s4-winbind: Use winbindd in the AD DC for fl2003dc and plugin_s4_dc
[Samba/wip.git] / ctdb / server / ctdb_vacuum.c
blob574ad87b126b242013675c4cba991c15f66aaa86
1 /*
2 ctdb vacuuming events
4 Copyright (C) Ronnie Sahlberg 2009
5 Copyright (C) Michael Adam 2010-2013
6 Copyright (C) Stefan Metzmacher 2010-2011
8 This program is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3 of the License, or
11 (at your option) any later version.
13 This program is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with this program; if not, see <http://www.gnu.org/licenses/>.
22 #include "includes.h"
23 #include "tdb.h"
24 #include "system/network.h"
25 #include "system/filesys.h"
26 #include "system/dir.h"
27 #include "../include/ctdb_private.h"
28 #include "db_wrap.h"
29 #include "lib/util/dlinklist.h"
30 #include "../include/ctdb_private.h"
31 #include "../common/rb_tree.h"
33 #define TIMELIMIT() timeval_current_ofs(10, 0)
35 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
37 struct ctdb_vacuum_child_context {
38 struct ctdb_vacuum_child_context *next, *prev;
39 struct ctdb_vacuum_handle *vacuum_handle;
40 /* fd child writes status to */
41 int fd[2];
42 pid_t child_pid;
43 enum vacuum_child_status status;
44 struct timeval start_time;
47 struct ctdb_vacuum_handle {
48 struct ctdb_db_context *ctdb_db;
49 struct ctdb_vacuum_child_context *child_ctx;
50 uint32_t fast_path_count;
54 /* a list of records to possibly delete */
55 struct vacuum_data {
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_list;
61 struct ctdb_marshall_buffer **vacuum_fetch_list;
62 struct timeval start;
63 bool traverse_error;
64 bool vacuum;
65 struct {
66 struct {
67 uint32_t added_to_vacuum_fetch_list;
68 uint32_t added_to_delete_list;
69 uint32_t deleted;
70 uint32_t skipped;
71 uint32_t error;
72 uint32_t total;
73 } delete_queue;
74 struct {
75 uint32_t scheduled;
76 uint32_t skipped;
77 uint32_t error;
78 uint32_t total;
79 } db_traverse;
80 struct {
81 uint32_t total;
82 uint32_t remote_error;
83 uint32_t local_error;
84 uint32_t deleted;
85 uint32_t skipped;
86 uint32_t left;
87 } delete_list;
88 struct {
89 uint32_t vacuumed;
90 uint32_t copied;
91 } repack;
92 } count;
95 /* this structure contains the information for one record to be deleted */
96 struct delete_record_data {
97 struct ctdb_context *ctdb;
98 struct ctdb_db_context *ctdb_db;
99 struct ctdb_ltdb_header hdr;
100 TDB_DATA key;
101 uint8_t keydata[1];
104 struct delete_records_list {
105 struct ctdb_marshall_buffer *records;
106 struct vacuum_data *vdata;
109 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
110 const struct ctdb_ltdb_header *hdr,
111 TDB_DATA key);
114 * Store key and header in a tree, indexed by the key hash.
116 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
117 struct ctdb_db_context *ctdb_db,
118 trbt_tree_t *tree,
119 const struct ctdb_ltdb_header *hdr,
120 TDB_DATA key)
122 struct delete_record_data *dd;
123 uint32_t hash;
124 size_t len;
126 len = offsetof(struct delete_record_data, keydata) + key.dsize;
128 dd = (struct delete_record_data *)talloc_size(tree, len);
129 if (dd == NULL) {
130 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
131 return -1;
133 talloc_set_name_const(dd, "struct delete_record_data");
135 dd->ctdb = ctdb;
136 dd->ctdb_db = ctdb_db;
137 dd->key.dsize = key.dsize;
138 dd->key.dptr = dd->keydata;
139 memcpy(dd->keydata, key.dptr, key.dsize);
141 dd->hdr = *hdr;
143 hash = ctdb_hash(&key);
145 trbt_insert32(tree, hash, dd);
147 return 0;
150 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
151 struct ctdb_ltdb_header *hdr)
153 struct ctdb_context *ctdb = vdata->ctdb;
154 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
155 uint32_t hash;
156 int ret;
158 hash = ctdb_hash(&key);
160 if (trbt_lookup32(vdata->delete_list, hash)) {
161 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
162 return 0;
165 ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
166 vdata->delete_list,
167 hdr, key);
168 if (ret != 0) {
169 return -1;
172 vdata->count.delete_list.total++;
174 return 0;
178 * Add a record to the list of records to be sent
179 * to their lmaster with VACUUM_FETCH.
181 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
182 TDB_DATA key)
184 struct ctdb_context *ctdb = vdata->ctdb;
185 struct ctdb_rec_data *rec;
186 uint32_t lmaster;
187 size_t old_size;
188 struct ctdb_marshall_buffer *vfl;
190 lmaster = ctdb_lmaster(ctdb, &key);
192 vfl = vdata->vacuum_fetch_list[lmaster];
194 rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
195 if (rec == NULL) {
196 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
197 vdata->traverse_error = true;
198 return -1;
201 old_size = talloc_get_size(vfl);
202 vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
203 if (vfl == NULL) {
204 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
205 vdata->traverse_error = true;
206 return -1;
208 vdata->vacuum_fetch_list[lmaster] = vfl;
210 vfl->count++;
211 memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
212 talloc_free(rec);
214 return 0;
218 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
219 struct timeval t, void *private_data);
221 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
223 struct ctdb_ltdb_header *header =
224 (struct ctdb_ltdb_header *)private_data;
226 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
227 return -1;
230 *header = *(struct ctdb_ltdb_header *)data.dptr;
232 return 0;
236 * traverse function for gathering the records that can be deleted
238 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
239 void *private_data)
241 struct vacuum_data *vdata = talloc_get_type(private_data,
242 struct vacuum_data);
243 struct ctdb_context *ctdb = vdata->ctdb;
244 struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
245 uint32_t lmaster;
246 struct ctdb_ltdb_header *hdr;
247 int res = 0;
249 vdata->count.db_traverse.total++;
251 lmaster = ctdb_lmaster(ctdb, &key);
252 if (lmaster >= ctdb->num_nodes) {
253 vdata->count.db_traverse.error++;
254 DEBUG(DEBUG_CRIT, (__location__
255 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
256 " with hash[%u]!\n",
257 (unsigned)lmaster,
258 (unsigned)ctdb->num_nodes,
259 (unsigned)ctdb_hash(&key)));
260 return -1;
263 if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
264 /* it is not a deleted record */
265 vdata->count.db_traverse.skipped++;
266 return 0;
269 hdr = (struct ctdb_ltdb_header *)data.dptr;
271 if (hdr->dmaster != ctdb->pnn) {
272 vdata->count.db_traverse.skipped++;
273 return 0;
277 * Add the record to this process's delete_queue for processing
278 * in the subsequent traverse in the fast vacuum run.
280 res = insert_record_into_delete_queue(ctdb_db, hdr, key);
281 if (res != 0) {
282 vdata->count.db_traverse.error++;
283 } else {
284 vdata->count.db_traverse.scheduled++;
287 return 0;
291 * traverse the tree of records to delete and marshall them into
292 * a blob
294 static int delete_marshall_traverse(void *param, void *data)
296 struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
297 struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
298 struct ctdb_rec_data *rec;
299 size_t old_size;
301 rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
302 if (rec == NULL) {
303 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
304 return 0;
307 old_size = talloc_get_size(recs->records);
308 recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
309 if (recs->records == NULL) {
310 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
311 return 0;
313 recs->records->count++;
314 memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
315 return 0;
319 * Variant of delete_marshall_traverse() that bumps the
320 * RSN of each traversed record in the database.
322 * This is needed to ensure that when rolling out our
323 * empty record copy before remote deletion, we as the
324 * record's dmaster keep a higher RSN than the non-dmaster
325 * nodes. This is needed to prevent old copies from
326 * resurrection in recoveries.
328 static int delete_marshall_traverse_first(void *param, void *data)
330 struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
331 struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
332 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
333 struct ctdb_context *ctdb = ctdb_db->ctdb;
334 struct ctdb_ltdb_header header;
335 uint32_t lmaster;
336 uint32_t hash = ctdb_hash(&(dd->key));
337 int res;
339 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
340 if (res != 0) {
341 DEBUG(DEBUG_ERR,
342 (__location__ " Error getting chainlock on record with "
343 "key hash [0x%08x] on database db[%s].\n",
344 hash, ctdb_db->db_name));
345 recs->vdata->count.delete_list.skipped++;
346 recs->vdata->count.delete_list.left--;
347 talloc_free(dd);
348 return 0;
352 * Verify that the record is still empty, its RSN has not
353 * changed and that we are still its lmaster and dmaster.
356 res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
357 vacuum_record_parser, &header);
358 if (res != 0) {
359 goto skip;
362 if (header.flags & CTDB_REC_RO_FLAGS) {
363 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
364 "on database db[%s] has read-only flags. "
365 "skipping.\n",
366 hash, ctdb_db->db_name));
367 goto skip;
370 if (header.dmaster != ctdb->pnn) {
371 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
372 "on database db[%s] has been migrated away. "
373 "skipping.\n",
374 hash, ctdb_db->db_name));
375 goto skip;
378 if (header.rsn != dd->hdr.rsn) {
379 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
380 "on database db[%s] seems to have been "
381 "migrated away and back again (with empty "
382 "data). skipping.\n",
383 hash, ctdb_db->db_name));
384 goto skip;
387 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
389 if (lmaster != ctdb->pnn) {
390 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
391 "delete list (key hash [0x%08x], db[%s]). "
392 "Strange! skipping.\n",
393 hash, ctdb_db->db_name));
394 goto skip;
398 * Increment the record's RSN to ensure the dmaster (i.e. the current
399 * node) has the highest RSN of the record in the cluster.
400 * This is to prevent old record copies from resurrecting in recoveries
401 * if something should fail during the deletion process.
402 * Note that ctdb_ltdb_store_server() increments the RSN if called
403 * on the record's dmaster.
406 res = ctdb_ltdb_store(ctdb_db, dd->key, &header, tdb_null);
407 if (res != 0) {
408 DEBUG(DEBUG_ERR, (__location__ ": Failed to store record with "
409 "key hash [0x%08x] on database db[%s].\n",
410 hash, ctdb_db->db_name));
411 goto skip;
414 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
416 goto done;
418 skip:
419 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
421 recs->vdata->count.delete_list.skipped++;
422 recs->vdata->count.delete_list.left--;
423 talloc_free(dd);
424 dd = NULL;
426 done:
427 if (dd == NULL) {
428 return 0;
431 return delete_marshall_traverse(param, data);
435 * traverse function for the traversal of the delete_queue,
436 * the fast-path vacuuming list.
438 * - If the record has been migrated off the node
439 * or has been revived (filled with data) on the node,
440 * then skip the record.
442 * - If the current node is the record's lmaster and it is
443 * a record that has never been migrated with data, then
444 * delete the record from the local tdb.
446 * - If the current node is the record's lmaster and it has
447 * been migrated with data, then schedule it for the normal
448 * vacuuming procedure (i.e. add it to the delete_list).
450 * - If the current node is NOT the record's lmaster then
451 * add it to the list of records that are to be sent to
452 * the lmaster with the VACUUM_FETCH message.
454 static int delete_queue_traverse(void *param, void *data)
456 struct delete_record_data *dd =
457 talloc_get_type(data, struct delete_record_data);
458 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
459 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
460 struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
461 int res;
462 struct ctdb_ltdb_header header;
463 uint32_t lmaster;
464 uint32_t hash = ctdb_hash(&(dd->key));
466 vdata->count.delete_queue.total++;
468 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
469 if (res != 0) {
470 DEBUG(DEBUG_ERR,
471 (__location__ " Error getting chainlock on record with "
472 "key hash [0x%08x] on database db[%s].\n",
473 hash, ctdb_db->db_name));
474 vdata->count.delete_queue.error++;
475 return 0;
478 res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
479 vacuum_record_parser, &header);
480 if (res != 0) {
481 goto skipped;
484 if (header.dmaster != ctdb->pnn) {
485 /* The record has been migrated off the node. Skip. */
486 goto skipped;
489 if (header.rsn != dd->hdr.rsn) {
491 * The record has been migrated off the node and back again.
492 * But not requeued for deletion. Skip it.
494 goto skipped;
498 * We are dmaster, and the record has no data, and it has
499 * not been migrated after it has been queued for deletion.
501 * At this stage, the record could still have been revived locally
502 * and last been written with empty data. This can only be
503 * fixed with the addition of an active or delete flag. (TODO)
506 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
508 if (lmaster != ctdb->pnn) {
509 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
511 if (res != 0) {
512 DEBUG(DEBUG_ERR,
513 (__location__ " Error adding record to list "
514 "of records to send to lmaster.\n"));
515 vdata->count.delete_queue.error++;
516 } else {
517 vdata->count.delete_queue.added_to_vacuum_fetch_list++;
519 goto done;
522 /* use header->flags or dd->hdr.flags ?? */
523 if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
524 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
526 if (res != 0) {
527 DEBUG(DEBUG_ERR,
528 (__location__ " Error adding record to list "
529 "of records for deletion on lmaster.\n"));
530 vdata->count.delete_queue.error++;
531 } else {
532 vdata->count.delete_queue.added_to_delete_list++;
534 } else {
535 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
537 if (res != 0) {
538 DEBUG(DEBUG_ERR,
539 (__location__ " Error deleting record with key "
540 "hash [0x%08x] from local data base db[%s].\n",
541 hash, ctdb_db->db_name));
542 vdata->count.delete_queue.error++;
543 goto done;
546 DEBUG(DEBUG_DEBUG,
547 (__location__ " Deleted record with key hash "
548 "[0x%08x] from local data base db[%s].\n",
549 hash, ctdb_db->db_name));
550 vdata->count.delete_queue.deleted++;
553 goto done;
555 skipped:
556 vdata->count.delete_queue.skipped++;
558 done:
559 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
561 return 0;
565 * Delete the records that we are lmaster and dmaster for and
566 * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
567 * control.
569 static int delete_record_traverse(void *param, void *data)
571 struct delete_record_data *dd =
572 talloc_get_type(data, struct delete_record_data);
573 struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
574 struct ctdb_db_context *ctdb_db = dd->ctdb_db;
575 struct ctdb_context *ctdb = ctdb_db->ctdb;
576 int res;
577 struct ctdb_ltdb_header header;
578 uint32_t lmaster;
579 uint32_t hash = ctdb_hash(&(dd->key));
581 res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
582 if (res != 0) {
583 DEBUG(DEBUG_ERR,
584 (__location__ " Error getting chainlock on record with "
585 "key hash [0x%08x] on database db[%s].\n",
586 hash, ctdb_db->db_name));
587 vdata->count.delete_list.local_error++;
588 vdata->count.delete_list.left--;
589 talloc_free(dd);
590 return 0;
594 * Verify that the record is still empty, its RSN has not
595 * changed and that we are still its lmaster and dmaster.
598 res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
599 vacuum_record_parser, &header);
600 if (res != 0) {
601 goto skip;
604 if (header.flags & CTDB_REC_RO_FLAGS) {
605 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
606 "on database db[%s] has read-only flags. "
607 "skipping.\n",
608 hash, ctdb_db->db_name));
609 goto skip;
612 if (header.dmaster != ctdb->pnn) {
613 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
614 "on database db[%s] has been migrated away. "
615 "skipping.\n",
616 hash, ctdb_db->db_name));
617 goto skip;
620 if (header.rsn != dd->hdr.rsn + 1) {
622 * The record has been migrated off the node and back again.
623 * But not requeued for deletion. Skip it.
624 * (Note that the first marshall traverse has bumped the RSN
625 * on disk.)
627 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
628 "on database db[%s] seems to have been "
629 "migrated away and back again (with empty "
630 "data). skipping.\n",
631 hash, ctdb_db->db_name));
632 goto skip;
635 lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
637 if (lmaster != ctdb->pnn) {
638 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
639 "delete list (key hash [0x%08x], db[%s]). "
640 "Strange! skipping.\n",
641 hash, ctdb_db->db_name));
642 goto skip;
645 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
647 if (res != 0) {
648 DEBUG(DEBUG_ERR,
649 (__location__ " Error deleting record with key hash "
650 "[0x%08x] from local data base db[%s].\n",
651 hash, ctdb_db->db_name));
652 vdata->count.delete_list.local_error++;
653 goto done;
656 DEBUG(DEBUG_DEBUG,
657 (__location__ " Deleted record with key hash [0x%08x] from "
658 "local data base db[%s].\n", hash, ctdb_db->db_name));
660 vdata->count.delete_list.deleted++;
661 goto done;
663 skip:
664 vdata->count.delete_list.skipped++;
666 done:
667 tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
669 talloc_free(dd);
670 vdata->count.delete_list.left--;
672 return 0;
676 * Traverse the delete_queue.
677 * Records are either deleted directly or filled
678 * into the delete list or the vacuum fetch lists
679 * for further processing.
681 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
682 struct vacuum_data *vdata)
684 uint32_t sum;
685 int ret;
687 ret = trbt_traversearray32(ctdb_db->delete_queue, 1,
688 delete_queue_traverse, vdata);
690 if (ret != 0) {
691 DEBUG(DEBUG_ERR, (__location__ " Error traversing "
692 "the delete queue.\n"));
695 sum = vdata->count.delete_queue.deleted
696 + vdata->count.delete_queue.skipped
697 + vdata->count.delete_queue.error
698 + vdata->count.delete_queue.added_to_delete_list
699 + vdata->count.delete_queue.added_to_vacuum_fetch_list;
701 if (vdata->count.delete_queue.total != sum) {
702 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
703 "counts for db[%s]: total[%u] != sum[%u]\n",
704 ctdb_db->db_name,
705 (unsigned)vdata->count.delete_queue.total,
706 (unsigned)sum));
709 if (vdata->count.delete_queue.total > 0) {
710 DEBUG(DEBUG_INFO,
711 (__location__
712 " fast vacuuming delete_queue traverse statistics: "
713 "db[%s] "
714 "total[%u] "
715 "del[%u] "
716 "skp[%u] "
717 "err[%u] "
718 "adl[%u] "
719 "avf[%u]\n",
720 ctdb_db->db_name,
721 (unsigned)vdata->count.delete_queue.total,
722 (unsigned)vdata->count.delete_queue.deleted,
723 (unsigned)vdata->count.delete_queue.skipped,
724 (unsigned)vdata->count.delete_queue.error,
725 (unsigned)vdata->count.delete_queue.added_to_delete_list,
726 (unsigned)vdata->count.delete_queue.added_to_vacuum_fetch_list));
729 return;
733 * read-only traverse of the database, looking for records that
734 * might be able to be vacuumed.
736 * This is not done each time but only every tunable
737 * VacuumFastPathCount times.
739 static void ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
740 struct vacuum_data *vdata)
742 int ret;
744 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
745 if (ret == -1 || vdata->traverse_error) {
746 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
747 "'%s'\n", ctdb_db->db_name));
750 if (vdata->count.db_traverse.total > 0) {
751 DEBUG(DEBUG_INFO,
752 (__location__
753 " full vacuuming db traverse statistics: "
754 "db[%s] "
755 "total[%u] "
756 "skp[%u] "
757 "err[%u] "
758 "sched[%u]\n",
759 ctdb_db->db_name,
760 (unsigned)vdata->count.db_traverse.total,
761 (unsigned)vdata->count.db_traverse.skipped,
762 (unsigned)vdata->count.db_traverse.error,
763 (unsigned)vdata->count.db_traverse.scheduled));
766 return;
770 * Process the vacuum fetch lists:
771 * For records for which we are not the lmaster, tell the lmaster to
772 * fetch the record.
774 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
775 struct vacuum_data *vdata)
777 int i;
778 struct ctdb_context *ctdb = ctdb_db->ctdb;
780 for (i = 0; i < ctdb->num_nodes; i++) {
781 TDB_DATA data;
782 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
784 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
785 continue;
788 if (vfl->count == 0) {
789 continue;
792 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
793 vfl->count, ctdb->nodes[i]->pnn,
794 ctdb_db->db_name));
796 data.dsize = talloc_get_size(vfl);
797 data.dptr = (void *)vfl;
798 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
799 CTDB_SRVID_VACUUM_FETCH,
800 data) != 0)
802 DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
803 "fetch message to %u\n",
804 ctdb->nodes[i]->pnn));
808 return;
812 * Process the delete list:
814 * This is the last step of vacuuming that consistently deletes
815 * those records that have been migrated with data and can hence
816 * not be deleted when leaving a node.
818 * In this step, the lmaster does the final deletion of those empty
819 * records that it is also dmaster for. It has ususally received
820 * at least some of these records previously from the former dmasters
821 * with the vacuum fetch message.
823 * This last step is implemented as a 3-phase process to protect from
824 * races leading to data corruption:
826 * 1) Send the lmaster's copy to all other active nodes with the
827 * RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
828 * 2) Send the records that could successfully be stored remotely
829 * in step #1 to all active nodes with the TRY_DELETE_RECORDS
830 * control. The remote notes delete their local copy.
831 * 3) The lmaster locally deletes its copies of all records that
832 * could successfully be deleted remotely in step #2.
834 static void ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
835 struct vacuum_data *vdata)
837 int ret, i;
838 struct ctdb_context *ctdb = ctdb_db->ctdb;
839 struct delete_records_list *recs;
840 TDB_DATA indata;
841 struct ctdb_node_map *nodemap;
842 uint32_t *active_nodes;
843 int num_active_nodes;
844 TALLOC_CTX *tmp_ctx;
845 uint32_t sum;
847 if (vdata->count.delete_list.total == 0) {
848 return;
851 tmp_ctx = talloc_new(vdata);
852 if (tmp_ctx == NULL) {
853 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
854 return;
857 vdata->count.delete_list.left = vdata->count.delete_list.total;
860 * get the list of currently active nodes
863 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
864 CTDB_CURRENT_NODE,
865 tmp_ctx,
866 &nodemap);
867 if (ret != 0) {
868 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
869 goto done;
872 active_nodes = list_of_active_nodes(ctdb, nodemap,
873 nodemap, /* talloc context */
874 false /* include self */);
875 /* yuck! ;-) */
876 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
879 * Now delete the records all active nodes in a three-phase process:
880 * 1) send all active remote nodes the current empty copy with this
881 * node as DMASTER
882 * 2) if all nodes could store the new copy,
883 * tell all the active remote nodes to delete all their copy
884 * 3) if all remote nodes deleted their record copy, delete it locally
888 * Step 1:
889 * Send currently empty record copy to all active nodes for storing.
892 recs = talloc_zero(tmp_ctx, struct delete_records_list);
893 if (recs == NULL) {
894 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
895 goto done;
897 recs->records = (struct ctdb_marshall_buffer *)
898 talloc_zero_size(recs,
899 offsetof(struct ctdb_marshall_buffer, data));
900 if (recs->records == NULL) {
901 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
902 goto done;
904 recs->records->db_id = ctdb_db->db_id;
905 recs->vdata = vdata;
908 * traverse the tree of all records we want to delete and
909 * create a blob we can send to the other nodes.
911 * We call delete_marshall_traverse_first() to bump the
912 * records' RSNs in the database, to ensure we (as dmaster)
913 * keep the highest RSN of the records in the cluster.
915 ret = trbt_traversearray32(vdata->delete_list, 1,
916 delete_marshall_traverse_first, recs);
917 if (ret != 0) {
918 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
919 "delete list for first marshalling.\n"));
922 indata.dsize = talloc_get_size(recs->records);
923 indata.dptr = (void *)recs->records;
925 for (i = 0; i < num_active_nodes; i++) {
926 struct ctdb_marshall_buffer *records;
927 struct ctdb_rec_data *rec;
928 int32_t res;
929 TDB_DATA outdata;
931 ret = ctdb_control(ctdb, active_nodes[i], 0,
932 CTDB_CONTROL_RECEIVE_RECORDS, 0,
933 indata, recs, &outdata, &res,
934 NULL, NULL);
935 if (ret != 0 || res != 0) {
936 DEBUG(DEBUG_ERR, ("Error storing record copies on "
937 "node %u: ret[%d] res[%d]\n",
938 active_nodes[i], ret, res));
939 goto done;
943 * outdata contains the list of records coming back
944 * from the node: These are the records that the
945 * remote node could not store. We remove these from
946 * the list to process further.
948 records = (struct ctdb_marshall_buffer *)outdata.dptr;
949 rec = (struct ctdb_rec_data *)&records->data[0];
950 while (records->count-- > 1) {
951 TDB_DATA reckey, recdata;
952 struct ctdb_ltdb_header *rechdr;
953 struct delete_record_data *dd;
955 reckey.dptr = &rec->data[0];
956 reckey.dsize = rec->keylen;
957 recdata.dptr = &rec->data[reckey.dsize];
958 recdata.dsize = rec->datalen;
960 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
961 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
962 goto done;
964 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
965 recdata.dptr += sizeof(*rechdr);
966 recdata.dsize -= sizeof(*rechdr);
968 dd = (struct delete_record_data *)trbt_lookup32(
969 vdata->delete_list,
970 ctdb_hash(&reckey));
971 if (dd != NULL) {
973 * The other node could not store the record
974 * copy and it is the first node that failed.
975 * So we should remove it from the tree and
976 * update statistics.
978 talloc_free(dd);
979 vdata->count.delete_list.remote_error++;
980 vdata->count.delete_list.left--;
981 } else {
982 DEBUG(DEBUG_ERR, (__location__ " Failed to "
983 "find record with hash 0x%08x coming "
984 "back from RECEIVE_RECORDS "
985 "control in delete list.\n",
986 ctdb_hash(&reckey)));
987 vdata->count.delete_list.local_error++;
988 vdata->count.delete_list.left--;
991 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
995 if (vdata->count.delete_list.left == 0) {
996 goto success;
1000 * Step 2:
1001 * Send the remaining records to all active nodes for deletion.
1003 * The lmaster's (i.e. our) copies of these records have been stored
1004 * successfully on the other nodes.
1008 * Create a marshall blob from the remaining list of records to delete.
1011 talloc_free(recs->records);
1013 recs->records = (struct ctdb_marshall_buffer *)
1014 talloc_zero_size(recs,
1015 offsetof(struct ctdb_marshall_buffer, data));
1016 if (recs->records == NULL) {
1017 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1018 goto done;
1020 recs->records->db_id = ctdb_db->db_id;
1022 ret = trbt_traversearray32(vdata->delete_list, 1,
1023 delete_marshall_traverse, recs);
1024 if (ret != 0) {
1025 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1026 "delete list for second marshalling.\n"));
1029 indata.dsize = talloc_get_size(recs->records);
1030 indata.dptr = (void *)recs->records;
1032 for (i = 0; i < num_active_nodes; i++) {
1033 struct ctdb_marshall_buffer *records;
1034 struct ctdb_rec_data *rec;
1035 int32_t res;
1036 TDB_DATA outdata;
1038 ret = ctdb_control(ctdb, active_nodes[i], 0,
1039 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
1040 indata, recs, &outdata, &res,
1041 NULL, NULL);
1042 if (ret != 0 || res != 0) {
1043 DEBUG(DEBUG_ERR, ("Failed to delete records on "
1044 "node %u: ret[%d] res[%d]\n",
1045 active_nodes[i], ret, res));
1046 goto done;
1050 * outdata contains the list of records coming back
1051 * from the node: These are the records that the
1052 * remote node could not delete. We remove these from
1053 * the list to delete locally.
1055 records = (struct ctdb_marshall_buffer *)outdata.dptr;
1056 rec = (struct ctdb_rec_data *)&records->data[0];
1057 while (records->count-- > 1) {
1058 TDB_DATA reckey, recdata;
1059 struct ctdb_ltdb_header *rechdr;
1060 struct delete_record_data *dd;
1062 reckey.dptr = &rec->data[0];
1063 reckey.dsize = rec->keylen;
1064 recdata.dptr = &rec->data[reckey.dsize];
1065 recdata.dsize = rec->datalen;
1067 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
1068 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
1069 goto done;
1071 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
1072 recdata.dptr += sizeof(*rechdr);
1073 recdata.dsize -= sizeof(*rechdr);
1075 dd = (struct delete_record_data *)trbt_lookup32(
1076 vdata->delete_list,
1077 ctdb_hash(&reckey));
1078 if (dd != NULL) {
1080 * The other node could not delete the
1081 * record and it is the first node that
1082 * failed. So we should remove it from
1083 * the tree and update statistics.
1085 talloc_free(dd);
1086 vdata->count.delete_list.remote_error++;
1087 vdata->count.delete_list.left--;
1088 } else {
1089 DEBUG(DEBUG_ERR, (__location__ " Failed to "
1090 "find record with hash 0x%08x coming "
1091 "back from TRY_DELETE_RECORDS "
1092 "control in delete list.\n",
1093 ctdb_hash(&reckey)));
1094 vdata->count.delete_list.local_error++;
1095 vdata->count.delete_list.left--;
1098 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
1102 if (vdata->count.delete_list.left == 0) {
1103 goto success;
1107 * Step 3:
1108 * Delete the remaining records locally.
1110 * These records have successfully been deleted on all
1111 * active remote nodes.
1114 ret = trbt_traversearray32(vdata->delete_list, 1,
1115 delete_record_traverse, vdata);
1116 if (ret != 0) {
1117 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1118 "delete list for deletion.\n"));
1121 success:
1123 if (vdata->count.delete_list.left != 0) {
1124 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
1125 "there are %u records left for deletion after "
1126 "processing delete list\n",
1127 ctdb_db->db_name,
1128 (unsigned)vdata->count.delete_list.left));
1131 sum = vdata->count.delete_list.deleted
1132 + vdata->count.delete_list.skipped
1133 + vdata->count.delete_list.remote_error
1134 + vdata->count.delete_list.local_error
1135 + vdata->count.delete_list.left;
1137 if (vdata->count.delete_list.total != sum) {
1138 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
1139 "delete list counts for db[%s]: total[%u] != sum[%u]\n",
1140 ctdb_db->db_name,
1141 (unsigned)vdata->count.delete_list.total,
1142 (unsigned)sum));
1145 if (vdata->count.delete_list.total > 0) {
1146 DEBUG(DEBUG_INFO,
1147 (__location__
1148 " vacuum delete list statistics: "
1149 "db[%s] "
1150 "total[%u] "
1151 "del[%u] "
1152 "skip[%u] "
1153 "rem.err[%u] "
1154 "loc.err[%u] "
1155 "left[%u]\n",
1156 ctdb_db->db_name,
1157 (unsigned)vdata->count.delete_list.total,
1158 (unsigned)vdata->count.delete_list.deleted,
1159 (unsigned)vdata->count.delete_list.skipped,
1160 (unsigned)vdata->count.delete_list.remote_error,
1161 (unsigned)vdata->count.delete_list.local_error,
1162 (unsigned)vdata->count.delete_list.left));
1165 done:
1166 talloc_free(tmp_ctx);
1168 return;
1172 * initialize the vacuum_data
1174 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
1175 struct vacuum_data *vdata)
1177 int i;
1178 struct ctdb_context *ctdb = ctdb_db->ctdb;
1180 vdata->count.delete_queue.added_to_delete_list = 0;
1181 vdata->count.delete_queue.added_to_vacuum_fetch_list = 0;
1182 vdata->count.delete_queue.deleted = 0;
1183 vdata->count.delete_queue.skipped = 0;
1184 vdata->count.delete_queue.error = 0;
1185 vdata->count.delete_queue.total = 0;
1186 vdata->count.db_traverse.scheduled = 0;
1187 vdata->count.db_traverse.skipped = 0;
1188 vdata->count.db_traverse.error = 0;
1189 vdata->count.db_traverse.total = 0;
1190 vdata->count.delete_list.total = 0;
1191 vdata->count.delete_list.left = 0;
1192 vdata->count.delete_list.remote_error = 0;
1193 vdata->count.delete_list.local_error = 0;
1194 vdata->count.delete_list.skipped = 0;
1195 vdata->count.delete_list.deleted = 0;
1197 /* the list needs to be of length num_nodes */
1198 vdata->vacuum_fetch_list = talloc_zero_array(vdata,
1199 struct ctdb_marshall_buffer *,
1200 ctdb->num_nodes);
1201 if (vdata->vacuum_fetch_list == NULL) {
1202 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1203 return -1;
1205 for (i = 0; i < ctdb->num_nodes; i++) {
1206 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
1207 talloc_zero_size(vdata->vacuum_fetch_list,
1208 offsetof(struct ctdb_marshall_buffer, data));
1209 if (vdata->vacuum_fetch_list[i] == NULL) {
1210 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1211 return -1;
1213 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
1216 return 0;
1220 * Vacuum a DB:
1221 * - Always do the fast vacuuming run, which traverses
1222 * the in-memory delete queue: these records have been
1223 * scheduled for deletion.
1224 * - Only if explicitly requested, the database is traversed
1225 * in order to use the traditional heuristics on empty records
1226 * to trigger deletion.
1227 * This is done only every VacuumFastPathCount'th vacuuming run.
1229 * The traverse runs fill two lists:
1231 * - The delete_list:
1232 * This is the list of empty records the current
1233 * node is lmaster and dmaster for. These records are later
1234 * deleted first on other nodes and then locally.
1236 * The fast vacuuming run has a short cut for those records
1237 * that have never been migrated with data: these records
1238 * are immediately deleted locally, since they have left
1239 * no trace on other nodes.
1241 * - The vacuum_fetch lists
1242 * (one for each other lmaster node):
1243 * The records in this list are sent for deletion to
1244 * their lmaster in a bulk VACUUM_FETCH message.
1246 * The lmaster then migrates all these records to itelf
1247 * so that they can be vacuumed there.
1249 * This executes in the child context.
1251 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1252 struct vacuum_data *vdata,
1253 bool full_vacuum_run)
1255 struct ctdb_context *ctdb = ctdb_db->ctdb;
1256 int ret, pnn;
1258 DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1259 "%s db_id[0x%08x]\n",
1260 full_vacuum_run ? "full" : "fast",
1261 ctdb_db->db_name, ctdb_db->db_id));
1263 ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1264 if (ret != 0) {
1265 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1266 return ret;
1269 pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1270 if (pnn == -1) {
1271 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1272 return -1;
1275 ctdb->pnn = pnn;
1277 ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
1278 if (ret != 0) {
1279 return ret;
1282 if (full_vacuum_run) {
1283 ctdb_vacuum_traverse_db(ctdb_db, vdata);
1286 ctdb_process_delete_queue(ctdb_db, vdata);
1288 ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1290 ctdb_process_delete_list(ctdb_db, vdata);
1292 /* this ensures we run our event queue */
1293 ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1295 return 0;
1300 * traverse function for repacking
1302 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
1303 void *private_data)
1305 struct vacuum_data *vdata = (struct vacuum_data *)private_data;
1307 if (vdata->vacuum) {
1308 uint32_t hash = ctdb_hash(&key);
1309 struct delete_record_data *kd;
1311 * check if we can ignore this record because it's in the delete_list
1313 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1315 * there might be hash collisions so we have to compare the keys here to be sure
1317 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1318 struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1320 * we have to check if the record hasn't changed in the meantime in order to
1321 * savely remove it from the database
1323 if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1324 hdr->dmaster == kd->ctdb->pnn &&
1325 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1326 kd->hdr.rsn == hdr->rsn) {
1327 vdata->count.repack.vacuumed++;
1328 return 0;
1332 if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1333 vdata->traverse_error = true;
1334 return -1;
1336 vdata->count.repack.copied++;
1337 return 0;
1341 * repack a tdb
1343 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1345 struct tdb_context *tmp_db;
1347 if (tdb_transaction_start(tdb) != 0) {
1348 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1349 return -1;
1352 tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1353 TDB_INTERNAL|TDB_DISALLOW_NESTING,
1354 O_RDWR|O_CREAT, 0);
1355 if (tmp_db == NULL) {
1356 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1357 tdb_transaction_cancel(tdb);
1358 return -1;
1361 vdata->traverse_error = false;
1362 vdata->dest_db = tmp_db;
1363 vdata->vacuum = true;
1364 vdata->count.repack.vacuumed = 0;
1365 vdata->count.repack.copied = 0;
1368 * repack and vacuum on-the-fly by not writing the records that are
1369 * no longer needed
1371 if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1372 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1373 tdb_transaction_cancel(tdb);
1374 tdb_close(tmp_db);
1375 return -1;
1378 DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n",
1379 vdata->count.repack.vacuumed));
1381 if (vdata->traverse_error) {
1382 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1383 tdb_transaction_cancel(tdb);
1384 tdb_close(tmp_db);
1385 return -1;
1388 if (tdb_wipe_all(tdb) != 0) {
1389 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1390 tdb_transaction_cancel(tdb);
1391 tdb_close(tmp_db);
1392 return -1;
1395 vdata->traverse_error = false;
1396 vdata->dest_db = tdb;
1397 vdata->vacuum = false;
1398 vdata->count.repack.copied = 0;
1400 if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1401 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1402 tdb_transaction_cancel(tdb);
1403 tdb_close(tmp_db);
1404 return -1;
1407 if (vdata->traverse_error) {
1408 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1409 tdb_transaction_cancel(tdb);
1410 tdb_close(tmp_db);
1411 return -1;
1414 tdb_close(tmp_db);
1417 if (tdb_transaction_commit(tdb) != 0) {
1418 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1419 return -1;
1421 DEBUG(DEBUG_INFO,(__location__ " %u records copied\n",
1422 vdata->count.repack.copied));
1424 return 0;
1428 * repack and vaccum a db
1429 * called from the child context
1431 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1432 TALLOC_CTX *mem_ctx,
1433 bool full_vacuum_run)
1435 uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1436 const char *name = ctdb_db->db_name;
1437 int freelist_size = 0;
1438 struct vacuum_data *vdata;
1440 vdata = talloc_zero(mem_ctx, struct vacuum_data);
1441 if (vdata == NULL) {
1442 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1443 return -1;
1446 vdata->ctdb = ctdb_db->ctdb;
1447 vdata->repack_limit = repack_limit;
1448 vdata->delete_list = trbt_create(vdata, 0);
1449 vdata->ctdb_db = ctdb_db;
1450 if (vdata->delete_list == NULL) {
1451 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1452 talloc_free(vdata);
1453 return -1;
1456 vdata->start = timeval_current();
1459 * gather all records that can be deleted in vdata
1461 if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1462 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1465 if (repack_limit != 0) {
1466 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1467 if (freelist_size == -1) {
1468 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1469 talloc_free(vdata);
1470 return -1;
1475 * decide if a repack is necessary
1477 if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1479 talloc_free(vdata);
1480 return 0;
1483 DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1484 name, freelist_size, vdata->count.delete_list.left));
1487 * repack and implicitely get rid of the records we can delete
1489 if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1490 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1491 talloc_free(vdata);
1492 return -1;
1494 talloc_free(vdata);
1496 return 0;
1499 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1501 uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1503 return interval;
1506 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1508 double l = timeval_elapsed(&child_ctx->start_time);
1509 struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1510 struct ctdb_context *ctdb = ctdb_db->ctdb;
1512 DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1514 if (child_ctx->child_pid != -1) {
1515 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1516 } else {
1517 /* Bump the number of successful fast-path runs. */
1518 child_ctx->vacuum_handle->fast_path_count++;
1521 DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1523 event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1524 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1525 ctdb_vacuum_event, child_ctx->vacuum_handle);
1527 return 0;
1531 * this event is generated when a vacuum child process times out
1533 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1534 struct timeval t, void *private_data)
1536 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1538 DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1540 child_ctx->status = VACUUM_TIMEOUT;
1542 talloc_free(child_ctx);
1547 * this event is generated when a vacuum child process has completed
1549 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1550 uint16_t flags, void *private_data)
1552 struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1553 char c = 0;
1554 int ret;
1556 DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1557 child_ctx->child_pid = -1;
1559 ret = read(child_ctx->fd[0], &c, 1);
1560 if (ret != 1 || c != 0) {
1561 child_ctx->status = VACUUM_ERROR;
1562 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));
1563 } else {
1564 child_ctx->status = VACUUM_OK;
1567 talloc_free(child_ctx);
1571 * this event is called every time we need to start a new vacuum process
1573 static void
1574 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1575 struct timeval t, void *private_data)
1577 struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1578 struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1579 struct ctdb_context *ctdb = ctdb_db->ctdb;
1580 struct ctdb_vacuum_child_context *child_ctx;
1581 struct tevent_fd *fde;
1582 int ret;
1584 /* we dont vacuum if we are in recovery mode, or db frozen */
1585 if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1586 ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1587 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1588 ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1589 : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1590 ? "freeze pending"
1591 : "frozen"));
1592 event_add_timed(ctdb->ev, vacuum_handle,
1593 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1594 ctdb_vacuum_event, vacuum_handle);
1595 return;
1598 child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1599 if (child_ctx == NULL) {
1600 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1601 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1605 ret = pipe(child_ctx->fd);
1606 if (ret != 0) {
1607 talloc_free(child_ctx);
1608 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1609 event_add_timed(ctdb->ev, vacuum_handle,
1610 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1611 ctdb_vacuum_event, vacuum_handle);
1612 return;
1615 if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1616 vacuum_handle->fast_path_count = 0;
1619 child_ctx->child_pid = ctdb_fork(ctdb);
1620 if (child_ctx->child_pid == (pid_t)-1) {
1621 close(child_ctx->fd[0]);
1622 close(child_ctx->fd[1]);
1623 talloc_free(child_ctx);
1624 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1625 event_add_timed(ctdb->ev, vacuum_handle,
1626 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1627 ctdb_vacuum_event, vacuum_handle);
1628 return;
1632 if (child_ctx->child_pid == 0) {
1633 char cc = 0;
1634 bool full_vacuum_run = false;
1635 close(child_ctx->fd[0]);
1637 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1638 ctdb_set_process_name("ctdb_vacuum");
1639 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1640 DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1641 _exit(1);
1645 * repack the db
1647 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1648 (vacuum_handle->fast_path_count == 0))
1650 full_vacuum_run = true;
1652 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1653 full_vacuum_run);
1655 write(child_ctx->fd[1], &cc, 1);
1656 _exit(0);
1659 set_close_on_exec(child_ctx->fd[0]);
1660 close(child_ctx->fd[1]);
1662 child_ctx->status = VACUUM_RUNNING;
1663 child_ctx->start_time = timeval_current();
1665 DLIST_ADD(ctdb->vacuumers, child_ctx);
1666 talloc_set_destructor(child_ctx, vacuum_child_destructor);
1669 * Clear the fastpath vacuuming list in the parent.
1671 talloc_free(ctdb_db->delete_queue);
1672 ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1673 if (ctdb_db->delete_queue == NULL) {
1674 /* fatal here? ... */
1675 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1676 "in parent context. Shutting down\n");
1679 event_add_timed(ctdb->ev, child_ctx,
1680 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1681 vacuum_child_timeout, child_ctx);
1683 DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1685 fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1686 EVENT_FD_READ, vacuum_child_handler, child_ctx);
1687 tevent_fd_set_auto_close(fde);
1689 vacuum_handle->child_ctx = child_ctx;
1690 child_ctx->vacuum_handle = vacuum_handle;
1693 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1695 /* Simply free them all. */
1696 while (ctdb->vacuumers) {
1697 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1698 ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1699 (int)ctdb->vacuumers->child_pid));
1700 /* vacuum_child_destructor kills it, removes from list */
1701 talloc_free(ctdb->vacuumers);
1705 /* this function initializes the vacuuming context for a database
1706 * starts the vacuuming events
1708 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1710 if (ctdb_db->persistent != 0) {
1711 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1712 return 0;
1715 ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1716 CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1718 ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1719 ctdb_db->vacuum_handle->fast_path_count = 0;
1721 event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1722 timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1723 ctdb_vacuum_event, ctdb_db->vacuum_handle);
1725 return 0;
1728 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1729 const struct ctdb_ltdb_header *hdr,
1730 const TDB_DATA key)
1732 struct delete_record_data *kd;
1733 uint32_t hash;
1735 hash = (uint32_t)ctdb_hash(&key);
1737 DEBUG(DEBUG_DEBUG, (__location__
1738 " remove_record_from_delete_queue: "
1739 "db[%s] "
1740 "db_id[0x%08x] "
1741 "key_hash[0x%08x] "
1742 "lmaster[%u] "
1743 "migrated_with_data[%s]\n",
1744 ctdb_db->db_name, ctdb_db->db_id,
1745 hash,
1746 ctdb_lmaster(ctdb_db->ctdb, &key),
1747 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1749 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1750 if (kd == NULL) {
1751 DEBUG(DEBUG_DEBUG, (__location__
1752 " remove_record_from_delete_queue: "
1753 "record not in queue (hash[0x%08x])\n.",
1754 hash));
1755 return;
1758 if ((kd->key.dsize != key.dsize) ||
1759 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1761 DEBUG(DEBUG_DEBUG, (__location__
1762 " remove_record_from_delete_queue: "
1763 "hash collision for key with hash[0x%08x] "
1764 "in db[%s] - skipping\n",
1765 hash, ctdb_db->db_name));
1766 return;
1769 DEBUG(DEBUG_DEBUG, (__location__
1770 " remove_record_from_delete_queue: "
1771 "removing key with hash[0x%08x]\n",
1772 hash));
1774 talloc_free(kd);
1776 return;
1780 * Insert a record into the ctdb_db context's delete queue,
1781 * handling hash collisions.
1783 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1784 const struct ctdb_ltdb_header *hdr,
1785 TDB_DATA key)
1787 struct delete_record_data *kd;
1788 uint32_t hash;
1789 int ret;
1791 hash = (uint32_t)ctdb_hash(&key);
1793 DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1794 "db_id[0x%08x] "
1795 "key_hash[0x%08x] "
1796 "lmaster[%u] "
1797 "migrated_with_data[%s]\n",
1798 ctdb_db->db_name, ctdb_db->db_id,
1799 hash,
1800 ctdb_lmaster(ctdb_db->ctdb, &key),
1801 hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1803 kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1804 if (kd != NULL) {
1805 if ((kd->key.dsize != key.dsize) ||
1806 (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1808 DEBUG(DEBUG_INFO,
1809 (__location__ " schedule for deletion: "
1810 "hash collision for key hash [0x%08x]. "
1811 "Skipping the record.\n", hash));
1812 return 0;
1813 } else {
1814 DEBUG(DEBUG_DEBUG,
1815 (__location__ " schedule for deletion: "
1816 "updating entry for key with hash [0x%08x].\n",
1817 hash));
1821 ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1822 ctdb_db->delete_queue,
1823 hdr, key);
1824 if (ret != 0) {
1825 DEBUG(DEBUG_INFO,
1826 (__location__ " schedule for deletion: error "
1827 "inserting key with hash [0x%08x] into delete queue\n",
1828 hash));
1829 return -1;
1832 return 0;
1836 * Schedule a record for deletetion.
1837 * Called from the parent context.
1839 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1840 TDB_DATA indata)
1842 struct ctdb_control_schedule_for_deletion *dd;
1843 struct ctdb_db_context *ctdb_db;
1844 int ret;
1845 TDB_DATA key;
1847 dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1849 ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1850 if (ctdb_db == NULL) {
1851 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1852 dd->db_id));
1853 return -1;
1856 key.dsize = dd->keylen;
1857 key.dptr = dd->key;
1859 ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1861 return ret;
1864 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1865 const struct ctdb_ltdb_header *hdr,
1866 TDB_DATA key)
1868 int ret;
1869 struct ctdb_control_schedule_for_deletion *dd;
1870 TDB_DATA indata;
1871 int32_t status;
1873 if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1874 /* main daemon - directly queue */
1875 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1877 return ret;
1880 /* if we dont have a connection to the daemon we can not send
1881 a control. For example sometimes from update_record control child
1882 process.
1884 if (!ctdb_db->ctdb->can_send_controls) {
1885 return -1;
1889 /* child process: send the main daemon a control */
1890 indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1891 indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1892 if (indata.dptr == NULL) {
1893 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1894 return -1;
1896 dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1897 dd->db_id = ctdb_db->db_id;
1898 dd->hdr = *hdr;
1899 dd->keylen = key.dsize;
1900 memcpy(dd->key, key.dptr, key.dsize);
1902 ret = ctdb_control(ctdb_db->ctdb,
1903 CTDB_CURRENT_NODE,
1904 ctdb_db->db_id,
1905 CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1906 CTDB_CTRL_FLAG_NOREPLY, /* flags */
1907 indata,
1908 NULL, /* mem_ctx */
1909 NULL, /* outdata */
1910 &status,
1911 NULL, /* timeout : NULL == wait forever */
1912 NULL); /* error message */
1914 talloc_free(indata.dptr);
1916 if (ret != 0 || status != 0) {
1917 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1918 "SCHEDULE_FOR_DELETION "
1919 "control.\n"));
1920 if (status != 0) {
1921 ret = -1;
1925 return ret;
1928 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1929 const struct ctdb_ltdb_header *hdr,
1930 const TDB_DATA key)
1932 if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1934 * Only remove the record from the delete queue if called
1935 * in the main daemon.
1937 return;
1940 remove_record_from_delete_queue(ctdb_db, hdr, key);
1942 return;