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/events/events.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/events/events.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_handle *vacuum_handle;
42 enum vacuum_child_status status;
43 struct timeval start_time;
46 struct ctdb_vacuum_handle {
47 struct ctdb_db_context *ctdb_db;
48 struct ctdb_vacuum_child_context *child_ctx;
49 uint32_t fast_path_count;
53 /* a list of records to possibly delete */
55 uint32_t vacuum_limit;
56 uint32_t repack_limit;
57 struct ctdb_context *ctdb;
58 struct ctdb_db_context *ctdb_db;
59 struct tdb_context *dest_db;
60 trbt_tree_t *delete_tree;
61 uint32_t delete_count;
62 struct ctdb_marshall_buffer **list;
71 /* tuning information stored for every db */
72 struct vacuum_tuning_data {
73 uint32_t last_num_repack;
74 uint32_t last_num_empty;
75 uint32_t last_interval;
76 uint32_t new_interval;
77 struct timeval last_start;
81 /* this structure contains the information for one record to be deleted */
82 struct delete_record_data {
83 struct ctdb_context *ctdb;
84 struct ctdb_db_context *ctdb_db;
85 struct ctdb_ltdb_header hdr;
89 struct delete_records_list {
90 struct ctdb_marshall_buffer *records;
94 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
95 struct ctdb_ltdb_header *hdr)
97 struct ctdb_context *ctdb = vdata->ctdb;
98 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
100 struct delete_record_data *dd;
102 hash = ctdb_hash(&key);
104 if (trbt_lookup32(vdata->delete_tree, hash)) {
105 DEBUG(DEBUG_DEBUG, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
109 /* store key and header indexed by the key hash */
110 dd = talloc_zero(vdata->delete_tree, struct delete_record_data);
112 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
116 dd->ctdb_db = ctdb_db;
117 dd->key.dsize = key.dsize;
118 dd->key.dptr = talloc_memdup(dd, key.dptr, key.dsize);
119 if (dd->key.dptr == NULL) {
120 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
126 trbt_insert32(vdata->delete_tree, hash, dd);
128 vdata->delete_count++;
134 * Add a record to the list of records to be sent
135 * to their lmaster with VACUUM_FETCH.
137 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
140 struct ctdb_context *ctdb = vdata->ctdb;
141 struct ctdb_rec_data *rec;
145 lmaster = ctdb_lmaster(ctdb, &key);
147 rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
149 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
150 vdata->traverse_error = true;
154 old_size = talloc_get_size(vdata->list[lmaster]);
155 vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
156 old_size + rec->length);
157 if (vdata->list[lmaster] == NULL) {
158 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
159 vdata->traverse_error = true;
163 vdata->list[lmaster]->count++;
164 memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
173 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
174 struct timeval t, void *private_data);
178 * traverse function for gathering the records that can be deleted
180 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
182 struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
183 struct ctdb_context *ctdb = vdata->ctdb;
185 struct ctdb_ltdb_header *hdr;
188 lmaster = ctdb_lmaster(ctdb, &key);
189 if (lmaster >= ctdb->num_nodes) {
190 DEBUG(DEBUG_CRIT, (__location__
191 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
194 (unsigned)ctdb->num_nodes,
195 (unsigned)ctdb_hash(&key)));
199 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
200 /* its not a deleted record */
204 hdr = (struct ctdb_ltdb_header *)data.dptr;
206 if (hdr->dmaster != ctdb->pnn) {
210 if (lmaster == ctdb->pnn) {
212 * We are both lmaster and dmaster, and the record * is empty.
213 * So we should be able to delete it.
215 res = add_record_to_delete_tree(vdata, key, hdr);
218 * We are not lmaster.
219 * Add the record to the blob ready to send to the nodes.
221 res = add_record_to_vacuum_fetch_list(vdata, key);
228 * traverse the tree of records to delete and marshall them into
231 static void delete_traverse(void *param, void *data)
233 struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
234 struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
235 struct ctdb_rec_data *rec;
238 rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
240 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
244 old_size = talloc_get_size(recs->records);
245 recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
246 if (recs->records == NULL) {
247 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
250 recs->records->count++;
251 memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
255 * traverse function for the traversal of the delete_queue,
256 * the fast-path vacuuming list.
258 * - If the record has been migrated off the node
259 * or has been revived (filled with data) on the node,
260 * then skip the record.
262 * - If the current node is the record's lmaster and it is
263 * a record that has never been migrated with data, then
264 * delete the record from the local tdb.
266 * - If the current node is the record's lmaster and it has
267 * been migrated with data, then schedule it for the normal
268 * vacuuming procedure (i.e. add it to the delete_list).
270 * - If the current node is NOT the record's lmaster then
271 * add it to the list of records that are to be sent to
272 * the lmaster with the VACUUM_FETCH message.
274 static void delete_queue_traverse(void *param, void *data)
276 struct delete_record_data *dd =
277 talloc_get_type(data, struct delete_record_data);
278 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
279 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
280 struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
282 struct ctdb_ltdb_header *header;
286 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
288 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
292 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
293 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
294 /* Does not exist or not a ctdb record. Skip. */
298 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
299 /* The record has been recycled (filled with data). Skip. */
303 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
305 if (header->dmaster != ctdb->pnn) {
306 /* The record has been migrated off the node. Skip. */
311 if (header->rsn != dd->hdr.rsn) {
313 * The record has been migrated off the node and back again.
314 * But not requeued for deletion. Skip it.
320 * We are dmaster, and the record has no data, and it has
321 * not been migrated after it has been queued for deletion.
323 * At this stage, the record could still have been revived locally
324 * and last been written with empty data. This can only be
325 * fixed with the addition of an active or delete flag. (TODO)
328 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
330 if (lmaster != ctdb->pnn) {
331 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
335 (__location__ " Error adding record to list "
336 "of records to send to lmaster.\n"));
342 /* use header->flags or dd->hdr.flags ?? */
343 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
344 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
348 (__location__ " Error adding record to list "
349 "of records for deletion on lmaster.\n"));
352 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
356 (__location__ " Error deleting record from local "
362 if (tdb_data.dptr != NULL) {
365 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
371 * read-only traverse the database in order to find
372 * records that can be deleted and try to delete these
373 * records on the other nodes
374 * this executes in the child context
376 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata)
378 struct ctdb_context *ctdb = ctdb_db->ctdb;
379 const char *name = ctdb_db->db_name;
382 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
384 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
388 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
390 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
395 /* the list needs to be of length num_nodes */
396 vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
397 if (vdata->list == NULL) {
398 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
401 for (i = 0; i < ctdb->num_nodes; i++) {
402 vdata->list[i] = (struct ctdb_marshall_buffer *)
403 talloc_zero_size(vdata->list,
404 offsetof(struct ctdb_marshall_buffer, data));
405 if (vdata->list[i] == NULL) {
406 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
409 vdata->list[i]->db_id = ctdb_db->db_id;
413 * Traverse the delete_queue.
414 * This builds the same lists as the db traverse.
416 trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
418 /* read-only traverse, looking for records that might be able to be vacuumed */
419 if (tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata) == -1 ||
420 vdata->traverse_error) {
421 DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
426 * For records where we are not the lmaster,
427 * tell the lmaster to fetch the record.
429 for (i = 0; i < ctdb->num_nodes; i++) {
432 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
436 if (vdata->list[i]->count == 0) {
440 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
441 vdata->list[i]->count, ctdb->nodes[i]->pnn,
444 data.dsize = talloc_get_size(vdata->list[i]);
445 data.dptr = (void *)vdata->list[i];
446 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
447 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
448 "fetch message to %u\n",
449 ctdb->nodes[i]->pnn));
454 /* Process all records we can delete (if any) */
455 if (vdata->delete_count > 0) {
456 struct delete_records_list *recs;
457 TDB_DATA indata, outdata;
459 struct ctdb_node_map *nodemap;
460 uint32_t *active_nodes;
461 int num_active_nodes;
463 recs = talloc_zero(vdata, struct delete_records_list);
465 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
468 recs->records = (struct ctdb_marshall_buffer *)
469 talloc_zero_size(vdata,
470 offsetof(struct ctdb_marshall_buffer, data));
471 if (recs->records == NULL) {
472 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
475 recs->records->db_id = ctdb_db->db_id;
478 * traverse the tree of all records we want to delete and
479 * create a blob we can send to the other nodes.
481 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
483 indata.dsize = talloc_get_size(recs->records);
484 indata.dptr = (void *)recs->records;
487 * now tell all the active nodes to delete all these records
491 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
493 recs, /* talloc context */
496 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
500 active_nodes = list_of_active_nodes(ctdb, nodemap,
501 nodemap, /* talloc context */
502 false /* include self */);
504 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
506 for (i = 0; i < num_active_nodes; i++) {
507 struct ctdb_marshall_buffer *records;
508 struct ctdb_rec_data *rec;
510 ret = ctdb_control(ctdb, active_nodes[i], 0,
511 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
512 indata, recs, &outdata, &res,
514 if (ret != 0 || res != 0) {
515 DEBUG(DEBUG_ERR, ("Failed to delete records on "
516 "node %u: ret[%d] res[%d]\n",
517 active_nodes[i], ret, res));
522 * outdata countains the list of records coming back
523 * from the node which the node could not delete
525 records = (struct ctdb_marshall_buffer *)outdata.dptr;
526 rec = (struct ctdb_rec_data *)&records->data[0];
527 while (records->count-- > 1) {
528 TDB_DATA reckey, recdata;
529 struct ctdb_ltdb_header *rechdr;
531 reckey.dptr = &rec->data[0];
532 reckey.dsize = rec->keylen;
533 recdata.dptr = &rec->data[reckey.dsize];
534 recdata.dsize = rec->datalen;
536 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
537 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
540 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
541 recdata.dptr += sizeof(*rechdr);
542 recdata.dsize -= sizeof(*rechdr);
545 * that other node couldnt delete the record
546 * so we should delete it and thereby remove it from the tree
548 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
550 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
554 /* free nodemap and active_nodes */
555 talloc_free(nodemap);
558 * The only records remaining in the tree would be those
559 * records where all other nodes could successfully
560 * delete them, so we can safely delete them on the
561 * lmaster as well. Deletion implictely happens while
562 * we repack the database. The repack algorithm revisits
563 * the tree in order to find the records that don't need
564 * to be copied / repacked.
568 /* this ensures we run our event queue */
569 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
576 * traverse function for repacking
578 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
580 struct vacuum_data *vdata = (struct vacuum_data *)private;
583 uint32_t hash = ctdb_hash(&key);
584 struct delete_record_data *kd;
586 * check if we can ignore this record because it's in the delete_tree
588 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
590 * there might be hash collisions so we have to compare the keys here to be sure
592 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
593 struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
595 * we have to check if the record hasn't changed in the meantime in order to
596 * savely remove it from the database
598 if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
599 hdr->dmaster == kd->ctdb->pnn &&
600 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
601 kd->hdr.rsn == hdr->rsn) {
607 if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
608 vdata->traverse_error = true;
618 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
620 struct tdb_context *tmp_db;
622 if (tdb_transaction_start(tdb) != 0) {
623 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
627 tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
628 TDB_INTERNAL|TDB_DISALLOW_NESTING,
630 if (tmp_db == NULL) {
631 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
632 tdb_transaction_cancel(tdb);
636 vdata->traverse_error = false;
637 vdata->dest_db = tmp_db;
638 vdata->vacuum = true;
643 * repack and vacuum on-the-fly by not writing the records that are
646 if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
647 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
648 tdb_transaction_cancel(tdb);
653 DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
655 if (vdata->traverse_error) {
656 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
657 tdb_transaction_cancel(tdb);
662 if (tdb_wipe_all(tdb) != 0) {
663 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
664 tdb_transaction_cancel(tdb);
669 vdata->traverse_error = false;
670 vdata->dest_db = tdb;
671 vdata->vacuum = false;
674 if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
675 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
676 tdb_transaction_cancel(tdb);
681 if (vdata->traverse_error) {
682 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
683 tdb_transaction_cancel(tdb);
691 if (tdb_transaction_commit(tdb) != 0) {
692 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
695 DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
700 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
702 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
703 TDB_CONTEXT *tune_tdb;
705 struct vacuum_tuning_data tdata;
706 struct vacuum_tuning_data *tptr;
710 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
711 ctdb_db->ctdb->db_directory_state,
712 TUNINGDBNAME, ctdb_db->ctdb->pnn);
713 if (vac_dbname == NULL) {
714 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
715 talloc_free(tmp_ctx);
719 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
720 flags |= TDB_DISALLOW_NESTING;
721 tune_tdb = tdb_open(vac_dbname, 0,
723 O_RDWR|O_CREAT, 0600);
724 if (tune_tdb == NULL) {
725 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
726 talloc_free(tmp_ctx);
730 if (tdb_transaction_start(tune_tdb) != 0) {
731 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
735 key.dptr = discard_const(ctdb_db->db_name);
736 key.dsize = strlen(ctdb_db->db_name);
737 value = tdb_fetch(tune_tdb, key);
739 if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
740 tptr = (struct vacuum_tuning_data *)value.dptr;
744 * re-calc new vacuum interval:
745 * in case no limit was reached we continously increase the interval
746 * until vacuum_max_interval is reached
747 * in case a limit was reached we divide the current interval by 2
748 * unless vacuum_min_interval is reached
750 if (freelist < vdata->repack_limit &&
751 vdata->delete_count < vdata->vacuum_limit) {
752 if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
753 tdata.new_interval = tdata.last_interval * 110 / 100;
754 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
755 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
758 tdata.new_interval = tdata.last_interval / 2;
759 if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
760 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
761 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
763 DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
764 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
766 tdata.last_interval = tdata.new_interval;
768 DEBUG(DEBUG_ERR,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
769 tdata.last_num_repack = freelist;
770 tdata.last_num_empty = vdata->delete_count;
771 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
774 if (value.dptr != NULL) {
778 tdata.last_start = vdata->start;
779 tdata.last_duration = timeval_elapsed(&vdata->start);
781 value.dptr = (unsigned char *)&tdata;
782 value.dsize = sizeof(tdata);
784 if (tdb_store(tune_tdb, key, value, 0) != 0) {
785 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
786 tdb_transaction_cancel(tune_tdb);
788 talloc_free(tmp_ctx);
791 tdb_transaction_commit(tune_tdb);
793 talloc_free(tmp_ctx);
799 * repack and vaccum a db
800 * called from the child context
802 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
805 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
806 uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
807 const char *name = ctdb_db->db_name;
809 struct vacuum_data *vdata;
811 size = tdb_freelist_size(ctdb_db->ltdb->tdb);
813 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
817 vdata = talloc_zero(mem_ctx, struct vacuum_data);
819 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
823 vdata->ctdb = ctdb_db->ctdb;
824 vdata->vacuum_limit = vacuum_limit;
825 vdata->repack_limit = repack_limit;
826 vdata->delete_tree = trbt_create(vdata, 0);
827 if (vdata->delete_tree == NULL) {
828 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
833 vdata->start = timeval_current();
836 * gather all records that can be deleted in vdata
838 if (ctdb_vacuum_db(ctdb_db, vdata) != 0) {
839 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
843 * decide if a repack is necessary
845 if (size < repack_limit && vdata->delete_count < vacuum_limit) {
846 update_tuning_db(ctdb_db, vdata, size);
851 DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
852 name, size, vdata->delete_count));
855 * repack and implicitely get rid of the records we can delete
857 if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
858 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
859 update_tuning_db(ctdb_db, vdata, size);
863 update_tuning_db(ctdb_db, vdata, size);
869 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
871 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
875 uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
876 struct ctdb_context *ctdb = ctdb_db->ctdb;
879 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
880 if (vac_dbname == NULL) {
881 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
882 talloc_free(tmp_ctx);
886 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
887 flags |= TDB_DISALLOW_NESTING;
888 tdb = tdb_open(vac_dbname, 0,
890 O_RDWR|O_CREAT, 0600);
892 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval\n", vac_dbname));
893 talloc_free(tmp_ctx);
897 key.dptr = discard_const(ctdb_db->db_name);
898 key.dsize = strlen(ctdb_db->db_name);
900 value = tdb_fetch(tdb, key);
902 if (value.dptr != NULL) {
903 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
904 struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
906 interval = tptr->new_interval;
908 if (interval < ctdb->tunable.vacuum_min_interval) {
909 interval = ctdb->tunable.vacuum_min_interval;
911 if (interval > ctdb->tunable.vacuum_max_interval) {
912 interval = ctdb->tunable.vacuum_max_interval;
919 talloc_free(tmp_ctx);
924 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
926 double l = timeval_elapsed(&child_ctx->start_time);
927 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
928 struct ctdb_context *ctdb = ctdb_db->ctdb;
930 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
932 if (child_ctx->child_pid != -1) {
933 kill(child_ctx->child_pid, SIGKILL);
935 /* Bump the number of successful fast-path runs. */
936 child_ctx->vacuum_handle->fast_path_count++;
939 event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
940 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
941 ctdb_vacuum_event, child_ctx->vacuum_handle);
947 * this event is generated when a vacuum child process times out
949 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
950 struct timeval t, void *private_data)
952 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
954 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
956 child_ctx->status = VACUUM_TIMEOUT;
958 talloc_free(child_ctx);
963 * this event is generated when a vacuum child process has completed
965 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
966 uint16_t flags, void *private_data)
968 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
972 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
973 child_ctx->child_pid = -1;
975 ret = read(child_ctx->fd[0], &c, 1);
976 if (ret != 1 || c != 0) {
977 child_ctx->status = VACUUM_ERROR;
978 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));
980 child_ctx->status = VACUUM_OK;
983 talloc_free(child_ctx);
987 * this event is called every time we need to start a new vacuum process
990 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
991 struct timeval t, void *private_data)
993 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
994 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
995 struct ctdb_context *ctdb = ctdb_db->ctdb;
996 struct ctdb_vacuum_child_context *child_ctx;
999 /* we dont vacuum if we are in recovery mode */
1000 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE) {
1001 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1005 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1006 if (child_ctx == NULL) {
1007 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1008 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1012 ret = pipe(child_ctx->fd);
1014 talloc_free(child_ctx);
1015 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1016 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1020 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1021 vacuum_handle->fast_path_count = 0;
1024 child_ctx->child_pid = fork();
1025 if (child_ctx->child_pid == (pid_t)-1) {
1026 close(child_ctx->fd[0]);
1027 close(child_ctx->fd[1]);
1028 talloc_free(child_ctx);
1029 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1030 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1035 if (child_ctx->child_pid == 0) {
1037 close(child_ctx->fd[0]);
1039 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1041 if (switch_from_server_to_client(ctdb) != 0) {
1042 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1049 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx);
1051 write(child_ctx->fd[1], &cc, 1);
1055 set_close_on_exec(child_ctx->fd[0]);
1056 close(child_ctx->fd[1]);
1058 child_ctx->status = VACUUM_RUNNING;
1059 child_ctx->start_time = timeval_current();
1061 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1064 * Clear the fastpath vacuuming list in the parent.
1066 talloc_free(ctdb_db->delete_queue);
1067 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1068 if (ctdb_db->delete_queue == NULL) {
1069 /* fatal here? ... */
1070 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1071 "in parent context. Shutting down\n");
1074 event_add_timed(ctdb->ev, child_ctx,
1075 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1076 vacuum_child_timeout, child_ctx);
1078 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1080 event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1081 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1082 vacuum_child_handler,
1085 vacuum_handle->child_ctx = child_ctx;
1086 child_ctx->vacuum_handle = vacuum_handle;
1090 /* this function initializes the vacuuming context for a database
1091 * starts the vacuuming events
1093 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1095 if (ctdb_db->persistent != 0) {
1096 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1100 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1101 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1103 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1105 event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1106 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1107 ctdb_vacuum_event, ctdb_db->vacuum_handle);