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;
336 uint32_t hash = ctdb_hash(&(dd->key));
340 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
343 (__location__ " Error getting chainlock on record with "
344 "key hash [0x%08x] on database db[%s].\n",
345 hash, ctdb_db->db_name));
350 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
351 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
352 /* Does not exist or not a ctdb record. Skip. */
356 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
357 /* The record has been recycled (filled with data). Skip. */
361 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
363 if (header->dmaster != ctdb->pnn) {
364 /* The record has been migrated off the node. Skip. */
368 if (header->rsn != dd->hdr.rsn) {
370 * The record has been migrated off the node and back again.
371 * But not requeued for deletion. Skip it.
377 * We are dmaster, and the record has no data, and it has
378 * not been migrated after it has been queued for deletion.
380 * At this stage, the record could still have been revived locally
381 * and last been written with empty data. This can only be
382 * fixed with the addition of an active or delete flag. (TODO)
385 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
387 if (lmaster != ctdb->pnn) {
388 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
392 (__location__ " Error adding record to list "
393 "of records to send to lmaster.\n"));
396 vdata->fast_added_to_vacuum_fetch_list++;
401 /* use header->flags or dd->hdr.flags ?? */
402 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
403 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
407 (__location__ " Error adding record to list "
408 "of records for deletion on lmaster.\n"));
411 vdata->fast_added_to_delete_list++;
414 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
418 (__location__ " Error deleting record with key "
419 "hash [0x%08x] from local data base db[%s].\n",
420 hash, ctdb_db->db_name));
424 (__location__ " Deleted record with key hash "
425 "[0x%08x] from local data base db[%s].\n",
426 hash, ctdb_db->db_name));
427 vdata->fast_deleted++;
434 vdata->fast_skipped++;
437 if (tdb_data.dptr != NULL) {
440 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
446 * Delete the records that we are lmaster and dmaster for and
447 * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
450 static int delete_record_traverse(void *param, void *data)
452 struct delete_record_data *dd =
453 talloc_get_type(data, struct delete_record_data);
454 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
455 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
456 struct ctdb_context *ctdb = ctdb_db->ctdb;
458 struct ctdb_ltdb_header *header;
461 bool deleted = false;
462 uint32_t hash = ctdb_hash(&(dd->key));
464 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
467 (__location__ " Error getting chainlock on record with "
468 "key hash [0x%08x] on database db[%s].\n",
469 hash, ctdb_db->db_name));
470 vdata->delete_local_error++;
475 * Verify that the record is still empty, its RSN has not
476 * changed and that we are still its lmaster and dmaster.
479 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
480 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
481 /* Does not exist or not a ctdb record. Skip. */
482 vdata->delete_skipped++;
486 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
487 /* The record has been recycled (filled with data). Skip. */
488 vdata->delete_skipped++;
492 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
494 if (header->dmaster != ctdb->pnn) {
495 /* The record has been migrated off the node. Skip. */
496 vdata->delete_skipped++;
501 if (header->rsn != dd->hdr.rsn) {
503 * The record has been migrated off the node and back again.
504 * But not requeued for deletion. Skip it.
506 vdata->delete_skipped++;
510 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
512 if (lmaster != ctdb->pnn) {
513 /* we are not lmaster - strange */
514 vdata->delete_skipped++;
518 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
522 (__location__ " Error deleting record with key hash "
523 "[0x%08x] from local data base db[%s].\n",
524 hash, ctdb_db->db_name));
525 vdata->delete_local_error++;
532 (__location__ " Deleted record with key hash [0x%08x] from "
533 "local data base db[%s].\n", hash, ctdb_db->db_name));
536 if (tdb_data.dptr != NULL) {
540 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
544 * successfully deleted the record locally.
545 * remove it from the list and update statistics.
548 vdata->delete_deleted++;
549 vdata->delete_left--;
556 * Fast vacuuming run:
557 * Traverse the delete_queue.
558 * This fills the same lists as the database traverse.
560 static void ctdb_vacuum_db_fast(struct ctdb_db_context *ctdb_db,
561 struct vacuum_data *vdata)
563 trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
565 if (vdata->fast_total > 0) {
568 " fast vacuuming delete_queue traverse statistics: "
577 (unsigned)vdata->fast_total,
578 (unsigned)vdata->fast_deleted,
579 (unsigned)vdata->fast_skipped,
580 (unsigned)vdata->fast_error,
581 (unsigned)vdata->fast_added_to_delete_list,
582 (unsigned)vdata->fast_added_to_vacuum_fetch_list));
590 * read-only traverse of the database, looking for records that
591 * might be able to be vacuumed.
593 * This is not done each time but only every tunable
594 * VacuumFastPathCount times.
596 static int ctdb_vacuum_db_full(struct ctdb_db_context *ctdb_db,
597 struct vacuum_data *vdata,
598 bool full_vacuum_run)
602 if (!full_vacuum_run) {
606 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
607 if (ret == -1 || vdata->traverse_error) {
608 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
609 "'%s'\n", ctdb_db->db_name));
613 if (vdata->full_total > 0) {
616 " full vacuuming db traverse statistics: "
624 (unsigned)vdata->full_total,
625 (unsigned)vdata->full_skipped,
626 (unsigned)vdata->full_error,
627 (unsigned)vdata->full_added_to_delete_list,
628 (unsigned)vdata->full_added_to_vacuum_fetch_list));
635 * Process the vacuum fetch lists:
636 * For records for which we are not the lmaster, tell the lmaster to
639 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
640 struct vacuum_data *vdata)
643 struct ctdb_context *ctdb = ctdb_db->ctdb;
645 for (i = 0; i < ctdb->num_nodes; i++) {
647 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
649 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
653 if (vfl->count == 0) {
657 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
658 vfl->count, ctdb->nodes[i]->pnn,
661 data.dsize = talloc_get_size(vfl);
662 data.dptr = (void *)vfl;
663 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn,
664 CTDB_SRVID_VACUUM_FETCH,
667 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
668 "fetch message to %u\n",
669 ctdb->nodes[i]->pnn));
678 * Proces the delete list:
679 * Send the records to delete to all other nodes with the
680 * try_delete_records control.
682 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
683 struct vacuum_data *vdata)
686 struct ctdb_context *ctdb = ctdb_db->ctdb;
688 vdata->delete_left = vdata->delete_count;
690 if (vdata->delete_count > 0) {
691 struct delete_records_list *recs;
692 TDB_DATA indata, outdata;
694 struct ctdb_node_map *nodemap;
695 uint32_t *active_nodes;
696 int num_active_nodes;
698 recs = talloc_zero(vdata, struct delete_records_list);
700 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
703 recs->records = (struct ctdb_marshall_buffer *)
704 talloc_zero_size(vdata,
705 offsetof(struct ctdb_marshall_buffer, data));
706 if (recs->records == NULL) {
707 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
710 recs->records->db_id = ctdb_db->db_id;
713 * traverse the tree of all records we want to delete and
714 * create a blob we can send to the other nodes.
716 trbt_traversearray32(vdata->delete_list, 1,
717 delete_marshall_traverse, recs);
719 indata.dsize = talloc_get_size(recs->records);
720 indata.dptr = (void *)recs->records;
723 * now tell all the active nodes to delete all these records
727 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
729 recs, /* talloc context */
732 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
736 active_nodes = list_of_active_nodes(ctdb, nodemap,
737 nodemap, /* talloc context */
738 false /* include self */);
740 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
742 for (i = 0; i < num_active_nodes; i++) {
743 struct ctdb_marshall_buffer *records;
744 struct ctdb_rec_data *rec;
746 ret = ctdb_control(ctdb, active_nodes[i], 0,
747 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
748 indata, recs, &outdata, &res,
750 if (ret != 0 || res != 0) {
751 DEBUG(DEBUG_ERR, ("Failed to delete records on "
752 "node %u: ret[%d] res[%d]\n",
753 active_nodes[i], ret, res));
758 * outdata contains the list of records coming back
759 * from the node: These are the records that the
760 * remote node could not delete.
762 * NOTE: There is a problem here:
764 * When a node failed to delete the record, but
765 * others succeeded, we may have created gaps in the
766 * history of the record. Hence when a node dies, an
767 * closed file handle might be resurrected or an open
768 * file handle might be lost, leading to blocked access
769 * or data corruption.
771 * TODO: This needs to be fixed!
773 records = (struct ctdb_marshall_buffer *)outdata.dptr;
774 rec = (struct ctdb_rec_data *)&records->data[0];
775 while (records->count-- > 1) {
776 TDB_DATA reckey, recdata;
777 struct ctdb_ltdb_header *rechdr;
778 struct delete_record_data *dd;
780 reckey.dptr = &rec->data[0];
781 reckey.dsize = rec->keylen;
782 recdata.dptr = &rec->data[reckey.dsize];
783 recdata.dsize = rec->datalen;
785 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
786 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
789 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
790 recdata.dptr += sizeof(*rechdr);
791 recdata.dsize -= sizeof(*rechdr);
793 dd = (struct delete_record_data *)trbt_lookup32(
795 (uint32_t)tdb_jenkins_hash(&reckey));
798 * The other node could not delete the
799 * record and it is the first node that
800 * failed. So we should remove it from
801 * the tree and update statistics.
804 vdata->delete_remote_error++;
805 vdata->delete_left--;
808 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
812 /* free nodemap and active_nodes */
813 talloc_free(nodemap);
816 if (vdata->delete_left > 0) {
818 * The only records remaining in the tree are those
819 * records which all other nodes could successfully
820 * delete, so we can safely delete them on the
823 trbt_traversearray32(vdata->delete_list, 1,
824 delete_record_traverse, vdata);
827 if (vdata->delete_count > 0) {
830 " vacuum delete list statistics: "
839 (unsigned)vdata->delete_count,
840 (unsigned)vdata->delete_remote_error,
841 (unsigned)vdata->delete_local_error,
842 (unsigned)vdata->delete_skipped,
843 (unsigned)vdata->delete_deleted,
844 (unsigned)vdata->delete_left));
851 * initialize the vacuum_data
853 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
854 struct vacuum_data *vdata)
857 struct ctdb_context *ctdb = ctdb_db->ctdb;
859 vdata->fast_added_to_delete_list = 0;
860 vdata->fast_added_to_vacuum_fetch_list = 0;
861 vdata->fast_deleted = 0;
862 vdata->fast_skipped = 0;
863 vdata->fast_error = 0;
864 vdata->fast_total = 0;
865 vdata->full_added_to_delete_list = 0;
866 vdata->full_added_to_vacuum_fetch_list = 0;
867 vdata->full_skipped = 0;
868 vdata->full_error = 0;
869 vdata->full_total = 0;
870 vdata->delete_count = 0;
871 vdata->delete_left = 0;
872 vdata->delete_remote_error = 0;
873 vdata->delete_local_error = 0;
874 vdata->delete_skipped = 0;
875 vdata->delete_deleted = 0;
877 /* the list needs to be of length num_nodes */
878 vdata->vacuum_fetch_list = talloc_zero_array(vdata,
879 struct ctdb_marshall_buffer *,
881 if (vdata->vacuum_fetch_list == NULL) {
882 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
885 for (i = 0; i < ctdb->num_nodes; i++) {
886 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
887 talloc_zero_size(vdata->vacuum_fetch_list,
888 offsetof(struct ctdb_marshall_buffer, data));
889 if (vdata->vacuum_fetch_list[i] == NULL) {
890 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
893 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
901 * - Always do the fast vacuuming run, which traverses
902 * the in-memory delete queue: these records have been
903 * scheduled for deletion.
904 * - Only if explicitly requested, the database is traversed
905 * in order to use the traditional heuristics on empty records
906 * to trigger deletion.
907 * This is done only every VacuumFastPathCount'th vacuuming run.
909 * The traverse runs fill two lists:
912 * This is the list of empty records the current
913 * node is lmaster and dmaster for. These records are later
914 * deleted first on other nodes and then locally.
916 * The fast vacuuming run has a short cut for those records
917 * that have never been migrated with data: these records
918 * are immediately deleted locally, since they have left
919 * no trace on other nodes.
921 * - The vacuum_fetch lists
922 * (one for each other lmaster node):
923 * The records in this list are sent for deletion to
924 * their lmaster in a bulk VACUUM_FETCH message.
926 * The lmaster then migrates all these records to itelf
927 * so that they can be vacuumed there.
929 * This executes in the child context.
931 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
932 struct vacuum_data *vdata,
933 bool full_vacuum_run)
935 struct ctdb_context *ctdb = ctdb_db->ctdb;
938 DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
939 "%s db_id[0x%08x]\n",
940 full_vacuum_run ? "full" : "fast",
941 ctdb_db->db_name, ctdb_db->db_id));
943 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
945 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
949 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
951 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
957 ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
962 ctdb_vacuum_db_fast(ctdb_db, vdata);
964 ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
969 ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
974 ret = ctdb_process_delete_list(ctdb_db, vdata);
979 /* this ensures we run our event queue */
980 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
987 * traverse function for repacking
989 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
991 struct vacuum_data *vdata = (struct vacuum_data *)private;
994 uint32_t hash = (uint32_t)tdb_jenkins_hash(&key);
995 struct delete_record_data *kd;
997 * check if we can ignore this record because it's in the delete_list
999 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1001 * there might be hash collisions so we have to compare the keys here to be sure
1003 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1004 struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1006 * we have to check if the record hasn't changed in the meantime in order to
1007 * savely remove it from the database
1009 if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1010 hdr->dmaster == kd->ctdb->pnn &&
1011 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1012 kd->hdr.rsn == hdr->rsn) {
1018 if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1019 vdata->traverse_error = true;
1029 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1031 struct tdb_context *tmp_db;
1033 if (tdb_transaction_start(tdb) != 0) {
1034 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1038 tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1039 TDB_INTERNAL|TDB_DISALLOW_NESTING,
1041 if (tmp_db == NULL) {
1042 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1043 tdb_transaction_cancel(tdb);
1047 vdata->traverse_error = false;
1048 vdata->dest_db = tmp_db;
1049 vdata->vacuum = true;
1050 vdata->vacuumed = 0;
1054 * repack and vacuum on-the-fly by not writing the records that are
1057 if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1058 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1059 tdb_transaction_cancel(tdb);
1064 DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1066 if (vdata->traverse_error) {
1067 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1068 tdb_transaction_cancel(tdb);
1073 if (tdb_wipe_all(tdb) != 0) {
1074 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1075 tdb_transaction_cancel(tdb);
1080 vdata->traverse_error = false;
1081 vdata->dest_db = tdb;
1082 vdata->vacuum = false;
1085 if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1086 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1087 tdb_transaction_cancel(tdb);
1092 if (vdata->traverse_error) {
1093 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1094 tdb_transaction_cancel(tdb);
1102 if (tdb_transaction_commit(tdb) != 0) {
1103 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1106 DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1112 * repack and vaccum a db
1113 * called from the child context
1115 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1116 TALLOC_CTX *mem_ctx,
1117 bool full_vacuum_run)
1119 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1120 uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1121 const char *name = ctdb_db->db_name;
1123 struct vacuum_data *vdata;
1125 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1126 if (freelist_size == -1) {
1127 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1131 vdata = talloc_zero(mem_ctx, struct vacuum_data);
1132 if (vdata == NULL) {
1133 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1137 vdata->ctdb = ctdb_db->ctdb;
1138 vdata->vacuum_limit = vacuum_limit;
1139 vdata->repack_limit = repack_limit;
1140 vdata->delete_list = trbt_create(vdata, 0);
1141 vdata->ctdb_db = ctdb_db;
1142 if (vdata->delete_list == NULL) {
1143 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1148 vdata->start = timeval_current();
1151 * gather all records that can be deleted in vdata
1153 if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1154 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1158 * decide if a repack is necessary
1160 if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1166 DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1167 name, freelist_size, vdata->delete_left));
1170 * repack and implicitely get rid of the records we can delete
1172 if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1173 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1182 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1184 uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1189 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1191 double l = timeval_elapsed(&child_ctx->start_time);
1192 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1193 struct ctdb_context *ctdb = ctdb_db->ctdb;
1195 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1197 if (child_ctx->child_pid != -1) {
1198 kill(child_ctx->child_pid, SIGKILL);
1200 /* Bump the number of successful fast-path runs. */
1201 child_ctx->vacuum_handle->fast_path_count++;
1204 DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1206 event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1207 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1208 ctdb_vacuum_event, child_ctx->vacuum_handle);
1214 * this event is generated when a vacuum child process times out
1216 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1217 struct timeval t, void *private_data)
1219 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1221 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1223 child_ctx->status = VACUUM_TIMEOUT;
1225 talloc_free(child_ctx);
1230 * this event is generated when a vacuum child process has completed
1232 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1233 uint16_t flags, void *private_data)
1235 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1239 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1240 child_ctx->child_pid = -1;
1242 ret = read(child_ctx->fd[0], &c, 1);
1243 if (ret != 1 || c != 0) {
1244 child_ctx->status = VACUUM_ERROR;
1245 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));
1247 child_ctx->status = VACUUM_OK;
1250 talloc_free(child_ctx);
1254 * this event is called every time we need to start a new vacuum process
1257 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1258 struct timeval t, void *private_data)
1260 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1261 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1262 struct ctdb_context *ctdb = ctdb_db->ctdb;
1263 struct ctdb_vacuum_child_context *child_ctx;
1266 /* we dont vacuum if we are in recovery mode, or db frozen */
1267 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1268 ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1269 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1270 ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1271 : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
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 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1281 if (child_ctx == NULL) {
1282 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1283 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1287 ret = pipe(child_ctx->fd);
1289 talloc_free(child_ctx);
1290 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1291 event_add_timed(ctdb->ev, vacuum_handle,
1292 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1293 ctdb_vacuum_event, vacuum_handle);
1297 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1298 vacuum_handle->fast_path_count = 0;
1301 child_ctx->child_pid = fork();
1302 if (child_ctx->child_pid == (pid_t)-1) {
1303 close(child_ctx->fd[0]);
1304 close(child_ctx->fd[1]);
1305 talloc_free(child_ctx);
1306 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1307 event_add_timed(ctdb->ev, vacuum_handle,
1308 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1309 ctdb_vacuum_event, vacuum_handle);
1314 if (child_ctx->child_pid == 0) {
1316 bool full_vacuum_run = false;
1317 close(child_ctx->fd[0]);
1319 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1321 if (switch_from_server_to_client(ctdb) != 0) {
1322 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1329 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1330 (vacuum_handle->fast_path_count == 0))
1332 full_vacuum_run = true;
1334 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1337 write(child_ctx->fd[1], &cc, 1);
1341 set_close_on_exec(child_ctx->fd[0]);
1342 close(child_ctx->fd[1]);
1344 child_ctx->status = VACUUM_RUNNING;
1345 child_ctx->start_time = timeval_current();
1347 DLIST_ADD(ctdb->vacuumers, child_ctx);
1348 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1351 * Clear the fastpath vacuuming list in the parent.
1353 talloc_free(ctdb_db->delete_queue);
1354 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1355 if (ctdb_db->delete_queue == NULL) {
1356 /* fatal here? ... */
1357 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1358 "in parent context. Shutting down\n");
1361 event_add_timed(ctdb->ev, child_ctx,
1362 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1363 vacuum_child_timeout, child_ctx);
1365 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1367 event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1368 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1369 vacuum_child_handler,
1372 vacuum_handle->child_ctx = child_ctx;
1373 child_ctx->vacuum_handle = vacuum_handle;
1376 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1378 /* Simply free them all. */
1379 while (ctdb->vacuumers) {
1380 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1381 ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1382 (int)ctdb->vacuumers->child_pid));
1383 /* vacuum_child_destructor kills it, removes from list */
1384 talloc_free(ctdb->vacuumers);
1388 /* this function initializes the vacuuming context for a database
1389 * starts the vacuuming events
1391 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1393 if (ctdb_db->persistent != 0) {
1394 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1398 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1399 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1401 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1402 ctdb_db->vacuum_handle->fast_path_count = 0;
1404 event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1405 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1406 ctdb_vacuum_event, ctdb_db->vacuum_handle);
1411 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1412 const struct ctdb_ltdb_header *hdr,
1415 struct delete_record_data *kd;
1418 hash = (uint32_t)tdb_jenkins_hash(discard_const(&key));
1420 DEBUG(DEBUG_DEBUG, (__location__
1421 " remove_record_from_delete_queue: db[%s] "
1425 "migrated_with_data[%s]\n",
1426 ctdb_db->db_name, ctdb_db->db_id,
1428 ctdb_lmaster(ctdb_db->ctdb, &key),
1429 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1431 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1435 if (kd->key.dsize != key.dsize) {
1438 if (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0) {
1448 * Insert a record into the ctdb_db context's delete queue,
1449 * handling hash collisions.
1451 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1452 const struct ctdb_ltdb_header *hdr,
1455 struct delete_record_data *kd;
1459 hash = (uint32_t)tdb_jenkins_hash(&key);
1461 DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1465 "migrated_with_data[%s]\n",
1466 ctdb_db->db_name, ctdb_db->db_id,
1468 ctdb_lmaster(ctdb_db->ctdb, &key),
1469 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1471 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1473 if ((kd->key.dsize != key.dsize) ||
1474 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1477 (__location__ " schedule for deletion: "
1478 "hash collision for key hash [0x%08x]. "
1479 "Skipping the record.\n", hash));
1483 (__location__ " schedule for deletion: "
1484 "updating entry for key with hash [0x%08x].\n",
1489 ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1490 ctdb_db->delete_queue,
1494 (__location__ " schedule for deletion: error "
1495 "inserting key with hash [0x%08x] into delete queue\n",
1504 * Schedule a record for deletetion.
1505 * Called from the parent context.
1507 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1510 struct ctdb_control_schedule_for_deletion *dd;
1511 struct ctdb_db_context *ctdb_db;
1515 dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1517 ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1518 if (ctdb_db == NULL) {
1519 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1524 key.dsize = dd->keylen;
1527 ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1532 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1533 const struct ctdb_ltdb_header *hdr,
1537 struct ctdb_control_schedule_for_deletion *dd;
1541 if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1542 /* main daemon - directly queue */
1543 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1548 /* child process: send the main daemon a control */
1550 indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1551 indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1552 if (indata.dptr == NULL) {
1553 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1556 dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1557 dd->db_id = ctdb_db->db_id;
1559 dd->keylen = key.dsize;
1560 memcpy(dd->key, key.dptr, key.dsize);
1562 ret = ctdb_control(ctdb_db->ctdb,
1565 CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1566 CTDB_CTRL_FLAG_NOREPLY, /* flags */
1571 NULL, /* timeout : NULL == wait forever */
1572 NULL); /* error message */
1574 talloc_free(indata.dptr);
1576 if (ret != 0 || status != 0) {
1577 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1578 "SCHEDULE_FOR_DELETION "
1588 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1589 const struct ctdb_ltdb_header *hdr,
1592 if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1594 * Only remove the record from the delete queue if called
1595 * in the main daemon.
1600 remove_record_from_delete_queue(ctdb_db, hdr, key);