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 /* its 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 void 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);
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 void delete_queue_traverse(void *param, void *data)
327 struct delete_record_data *dd =
328 talloc_get_type(data, struct delete_record_data);
329 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
330 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
331 struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
333 struct ctdb_ltdb_header *header;
339 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
341 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
346 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
347 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
348 /* Does not exist or not a ctdb record. Skip. */
352 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
353 /* The record has been recycled (filled with data). Skip. */
357 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
359 if (header->dmaster != ctdb->pnn) {
360 /* The record has been migrated off the node. Skip. */
365 if (header->rsn != dd->hdr.rsn) {
367 * The record has been migrated off the node and back again.
368 * But not requeued for deletion. Skip it.
374 * We are dmaster, and the record has no data, and it has
375 * not been migrated after it has been queued for deletion.
377 * At this stage, the record could still have been revived locally
378 * and last been written with empty data. This can only be
379 * fixed with the addition of an active or delete flag. (TODO)
382 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
384 if (lmaster != ctdb->pnn) {
385 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
389 (__location__ " Error adding record to list "
390 "of records to send to lmaster.\n"));
393 vdata->fast_added_to_vacuum_fetch_list++;
398 /* use header->flags or dd->hdr.flags ?? */
399 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
400 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
404 (__location__ " Error adding record to list "
405 "of records for deletion on lmaster.\n"));
408 vdata->fast_added_to_delete_tree++;
411 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
415 (__location__ " Error deleting record from local "
419 vdata->fast_deleted++;
426 vdata->fast_skipped++;
429 if (tdb_data.dptr != NULL) {
432 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
438 * read-only traverse the database in order to find
439 * records that can be deleted and try to delete these
440 * records on the other nodes
441 * this executes in the child context
443 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
444 struct vacuum_data *vdata,
445 bool full_vacuum_run)
447 struct ctdb_context *ctdb = ctdb_db->ctdb;
448 const char *name = ctdb_db->db_name;
451 DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
452 "%s db_id[0x%08x]\n",
453 full_vacuum_run ? "full" : "fast",
454 ctdb_db->db_name, ctdb_db->db_id));
456 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
458 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
462 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
464 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
470 vdata->fast_added_to_delete_tree = 0;
471 vdata->fast_added_to_vacuum_fetch_list = 0;
472 vdata->fast_deleted = 0;
473 vdata->fast_skipped = 0;
474 vdata->fast_error = 0;
475 vdata->fast_total = 0;
476 vdata->full_added_to_delete_tree = 0;
477 vdata->full_added_to_vacuum_fetch_list = 0;
478 vdata->full_skipped = 0;
479 vdata->full_error = 0;
480 vdata->full_total = 0;
482 /* the list needs to be of length num_nodes */
483 vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
484 if (vdata->list == NULL) {
485 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
488 for (i = 0; i < ctdb->num_nodes; i++) {
489 vdata->list[i] = (struct ctdb_marshall_buffer *)
490 talloc_zero_size(vdata->list,
491 offsetof(struct ctdb_marshall_buffer, data));
492 if (vdata->list[i] == NULL) {
493 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
496 vdata->list[i]->db_id = ctdb_db->db_id;
500 * Traverse the delete_queue.
501 * This builds the same lists as the db traverse.
503 trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
505 if (vdata->fast_total > 0) {
508 " fast vacuuming delete_queue traverse statistics: "
517 (unsigned)vdata->fast_total,
518 (unsigned)vdata->fast_deleted,
519 (unsigned)vdata->fast_skipped,
520 (unsigned)vdata->fast_error,
521 (unsigned)vdata->fast_added_to_delete_tree,
522 (unsigned)vdata->fast_added_to_vacuum_fetch_list));
526 * read-only traverse of the database, looking for records that
527 * might be able to be vacuumed.
529 * This is not done each time but only every tunable
530 * VacuumFastPathCount times.
532 if (full_vacuum_run) {
533 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
534 if (ret == -1 || vdata->traverse_error) {
535 DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
538 if (vdata->full_total > 0) {
541 " full vacuuming db traverse statistics: "
549 (unsigned)vdata->full_total,
550 (unsigned)vdata->full_skipped,
551 (unsigned)vdata->full_error,
552 (unsigned)vdata->full_added_to_delete_tree,
553 (unsigned)vdata->full_added_to_vacuum_fetch_list));
558 * For records where we are not the lmaster,
559 * tell the lmaster to fetch the record.
561 for (i = 0; i < ctdb->num_nodes; i++) {
564 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
568 if (vdata->list[i]->count == 0) {
572 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
573 vdata->list[i]->count, ctdb->nodes[i]->pnn,
576 data.dsize = talloc_get_size(vdata->list[i]);
577 data.dptr = (void *)vdata->list[i];
578 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
579 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
580 "fetch message to %u\n",
581 ctdb->nodes[i]->pnn));
586 /* Process all records we can delete (if any) */
587 if (vdata->delete_count > 0) {
588 struct delete_records_list *recs;
589 TDB_DATA indata, outdata;
591 struct ctdb_node_map *nodemap;
592 uint32_t *active_nodes;
593 int num_active_nodes;
595 recs = talloc_zero(vdata, struct delete_records_list);
597 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
600 recs->records = (struct ctdb_marshall_buffer *)
601 talloc_zero_size(vdata,
602 offsetof(struct ctdb_marshall_buffer, data));
603 if (recs->records == NULL) {
604 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
607 recs->records->db_id = ctdb_db->db_id;
610 * traverse the tree of all records we want to delete and
611 * create a blob we can send to the other nodes.
613 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
615 indata.dsize = talloc_get_size(recs->records);
616 indata.dptr = (void *)recs->records;
619 * now tell all the active nodes to delete all these records
623 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
625 recs, /* talloc context */
628 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
632 active_nodes = list_of_active_nodes(ctdb, nodemap,
633 nodemap, /* talloc context */
634 false /* include self */);
636 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
638 for (i = 0; i < num_active_nodes; i++) {
639 struct ctdb_marshall_buffer *records;
640 struct ctdb_rec_data *rec;
642 ret = ctdb_control(ctdb, active_nodes[i], 0,
643 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
644 indata, recs, &outdata, &res,
646 if (ret != 0 || res != 0) {
647 DEBUG(DEBUG_ERR, ("Failed to delete records on "
648 "node %u: ret[%d] res[%d]\n",
649 active_nodes[i], ret, res));
654 * outdata countains the list of records coming back
655 * from the node which the node could not delete
657 records = (struct ctdb_marshall_buffer *)outdata.dptr;
658 rec = (struct ctdb_rec_data *)&records->data[0];
659 while (records->count-- > 1) {
660 TDB_DATA reckey, recdata;
661 struct ctdb_ltdb_header *rechdr;
663 reckey.dptr = &rec->data[0];
664 reckey.dsize = rec->keylen;
665 recdata.dptr = &rec->data[reckey.dsize];
666 recdata.dsize = rec->datalen;
668 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
669 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
672 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
673 recdata.dptr += sizeof(*rechdr);
674 recdata.dsize -= sizeof(*rechdr);
677 * that other node couldnt delete the record
678 * so we should delete it and thereby remove it from the tree
680 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
682 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
686 /* free nodemap and active_nodes */
687 talloc_free(nodemap);
690 * The only records remaining in the tree would be those
691 * records where all other nodes could successfully
692 * delete them, so we can safely delete them on the
693 * lmaster as well. Deletion implictely happens while
694 * we repack the database. The repack algorithm revisits
695 * the tree in order to find the records that don't need
696 * to be copied / repacked.
700 /* this ensures we run our event queue */
701 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
708 * traverse function for repacking
710 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
712 struct vacuum_data *vdata = (struct vacuum_data *)private;
715 uint32_t hash = ctdb_hash(&key);
716 struct delete_record_data *kd;
718 * check if we can ignore this record because it's in the delete_tree
720 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
722 * there might be hash collisions so we have to compare the keys here to be sure
724 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
725 struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
727 * we have to check if the record hasn't changed in the meantime in order to
728 * savely remove it from the database
730 if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
731 hdr->dmaster == kd->ctdb->pnn &&
732 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
733 kd->hdr.rsn == hdr->rsn) {
739 if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
740 vdata->traverse_error = true;
750 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
752 struct tdb_context *tmp_db;
754 if (tdb_transaction_start(tdb) != 0) {
755 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
759 tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
760 TDB_INTERNAL|TDB_DISALLOW_NESTING,
762 if (tmp_db == NULL) {
763 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
764 tdb_transaction_cancel(tdb);
768 vdata->traverse_error = false;
769 vdata->dest_db = tmp_db;
770 vdata->vacuum = true;
775 * repack and vacuum on-the-fly by not writing the records that are
778 if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
779 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
780 tdb_transaction_cancel(tdb);
785 DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
787 if (vdata->traverse_error) {
788 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
789 tdb_transaction_cancel(tdb);
794 if (tdb_wipe_all(tdb) != 0) {
795 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
796 tdb_transaction_cancel(tdb);
801 vdata->traverse_error = false;
802 vdata->dest_db = tdb;
803 vdata->vacuum = false;
806 if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
807 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
808 tdb_transaction_cancel(tdb);
813 if (vdata->traverse_error) {
814 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
815 tdb_transaction_cancel(tdb);
823 if (tdb_transaction_commit(tdb) != 0) {
824 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
827 DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
832 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
834 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
835 TDB_CONTEXT *tune_tdb;
837 struct vacuum_tuning_data tdata;
838 struct vacuum_tuning_data *tptr;
842 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
843 ctdb_db->ctdb->db_directory_state,
844 TUNINGDBNAME, ctdb_db->ctdb->pnn);
845 if (vac_dbname == NULL) {
846 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
847 talloc_free(tmp_ctx);
851 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
852 flags |= TDB_DISALLOW_NESTING;
853 tune_tdb = tdb_open(vac_dbname, 0,
855 O_RDWR|O_CREAT, 0600);
856 if (tune_tdb == NULL) {
857 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
858 talloc_free(tmp_ctx);
862 if (tdb_transaction_start(tune_tdb) != 0) {
863 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
867 key.dptr = discard_const(ctdb_db->db_name);
868 key.dsize = strlen(ctdb_db->db_name);
869 value = tdb_fetch(tune_tdb, key);
871 if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
872 tptr = (struct vacuum_tuning_data *)value.dptr;
876 * re-calc new vacuum interval:
877 * in case no limit was reached we continuously increase the interval
878 * until vacuum_max_interval is reached
879 * in case a limit was reached we divide the current interval by 2
880 * unless vacuum_min_interval is reached
882 if (freelist < vdata->repack_limit &&
883 vdata->delete_count < vdata->vacuum_limit) {
884 if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
885 tdata.new_interval = tdata.last_interval * 110 / 100;
886 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
887 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
890 tdata.new_interval = tdata.last_interval / 2;
891 if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
892 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
893 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
895 DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
896 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
898 tdata.last_interval = tdata.new_interval;
900 DEBUG(DEBUG_DEBUG,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
901 tdata.last_num_repack = freelist;
902 tdata.last_num_empty = vdata->delete_count;
903 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
906 if (value.dptr != NULL) {
910 tdata.last_start = vdata->start;
911 tdata.last_duration = timeval_elapsed(&vdata->start);
913 value.dptr = (unsigned char *)&tdata;
914 value.dsize = sizeof(tdata);
916 if (tdb_store(tune_tdb, key, value, 0) != 0) {
917 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
918 tdb_transaction_cancel(tune_tdb);
920 talloc_free(tmp_ctx);
923 tdb_transaction_commit(tune_tdb);
925 talloc_free(tmp_ctx);
931 * repack and vaccum a db
932 * called from the child context
934 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
936 bool full_vacuum_run)
938 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
939 uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
940 const char *name = ctdb_db->db_name;
942 struct vacuum_data *vdata;
944 size = tdb_freelist_size(ctdb_db->ltdb->tdb);
946 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
950 vdata = talloc_zero(mem_ctx, struct vacuum_data);
952 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
956 vdata->ctdb = ctdb_db->ctdb;
957 vdata->vacuum_limit = vacuum_limit;
958 vdata->repack_limit = repack_limit;
959 vdata->delete_tree = trbt_create(vdata, 0);
960 vdata->ctdb_db = ctdb_db;
961 if (vdata->delete_tree == NULL) {
962 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
967 vdata->start = timeval_current();
970 * gather all records that can be deleted in vdata
972 if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
973 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
977 * decide if a repack is necessary
979 if (size < repack_limit && vdata->delete_count < vacuum_limit) {
980 update_tuning_db(ctdb_db, vdata, size);
985 DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
986 name, size, vdata->delete_count));
989 * repack and implicitely get rid of the records we can delete
991 if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
992 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
993 update_tuning_db(ctdb_db, vdata, size);
997 update_tuning_db(ctdb_db, vdata, size);
1003 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1005 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1007 TDB_DATA key, value;
1009 uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
1010 struct ctdb_context *ctdb = ctdb_db->ctdb;
1013 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
1014 if (vac_dbname == NULL) {
1015 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
1016 talloc_free(tmp_ctx);
1020 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
1021 flags |= TDB_DISALLOW_NESTING;
1022 tdb = tdb_open(vac_dbname, 0,
1024 O_RDWR|O_CREAT, 0600);
1026 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval. Errno : %s (%d)\n", vac_dbname, strerror(errno), errno));
1027 talloc_free(tmp_ctx);
1031 key.dptr = discard_const(ctdb_db->db_name);
1032 key.dsize = strlen(ctdb_db->db_name);
1034 value = tdb_fetch(tdb, key);
1036 if (value.dptr != NULL) {
1037 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
1038 struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
1040 interval = tptr->new_interval;
1042 if (interval < ctdb->tunable.vacuum_min_interval) {
1043 interval = ctdb->tunable.vacuum_min_interval;
1045 if (interval > ctdb->tunable.vacuum_max_interval) {
1046 interval = ctdb->tunable.vacuum_max_interval;
1053 talloc_free(tmp_ctx);
1058 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1060 double l = timeval_elapsed(&child_ctx->start_time);
1061 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1062 struct ctdb_context *ctdb = ctdb_db->ctdb;
1064 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1066 if (child_ctx->child_pid != -1) {
1067 kill(child_ctx->child_pid, SIGKILL);
1069 /* Bump the number of successful fast-path runs. */
1070 child_ctx->vacuum_handle->fast_path_count++;
1073 DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1075 event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1076 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1077 ctdb_vacuum_event, child_ctx->vacuum_handle);
1083 * this event is generated when a vacuum child process times out
1085 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1086 struct timeval t, void *private_data)
1088 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1090 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1092 child_ctx->status = VACUUM_TIMEOUT;
1094 talloc_free(child_ctx);
1099 * this event is generated when a vacuum child process has completed
1101 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1102 uint16_t flags, void *private_data)
1104 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1108 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1109 child_ctx->child_pid = -1;
1111 ret = read(child_ctx->fd[0], &c, 1);
1112 if (ret != 1 || c != 0) {
1113 child_ctx->status = VACUUM_ERROR;
1114 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));
1116 child_ctx->status = VACUUM_OK;
1119 talloc_free(child_ctx);
1123 * this event is called every time we need to start a new vacuum process
1126 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1127 struct timeval t, void *private_data)
1129 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1130 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1131 struct ctdb_context *ctdb = ctdb_db->ctdb;
1132 struct ctdb_vacuum_child_context *child_ctx;
1133 struct tevent_fd *fde;
1136 /* we dont vacuum if we are in recovery mode, or db frozen */
1137 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1138 ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1139 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1140 ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1141 : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1144 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1148 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1149 if (child_ctx == NULL) {
1150 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1151 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1155 ret = pipe(child_ctx->fd);
1157 talloc_free(child_ctx);
1158 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1159 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1163 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1164 vacuum_handle->fast_path_count = 0;
1167 child_ctx->child_pid = ctdb_fork(ctdb);
1168 if (child_ctx->child_pid == (pid_t)-1) {
1169 close(child_ctx->fd[0]);
1170 close(child_ctx->fd[1]);
1171 talloc_free(child_ctx);
1172 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1173 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1178 if (child_ctx->child_pid == 0) {
1180 bool full_vacuum_run = false;
1181 close(child_ctx->fd[0]);
1183 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1185 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1186 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1193 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1194 (vacuum_handle->fast_path_count == 0))
1196 full_vacuum_run = true;
1198 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1201 write(child_ctx->fd[1], &cc, 1);
1205 set_close_on_exec(child_ctx->fd[0]);
1206 close(child_ctx->fd[1]);
1208 child_ctx->status = VACUUM_RUNNING;
1209 child_ctx->start_time = timeval_current();
1211 DLIST_ADD(ctdb->vacuumers, child_ctx);
1212 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1215 * Clear the fastpath vacuuming list in the parent.
1217 talloc_free(ctdb_db->delete_queue);
1218 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1219 if (ctdb_db->delete_queue == NULL) {
1220 /* fatal here? ... */
1221 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1222 "in parent context. Shutting down\n");
1225 event_add_timed(ctdb->ev, child_ctx,
1226 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1227 vacuum_child_timeout, child_ctx);
1229 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1231 fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1232 EVENT_FD_READ, vacuum_child_handler, child_ctx);
1233 tevent_fd_set_auto_close(fde);
1235 vacuum_handle->child_ctx = child_ctx;
1236 child_ctx->vacuum_handle = vacuum_handle;
1239 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1241 /* Simply free them all. */
1242 while (ctdb->vacuumers) {
1243 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1244 ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1245 (int)ctdb->vacuumers->child_pid));
1246 /* vacuum_child_destructor kills it, removes from list */
1247 talloc_free(ctdb->vacuumers);
1251 /* this function initializes the vacuuming context for a database
1252 * starts the vacuuming events
1254 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1256 if (ctdb_db->persistent != 0) {
1257 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1261 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1262 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1264 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1265 ctdb_db->vacuum_handle->fast_path_count = 0;
1267 event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1268 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1269 ctdb_vacuum_event, ctdb_db->vacuum_handle);
1275 * Insert a record into the ctdb_db context's delete queue,
1276 * handling hash collisions.
1278 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1279 const struct ctdb_ltdb_header *hdr,
1282 struct delete_record_data *kd;
1286 hash = (uint32_t)ctdb_hash(&key);
1288 DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1292 "migrated_with_data[%s]\n",
1293 ctdb_db->db_name, ctdb_db->db_id,
1295 ctdb_lmaster(ctdb_db->ctdb, &key),
1296 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1298 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1300 if ((kd->key.dsize != key.dsize) ||
1301 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1304 ("schedule for deletion: Hash collision (0x%08x)."
1305 " Skipping the record.\n", hash));
1309 ("schedule for deletion: Overwriting entry for "
1310 "key with hash 0x%08x.\n", hash));
1314 ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1315 ctdb_db->delete_queue,
1325 * Schedule a record for deletetion.
1326 * Called from the parent context.
1328 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1331 struct ctdb_control_schedule_for_deletion *dd;
1332 struct ctdb_db_context *ctdb_db;
1336 dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1338 ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1339 if (ctdb_db == NULL) {
1340 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1345 key.dsize = dd->keylen;
1348 ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1353 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1354 const struct ctdb_ltdb_header *hdr,
1358 struct ctdb_control_schedule_for_deletion *dd;
1362 if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1363 /* main daemon - directly queue */
1364 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1369 /* child process: send the main daemon a control */
1371 indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1372 indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1373 if (indata.dptr == NULL) {
1374 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1377 dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1378 dd->db_id = ctdb_db->db_id;
1380 dd->keylen = key.dsize;
1381 memcpy(dd->key, key.dptr, key.dsize);
1383 ret = ctdb_control(ctdb_db->ctdb,
1386 CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1387 CTDB_CTRL_FLAG_NOREPLY, /* flags */
1392 NULL, /* timeout : NULL == wait forever */
1393 NULL); /* error message */
1395 talloc_free(indata.dptr);
1397 if (ret != 0 || status != 0) {
1398 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1399 "SCHEDULE_FOR_DELETION "