 * or data corruption.
754 * TODO: This needs to be fixed!
756 records = (struct ctdb_marshall_buffer *)outdata.dptr;
757 rec = (struct ctdb_rec_data *)&records->data[0];
758 while (records->count-- > 1) {
759 TDB_DATA reckey, recdata;
760 struct ctdb_ltdb_header *rechdr;
761 struct delete_record_data *dd;
763 reckey.dptr = &rec->data[0];
764 reckey.dsize = rec->keylen;
765 recdata.dptr = &rec->data[reckey.dsize];
766 recdata.dsize = rec->datalen;
768 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
769 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
772 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
773 recdata.dptr += sizeof(*rechdr);
774 recdata.dsize -= sizeof(*rechdr);
776 dd = (struct delete_record_data *)trbt_lookup32(
778 (uint32_t)tdb_jenkins_hash(&reckey));
781 * The other node could not delete the
782 * record and it is the first node that
783 * failed. So we should remove it from
784 * the tree and update statistics.
787 vdata->delete_remote_error++;
788 vdata->delete_left--;
791 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
795 /* free nodemap and active_nodes */
796 talloc_free(nodemap);
799 if (vdata->delete_left > 0) {
801 * The only records remaining in the tree are those
802 * records which all other nodes could successfully
803 * delete, so we can safely delete them on the
806 trbt_traversearray32(vdata->delete_list, 1,
807 delete_record_traverse, vdata);
810 if (vdata->delete_count > 0) {
813 " vacuum delete list statistics: "
822 (unsigned)vdata->delete_count,
823 (unsigned)vdata->delete_remote_error,
824 (unsigned)vdata->delete_local_error,
825 (unsigned)vdata->delete_skipped,
826 (unsigned)vdata->delete_deleted,
827 (unsigned)vdata->delete_left));
834 * initialize the vacuum_data
836 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
837 struct vacuum_data *vdata)
840 struct ctdb_context *ctdb = ctdb_db->ctdb;
842 vdata->fast_added_to_delete_list = 0;
843 vdata->fast_added_to_vacuum_fetch_list = 0;
844 vdata->fast_deleted = 0;
845 vdata->fast_skipped = 0;
846 vdata->fast_error = 0;
847 vdata->fast_total = 0;
848 vdata->full_added_to_delete_list = 0;
849 vdata->full_added_to_vacuum_fetch_list = 0;
850 vdata->full_skipped = 0;
851 vdata->full_error = 0;
852 vdata->full_total = 0;
853 vdata->delete_count = 0;
854 vdata->delete_left = 0;
855 vdata->delete_remote_error = 0;
856 vdata->delete_local_error = 0;
857 vdata->delete_skipped = 0;
858 vdata->delete_deleted = 0;
860 /* the list needs to be of length num_nodes */
861 vdata->vacuum_fetch_list = talloc_array(vdata,
862 struct ctdb_marshall_buffer *,
864 if (vdata->vacuum_fetch_list == NULL) {
865 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
868 for (i = 0; i < ctdb->num_nodes; i++) {
869 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
870 talloc_zero_size(vdata->vacuum_fetch_list,
871 offsetof(struct ctdb_marshall_buffer, data));
872 if (vdata->vacuum_fetch_list[i] == NULL) {
873 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
876 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
884 * - Always do the fast vacuuming run, which traverses
885 * the in-memory delete queue: these records have been
886 * scheduled for deletion.
887 * - Only if explicitly requested, the database is traversed
888 * in order to use the traditional heuristics on empty records
889 * to trigger deletion.
890 * This is done only every VacuumFastPathCount'th vacuuming run.
892 * The traverse runs fill two lists:
895 * This is the list of empty records the current
896 * node is lmaster and dmaster for. These records are later
897 * deleted first on other nodes and then locally.
899 * The fast vacuuming run has a short cut for those records
900 * that have never been migrated with data: these records
901 * are immediately deleted locally, since they have left
902 * no trace on other nodes.
904 * - The vacuum_fetch lists
905 * (one for each other lmaster node):
906 * The records in this list are sent for deletion to
907 * their lmaster in a bulk VACUUM_FETCH message.
909 * The lmaster then migrates all these records to itelf
910 * so that they can be vacuumed there.
912 * This executes in the child context.
914 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
915 struct vacuum_data *vdata,
916 bool full_vacuum_run)
918 struct ctdb_context *ctdb = ctdb_db->ctdb;
921 DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
922 "%s db_id[0x%08x]\n",
923 full_vacuum_run ? "full" : "fast",
924 ctdb_db->db_name, ctdb_db->db_id));
926 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
928 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
932 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
934 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
940 ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
945 ctdb_vacuum_db_fast(ctdb_db, vdata);
947 ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
952 ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
957 ret = ctdb_process_delete_list(ctdb_db, vdata);
962 /* this ensures we run our event queue */
963 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
970 * traverse function for repacking
972 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
974 struct vacuum_data *vdata = (struct vacuum_data *)private;
977 uint32_t hash = (uint32_t)tdb_jenkins_hash(&key);
978 struct delete_record_data *kd;
980 * check if we can ignore this record because it's in the delete_list
982 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
984 * there might be hash collisions so we have to compare the keys here to be sure
986 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
987 struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
989 * we have to check if the record hasn't changed in the meantime in order to
990 * savely remove it from the database
992 if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
993 hdr->dmaster == kd->ctdb->pnn &&
994 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
995 kd->hdr.rsn == hdr->rsn) {
1001 if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1002 vdata->traverse_error = true;
1012 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1014 struct tdb_context *tmp_db;
1016 if (tdb_transaction_start(tdb) != 0) {
1017 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1021 tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1022 TDB_INTERNAL|TDB_DISALLOW_NESTING,
1024 if (tmp_db == NULL) {
1025 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1026 tdb_transaction_cancel(tdb);
1030 vdata->traverse_error = false;
1031 vdata->dest_db = tmp_db;
1032 vdata->vacuum = true;
1033 vdata->vacuumed = 0;
1037 * repack and vacuum on-the-fly by not writing the records that are
1040 if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1041 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1042 tdb_transaction_cancel(tdb);
1047 DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1049 if (vdata->traverse_error) {
1050 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1051 tdb_transaction_cancel(tdb);
1056 if (tdb_wipe_all(tdb) != 0) {
1057 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1058 tdb_transaction_cancel(tdb);
1063 vdata->traverse_error = false;
1064 vdata->dest_db = tdb;
1065 vdata->vacuum = false;
1068 if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1069 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1070 tdb_transaction_cancel(tdb);
1075 if (vdata->traverse_error) {
1076 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1077 tdb_transaction_cancel(tdb);
1085 if (tdb_transaction_commit(tdb) != 0) {
1086 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1089 DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1095 * repack and vaccum a db
1096 * called from the child context
1098 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1099 TALLOC_CTX *mem_ctx,
1100 bool full_vacuum_run)
1102 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1103 uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1104 const char *name = ctdb_db->db_name;
1106 struct vacuum_data *vdata;
1108 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1109 if (freelist_size == -1) {
1110 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1114 vdata = talloc_zero(mem_ctx, struct vacuum_data);
1115 if (vdata == NULL) {
1116 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1120 vdata->ctdb = ctdb_db->ctdb;
1121 vdata->vacuum_limit = vacuum_limit;
1122 vdata->repack_limit = repack_limit;
1123 vdata->delete_list = trbt_create(vdata, 0);
1124 vdata->ctdb_db = ctdb_db;
1125 if (vdata->delete_list == NULL) {
1126 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1131 vdata->start = timeval_current();
1134 * gather all records that can be deleted in vdata
1136 if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1137 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1141 * decide if a repack is necessary
1143 if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1149 DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1150 name, freelist_size, vdata->delete_left));
1153 * repack and implicitely get rid of the records we can delete
1155 if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1156 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1165 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1167 uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1172 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1174 double l = timeval_elapsed(&child_ctx->start_time);
1175 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1176 struct ctdb_context *ctdb = ctdb_db->ctdb;
1178 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1180 if (child_ctx->child_pid != -1) {
1181 kill(child_ctx->child_pid, SIGKILL);
1183 /* Bump the number of successful fast-path runs. */
1184 child_ctx->vacuum_handle->fast_path_count++;
1187 DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1189 event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1190 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1191 ctdb_vacuum_event, child_ctx->vacuum_handle);
1197 * this event is generated when a vacuum child process times out
1199 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1200 struct timeval t, void *private_data)
1202 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1204 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1206 child_ctx->status = VACUUM_TIMEOUT;
1208 talloc_free(child_ctx);
1213 * this event is generated when a vacuum child process has completed
1215 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1216 uint16_t flags, void *private_data)
1218 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1222 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1223 child_ctx->child_pid = -1;
1225 ret = read(child_ctx->fd[0], &c, 1);
1226 if (ret != 1 || c != 0) {
1227 child_ctx->status = VACUUM_ERROR;
1228 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));
1230 child_ctx->status = VACUUM_OK;
1233 talloc_free(child_ctx);
1237 * this event is called every time we need to start a new vacuum process
1240 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1241 struct timeval t, void *private_data)
1243 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1244 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1245 struct ctdb_context *ctdb = ctdb_db->ctdb;
1246 struct ctdb_vacuum_child_context *child_ctx;
1249 /* we dont vacuum if we are in recovery mode, or db frozen */
1250 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1251 ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1252 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1253 ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1254 : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1257 event_add_timed(ctdb->ev, vacuum_handle,
1258 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1259 ctdb_vacuum_event, vacuum_handle);
1263 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1264 if (child_ctx == NULL) {
1265 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1266 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1270 ret = pipe(child_ctx->fd);
1272 talloc_free(child_ctx);
1273 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1274 event_add_timed(ctdb->ev, vacuum_handle,
1275 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1276 ctdb_vacuum_event, vacuum_handle);
1280 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1281 vacuum_handle->fast_path_count = 0;
1284 child_ctx->child_pid = fork();
1285 if (child_ctx->child_pid == (pid_t)-1) {
1286 close(child_ctx->fd[0]);
1287 close(child_ctx->fd[1]);
1288 talloc_free(child_ctx);
1289 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1290 event_add_timed(ctdb->ev, vacuum_handle,
1291 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1292 ctdb_vacuum_event, vacuum_handle);
1297 if (child_ctx->child_pid == 0) {
1299 bool full_vacuum_run = false;
1300 close(child_ctx->fd[0]);
1302 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1304 if (switch_from_server_to_client(ctdb) != 0) {
1305 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1312 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1313 (vacuum_handle->fast_path_count == 0))
1315 full_vacuum_run = true;
1317 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1320 write(child_ctx->fd[1], &cc, 1);
1324 set_close_on_exec(child_ctx->fd[0]);
1325 close(child_ctx->fd[1]);
1327 child_ctx->status = VACUUM_RUNNING;
1328 child_ctx->start_time = timeval_current();
1330 DLIST_ADD(ctdb->vacuumers, child_ctx);
1331 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1334 * Clear the fastpath vacuuming list in the parent.
1336 talloc_free(ctdb_db->delete_queue);
1337 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1338 if (ctdb_db->delete_queue == NULL) {
1339 /* fatal here? ... */
1340 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1341 "in parent context. Shutting down\n");
1344 event_add_timed(ctdb->ev, child_ctx,
1345 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1346 vacuum_child_timeout, child_ctx);
1348 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1350 event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1351 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1352 vacuum_child_handler,
1355 vacuum_handle->child_ctx = child_ctx;
1356 child_ctx->vacuum_handle = vacuum_handle;
1359 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1361 /* Simply free them all. */
1362 while (ctdb->vacuumers) {
1363 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1364 ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1365 (int)ctdb->vacuumers->child_pid));
1366 /* vacuum_child_destructor kills it, removes from list */
1367 talloc_free(ctdb->vacuumers);
1371 /* this function initializes the vacuuming context for a database
1372 * starts the vacuuming events
1374 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1376 if (ctdb_db->persistent != 0) {
1377 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1381 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1382 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1384 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1385 ctdb_db->vacuum_handle->fast_path_count = 0;
1387 event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1388 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1389 ctdb_vacuum_event, ctdb_db->vacuum_handle);
1394 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1395 const struct ctdb_ltdb_header *hdr,
1398 struct delete_record_data *kd;
1401 hash = (uint32_t)tdb_jenkins_hash(discard_const(&key));
1403 DEBUG(DEBUG_DEBUG, (__location__
1404 " remove_record_from_delete_queue: db[%s] "
1408 "migrated_with_data[%s]\n",
1409 ctdb_db->db_name, ctdb_db->db_id,
1411 ctdb_lmaster(ctdb_db->ctdb, &key),
1412 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1414 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1418 if (kd->key.dsize != key.dsize) {
1421 if (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0) {
1431 * Insert a record into the ctdb_db context's delete queue,
1432 * handling hash collisions.
1434 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1435 const struct ctdb_ltdb_header *hdr,
1438 struct delete_record_data *kd;
1442 hash = (uint32_t)tdb_jenkins_hash(&key);
1444 DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1448 "migrated_with_data[%s]\n",
1449 ctdb_db->db_name, ctdb_db->db_id,
1451 ctdb_lmaster(ctdb_db->ctdb, &key),
1452 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1454 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1456 if ((kd->key.dsize != key.dsize) ||
1457 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1460 ("schedule for deletion: Hash collision (0x%08x)."
1461 " Skipping the record.\n", hash));
1465 ("schedule for deletion: Overwriting entry for "
1466 "key with hash 0x%08x.\n", hash));
1470 ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1471 ctdb_db->delete_queue,
1481 * Schedule a record for deletetion.
1482 * Called from the parent context.
1484 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1487 struct ctdb_control_schedule_for_deletion *dd;
1488 struct ctdb_db_context *ctdb_db;
1492 dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1494 ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1495 if (ctdb_db == NULL) {
1496 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1501 key.dsize = dd->keylen;
1504 ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1509 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1510 const struct ctdb_ltdb_header *hdr,
1514 struct ctdb_control_schedule_for_deletion *dd;
1518 if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1519 /* main daemon - directly queue */
1520 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1525 /* child process: send the main daemon a control */
1527 indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1528 indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1529 if (indata.dptr == NULL) {
1530 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1533 dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1534 dd->db_id = ctdb_db->db_id;
1536 dd->keylen = key.dsize;
1537 memcpy(dd->key, key.dptr, key.dsize);
1539 ret = ctdb_control(ctdb_db->ctdb,
1542 CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1543 CTDB_CTRL_FLAG_NOREPLY, /* flags */
1548 NULL, /* timeout : NULL == wait forever */
1549 NULL); /* error message */
1551 talloc_free(indata.dptr);
1553 if (ret != 0 || status != 0) {
1554 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1555 "SCHEDULE_FOR_DELETION "
1565 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1566 const struct ctdb_ltdb_header *hdr,
1569 if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1571 * Only remove the record from the delete queue if called
1572 * in the main daemon.
1577 remove_record_from_delete_queue(ctdb_db, hdr, key);