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 * Store key and header in a tree, indexed by the key hash.
96 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
97 struct ctdb_db_context *ctdb_db,
99 const struct ctdb_ltdb_header *hdr,
102 struct delete_record_data *dd;
105 dd = talloc_zero(tree, struct delete_record_data);
107 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
112 dd->ctdb_db = ctdb_db;
113 dd->key.dsize = key.dsize;
114 dd->key.dptr = talloc_memdup(dd, key.dptr, key.dsize);
115 if (dd->key.dptr == NULL) {
116 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
122 hash = ctdb_hash(&key);
124 trbt_insert32(tree, hash, dd);
129 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
130 struct ctdb_ltdb_header *hdr)
132 struct ctdb_context *ctdb = vdata->ctdb;
133 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
137 hash = ctdb_hash(&key);
139 if (trbt_lookup32(vdata->delete_tree, hash)) {
140 DEBUG(DEBUG_DEBUG, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
144 ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
151 vdata->delete_count++;
157 * Add a record to the list of records to be sent
158 * to their lmaster with VACUUM_FETCH.
160 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
163 struct ctdb_context *ctdb = vdata->ctdb;
164 struct ctdb_rec_data *rec;
168 lmaster = ctdb_lmaster(ctdb, &key);
170 rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
172 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
173 vdata->traverse_error = true;
177 old_size = talloc_get_size(vdata->list[lmaster]);
178 vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
179 old_size + rec->length);
180 if (vdata->list[lmaster] == NULL) {
181 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
182 vdata->traverse_error = true;
186 vdata->list[lmaster]->count++;
187 memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
196 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
197 struct timeval t, void *private_data);
201 * traverse function for gathering the records that can be deleted
203 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
205 struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
206 struct ctdb_context *ctdb = vdata->ctdb;
208 struct ctdb_ltdb_header *hdr;
211 lmaster = ctdb_lmaster(ctdb, &key);
212 if (lmaster >= ctdb->num_nodes) {
213 DEBUG(DEBUG_CRIT, (__location__
214 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
217 (unsigned)ctdb->num_nodes,
218 (unsigned)ctdb_hash(&key)));
222 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
223 /* its not a deleted record */
227 hdr = (struct ctdb_ltdb_header *)data.dptr;
229 if (hdr->dmaster != ctdb->pnn) {
233 if (lmaster == ctdb->pnn) {
235 * We are both lmaster and dmaster, and the record * is empty.
236 * So we should be able to delete it.
238 res = add_record_to_delete_tree(vdata, key, hdr);
241 * We are not lmaster.
242 * Add the record to the blob ready to send to the nodes.
244 res = add_record_to_vacuum_fetch_list(vdata, key);
251 * traverse the tree of records to delete and marshall them into
254 static void delete_traverse(void *param, void *data)
256 struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
257 struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
258 struct ctdb_rec_data *rec;
261 rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
263 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
267 old_size = talloc_get_size(recs->records);
268 recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
269 if (recs->records == NULL) {
270 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
273 recs->records->count++;
274 memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
278 * traverse function for the traversal of the delete_queue,
279 * the fast-path vacuuming list.
281 * - If the record has been migrated off the node
282 * or has been revived (filled with data) on the node,
283 * then skip the record.
285 * - If the current node is the record's lmaster and it is
286 * a record that has never been migrated with data, then
287 * delete the record from the local tdb.
289 * - If the current node is the record's lmaster and it has
290 * been migrated with data, then schedule it for the normal
291 * vacuuming procedure (i.e. add it to the delete_list).
293 * - If the current node is NOT the record's lmaster then
294 * add it to the list of records that are to be sent to
295 * the lmaster with the VACUUM_FETCH message.
297 static void delete_queue_traverse(void *param, void *data)
299 struct delete_record_data *dd =
300 talloc_get_type(data, struct delete_record_data);
301 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
302 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
303 struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
305 struct ctdb_ltdb_header *header;
309 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
311 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
315 tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
316 if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
317 /* Does not exist or not a ctdb record. Skip. */
321 if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
322 /* The record has been recycled (filled with data). Skip. */
326 header = (struct ctdb_ltdb_header *)tdb_data.dptr;
328 if (header->dmaster != ctdb->pnn) {
329 /* The record has been migrated off the node. Skip. */
334 if (header->rsn != dd->hdr.rsn) {
336 * The record has been migrated off the node and back again.
337 * But not requeued for deletion. Skip it.
343 * We are dmaster, and the record has no data, and it has
344 * not been migrated after it has been queued for deletion.
346 * At this stage, the record could still have been revived locally
347 * and last been written with empty data. This can only be
348 * fixed with the addition of an active or delete flag. (TODO)
351 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
353 if (lmaster != ctdb->pnn) {
354 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
358 (__location__ " Error adding record to list "
359 "of records to send to lmaster.\n"));
365 /* use header->flags or dd->hdr.flags ?? */
366 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
367 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
371 (__location__ " Error adding record to list "
372 "of records for deletion on lmaster.\n"));
375 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
379 (__location__ " Error deleting record from local "
385 if (tdb_data.dptr != NULL) {
388 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
394 * read-only traverse the database in order to find
395 * records that can be deleted and try to delete these
396 * records on the other nodes
397 * this executes in the child context
399 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
400 struct vacuum_data *vdata,
401 bool full_vacuum_run)
403 struct ctdb_context *ctdb = ctdb_db->ctdb;
404 const char *name = ctdb_db->db_name;
407 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
409 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
413 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
415 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
420 /* the list needs to be of length num_nodes */
421 vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
422 if (vdata->list == NULL) {
423 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
426 for (i = 0; i < ctdb->num_nodes; i++) {
427 vdata->list[i] = (struct ctdb_marshall_buffer *)
428 talloc_zero_size(vdata->list,
429 offsetof(struct ctdb_marshall_buffer, data));
430 if (vdata->list[i] == NULL) {
431 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
434 vdata->list[i]->db_id = ctdb_db->db_id;
438 * Traverse the delete_queue.
439 * This builds the same lists as the db traverse.
441 trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
444 * read-only traverse of the database, looking for records that
445 * might be able to be vacuumed.
447 * This is not done each time but only every tunable
448 * VacuumFastPathCount times.
450 if (full_vacuum_run) {
451 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
452 if (ret == -1 || vdata->traverse_error) {
453 DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
459 * For records where we are not the lmaster,
460 * tell the lmaster to fetch the record.
462 for (i = 0; i < ctdb->num_nodes; i++) {
465 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
469 if (vdata->list[i]->count == 0) {
473 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
474 vdata->list[i]->count, ctdb->nodes[i]->pnn,
477 data.dsize = talloc_get_size(vdata->list[i]);
478 data.dptr = (void *)vdata->list[i];
479 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
480 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
481 "fetch message to %u\n",
482 ctdb->nodes[i]->pnn));
487 /* Process all records we can delete (if any) */
488 if (vdata->delete_count > 0) {
489 struct delete_records_list *recs;
490 TDB_DATA indata, outdata;
492 struct ctdb_node_map *nodemap;
493 uint32_t *active_nodes;
494 int num_active_nodes;
496 recs = talloc_zero(vdata, struct delete_records_list);
498 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
501 recs->records = (struct ctdb_marshall_buffer *)
502 talloc_zero_size(vdata,
503 offsetof(struct ctdb_marshall_buffer, data));
504 if (recs->records == NULL) {
505 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
508 recs->records->db_id = ctdb_db->db_id;
511 * traverse the tree of all records we want to delete and
512 * create a blob we can send to the other nodes.
514 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
516 indata.dsize = talloc_get_size(recs->records);
517 indata.dptr = (void *)recs->records;
520 * now tell all the active nodes to delete all these records
524 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
526 recs, /* talloc context */
529 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
533 active_nodes = list_of_active_nodes(ctdb, nodemap,
534 nodemap, /* talloc context */
535 false /* include self */);
537 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
539 for (i = 0; i < num_active_nodes; i++) {
540 struct ctdb_marshall_buffer *records;
541 struct ctdb_rec_data *rec;
543 ret = ctdb_control(ctdb, active_nodes[i], 0,
544 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
545 indata, recs, &outdata, &res,
547 if (ret != 0 || res != 0) {
548 DEBUG(DEBUG_ERR, ("Failed to delete records on "
549 "node %u: ret[%d] res[%d]\n",
550 active_nodes[i], ret, res));
555 * outdata countains the list of records coming back
556 * from the node which the node could not delete
558 records = (struct ctdb_marshall_buffer *)outdata.dptr;
559 rec = (struct ctdb_rec_data *)&records->data[0];
560 while (records->count-- > 1) {
561 TDB_DATA reckey, recdata;
562 struct ctdb_ltdb_header *rechdr;
564 reckey.dptr = &rec->data[0];
565 reckey.dsize = rec->keylen;
566 recdata.dptr = &rec->data[reckey.dsize];
567 recdata.dsize = rec->datalen;
569 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
570 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
573 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
574 recdata.dptr += sizeof(*rechdr);
575 recdata.dsize -= sizeof(*rechdr);
578 * that other node couldnt delete the record
579 * so we should delete it and thereby remove it from the tree
581 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
583 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
587 /* free nodemap and active_nodes */
588 talloc_free(nodemap);
591 * The only records remaining in the tree would be those
592 * records where all other nodes could successfully
593 * delete them, so we can safely delete them on the
594 * lmaster as well. Deletion implictely happens while
595 * we repack the database. The repack algorithm revisits
596 * the tree in order to find the records that don't need
597 * to be copied / repacked.
601 /* this ensures we run our event queue */
602 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
609 * traverse function for repacking
611 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
613 struct vacuum_data *vdata = (struct vacuum_data *)private;
616 uint32_t hash = ctdb_hash(&key);
617 struct delete_record_data *kd;
619 * check if we can ignore this record because it's in the delete_tree
621 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
623 * there might be hash collisions so we have to compare the keys here to be sure
625 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
626 struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
628 * we have to check if the record hasn't changed in the meantime in order to
629 * savely remove it from the database
631 if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
632 hdr->dmaster == kd->ctdb->pnn &&
633 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
634 kd->hdr.rsn == hdr->rsn) {
640 if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
641 vdata->traverse_error = true;
651 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
653 struct tdb_context *tmp_db;
655 if (tdb_transaction_start(tdb) != 0) {
656 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
660 tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
661 TDB_INTERNAL|TDB_DISALLOW_NESTING,
663 if (tmp_db == NULL) {
664 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
665 tdb_transaction_cancel(tdb);
669 vdata->traverse_error = false;
670 vdata->dest_db = tmp_db;
671 vdata->vacuum = true;
676 * repack and vacuum on-the-fly by not writing the records that are
679 if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
680 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
681 tdb_transaction_cancel(tdb);
686 DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
688 if (vdata->traverse_error) {
689 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
690 tdb_transaction_cancel(tdb);
695 if (tdb_wipe_all(tdb) != 0) {
696 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
697 tdb_transaction_cancel(tdb);
702 vdata->traverse_error = false;
703 vdata->dest_db = tdb;
704 vdata->vacuum = false;
707 if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
708 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
709 tdb_transaction_cancel(tdb);
714 if (vdata->traverse_error) {
715 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
716 tdb_transaction_cancel(tdb);
724 if (tdb_transaction_commit(tdb) != 0) {
725 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
728 DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
733 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
735 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
736 TDB_CONTEXT *tune_tdb;
738 struct vacuum_tuning_data tdata;
739 struct vacuum_tuning_data *tptr;
743 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
744 ctdb_db->ctdb->db_directory_state,
745 TUNINGDBNAME, ctdb_db->ctdb->pnn);
746 if (vac_dbname == NULL) {
747 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
748 talloc_free(tmp_ctx);
752 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
753 flags |= TDB_DISALLOW_NESTING;
754 tune_tdb = tdb_open(vac_dbname, 0,
756 O_RDWR|O_CREAT, 0600);
757 if (tune_tdb == NULL) {
758 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
759 talloc_free(tmp_ctx);
763 if (tdb_transaction_start(tune_tdb) != 0) {
764 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
768 key.dptr = discard_const(ctdb_db->db_name);
769 key.dsize = strlen(ctdb_db->db_name);
770 value = tdb_fetch(tune_tdb, key);
772 if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
773 tptr = (struct vacuum_tuning_data *)value.dptr;
777 * re-calc new vacuum interval:
778 * in case no limit was reached we continously increase the interval
779 * until vacuum_max_interval is reached
780 * in case a limit was reached we divide the current interval by 2
781 * unless vacuum_min_interval is reached
783 if (freelist < vdata->repack_limit &&
784 vdata->delete_count < vdata->vacuum_limit) {
785 if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
786 tdata.new_interval = tdata.last_interval * 110 / 100;
787 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
788 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
791 tdata.new_interval = tdata.last_interval / 2;
792 if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
793 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
794 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
796 DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
797 tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
799 tdata.last_interval = tdata.new_interval;
801 DEBUG(DEBUG_ERR,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
802 tdata.last_num_repack = freelist;
803 tdata.last_num_empty = vdata->delete_count;
804 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
807 if (value.dptr != NULL) {
811 tdata.last_start = vdata->start;
812 tdata.last_duration = timeval_elapsed(&vdata->start);
814 value.dptr = (unsigned char *)&tdata;
815 value.dsize = sizeof(tdata);
817 if (tdb_store(tune_tdb, key, value, 0) != 0) {
818 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
819 tdb_transaction_cancel(tune_tdb);
821 talloc_free(tmp_ctx);
824 tdb_transaction_commit(tune_tdb);
826 talloc_free(tmp_ctx);
832 * repack and vaccum a db
833 * called from the child context
835 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
837 bool full_vacuum_run)
839 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
840 uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
841 const char *name = ctdb_db->db_name;
843 struct vacuum_data *vdata;
845 size = tdb_freelist_size(ctdb_db->ltdb->tdb);
847 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
851 vdata = talloc_zero(mem_ctx, struct vacuum_data);
853 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
857 vdata->ctdb = ctdb_db->ctdb;
858 vdata->vacuum_limit = vacuum_limit;
859 vdata->repack_limit = repack_limit;
860 vdata->delete_tree = trbt_create(vdata, 0);
861 if (vdata->delete_tree == NULL) {
862 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
867 vdata->start = timeval_current();
870 * gather all records that can be deleted in vdata
872 if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
873 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
877 * decide if a repack is necessary
879 if (size < repack_limit && vdata->delete_count < vacuum_limit) {
880 update_tuning_db(ctdb_db, vdata, size);
885 DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
886 name, size, vdata->delete_count));
889 * repack and implicitely get rid of the records we can delete
891 if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
892 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
893 update_tuning_db(ctdb_db, vdata, size);
897 update_tuning_db(ctdb_db, vdata, size);
903 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
905 TALLOC_CTX *tmp_ctx = talloc_new(NULL);
909 uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
910 struct ctdb_context *ctdb = ctdb_db->ctdb;
913 vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
914 if (vac_dbname == NULL) {
915 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
916 talloc_free(tmp_ctx);
920 flags = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
921 flags |= TDB_DISALLOW_NESTING;
922 tdb = tdb_open(vac_dbname, 0,
924 O_RDWR|O_CREAT, 0600);
926 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval\n", vac_dbname));
927 talloc_free(tmp_ctx);
931 key.dptr = discard_const(ctdb_db->db_name);
932 key.dsize = strlen(ctdb_db->db_name);
934 value = tdb_fetch(tdb, key);
936 if (value.dptr != NULL) {
937 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
938 struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
940 interval = tptr->new_interval;
942 if (interval < ctdb->tunable.vacuum_min_interval) {
943 interval = ctdb->tunable.vacuum_min_interval;
945 if (interval > ctdb->tunable.vacuum_max_interval) {
946 interval = ctdb->tunable.vacuum_max_interval;
953 talloc_free(tmp_ctx);
958 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
960 double l = timeval_elapsed(&child_ctx->start_time);
961 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
962 struct ctdb_context *ctdb = ctdb_db->ctdb;
964 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
966 if (child_ctx->child_pid != -1) {
967 kill(child_ctx->child_pid, SIGKILL);
969 /* Bump the number of successful fast-path runs. */
970 child_ctx->vacuum_handle->fast_path_count++;
973 event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
974 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
975 ctdb_vacuum_event, child_ctx->vacuum_handle);
981 * this event is generated when a vacuum child process times out
983 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
984 struct timeval t, void *private_data)
986 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
988 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
990 child_ctx->status = VACUUM_TIMEOUT;
992 talloc_free(child_ctx);
997 * this event is generated when a vacuum child process has completed
999 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1000 uint16_t flags, void *private_data)
1002 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1006 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1007 child_ctx->child_pid = -1;
1009 ret = read(child_ctx->fd[0], &c, 1);
1010 if (ret != 1 || c != 0) {
1011 child_ctx->status = VACUUM_ERROR;
1012 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));
1014 child_ctx->status = VACUUM_OK;
1017 talloc_free(child_ctx);
1021 * this event is called every time we need to start a new vacuum process
1024 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1025 struct timeval t, void *private_data)
1027 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1028 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1029 struct ctdb_context *ctdb = ctdb_db->ctdb;
1030 struct ctdb_vacuum_child_context *child_ctx;
1033 /* we dont vacuum if we are in recovery mode */
1034 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE) {
1035 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1039 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1040 if (child_ctx == NULL) {
1041 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1042 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1046 ret = pipe(child_ctx->fd);
1048 talloc_free(child_ctx);
1049 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1050 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1054 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1055 vacuum_handle->fast_path_count = 0;
1058 child_ctx->child_pid = fork();
1059 if (child_ctx->child_pid == (pid_t)-1) {
1060 close(child_ctx->fd[0]);
1061 close(child_ctx->fd[1]);
1062 talloc_free(child_ctx);
1063 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1064 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1069 if (child_ctx->child_pid == 0) {
1071 bool full_vacuum_run = false;
1072 close(child_ctx->fd[0]);
1074 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1076 if (switch_from_server_to_client(ctdb) != 0) {
1077 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1084 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1085 (vacuum_handle->fast_path_count == 0))
1087 full_vacuum_run = true;
1089 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1092 write(child_ctx->fd[1], &cc, 1);
1096 set_close_on_exec(child_ctx->fd[0]);
1097 close(child_ctx->fd[1]);
1099 child_ctx->status = VACUUM_RUNNING;
1100 child_ctx->start_time = timeval_current();
1102 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1105 * Clear the fastpath vacuuming list in the parent.
1107 talloc_free(ctdb_db->delete_queue);
1108 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1109 if (ctdb_db->delete_queue == NULL) {
1110 /* fatal here? ... */
1111 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1112 "in parent context. Shutting down\n");
1115 event_add_timed(ctdb->ev, child_ctx,
1116 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1117 vacuum_child_timeout, child_ctx);
1119 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1121 event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1122 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1123 vacuum_child_handler,
1126 vacuum_handle->child_ctx = child_ctx;
1127 child_ctx->vacuum_handle = vacuum_handle;
1131 /* this function initializes the vacuuming context for a database
1132 * starts the vacuuming events
1134 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1136 if (ctdb_db->persistent != 0) {
1137 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1141 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1142 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1144 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1146 event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1147 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1148 ctdb_vacuum_event, ctdb_db->vacuum_handle);