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/>.
24 #include "system/network.h"
25 #include "system/filesys.h"
26 #include "system/dir.h"
27 #include "../include/ctdb_private.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 */
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 */
56 uint32_t vacuum_limit
;
57 uint32_t repack_limit
;
58 struct ctdb_context
*ctdb
;
59 struct ctdb_db_context
*ctdb_db
;
60 struct tdb_context
*dest_db
;
61 trbt_tree_t
*delete_list
;
62 uint32_t delete_count
;
63 struct ctdb_marshall_buffer
**vacuum_fetch_list
;
70 uint32_t fast_added_to_vacuum_fetch_list
;
71 uint32_t fast_added_to_delete_list
;
72 uint32_t fast_deleted
;
73 uint32_t fast_skipped
;
76 uint32_t full_scheduled
;
77 uint32_t full_skipped
;
81 uint32_t delete_remote_error
;
82 uint32_t delete_local_error
;
83 uint32_t delete_deleted
;
84 uint32_t delete_skipped
;
87 /* this structure contains the information for one record to be deleted */
88 struct delete_record_data
{
89 struct ctdb_context
*ctdb
;
90 struct ctdb_db_context
*ctdb_db
;
91 struct ctdb_ltdb_header hdr
;
96 struct delete_records_list
{
97 struct ctdb_marshall_buffer
*records
;
98 struct vacuum_data
*vdata
;
101 static int insert_record_into_delete_queue(struct ctdb_db_context
*ctdb_db
,
102 const struct ctdb_ltdb_header
*hdr
,
106 * Store key and header in a tree, indexed by the key hash.
108 static int insert_delete_record_data_into_tree(struct ctdb_context
*ctdb
,
109 struct ctdb_db_context
*ctdb_db
,
111 const struct ctdb_ltdb_header
*hdr
,
114 struct delete_record_data
*dd
;
118 len
= offsetof(struct delete_record_data
, keydata
) + key
.dsize
;
120 dd
= (struct delete_record_data
*)talloc_size(tree
, len
);
122 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
125 talloc_set_name_const(dd
, "struct delete_record_data");
128 dd
->ctdb_db
= ctdb_db
;
129 dd
->key
.dsize
= key
.dsize
;
130 dd
->key
.dptr
= dd
->keydata
;
131 memcpy(dd
->keydata
, key
.dptr
, key
.dsize
);
135 hash
= ctdb_hash(&key
);
137 trbt_insert32(tree
, hash
, dd
);
142 static int add_record_to_delete_list(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_list
, hash
)) {
153 DEBUG(DEBUG_INFO
, (__location__
" Hash collision 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
;
180 struct ctdb_marshall_buffer
*vfl
;
182 lmaster
= ctdb_lmaster(ctdb
, &key
);
184 vfl
= vdata
->vacuum_fetch_list
[lmaster
];
186 rec
= ctdb_marshall_record(vfl
, ctdb
->pnn
, key
, NULL
, tdb_null
);
188 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
189 vdata
->traverse_error
= true;
193 old_size
= talloc_get_size(vfl
);
194 vfl
= talloc_realloc_size(NULL
, vfl
, old_size
+ rec
->length
);
196 DEBUG(DEBUG_ERR
,(__location__
" Failed to expand\n"));
197 vdata
->traverse_error
= true;
200 vdata
->vacuum_fetch_list
[lmaster
] = vfl
;
203 memcpy(old_size
+(uint8_t *)vfl
, rec
, rec
->length
);
212 static void ctdb_vacuum_event(struct event_context
*ev
, struct timed_event
*te
,
213 struct timeval t
, void *private_data
);
215 static int vacuum_record_parser(TDB_DATA key
, TDB_DATA data
, void *private_data
)
217 struct ctdb_ltdb_header
*header
=
218 (struct ctdb_ltdb_header
*)private_data
;
220 if (data
.dsize
!= sizeof(struct ctdb_ltdb_header
)) {
224 *header
= *(struct ctdb_ltdb_header
*)data
.dptr
;
230 * traverse function for gathering the records that can be deleted
232 static int vacuum_traverse(struct tdb_context
*tdb
, TDB_DATA key
, TDB_DATA data
,
235 struct vacuum_data
*vdata
= talloc_get_type(private_data
,
237 struct ctdb_context
*ctdb
= vdata
->ctdb
;
238 struct ctdb_db_context
*ctdb_db
= vdata
->ctdb_db
;
240 struct ctdb_ltdb_header
*hdr
;
245 lmaster
= ctdb_lmaster(ctdb
, &key
);
246 if (lmaster
>= ctdb
->num_nodes
) {
248 DEBUG(DEBUG_CRIT
, (__location__
249 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
252 (unsigned)ctdb
->num_nodes
,
253 (unsigned)ctdb_hash(&key
)));
257 if (data
.dsize
!= sizeof(struct ctdb_ltdb_header
)) {
258 /* it is not a deleted record */
259 vdata
->full_skipped
++;
263 hdr
= (struct ctdb_ltdb_header
*)data
.dptr
;
265 if (hdr
->dmaster
!= ctdb
->pnn
) {
266 vdata
->full_skipped
++;
271 * Add the record to this process's delete_queue for processing
272 * in the subsequent traverse in the fast vacuum run.
274 res
= insert_record_into_delete_queue(ctdb_db
, hdr
, key
);
278 vdata
->full_scheduled
++;
285 * traverse the tree of records to delete and marshall them into
288 static int delete_marshall_traverse(void *param
, void *data
)
290 struct delete_record_data
*dd
= talloc_get_type(data
, struct delete_record_data
);
291 struct delete_records_list
*recs
= talloc_get_type(param
, struct delete_records_list
);
292 struct ctdb_rec_data
*rec
;
295 rec
= ctdb_marshall_record(dd
, recs
->records
->db_id
, dd
->key
, &dd
->hdr
, tdb_null
);
297 DEBUG(DEBUG_ERR
, (__location__
" failed to marshall record\n"));
301 old_size
= talloc_get_size(recs
->records
);
302 recs
->records
= talloc_realloc_size(NULL
, recs
->records
, old_size
+ rec
->length
);
303 if (recs
->records
== NULL
) {
304 DEBUG(DEBUG_ERR
,(__location__
" Failed to expand\n"));
307 recs
->records
->count
++;
308 memcpy(old_size
+(uint8_t *)(recs
->records
), rec
, rec
->length
);
313 * Variant of delete_marshall_traverse() that bumps the
314 * RSN of each traversed record in the database.
316 * This is needed to ensure that when rolling out our
317 * empty record copy before remote deletion, we as the
318 * record's dmaster keep a higher RSN than the non-dmaster
319 * nodes. This is needed to prevent old copies from
320 * resurrection in recoveries.
322 static int delete_marshall_traverse_first(void *param
, void *data
)
324 struct delete_record_data
*dd
= talloc_get_type(data
, struct delete_record_data
);
325 struct delete_records_list
*recs
= talloc_get_type(param
, struct delete_records_list
);
326 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
327 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
328 struct ctdb_ltdb_header
*header
;
329 TDB_DATA tdb_data
, ctdb_data
;
331 uint32_t hash
= ctdb_hash(&(dd
->key
));
334 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
337 (__location__
" Error getting chainlock on record with "
338 "key hash [0x%08x] on database db[%s].\n",
339 hash
, ctdb_db
->db_name
));
340 recs
->vdata
->delete_skipped
++;
346 * Verify that the record is still empty, its RSN has not
347 * changed and that we are still its lmaster and dmaster.
350 tdb_data
= tdb_fetch(ctdb_db
->ltdb
->tdb
, dd
->key
);
351 if (tdb_data
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
352 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
353 "on database db[%s] does not exist or is not"
354 " a ctdb-record. skipping.\n",
355 hash
, ctdb_db
->db_name
));
359 if (tdb_data
.dsize
> sizeof(struct ctdb_ltdb_header
)) {
360 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
361 "on database db[%s] has been recycled. "
363 hash
, ctdb_db
->db_name
));
367 header
= (struct ctdb_ltdb_header
*)tdb_data
.dptr
;
369 if (header
->flags
& CTDB_REC_RO_FLAGS
) {
370 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
371 "on database db[%s] has read-only flags. "
373 hash
, ctdb_db
->db_name
));
377 if (header
->dmaster
!= ctdb
->pnn
) {
378 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
379 "on database db[%s] has been migrated away. "
381 hash
, ctdb_db
->db_name
));
385 if (header
->rsn
!= dd
->hdr
.rsn
) {
386 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
387 "on database db[%s] seems to have been "
388 "migrated away and back again (with empty "
389 "data). skipping.\n",
390 hash
, ctdb_db
->db_name
));
394 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
396 if (lmaster
!= ctdb
->pnn
) {
397 DEBUG(DEBUG_INFO
, (__location__
": not lmaster for record in "
398 "delete list (key hash [0x%08x], db[%s]). "
399 "Strange! skipping.\n",
400 hash
, ctdb_db
->db_name
));
405 * Increment the record's RSN to ensure the dmaster (i.e. the current
406 * node) has the highest RSN of the record in the cluster.
407 * This is to prevent old record copies from resurrecting in recoveries
408 * if something should fail during the deletion process.
409 * Note that ctdb_ltdb_store_server() increments the RSN if called
410 * on the record's dmaster.
413 ctdb_data
.dptr
= tdb_data
.dptr
+ sizeof(struct ctdb_ltdb_header
);
414 ctdb_data
.dsize
= tdb_data
.dsize
- sizeof(struct ctdb_ltdb_header
);
416 res
= ctdb_ltdb_store(ctdb_db
, dd
->key
, header
, ctdb_data
);
418 DEBUG(DEBUG_ERR
, (__location__
": Failed to store record with "
419 "key hash [0x%08x] on database db[%s].\n",
420 hash
, ctdb_db
->db_name
));
424 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
429 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
431 recs
->vdata
->delete_skipped
++;
436 if (tdb_data
.dptr
!= NULL
) {
444 return delete_marshall_traverse(param
, data
);
448 * traverse function for the traversal of the delete_queue,
449 * the fast-path vacuuming list.
451 * - If the record has been migrated off the node
452 * or has been revived (filled with data) on the node,
453 * then skip the record.
455 * - If the current node is the record's lmaster and it is
456 * a record that has never been migrated with data, then
457 * delete the record from the local tdb.
459 * - If the current node is the record's lmaster and it has
460 * been migrated with data, then schedule it for the normal
461 * vacuuming procedure (i.e. add it to the delete_list).
463 * - If the current node is NOT the record's lmaster then
464 * add it to the list of records that are to be sent to
465 * the lmaster with the VACUUM_FETCH message.
467 static int delete_queue_traverse(void *param
, void *data
)
469 struct delete_record_data
*dd
=
470 talloc_get_type(data
, struct delete_record_data
);
471 struct vacuum_data
*vdata
= talloc_get_type(param
, struct vacuum_data
);
472 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
473 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
; /* or dd->ctdb ??? */
475 struct ctdb_ltdb_header header
;
477 uint32_t hash
= ctdb_hash(&(dd
->key
));
481 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
484 (__location__
" Error getting chainlock on record with "
485 "key hash [0x%08x] on database db[%s].\n",
486 hash
, ctdb_db
->db_name
));
491 res
= tdb_parse_record(ctdb_db
->ltdb
->tdb
, dd
->key
,
492 vacuum_record_parser
, &header
);
497 if (header
.dmaster
!= ctdb
->pnn
) {
498 /* The record has been migrated off the node. Skip. */
502 if (header
.rsn
!= dd
->hdr
.rsn
) {
504 * The record has been migrated off the node and back again.
505 * But not requeued for deletion. Skip it.
511 * We are dmaster, and the record has no data, and it has
512 * not been migrated after it has been queued for deletion.
514 * At this stage, the record could still have been revived locally
515 * and last been written with empty data. This can only be
516 * fixed with the addition of an active or delete flag. (TODO)
519 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
521 if (lmaster
!= ctdb
->pnn
) {
522 res
= add_record_to_vacuum_fetch_list(vdata
, dd
->key
);
526 (__location__
" Error adding record to list "
527 "of records to send to lmaster.\n"));
530 vdata
->fast_added_to_vacuum_fetch_list
++;
535 /* use header->flags or dd->hdr.flags ?? */
536 if (dd
->hdr
.flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
) {
537 res
= add_record_to_delete_list(vdata
, dd
->key
, &dd
->hdr
);
541 (__location__
" Error adding record to list "
542 "of records for deletion on lmaster.\n"));
545 vdata
->fast_added_to_delete_list
++;
548 res
= tdb_delete(ctdb_db
->ltdb
->tdb
, dd
->key
);
552 (__location__
" Error deleting record with key "
553 "hash [0x%08x] from local data base db[%s].\n",
554 hash
, ctdb_db
->db_name
));
560 (__location__
" Deleted record with key hash "
561 "[0x%08x] from local data base db[%s].\n",
562 hash
, ctdb_db
->db_name
));
563 vdata
->fast_deleted
++;
569 vdata
->fast_skipped
++;
572 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
578 * Delete the records that we are lmaster and dmaster for and
579 * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
582 static int delete_record_traverse(void *param
, void *data
)
584 struct delete_record_data
*dd
=
585 talloc_get_type(data
, struct delete_record_data
);
586 struct vacuum_data
*vdata
= talloc_get_type(param
, struct vacuum_data
);
587 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
588 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
590 struct ctdb_ltdb_header header
;
592 uint32_t hash
= ctdb_hash(&(dd
->key
));
594 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
597 (__location__
" Error getting chainlock on record with "
598 "key hash [0x%08x] on database db[%s].\n",
599 hash
, ctdb_db
->db_name
));
600 vdata
->delete_local_error
++;
601 vdata
->delete_left
--;
607 * Verify that the record is still empty, its RSN has not
608 * changed and that we are still its lmaster and dmaster.
611 res
= tdb_parse_record(ctdb_db
->ltdb
->tdb
, dd
->key
,
612 vacuum_record_parser
, &header
);
617 if (header
.flags
& CTDB_REC_RO_FLAGS
) {
618 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
619 "on database db[%s] has read-only flags. "
621 hash
, ctdb_db
->db_name
));
625 if (header
.dmaster
!= ctdb
->pnn
) {
626 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
627 "on database db[%s] has been migrated away. "
629 hash
, ctdb_db
->db_name
));
633 if (header
.rsn
!= dd
->hdr
.rsn
+ 1) {
635 * The record has been migrated off the node and back again.
636 * But not requeued for deletion. Skip it.
637 * (Note that the first marshall traverse has bumped the RSN
640 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
641 "on database db[%s] seems to have been "
642 "migrated away and back again (with empty "
643 "data). skipping.\n",
644 hash
, ctdb_db
->db_name
));
648 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
650 if (lmaster
!= ctdb
->pnn
) {
651 DEBUG(DEBUG_INFO
, (__location__
": not lmaster for record in "
652 "delete list (key hash [0x%08x], db[%s]). "
653 "Strange! skipping.\n",
654 hash
, ctdb_db
->db_name
));
658 res
= tdb_delete(ctdb_db
->ltdb
->tdb
, dd
->key
);
662 (__location__
" Error deleting record with key hash "
663 "[0x%08x] from local data base db[%s].\n",
664 hash
, ctdb_db
->db_name
));
665 vdata
->delete_local_error
++;
670 (__location__
" Deleted record with key hash [0x%08x] from "
671 "local data base db[%s].\n", hash
, ctdb_db
->db_name
));
673 vdata
->delete_deleted
++;
677 vdata
->delete_skipped
++;
680 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
683 vdata
->delete_left
--;
689 * Traverse the delete_queue.
690 * Records are either deleted directly or filled
691 * into the delete list or the vacuum fetch lists
692 * for further processing.
694 static void ctdb_process_delete_queue(struct ctdb_db_context
*ctdb_db
,
695 struct vacuum_data
*vdata
)
699 trbt_traversearray32(ctdb_db
->delete_queue
, 1, delete_queue_traverse
, vdata
);
701 sum
= vdata
->fast_deleted
702 + vdata
->fast_skipped
704 + vdata
->fast_added_to_delete_list
705 + vdata
->fast_added_to_vacuum_fetch_list
;
707 if (vdata
->fast_total
!= sum
) {
708 DEBUG(DEBUG_ERR
, (__location__
" Inconsistency in fast vacuum "
709 "counts for db[%s]: total[%u] != sum[%u]\n",
710 ctdb_db
->db_name
, (unsigned)vdata
->fast_total
,
714 if (vdata
->fast_total
> 0) {
717 " fast vacuuming delete_queue traverse statistics: "
726 (unsigned)vdata
->fast_total
,
727 (unsigned)vdata
->fast_deleted
,
728 (unsigned)vdata
->fast_skipped
,
729 (unsigned)vdata
->fast_error
,
730 (unsigned)vdata
->fast_added_to_delete_list
,
731 (unsigned)vdata
->fast_added_to_vacuum_fetch_list
));
738 * read-only traverse of the database, looking for records that
739 * might be able to be vacuumed.
741 * This is not done each time but only every tunable
742 * VacuumFastPathCount times.
744 static int ctdb_vacuum_traverse_db(struct ctdb_db_context
*ctdb_db
,
745 struct vacuum_data
*vdata
)
749 ret
= tdb_traverse_read(ctdb_db
->ltdb
->tdb
, vacuum_traverse
, vdata
);
750 if (ret
== -1 || vdata
->traverse_error
) {
751 DEBUG(DEBUG_ERR
, (__location__
" Traverse error in vacuuming "
752 "'%s'\n", ctdb_db
->db_name
));
756 if (vdata
->full_total
> 0) {
759 " full vacuuming db traverse statistics: "
766 (unsigned)vdata
->full_total
,
767 (unsigned)vdata
->full_skipped
,
768 (unsigned)vdata
->full_error
,
769 (unsigned)vdata
->full_scheduled
));
776 * Process the vacuum fetch lists:
777 * For records for which we are not the lmaster, tell the lmaster to
780 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context
*ctdb_db
,
781 struct vacuum_data
*vdata
)
784 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
786 for (i
= 0; i
< ctdb
->num_nodes
; i
++) {
788 struct ctdb_marshall_buffer
*vfl
= vdata
->vacuum_fetch_list
[i
];
790 if (ctdb
->nodes
[i
]->pnn
== ctdb
->pnn
) {
794 if (vfl
->count
== 0) {
798 DEBUG(DEBUG_INFO
, ("Found %u records for lmaster %u in '%s'\n",
799 vfl
->count
, ctdb
->nodes
[i
]->pnn
,
802 data
.dsize
= talloc_get_size(vfl
);
803 data
.dptr
= (void *)vfl
;
804 if (ctdb_client_send_message(ctdb
, ctdb
->nodes
[i
]->pnn
,
805 CTDB_SRVID_VACUUM_FETCH
,
808 DEBUG(DEBUG_ERR
, (__location__
" Failed to send vacuum "
809 "fetch message to %u\n",
810 ctdb
->nodes
[i
]->pnn
));
819 * Process the delete list:
821 * This is the last step of vacuuming that consistently deletes
822 * those records that have been migrated with data and can hence
823 * not be deleted when leaving a node.
825 * In this step, the lmaster does the final deletion of those empty
826 * records that it is also dmaster for. It has ususally received
827 * at least some of these records previously from the former dmasters
828 * with the vacuum fetch message.
830 * This last step is implemented as a 3-phase process to protect from
831 * races leading to data corruption:
833 * 1) Send the lmaster's copy to all other active nodes with the
834 * RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
835 * 2) Send the records that could successfully be stored remotely
836 * in step #1 to all active nodes with the TRY_DELETE_RECORDS
837 * control. The remote notes delete their local copy.
838 * 3) The lmaster locally deletes its copies of all records that
839 * could successfully be deleted remotely in step #2.
841 static int ctdb_process_delete_list(struct ctdb_db_context
*ctdb_db
,
842 struct vacuum_data
*vdata
)
845 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
846 struct delete_records_list
*recs
;
848 struct ctdb_node_map
*nodemap
;
849 uint32_t *active_nodes
;
850 int num_active_nodes
;
853 if (vdata
->delete_count
== 0) {
857 tmp_ctx
= talloc_new(vdata
);
858 if (tmp_ctx
== NULL
) {
859 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
863 vdata
->delete_left
= vdata
->delete_count
;
866 * get the list of currently active nodes
869 ret
= ctdb_ctrl_getnodemap(ctdb
, TIMELIMIT(),
874 DEBUG(DEBUG_ERR
,(__location__
" unable to get node map\n"));
879 active_nodes
= list_of_active_nodes(ctdb
, nodemap
,
880 nodemap
, /* talloc context */
881 false /* include self */);
883 num_active_nodes
= talloc_get_size(active_nodes
)/sizeof(*active_nodes
);
886 * Now delete the records all active nodes in a three-phase process:
887 * 1) send all active remote nodes the current empty copy with this
889 * 2) if all nodes could store the new copy,
890 * tell all the active remote nodes to delete all their copy
891 * 3) if all remote nodes deleted their record copy, delete it locally
896 * Send currently empty record copy to all active nodes for storing.
899 recs
= talloc_zero(tmp_ctx
, struct delete_records_list
);
901 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
905 recs
->records
= (struct ctdb_marshall_buffer
*)
906 talloc_zero_size(recs
,
907 offsetof(struct ctdb_marshall_buffer
, data
));
908 if (recs
->records
== NULL
) {
909 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
913 recs
->records
->db_id
= ctdb_db
->db_id
;
917 * traverse the tree of all records we want to delete and
918 * create a blob we can send to the other nodes.
920 * We call delete_marshall_traverse_first() to bump the
921 * records' RSNs in the database, to ensure we (as dmaster)
922 * keep the highest RSN of the records in the cluster.
924 trbt_traversearray32(vdata
->delete_list
, 1,
925 delete_marshall_traverse_first
, recs
);
927 indata
.dsize
= talloc_get_size(recs
->records
);
928 indata
.dptr
= (void *)recs
->records
;
930 for (i
= 0; i
< num_active_nodes
; i
++) {
931 struct ctdb_marshall_buffer
*records
;
932 struct ctdb_rec_data
*rec
;
936 ret
= ctdb_control(ctdb
, active_nodes
[i
], 0,
937 CTDB_CONTROL_RECEIVE_RECORDS
, 0,
938 indata
, recs
, &outdata
, &res
,
940 if (ret
!= 0 || res
!= 0) {
941 DEBUG(DEBUG_ERR
, ("Error storing record copies on "
942 "node %u: ret[%d] res[%d]\n",
943 active_nodes
[i
], ret
, res
));
949 * outdata contains the list of records coming back
950 * from the node: These are the records that the
951 * remote node could not store. We remove these from
952 * the list to process further.
954 records
= (struct ctdb_marshall_buffer
*)outdata
.dptr
;
955 rec
= (struct ctdb_rec_data
*)&records
->data
[0];
956 while (records
->count
-- > 1) {
957 TDB_DATA reckey
, recdata
;
958 struct ctdb_ltdb_header
*rechdr
;
959 struct delete_record_data
*dd
;
961 reckey
.dptr
= &rec
->data
[0];
962 reckey
.dsize
= rec
->keylen
;
963 recdata
.dptr
= &rec
->data
[reckey
.dsize
];
964 recdata
.dsize
= rec
->datalen
;
966 if (recdata
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
967 DEBUG(DEBUG_CRIT
,(__location__
" bad ltdb record\n"));
971 rechdr
= (struct ctdb_ltdb_header
*)recdata
.dptr
;
972 recdata
.dptr
+= sizeof(*rechdr
);
973 recdata
.dsize
-= sizeof(*rechdr
);
975 dd
= (struct delete_record_data
*)trbt_lookup32(
980 * The other node could not store the record
981 * copy and it is the first node that failed.
982 * So we should remove it from the tree and
986 vdata
->delete_remote_error
++;
987 vdata
->delete_left
--;
990 rec
= (struct ctdb_rec_data
*)(rec
->length
+ (uint8_t *)rec
);
994 if (vdata
->delete_left
== 0) {
1000 * Send the remaining records to all active nodes for deletion.
1002 * The lmaster's (i.e. our) copies of these records have been stored
1003 * successfully on the other nodes.
1007 * Create a marshall blob from the remaining list of records to delete.
1010 talloc_free(recs
->records
);
1012 recs
->records
= (struct ctdb_marshall_buffer
*)
1013 talloc_zero_size(recs
,
1014 offsetof(struct ctdb_marshall_buffer
, data
));
1015 if (recs
->records
== NULL
) {
1016 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1020 recs
->records
->db_id
= ctdb_db
->db_id
;
1022 trbt_traversearray32(vdata
->delete_list
, 1,
1023 delete_marshall_traverse
, recs
);
1025 indata
.dsize
= talloc_get_size(recs
->records
);
1026 indata
.dptr
= (void *)recs
->records
;
1028 for (i
= 0; i
< num_active_nodes
; i
++) {
1029 struct ctdb_marshall_buffer
*records
;
1030 struct ctdb_rec_data
*rec
;
1034 ret
= ctdb_control(ctdb
, active_nodes
[i
], 0,
1035 CTDB_CONTROL_TRY_DELETE_RECORDS
, 0,
1036 indata
, recs
, &outdata
, &res
,
1038 if (ret
!= 0 || res
!= 0) {
1039 DEBUG(DEBUG_ERR
, ("Failed to delete records on "
1040 "node %u: ret[%d] res[%d]\n",
1041 active_nodes
[i
], ret
, res
));
1047 * outdata contains the list of records coming back
1048 * from the node: These are the records that the
1049 * remote node could not delete. We remove these from
1050 * the list to delete locally.
1052 records
= (struct ctdb_marshall_buffer
*)outdata
.dptr
;
1053 rec
= (struct ctdb_rec_data
*)&records
->data
[0];
1054 while (records
->count
-- > 1) {
1055 TDB_DATA reckey
, recdata
;
1056 struct ctdb_ltdb_header
*rechdr
;
1057 struct delete_record_data
*dd
;
1059 reckey
.dptr
= &rec
->data
[0];
1060 reckey
.dsize
= rec
->keylen
;
1061 recdata
.dptr
= &rec
->data
[reckey
.dsize
];
1062 recdata
.dsize
= rec
->datalen
;
1064 if (recdata
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
1065 DEBUG(DEBUG_CRIT
,(__location__
" bad ltdb record\n"));
1069 rechdr
= (struct ctdb_ltdb_header
*)recdata
.dptr
;
1070 recdata
.dptr
+= sizeof(*rechdr
);
1071 recdata
.dsize
-= sizeof(*rechdr
);
1073 dd
= (struct delete_record_data
*)trbt_lookup32(
1075 ctdb_hash(&reckey
));
1078 * The other node could not delete the
1079 * record and it is the first node that
1080 * failed. So we should remove it from
1081 * the tree and update statistics.
1084 vdata
->delete_remote_error
++;
1085 vdata
->delete_left
--;
1088 rec
= (struct ctdb_rec_data
*)(rec
->length
+ (uint8_t *)rec
);
1092 if (vdata
->delete_left
== 0) {
1098 * Delete the remaining records locally.
1100 * These records have successfully been deleted on all
1101 * active remote nodes.
1104 trbt_traversearray32(vdata
->delete_list
, 1,
1105 delete_record_traverse
, vdata
);
1109 if (vdata
->delete_count
> 0) {
1112 " vacuum delete list statistics: "
1121 (unsigned)vdata
->delete_count
,
1122 (unsigned)vdata
->delete_deleted
,
1123 (unsigned)vdata
->delete_skipped
,
1124 (unsigned)vdata
->delete_remote_error
,
1125 (unsigned)vdata
->delete_local_error
,
1126 (unsigned)vdata
->delete_left
));
1132 talloc_free(tmp_ctx
);
1138 * initialize the vacuum_data
1140 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context
*ctdb_db
,
1141 struct vacuum_data
*vdata
)
1144 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1146 vdata
->fast_added_to_delete_list
= 0;
1147 vdata
->fast_added_to_vacuum_fetch_list
= 0;
1148 vdata
->fast_deleted
= 0;
1149 vdata
->fast_skipped
= 0;
1150 vdata
->fast_error
= 0;
1151 vdata
->fast_total
= 0;
1152 vdata
->full_scheduled
= 0;
1153 vdata
->full_skipped
= 0;
1154 vdata
->full_error
= 0;
1155 vdata
->full_total
= 0;
1156 vdata
->delete_count
= 0;
1157 vdata
->delete_left
= 0;
1158 vdata
->delete_remote_error
= 0;
1159 vdata
->delete_local_error
= 0;
1160 vdata
->delete_skipped
= 0;
1161 vdata
->delete_deleted
= 0;
1163 /* the list needs to be of length num_nodes */
1164 vdata
->vacuum_fetch_list
= talloc_zero_array(vdata
,
1165 struct ctdb_marshall_buffer
*,
1167 if (vdata
->vacuum_fetch_list
== NULL
) {
1168 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1171 for (i
= 0; i
< ctdb
->num_nodes
; i
++) {
1172 vdata
->vacuum_fetch_list
[i
] = (struct ctdb_marshall_buffer
*)
1173 talloc_zero_size(vdata
->vacuum_fetch_list
,
1174 offsetof(struct ctdb_marshall_buffer
, data
));
1175 if (vdata
->vacuum_fetch_list
[i
] == NULL
) {
1176 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1179 vdata
->vacuum_fetch_list
[i
]->db_id
= ctdb_db
->db_id
;
1187 * - Always do the fast vacuuming run, which traverses
1188 * the in-memory delete queue: these records have been
1189 * scheduled for deletion.
1190 * - Only if explicitly requested, the database is traversed
1191 * in order to use the traditional heuristics on empty records
1192 * to trigger deletion.
1193 * This is done only every VacuumFastPathCount'th vacuuming run.
1195 * The traverse runs fill two lists:
1197 * - The delete_list:
1198 * This is the list of empty records the current
1199 * node is lmaster and dmaster for. These records are later
1200 * deleted first on other nodes and then locally.
1202 * The fast vacuuming run has a short cut for those records
1203 * that have never been migrated with data: these records
1204 * are immediately deleted locally, since they have left
1205 * no trace on other nodes.
1207 * - The vacuum_fetch lists
1208 * (one for each other lmaster node):
1209 * The records in this list are sent for deletion to
1210 * their lmaster in a bulk VACUUM_FETCH message.
1212 * The lmaster then migrates all these records to itelf
1213 * so that they can be vacuumed there.
1215 * This executes in the child context.
1217 static int ctdb_vacuum_db(struct ctdb_db_context
*ctdb_db
,
1218 struct vacuum_data
*vdata
,
1219 bool full_vacuum_run
)
1221 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1224 DEBUG(DEBUG_INFO
, (__location__
" Entering %s vacuum run for db "
1225 "%s db_id[0x%08x]\n",
1226 full_vacuum_run
? "full" : "fast",
1227 ctdb_db
->db_name
, ctdb_db
->db_id
));
1229 ret
= ctdb_ctrl_getvnnmap(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
, ctdb
, &ctdb
->vnn_map
);
1231 DEBUG(DEBUG_ERR
, ("Unable to get vnnmap from local node\n"));
1235 pnn
= ctdb_ctrl_getpnn(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
);
1237 DEBUG(DEBUG_ERR
, ("Unable to get pnn from local node\n"));
1243 ret
= ctdb_vacuum_init_vacuum_data(ctdb_db
, vdata
);
1248 if (full_vacuum_run
) {
1249 ret
= ctdb_vacuum_traverse_db(ctdb_db
, vdata
);
1255 ctdb_process_delete_queue(ctdb_db
, vdata
);
1257 ret
= ctdb_process_vacuum_fetch_lists(ctdb_db
, vdata
);
1262 ret
= ctdb_process_delete_list(ctdb_db
, vdata
);
1267 /* this ensures we run our event queue */
1268 ctdb_ctrl_getpnn(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
);
1275 * traverse function for repacking
1277 static int repack_traverse(struct tdb_context
*tdb
, TDB_DATA key
, TDB_DATA data
,
1280 struct vacuum_data
*vdata
= (struct vacuum_data
*)private_data
;
1282 if (vdata
->vacuum
) {
1283 uint32_t hash
= ctdb_hash(&key
);
1284 struct delete_record_data
*kd
;
1286 * check if we can ignore this record because it's in the delete_list
1288 kd
= (struct delete_record_data
*)trbt_lookup32(vdata
->delete_list
, hash
);
1290 * there might be hash collisions so we have to compare the keys here to be sure
1292 if (kd
&& kd
->key
.dsize
== key
.dsize
&& memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) == 0) {
1293 struct ctdb_ltdb_header
*hdr
= (struct ctdb_ltdb_header
*)data
.dptr
;
1295 * we have to check if the record hasn't changed in the meantime in order to
1296 * savely remove it from the database
1298 if (data
.dsize
== sizeof(struct ctdb_ltdb_header
) &&
1299 hdr
->dmaster
== kd
->ctdb
->pnn
&&
1300 ctdb_lmaster(kd
->ctdb
, &(kd
->key
)) == kd
->ctdb
->pnn
&&
1301 kd
->hdr
.rsn
== hdr
->rsn
) {
1307 if (tdb_store(vdata
->dest_db
, key
, data
, TDB_INSERT
) != 0) {
1308 vdata
->traverse_error
= true;
1318 static int ctdb_repack_tdb(struct tdb_context
*tdb
, TALLOC_CTX
*mem_ctx
, struct vacuum_data
*vdata
)
1320 struct tdb_context
*tmp_db
;
1322 if (tdb_transaction_start(tdb
) != 0) {
1323 DEBUG(DEBUG_ERR
,(__location__
" Failed to start transaction\n"));
1327 tmp_db
= tdb_open("tmpdb", tdb_hash_size(tdb
),
1328 TDB_INTERNAL
|TDB_DISALLOW_NESTING
,
1330 if (tmp_db
== NULL
) {
1331 DEBUG(DEBUG_ERR
,(__location__
" Failed to create tmp_db\n"));
1332 tdb_transaction_cancel(tdb
);
1336 vdata
->traverse_error
= false;
1337 vdata
->dest_db
= tmp_db
;
1338 vdata
->vacuum
= true;
1339 vdata
->vacuumed
= 0;
1343 * repack and vacuum on-the-fly by not writing the records that are
1346 if (tdb_traverse_read(tdb
, repack_traverse
, vdata
) == -1) {
1347 DEBUG(DEBUG_ERR
,(__location__
" Failed to traverse copying out\n"));
1348 tdb_transaction_cancel(tdb
);
1353 DEBUG(DEBUG_INFO
,(__location__
" %u records vacuumed\n", vdata
->vacuumed
));
1355 if (vdata
->traverse_error
) {
1356 DEBUG(DEBUG_ERR
,(__location__
" Error during traversal\n"));
1357 tdb_transaction_cancel(tdb
);
1362 if (tdb_wipe_all(tdb
) != 0) {
1363 DEBUG(DEBUG_ERR
,(__location__
" Failed to wipe database\n"));
1364 tdb_transaction_cancel(tdb
);
1369 vdata
->traverse_error
= false;
1370 vdata
->dest_db
= tdb
;
1371 vdata
->vacuum
= false;
1374 if (tdb_traverse_read(tmp_db
, repack_traverse
, vdata
) == -1) {
1375 DEBUG(DEBUG_ERR
,(__location__
" Failed to traverse copying back\n"));
1376 tdb_transaction_cancel(tdb
);
1381 if (vdata
->traverse_error
) {
1382 DEBUG(DEBUG_ERR
,(__location__
" Error during second traversal\n"));
1383 tdb_transaction_cancel(tdb
);
1391 if (tdb_transaction_commit(tdb
) != 0) {
1392 DEBUG(DEBUG_ERR
,(__location__
" Failed to commit\n"));
1395 DEBUG(DEBUG_INFO
,(__location__
" %u records copied\n", vdata
->copied
));
1401 * repack and vaccum a db
1402 * called from the child context
1404 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context
*ctdb_db
,
1405 TALLOC_CTX
*mem_ctx
,
1406 bool full_vacuum_run
)
1408 uint32_t repack_limit
= ctdb_db
->ctdb
->tunable
.repack_limit
;
1409 uint32_t vacuum_limit
= ctdb_db
->ctdb
->tunable
.vacuum_limit
;
1410 const char *name
= ctdb_db
->db_name
;
1411 int freelist_size
= 0;
1412 struct vacuum_data
*vdata
;
1414 vdata
= talloc_zero(mem_ctx
, struct vacuum_data
);
1415 if (vdata
== NULL
) {
1416 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1420 vdata
->ctdb
= ctdb_db
->ctdb
;
1421 vdata
->vacuum_limit
= vacuum_limit
;
1422 vdata
->repack_limit
= repack_limit
;
1423 vdata
->delete_list
= trbt_create(vdata
, 0);
1424 vdata
->ctdb_db
= ctdb_db
;
1425 if (vdata
->delete_list
== NULL
) {
1426 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1431 vdata
->start
= timeval_current();
1434 * gather all records that can be deleted in vdata
1436 if (ctdb_vacuum_db(ctdb_db
, vdata
, full_vacuum_run
) != 0) {
1437 DEBUG(DEBUG_ERR
,(__location__
" Failed to vacuum '%s'\n", name
));
1440 if (repack_limit
!= 0) {
1441 freelist_size
= tdb_freelist_size(ctdb_db
->ltdb
->tdb
);
1442 if (freelist_size
== -1) {
1443 DEBUG(DEBUG_ERR
,(__location__
" Failed to get freelist size for '%s'\n", name
));
1450 * decide if a repack is necessary
1452 if ((repack_limit
== 0 || (uint32_t)freelist_size
< repack_limit
) &&
1453 vdata
->delete_left
< vacuum_limit
)
1459 DEBUG(DEBUG_INFO
,("Repacking %s with %u freelist entries and %u records to delete\n",
1460 name
, freelist_size
, vdata
->delete_left
));
1463 * repack and implicitely get rid of the records we can delete
1465 if (ctdb_repack_tdb(ctdb_db
->ltdb
->tdb
, mem_ctx
, vdata
) != 0) {
1466 DEBUG(DEBUG_ERR
,(__location__
" Failed to repack '%s'\n", name
));
1475 static uint32_t get_vacuum_interval(struct ctdb_db_context
*ctdb_db
)
1477 uint32_t interval
= ctdb_db
->ctdb
->tunable
.vacuum_interval
;
1482 static int vacuum_child_destructor(struct ctdb_vacuum_child_context
*child_ctx
)
1484 double l
= timeval_elapsed(&child_ctx
->start_time
);
1485 struct ctdb_db_context
*ctdb_db
= child_ctx
->vacuum_handle
->ctdb_db
;
1486 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1488 DEBUG(DEBUG_INFO
,("Vacuuming took %.3f seconds for database %s\n", l
, ctdb_db
->db_name
));
1490 if (child_ctx
->child_pid
!= -1) {
1491 ctdb_kill(ctdb
, child_ctx
->child_pid
, SIGKILL
);
1493 /* Bump the number of successful fast-path runs. */
1494 child_ctx
->vacuum_handle
->fast_path_count
++;
1497 DLIST_REMOVE(ctdb
->vacuumers
, child_ctx
);
1499 event_add_timed(ctdb
->ev
, child_ctx
->vacuum_handle
,
1500 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1501 ctdb_vacuum_event
, child_ctx
->vacuum_handle
);
1507 * this event is generated when a vacuum child process times out
1509 static void vacuum_child_timeout(struct event_context
*ev
, struct timed_event
*te
,
1510 struct timeval t
, void *private_data
)
1512 struct ctdb_vacuum_child_context
*child_ctx
= talloc_get_type(private_data
, struct ctdb_vacuum_child_context
);
1514 DEBUG(DEBUG_ERR
,("Vacuuming child process timed out for db %s\n", child_ctx
->vacuum_handle
->ctdb_db
->db_name
));
1516 child_ctx
->status
= VACUUM_TIMEOUT
;
1518 talloc_free(child_ctx
);
1523 * this event is generated when a vacuum child process has completed
1525 static void vacuum_child_handler(struct event_context
*ev
, struct fd_event
*fde
,
1526 uint16_t flags
, void *private_data
)
1528 struct ctdb_vacuum_child_context
*child_ctx
= talloc_get_type(private_data
, struct ctdb_vacuum_child_context
);
1532 DEBUG(DEBUG_INFO
,("Vacuuming child process %d finished for db %s\n", child_ctx
->child_pid
, child_ctx
->vacuum_handle
->ctdb_db
->db_name
));
1533 child_ctx
->child_pid
= -1;
1535 ret
= read(child_ctx
->fd
[0], &c
, 1);
1536 if (ret
!= 1 || c
!= 0) {
1537 child_ctx
->status
= VACUUM_ERROR
;
1538 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
));
1540 child_ctx
->status
= VACUUM_OK
;
1543 talloc_free(child_ctx
);
1547 * this event is called every time we need to start a new vacuum process
1550 ctdb_vacuum_event(struct event_context
*ev
, struct timed_event
*te
,
1551 struct timeval t
, void *private_data
)
1553 struct ctdb_vacuum_handle
*vacuum_handle
= talloc_get_type(private_data
, struct ctdb_vacuum_handle
);
1554 struct ctdb_db_context
*ctdb_db
= vacuum_handle
->ctdb_db
;
1555 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1556 struct ctdb_vacuum_child_context
*child_ctx
;
1557 struct tevent_fd
*fde
;
1560 /* we dont vacuum if we are in recovery mode, or db frozen */
1561 if (ctdb
->recovery_mode
== CTDB_RECOVERY_ACTIVE
||
1562 ctdb
->freeze_mode
[ctdb_db
->priority
] != CTDB_FREEZE_NONE
) {
1563 DEBUG(DEBUG_INFO
, ("Not vacuuming %s (%s)\n", ctdb_db
->db_name
,
1564 ctdb
->recovery_mode
== CTDB_RECOVERY_ACTIVE
? "in recovery"
1565 : ctdb
->freeze_mode
[ctdb_db
->priority
] == CTDB_FREEZE_PENDING
1568 event_add_timed(ctdb
->ev
, vacuum_handle
,
1569 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1570 ctdb_vacuum_event
, vacuum_handle
);
1574 child_ctx
= talloc(vacuum_handle
, struct ctdb_vacuum_child_context
);
1575 if (child_ctx
== NULL
) {
1576 DEBUG(DEBUG_CRIT
, (__location__
" Failed to allocate child context for vacuuming of %s\n", ctdb_db
->db_name
));
1577 ctdb_fatal(ctdb
, "Out of memory when crating vacuum child context. Shutting down\n");
1581 ret
= pipe(child_ctx
->fd
);
1583 talloc_free(child_ctx
);
1584 DEBUG(DEBUG_ERR
, ("Failed to create pipe for vacuum child process.\n"));
1585 event_add_timed(ctdb
->ev
, vacuum_handle
,
1586 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1587 ctdb_vacuum_event
, vacuum_handle
);
1591 if (vacuum_handle
->fast_path_count
> ctdb
->tunable
.vacuum_fast_path_count
) {
1592 vacuum_handle
->fast_path_count
= 0;
1595 child_ctx
->child_pid
= ctdb_fork(ctdb
);
1596 if (child_ctx
->child_pid
== (pid_t
)-1) {
1597 close(child_ctx
->fd
[0]);
1598 close(child_ctx
->fd
[1]);
1599 talloc_free(child_ctx
);
1600 DEBUG(DEBUG_ERR
, ("Failed to fork vacuum child process.\n"));
1601 event_add_timed(ctdb
->ev
, vacuum_handle
,
1602 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1603 ctdb_vacuum_event
, vacuum_handle
);
1608 if (child_ctx
->child_pid
== 0) {
1610 bool full_vacuum_run
= false;
1611 close(child_ctx
->fd
[0]);
1613 DEBUG(DEBUG_INFO
,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db
->db_name
));
1614 ctdb_set_process_name("ctdb_vacuum");
1615 if (switch_from_server_to_client(ctdb
, "vacuum-%s", ctdb_db
->db_name
) != 0) {
1616 DEBUG(DEBUG_CRIT
, (__location__
"ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1623 if ((ctdb
->tunable
.vacuum_fast_path_count
> 0) &&
1624 (vacuum_handle
->fast_path_count
== 0))
1626 full_vacuum_run
= true;
1628 cc
= ctdb_vacuum_and_repack_db(ctdb_db
, child_ctx
,
1631 write(child_ctx
->fd
[1], &cc
, 1);
1635 set_close_on_exec(child_ctx
->fd
[0]);
1636 close(child_ctx
->fd
[1]);
1638 child_ctx
->status
= VACUUM_RUNNING
;
1639 child_ctx
->start_time
= timeval_current();
1641 DLIST_ADD(ctdb
->vacuumers
, child_ctx
);
1642 talloc_set_destructor(child_ctx
, vacuum_child_destructor
);
1645 * Clear the fastpath vacuuming list in the parent.
1647 talloc_free(ctdb_db
->delete_queue
);
1648 ctdb_db
->delete_queue
= trbt_create(ctdb_db
, 0);
1649 if (ctdb_db
->delete_queue
== NULL
) {
1650 /* fatal here? ... */
1651 ctdb_fatal(ctdb
, "Out of memory when re-creating vacuum tree "
1652 "in parent context. Shutting down\n");
1655 event_add_timed(ctdb
->ev
, child_ctx
,
1656 timeval_current_ofs(ctdb
->tunable
.vacuum_max_run_time
, 0),
1657 vacuum_child_timeout
, child_ctx
);
1659 DEBUG(DEBUG_DEBUG
, (__location__
" Created PIPE FD:%d to child vacuum process\n", child_ctx
->fd
[0]));
1661 fde
= event_add_fd(ctdb
->ev
, child_ctx
, child_ctx
->fd
[0],
1662 EVENT_FD_READ
, vacuum_child_handler
, child_ctx
);
1663 tevent_fd_set_auto_close(fde
);
1665 vacuum_handle
->child_ctx
= child_ctx
;
1666 child_ctx
->vacuum_handle
= vacuum_handle
;
1669 void ctdb_stop_vacuuming(struct ctdb_context
*ctdb
)
1671 /* Simply free them all. */
1672 while (ctdb
->vacuumers
) {
1673 DEBUG(DEBUG_INFO
, ("Aborting vacuuming for %s (%i)\n",
1674 ctdb
->vacuumers
->vacuum_handle
->ctdb_db
->db_name
,
1675 (int)ctdb
->vacuumers
->child_pid
));
1676 /* vacuum_child_destructor kills it, removes from list */
1677 talloc_free(ctdb
->vacuumers
);
1681 /* this function initializes the vacuuming context for a database
1682 * starts the vacuuming events
1684 int ctdb_vacuum_init(struct ctdb_db_context
*ctdb_db
)
1686 if (ctdb_db
->persistent
!= 0) {
1687 DEBUG(DEBUG_ERR
,("Vacuuming is disabled for persistent database %s\n", ctdb_db
->db_name
));
1691 ctdb_db
->vacuum_handle
= talloc(ctdb_db
, struct ctdb_vacuum_handle
);
1692 CTDB_NO_MEMORY(ctdb_db
->ctdb
, ctdb_db
->vacuum_handle
);
1694 ctdb_db
->vacuum_handle
->ctdb_db
= ctdb_db
;
1695 ctdb_db
->vacuum_handle
->fast_path_count
= 0;
1697 event_add_timed(ctdb_db
->ctdb
->ev
, ctdb_db
->vacuum_handle
,
1698 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1699 ctdb_vacuum_event
, ctdb_db
->vacuum_handle
);
1704 static void remove_record_from_delete_queue(struct ctdb_db_context
*ctdb_db
,
1705 const struct ctdb_ltdb_header
*hdr
,
1708 struct delete_record_data
*kd
;
1711 hash
= (uint32_t)ctdb_hash(&key
);
1713 DEBUG(DEBUG_DEBUG
, (__location__
1714 " remove_record_from_delete_queue: "
1719 "migrated_with_data[%s]\n",
1720 ctdb_db
->db_name
, ctdb_db
->db_id
,
1722 ctdb_lmaster(ctdb_db
->ctdb
, &key
),
1723 hdr
->flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
? "yes" : "no"));
1725 kd
= (struct delete_record_data
*)trbt_lookup32(ctdb_db
->delete_queue
, hash
);
1727 DEBUG(DEBUG_DEBUG
, (__location__
1728 " remove_record_from_delete_queue: "
1729 "record not in queue (hash[0x%08x])\n.",
1734 if ((kd
->key
.dsize
!= key
.dsize
) ||
1735 (memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) != 0))
1737 DEBUG(DEBUG_DEBUG
, (__location__
1738 " remove_record_from_delete_queue: "
1739 "hash collision for key with hash[0x%08x] "
1740 "in db[%s] - skipping\n",
1741 hash
, ctdb_db
->db_name
));
1745 DEBUG(DEBUG_DEBUG
, (__location__
1746 " remove_record_from_delete_queue: "
1747 "removing key with hash[0x%08x]\n",
1756 * Insert a record into the ctdb_db context's delete queue,
1757 * handling hash collisions.
1759 static int insert_record_into_delete_queue(struct ctdb_db_context
*ctdb_db
,
1760 const struct ctdb_ltdb_header
*hdr
,
1763 struct delete_record_data
*kd
;
1767 hash
= (uint32_t)ctdb_hash(&key
);
1769 DEBUG(DEBUG_INFO
, (__location__
" schedule for deletion: db[%s] "
1773 "migrated_with_data[%s]\n",
1774 ctdb_db
->db_name
, ctdb_db
->db_id
,
1776 ctdb_lmaster(ctdb_db
->ctdb
, &key
),
1777 hdr
->flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
? "yes" : "no"));
1779 kd
= (struct delete_record_data
*)trbt_lookup32(ctdb_db
->delete_queue
, hash
);
1781 if ((kd
->key
.dsize
!= key
.dsize
) ||
1782 (memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) != 0))
1785 (__location__
" schedule for deletion: "
1786 "hash collision for key hash [0x%08x]. "
1787 "Skipping the record.\n", hash
));
1791 (__location__
" schedule for deletion: "
1792 "updating entry for key with hash [0x%08x].\n",
1797 ret
= insert_delete_record_data_into_tree(ctdb_db
->ctdb
, ctdb_db
,
1798 ctdb_db
->delete_queue
,
1802 (__location__
" schedule for deletion: error "
1803 "inserting key with hash [0x%08x] into delete queue\n",
1812 * Schedule a record for deletetion.
1813 * Called from the parent context.
1815 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context
*ctdb
,
1818 struct ctdb_control_schedule_for_deletion
*dd
;
1819 struct ctdb_db_context
*ctdb_db
;
1823 dd
= (struct ctdb_control_schedule_for_deletion
*)indata
.dptr
;
1825 ctdb_db
= find_ctdb_db(ctdb
, dd
->db_id
);
1826 if (ctdb_db
== NULL
) {
1827 DEBUG(DEBUG_ERR
, (__location__
" Unknown db id 0x%08x\n",
1832 key
.dsize
= dd
->keylen
;
1835 ret
= insert_record_into_delete_queue(ctdb_db
, &dd
->hdr
, key
);
1840 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context
*ctdb_db
,
1841 const struct ctdb_ltdb_header
*hdr
,
1845 struct ctdb_control_schedule_for_deletion
*dd
;
1849 if (ctdb_db
->ctdb
->ctdbd_pid
== getpid()) {
1850 /* main daemon - directly queue */
1851 ret
= insert_record_into_delete_queue(ctdb_db
, hdr
, key
);
1856 /* if we dont have a connection to the daemon we can not send
1857 a control. For example sometimes from update_record control child
1860 if (!ctdb_db
->ctdb
->can_send_controls
) {
1865 /* child process: send the main daemon a control */
1866 indata
.dsize
= offsetof(struct ctdb_control_schedule_for_deletion
, key
) + key
.dsize
;
1867 indata
.dptr
= talloc_zero_array(ctdb_db
, uint8_t, indata
.dsize
);
1868 if (indata
.dptr
== NULL
) {
1869 DEBUG(DEBUG_ERR
, (__location__
" out of memory\n"));
1872 dd
= (struct ctdb_control_schedule_for_deletion
*)(void *)indata
.dptr
;
1873 dd
->db_id
= ctdb_db
->db_id
;
1875 dd
->keylen
= key
.dsize
;
1876 memcpy(dd
->key
, key
.dptr
, key
.dsize
);
1878 ret
= ctdb_control(ctdb_db
->ctdb
,
1881 CTDB_CONTROL_SCHEDULE_FOR_DELETION
,
1882 CTDB_CTRL_FLAG_NOREPLY
, /* flags */
1887 NULL
, /* timeout : NULL == wait forever */
1888 NULL
); /* error message */
1890 talloc_free(indata
.dptr
);
1892 if (ret
!= 0 || status
!= 0) {
1893 DEBUG(DEBUG_ERR
, (__location__
" Error sending "
1894 "SCHEDULE_FOR_DELETION "
1904 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context
*ctdb_db
,
1905 const struct ctdb_ltdb_header
*hdr
,
1908 if (ctdb_db
->ctdb
->ctdbd_pid
!= getpid()) {
1910 * Only remove the record from the delete queue if called
1911 * in the main daemon.
1916 remove_record_from_delete_queue(ctdb_db
, hdr
, key
);