4 Copyright (C) Ronnie Sahlberg 2009
6 This program is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 3 of the License, or
9 (at your option) any later version.
11 This program is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with this program; if not, see <http://www.gnu.org/licenses/>.
21 #include "lib/tevent/tevent.h"
22 #include "lib/tdb/include/tdb.h"
23 #include "system/network.h"
24 #include "system/filesys.h"
25 #include "system/dir.h"
26 #include "../include/ctdb_private.h"
28 #include "lib/util/dlinklist.h"
29 #include "lib/tevent/tevent.h"
30 #include "../include/ctdb_private.h"
31 #include "../common/rb_tree.h"
33 #define TIMELIMIT() timeval_current_ofs(10, 0)
34 #define TUNINGDBNAME "vactune.tdb"
36 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
38 struct ctdb_vacuum_child_context {
39 struct ctdb_vacuum_child_context *next, *prev;
40 struct ctdb_vacuum_handle *vacuum_handle;
41 /* fd child writes status to */
44 enum vacuum_child_status status;
45 struct timeval start_time;
48 struct ctdb_vacuum_handle {
49 struct ctdb_db_context *ctdb_db;
50 struct ctdb_vacuum_child_context *child_ctx;
51 uint32_t fast_path_count;
55 /* a list of records to possibly delete */
57 uint32_t vacuum_limit;
58 uint32_t repack_limit;
59 struct ctdb_context *ctdb;
60 struct ctdb_db_context *ctdb_db;
61 struct tdb_context *dest_db;
62 trbt_tree_t *delete_tree;
63 uint32_t delete_count;
64 struct ctdb_marshall_buffer **list;
71 uint32_t fast_added_to_vacuum_fetch_list;
72 uint32_t fast_added_to_delete_tree;
73 uint32_t fast_deleted;
74 uint32_t fast_skipped;
77 uint32_t full_added_to_vacuum_fetch_list;
78 uint32_t full_added_to_delete_tree;
79 uint32_t full_skipped;
84 /* tuning information stored for every db */
85 struct vacuum_tuning_data {
86 uint32_t last_num_repack;
87 uint32_t last_num_empty;
88 uint32_t last_interval;
89 uint32_t new_interval;
90 struct timeval last_start;
94 /* this structure contains the information for one record to be deleted */
95 struct delete_record_data {
96 struct ctdb_context *ctdb;
97 struct ctdb_db_context *ctdb_db;
98 struct ctdb_ltdb_header hdr;
102 struct delete_records_list {
103 struct ctdb_marshall_buffer *records;
107 * Store key and header in a tree, indexed by the key hash.
109 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
110 struct ctdb_db_context *ctdb_db,
112 const struct ctdb_ltdb_header *hdr,
115 struct delete_record_data *dd;
118 dd = talloc_zero(tree, struct delete_record_data);
120 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
125 dd->ctdb_db = ctdb_db;
126 dd->key.dsize = key.dsize;
127 dd->key.dptr = talloc_memdup(dd, key.dptr, key.dsize);
128 if (dd->key.dptr == NULL) {
129 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
135 hash = ctdb_hash(&key);
137 trbt_insert32(tree, hash, dd);
142 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
143 struct ctdb_ltdb_header *hdr)
145 struct ctdb_context *ctdb = vdata->ctdb;
146 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
150 hash = ctdb_hash(&key);
152 if (trbt_lookup32(vdata->delete_tree, hash)) {
153 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
157 ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
164 vdata->delete_count++;
170 * Add a record to the list of records to be sent
171 * to their lmaster with VACUUM_FETCH.
173 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
176 struct ctdb_context *ctdb = vdata->ctdb;
177 struct ctdb_rec_data *rec;
181 lmaster = ctdb_lmaster(ctdb, &key);
183 rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
185 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
186 vdata->traverse_error = true;
190 old_size = talloc_get_size(vdata->list[lmaster]);
191 vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
192 old_size + rec->length);
193 if (vdata->list[lmaster] == NULL) {
194 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
195 vdata->traverse_error = true;
199 vdata->list[lmaster]->count++;
200 memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
209 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
210 struct timeval t, void *private_data);
214 * traverse function for gathering the records that can be deleted
216 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
218 struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
219 struct ctdb_context *ctdb = vdata->ctdb;
221 struct ctdb_ltdb_header *hdr;
226 lmaster = ctdb_lmaster(ctdb, &key);
227 if (lmaster >= ctdb->num_nodes) {
229 DEBUG(DEBUG_CRIT, (__location__
230 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
233 (unsigned)ctdb->num_nodes,
234 (unsigned)ctdb_hash(&key)));
238 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
239 /* it is not a deleted record */
240 vdata->full_skipped++;
244 hdr = (struct ctdb_ltdb_header *)data.dptr;
246 if (hdr->dmaster != ctdb->pnn) {
247 vdata->full_skipped++;
251 if (lmaster == ctdb->pnn) {
253 * We are both lmaster and dmaster, and the record is empty.
254 * So we should be able to delete it.
256 res = add_record_to_delete_tree(vdata, key, hdr);
260 vdata->full_added_to_delete_tree++;
264 * We are not lmaster.
265 * Add the record to the blob ready to send to the nodes.
267 res = add_record_to_vacuum_fetch_list(vdata, key);
271 vdata->full_added_to_vacuum_fetch_list++;
279 * traverse the tree of records to delete and marshall them into
282 static int delete_traverse(void *param, void *data)
284 struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
285 struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
286 struct ctdb_rec_data *rec;
289 rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
291 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
295 old_size = talloc_get_size(recs->records);
296 recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
297 if (recs->records == NULL) {
298 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
301 recs->records->count++;
302 memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
307 * traverse function for the traversal of the delete_queue,
308 * the fast-path vacuuming list.
310 * - If the record has been migrated off the node
311 * or has been revived (filled with data) on the node,
312 * then skip the record.
314 * - If the current node is the record's lmaster and it is
315 * a record that has never been migrated with data, then
316 * delete the record from the local tdb.
318 * - If the current node is the record's lmaster and it has
319 * been migrated with data, then schedule it for the normal
320 * vacuuming procedure (i.e. add it to the delete_list).
322 * - If the current node is NOT the record's lmaster then
323 * add it to the list of records that are to be sent to
324 * the lmaster with the VACUUM_FETCH message.
326 static int delete_queue_traverse(void *param, void *data)
328 struct delete_record_data *dd =
329 talloc_get_type(data, struct delete_record_data);
330 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
331 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
332 struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
334 struct ctdb_ltdb_header *header;
340 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
342 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
347 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
348 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
349 /* Does not exist or not a ctdb record. Skip. */
353 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
354 /* The record has been recycled (filled with data). Skip. */
358 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
360 if (header->dmaster != ctdb->pnn) {
361 /* The record has been migrated off the node. Skip. */
366 if (header->rsn != dd->hdr.rsn) {
368 * The record has been migrated off the node and back again.
369 * But not requeued for deletion. Skip it.
375 * We are dmaster, and the record has no data, and it has
376 * not been migrated after it has been queued for deletion.
378 * At this stage, the record could still have been revived locally
379 * and last been written with empty data. This can only be
380 * fixed with the addition of an active or delete flag. (TODO)
383 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
385 if (lmaster != ctdb->pnn) {
386 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
390 (__location__ " Error adding record to list "
391 "of records to send to lmaster.\n"));
394 vdata->fast_added_to_vacuum_fetch_list++;
399 /* use header->flags or dd->hdr.flags ?? */
400 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
401 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
405 (__location__ " Error adding record to list "
406 "of records for deletion on lmaster.\n"));
409 vdata->fast_added_to_delete_tree++;
412 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
416 (__location__ " Error deleting record from local "
420 vdata->fast_deleted++;
427 vdata->fast_skipped++;
430 if (tdb_data.dptr != NULL) {
433 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
440 * - Always do the fast vacuuming run, which traverses
441 * the in-memory delete queue: these records have been
442 * scheduled for deletion.
443 * - Only if explicitly requested, the database is traversed
444 * in order to use the traditional heuristics on empty records
445 * to trigger deletion.
446 * This is done only every VacuumFastPathCount'th vacuuming run.
448 * The traverse runs fill two lists:
451 * This is the list of empty records the current
452 * node is lmaster and dmaster for. These records are later
453 * deleted first on other nodes and then locally.
455 * The fast vacuuming run has a short cut for those records
456 * that have never been migrated with data: these records
457 * are immediately deleted locally, since they have left
458 * no trace on other nodes.
460 * - The vacuum_fetch lists
461 * (one for each other lmaster node):
462 * The records in this list are sent for deletion to
463 * their lmaster in a bulk VACUUM_FETCH message.
465 * The lmaster then migrates all these records to itelf
466 * so that they can be vacuumed there.
468 * This executes in the child context.
470 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
471 struct vacuum_data *vdata,
472 bool full_vacuum_run)
474 struct ctdb_context *ctdb = ctdb_db->ctdb;
475 const char *name = ctdb_db->db_name;
478 DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
479 "%s db_id[0x%08x]\n",
480 full_vacuum_run ? "full" : "fast",
481 ctdb_db->db_name, ctdb_db->db_id));
483 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
485 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
489 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
491 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
497 vdata->fast_added_to_delete_tree = 0;
498 vdata->fast_added_to_vacuum_fetch_list = 0;
499 vdata->fast_deleted = 0;
500 vdata->fast_skipped = 0;
501 vdata->fast_error = 0;
502 vdata->fast_total = 0;
503 vdata->full_added_to_delete_tree = 0;
504 vdata->full_added_to_vacuum_fetch_list = 0;
505 vdata->full_skipped = 0;
506 vdata->full_error = 0;
507 vdata->full_total = 0;
509 /* the list needs to be of length num_nodes */
510 vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
511 if (vdata->list == NULL) {
512 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
515 for (i = 0; i < ctdb->num_nodes; i++) {
516 vdata->list[i] = (struct ctdb_marshall_buffer *)
517 talloc_zero_size(vdata->list,
518 offsetof(struct ctdb_marshall_buffer, data));
519 if (vdata->list[i] == NULL) {
520 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
523 vdata->list[i]->db_id = ctdb_db->db_id;
527 * Traverse the delete_queue.
528 * This builds the same lists as the db traverse.
530 trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
532 if (vdata->fast_total > 0) {
535 " fast vacuuming delete_queue traverse statistics: "
544 (unsigned)vdata->fast_total,
545 (unsigned)vdata->fast_deleted,
546 (unsigned)vdata->fast_skipped,
547 (unsigned)vdata->fast_error,
548 (unsigned)vdata->fast_added_to_delete_tree,
549 (unsigned)vdata->fast_added_to_vacuum_fetch_list));
553 * read-only traverse of the database, looking for records that
554 * might be able to be vacuumed.
556 * This is not done each time but only every tunable
557 * VacuumFastPathCount times.
559 if (full_vacuum_run) {
560 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
561 if (ret == -1 || vdata->traverse_error) {
562 DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
565 if (vdata->full_total > 0) {
568 " full vacuuming db traverse statistics: "
576 (unsigned)vdata->full_total,
577 (unsigned)vdata->full_skipped,
578 (unsigned)vdata->full_error,
579 (unsigned)vdata->full_added_to_delete_tree,
580 (unsigned)vdata->full_added_to_vacuum_fetch_list));
585 * For records where we are not the lmaster,
586 * tell the lmaster to fetch the record.
588 for (i = 0; i < ctdb->num_nodes; i++) {
591 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
595 if (vdata->list[i]->count == 0) {
599 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
600 vdata->list[i]->count, ctdb->nodes[i]->pnn,
603 data.dsize = talloc_get_size(vdata->list[i]);
604 data.dptr = (void *)vdata->list[i];
605 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
606 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
607 "fetch message to %u\n",
608 ctdb->nodes[i]->pnn));
613 /* Process all records we can delete (if any) */
614 if (vdata->delete_count > 0) {
615 struct delete_records_list *recs;
616 TDB_DATA indata, outdata;
618 struct ctdb_node_map *nodemap;
619 uint32_t *active_nodes;
620 int num_active_nodes;
622 recs = talloc_zero(vdata, struct delete_records_list);
624 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
627 recs->records = (struct ctdb_marshall_buffer *)
628 talloc_zero_size(vdata,
629 offsetof(struct ctdb_marshall_buffer, data));
630 if (recs->records == NULL) {
631 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
634 recs->records->db_id = ctdb_db->db_id;
637 * traverse the tree of all records we want to delete and
638 * create a blob we can send to the other nodes.
640 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
642 indata.dsize = talloc_get_size(recs->records);
643 indata.dptr = (void *)recs->records;
646 * now tell all the active nodes to delete all these records
650 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
652 recs, /* talloc context */
655 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
659 active_nodes = list_of_active_nodes(ctdb, nodemap,
660 nodemap, /* talloc context */
661 false /* include self */);
663 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
665 for (i = 0; i < num_active_nodes; i++) {
666 struct ctdb_marshall_buffer *records;
667 struct ctdb_rec_data *rec;
669 ret = ctdb_control(ctdb, active_nodes[i], 0,
670 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
671 indata, recs, &outdata, &res,
673 if (ret != 0 || res != 0) {
674 DEBUG(DEBUG_ERR, ("Failed to delete records on "
675 "node %u: ret[%d] res[%d]\n",
676 active_nodes[i], ret, res));
681 * outdata countains the list of records coming back
682 * from the node which the node could not delete
684 records = (struct ctdb_marshall_buffer *)outdata.dptr;
685 rec = (struct ctdb_rec_data *)&records->data[0];
686 while (records->count-- > 1) {
687 TDB_DATA reckey, recdata;
688 struct ctdb_ltdb_header *rechdr;
690 reckey.dptr = &rec->data[0];
691 reckey.dsize = rec->keylen;
692 recdata.dptr = &rec->data[reckey.dsize];
693 recdata.dsize = rec->datalen;
695 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
696 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
699 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
700 recdata.dptr += sizeof(*rechdr);
701 recdata.dsize -= sizeof(*rechdr);
704 * that other node couldnt delete the record
705 * so we should delete it and thereby remove it from the tree
707 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
709 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
713 /* free nodemap and active_nodes */
714 talloc_free(nodemap);
717 * The only records remaining in the tree would be those
718 * records where all other nodes could successfully
719 * delete them, so we can safely delete them on the
720 * lmaster as well. Deletion implictely happens while
721 * we repack the database. The repack algorithm revisits
722 * the tree in order to find the records that don't need
723 * to be copied / repacked.
727 /* this ensures we run our event queue */
728 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
735 * traverse function for repacking
737 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
739 struct vacuum_data *vdata = (struct vacuum_data *)private;
742 uint32_t hash = ctdb_hash(&key);
743 struct delete_record_data *kd;
745 * check if we can ignore this record because it's in the delete_tree
747 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
749 * there might be hash collisions so we have to compare the keys here to be sure
751 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
752 struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
754 * we have to check if the record hasn't changed in the meantime in order to
755 * savely remove it from the database
757 if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
758 hdr->dmaster == kd->ctdb->pnn &&
759 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
760 kd->hdr.rsn == hdr->rsn) {
766 if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
767 vdata->traverse_error = true;
777 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
779 struct tdb_context *tmp_db;
781 if (tdb_transaction_start(tdb) != 0) {
782 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
786 tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
787 TDB_INTERNAL|TDB_DISALLOW_NESTING,
789 if (tmp_db == NULL) {
790 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
791 tdb_transaction_cancel(tdb);
795 vdata->traverse_error = false;
796 vdata->dest_db = tmp_db;
797 vdata->vacuum = true;
802 * repack and vacuum on-the-fly by not writing the records that are
805 if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
806 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
807 tdb_transaction_cancel(tdb);
812 DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
814 if (vdata->traverse_error) {
815 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
816 tdb_transaction_cancel(tdb);
821 if (tdb_wipe_all(tdb) != 0) {
822 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
823 tdb_transaction_cancel(tdb);
828 vdata->traverse_error = false;
829 vdata->dest_db = tdb;
830 vdata->vacuum = false;
833 if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
834 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
835 tdb_transaction_cancel(tdb);
840 if (vdata->traverse_error) {
841 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
842 tdb_transaction_cancel(tdb);
850 if (tdb_transaction_commit(tdb) != 0) {
851 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
854 DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
859 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
861 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
862 TDB_CONTEXT *tune_tdb;
864 struct vacuum_tuning_data tdata;
865 struct vacuum_tuning_data *tptr;
869 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
870 ctdb_db->ctdb->db_directory_state,
871 TUNINGDBNAME, ctdb_db->ctdb->pnn);
872 if (vac_dbname == NULL) {
873 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
874 talloc_free(tmp_ctx);
878 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
879 flags |= TDB_DISALLOW_NESTING;
880 tune_tdb = tdb_open(vac_dbname, 0,
882 O_RDWR|O_CREAT, 0600);
883 if (tune_tdb == NULL) {
884 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
885 talloc_free(tmp_ctx);
889 if (tdb_transaction_start(tune_tdb) != 0) {
890 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
894 key.dptr = discard_const(ctdb_db->db_name);
895 key.dsize = strlen(ctdb_db->db_name);
896 value = tdb_fetch(tune_tdb, key);
898 if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
899 tptr = (struct vacuum_tuning_data *)value.dptr;
903 * re-calc new vacuum interval:
904 * in case no limit was reached we continuously increase the interval
905 * until vacuum_max_interval is reached
906 * in case a limit was reached we divide the current interval by 2
907 * unless vacuum_min_interval is reached
909 if (freelist < vdata->repack_limit &&
910 vdata->delete_count < vdata->vacuum_limit) {
911 if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
912 tdata.new_interval = tdata.last_interval * 110 / 100;
913 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
914 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
917 tdata.new_interval = tdata.last_interval / 2;
918 if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
919 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
920 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
922 DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
923 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
925 tdata.last_interval = tdata.new_interval;
927 DEBUG(DEBUG_DEBUG,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
928 tdata.last_num_repack = freelist;
929 tdata.last_num_empty = vdata->delete_count;
930 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
933 if (value.dptr != NULL) {
937 tdata.last_start = vdata->start;
938 tdata.last_duration = timeval_elapsed(&vdata->start);
940 value.dptr = (unsigned char *)&tdata;
941 value.dsize = sizeof(tdata);
943 if (tdb_store(tune_tdb, key, value, 0) != 0) {
944 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
945 tdb_transaction_cancel(tune_tdb);
947 talloc_free(tmp_ctx);
950 tdb_transaction_commit(tune_tdb);
952 talloc_free(tmp_ctx);
958 * repack and vaccum a db
959 * called from the child context
961 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
963 bool full_vacuum_run)
965 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
966 uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
967 const char *name = ctdb_db->db_name;
969 struct vacuum_data *vdata;
971 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
972 if (freelist_size == -1) {
973 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
977 vdata = talloc_zero(mem_ctx, struct vacuum_data);
979 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
983 vdata->ctdb = ctdb_db->ctdb;
984 vdata->vacuum_limit = vacuum_limit;
985 vdata->repack_limit = repack_limit;
986 vdata->delete_tree = trbt_create(vdata, 0);
987 vdata->ctdb_db = ctdb_db;
988 if (vdata->delete_tree == NULL) {
989 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
994 vdata->start = timeval_current();
997 * gather all records that can be deleted in vdata
999 if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1000 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1004 * decide if a repack is necessary
1006 if (freelist_size < repack_limit && vdata->delete_count < vacuum_limit)
1008 update_tuning_db(ctdb_db, vdata, freelist_size);
1013 DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1014 name, freelist_size, vdata->delete_count));
1017 * repack and implicitely get rid of the records we can delete
1019 if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1020 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1021 update_tuning_db(ctdb_db, vdata, freelist_size);
1025 update_tuning_db(ctdb_db, vdata, freelist_size);
1031 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1033 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1035 TDB_DATA key, value;
1037 uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
1038 struct ctdb_context *ctdb = ctdb_db->ctdb;
1041 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
1042 if (vac_dbname == NULL) {
1043 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
1044 talloc_free(tmp_ctx);
1048 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
1049 flags |= TDB_DISALLOW_NESTING;
1050 tdb = tdb_open(vac_dbname, 0,
1052 O_RDWR|O_CREAT, 0600);
1054 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval. Errno : %s (%d)\n", vac_dbname, strerror(errno), errno));
1055 talloc_free(tmp_ctx);
1059 key.dptr = discard_const(ctdb_db->db_name);
1060 key.dsize = strlen(ctdb_db->db_name);
1062 value = tdb_fetch(tdb, key);
1064 if (value.dptr != NULL) {
1065 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
1066 struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
1068 interval = tptr->new_interval;
1070 if (interval < ctdb->tunable.vacuum_min_interval) {
1071 interval = ctdb->tunable.vacuum_min_interval;
1073 if (interval > ctdb->tunable.vacuum_max_interval) {
1074 interval = ctdb->tunable.vacuum_max_interval;
1081 talloc_free(tmp_ctx);
1086 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1088 double l = timeval_elapsed(&child_ctx->start_time);
1089 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1090 struct ctdb_context *ctdb = ctdb_db->ctdb;
1092 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1094 if (child_ctx->child_pid != -1) {
1095 kill(child_ctx->child_pid, SIGKILL);
1097 /* Bump the number of successful fast-path runs. */
1098 child_ctx->vacuum_handle->fast_path_count++;
1101 DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1103 event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1104 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1105 ctdb_vacuum_event, child_ctx->vacuum_handle);
1111 * this event is generated when a vacuum child process times out
1113 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1114 struct timeval t, void *private_data)
1116 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1118 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1120 child_ctx->status = VACUUM_TIMEOUT;
1122 talloc_free(child_ctx);
1127 * this event is generated when a vacuum child process has completed
1129 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1130 uint16_t flags, void *private_data)
1132 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1136 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1137 child_ctx->child_pid = -1;
1139 ret = read(child_ctx->fd[0], &c, 1);
1140 if (ret != 1 || c != 0) {
1141 child_ctx->status = VACUUM_ERROR;
1142 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));
1144 child_ctx->status = VACUUM_OK;
1147 talloc_free(child_ctx);
1151 * this event is called every time we need to start a new vacuum process
1154 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1155 struct timeval t, void *private_data)
1157 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1158 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1159 struct ctdb_context *ctdb = ctdb_db->ctdb;
1160 struct ctdb_vacuum_child_context *child_ctx;
1161 struct tevent_fd *fde;
1164 /* we dont vacuum if we are in recovery mode, or db frozen */
1165 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1166 ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1167 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1168 ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1169 : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1172 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1176 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1177 if (child_ctx == NULL) {
1178 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1179 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1183 ret = pipe(child_ctx->fd);
1185 talloc_free(child_ctx);
1186 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1187 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1191 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1192 vacuum_handle->fast_path_count = 0;
1195 child_ctx->child_pid = ctdb_fork(ctdb);
1196 if (child_ctx->child_pid == (pid_t)-1) {
1197 close(child_ctx->fd[0]);
1198 close(child_ctx->fd[1]);
1199 talloc_free(child_ctx);
1200 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1201 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1206 if (child_ctx->child_pid == 0) {
1208 bool full_vacuum_run = false;
1209 close(child_ctx->fd[0]);
1211 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1213 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1214 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1221 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1222 (vacuum_handle->fast_path_count == 0))
1224 full_vacuum_run = true;
1226 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1229 write(child_ctx->fd[1], &cc, 1);
1233 set_close_on_exec(child_ctx->fd[0]);
1234 close(child_ctx->fd[1]);
1236 child_ctx->status = VACUUM_RUNNING;
1237 child_ctx->start_time = timeval_current();
1239 DLIST_ADD(ctdb->vacuumers, child_ctx);
1240 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1243 * Clear the fastpath vacuuming list in the parent.
1245 talloc_free(ctdb_db->delete_queue);
1246 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1247 if (ctdb_db->delete_queue == NULL) {
1248 /* fatal here? ... */
1249 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1250 "in parent context. Shutting down\n");
1253 event_add_timed(ctdb->ev, child_ctx,
1254 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1255 vacuum_child_timeout, child_ctx);
1257 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1259 fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1260 EVENT_FD_READ, vacuum_child_handler, child_ctx);
1261 tevent_fd_set_auto_close(fde);
1263 vacuum_handle->child_ctx = child_ctx;
1264 child_ctx->vacuum_handle = vacuum_handle;
1267 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1269 /* Simply free them all. */
1270 while (ctdb->vacuumers) {
1271 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1272 ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1273 (int)ctdb->vacuumers->child_pid));
1274 /* vacuum_child_destructor kills it, removes from list */
1275 talloc_free(ctdb->vacuumers);
1279 /* this function initializes the vacuuming context for a database
1280 * starts the vacuuming events
1282 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1284 if (ctdb_db->persistent != 0) {
1285 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1289 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1290 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1292 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1293 ctdb_db->vacuum_handle->fast_path_count = 0;
1295 event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1296 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1297 ctdb_vacuum_event, ctdb_db->vacuum_handle);
1303 * Insert a record into the ctdb_db context's delete queue,
1304 * handling hash collisions.
1306 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1307 const struct ctdb_ltdb_header *hdr,
1310 struct delete_record_data *kd;
1314 hash = (uint32_t)ctdb_hash(&key);
1316 DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1320 "migrated_with_data[%s]\n",
1321 ctdb_db->db_name, ctdb_db->db_id,
1323 ctdb_lmaster(ctdb_db->ctdb, &key),
1324 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1326 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1328 if ((kd->key.dsize != key.dsize) ||
1329 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1332 ("schedule for deletion: Hash collision (0x%08x)."
1333 " Skipping the record.\n", hash));
1337 ("schedule for deletion: Overwriting entry for "
1338 "key with hash 0x%08x.\n", hash));
1342 ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1343 ctdb_db->delete_queue,
1353 * Schedule a record for deletetion.
1354 * Called from the parent context.
1356 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1359 struct ctdb_control_schedule_for_deletion *dd;
1360 struct ctdb_db_context *ctdb_db;
1364 dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1366 ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1367 if (ctdb_db == NULL) {
1368 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1373 key.dsize = dd->keylen;
1376 ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1381 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1382 const struct ctdb_ltdb_header *hdr,
1386 struct ctdb_control_schedule_for_deletion *dd;
1390 if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1391 /* main daemon - directly queue */
1392 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1397 /* child process: send the main daemon a control */
1399 indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1400 indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1401 if (indata.dptr == NULL) {
1402 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1405 dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1406 dd->db_id = ctdb_db->db_id;
1408 dd->keylen = key.dsize;
1409 memcpy(dd->key, key.dptr, key.dsize);
1411 ret = ctdb_control(ctdb_db->ctdb,
1414 CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1415 CTDB_CTRL_FLAG_NOREPLY, /* flags */
1420 NULL, /* timeout : NULL == wait forever */
1421 NULL); /* error message */
1423 talloc_free(indata.dptr);
1425 if (ret != 0 || status != 0) {
1426 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1427 "SCHEDULE_FOR_DELETION "