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_added_to_vacuum_fetch_list
;
77 uint32_t full_added_to_delete_list
;
78 uint32_t full_skipped
;
82 uint32_t delete_remote_error
;
83 uint32_t delete_local_error
;
84 uint32_t delete_deleted
;
85 uint32_t delete_skipped
;
88 /* this structure contains the information for one record to be deleted */
89 struct delete_record_data
{
90 struct ctdb_context
*ctdb
;
91 struct ctdb_db_context
*ctdb_db
;
92 struct ctdb_ltdb_header hdr
;
97 struct delete_records_list
{
98 struct ctdb_marshall_buffer
*records
;
99 struct vacuum_data
*vdata
;
103 * Store key and header in a tree, indexed by the key hash.
105 static int insert_delete_record_data_into_tree(struct ctdb_context
*ctdb
,
106 struct ctdb_db_context
*ctdb_db
,
108 const struct ctdb_ltdb_header
*hdr
,
111 struct delete_record_data
*dd
;
115 len
= offsetof(struct delete_record_data
, keydata
) + key
.dsize
;
117 dd
= (struct delete_record_data
*)talloc_size(tree
, len
);
119 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
122 talloc_set_name_const(dd
, "struct delete_record_data");
125 dd
->ctdb_db
= ctdb_db
;
126 dd
->key
.dsize
= key
.dsize
;
127 dd
->key
.dptr
= dd
->keydata
;
128 memcpy(dd
->keydata
, key
.dptr
, key
.dsize
);
132 hash
= ctdb_hash(&key
);
134 trbt_insert32(tree
, hash
, dd
);
139 static int add_record_to_delete_list(struct vacuum_data
*vdata
, TDB_DATA key
,
140 struct ctdb_ltdb_header
*hdr
)
142 struct ctdb_context
*ctdb
= vdata
->ctdb
;
143 struct ctdb_db_context
*ctdb_db
= vdata
->ctdb_db
;
147 hash
= ctdb_hash(&key
);
149 if (trbt_lookup32(vdata
->delete_list
, hash
)) {
150 DEBUG(DEBUG_INFO
, (__location__
" Hash collision when vacuuming, skipping this record.\n"));
154 ret
= insert_delete_record_data_into_tree(ctdb
, ctdb_db
,
161 vdata
->delete_count
++;
167 * Add a record to the list of records to be sent
168 * to their lmaster with VACUUM_FETCH.
170 static int add_record_to_vacuum_fetch_list(struct vacuum_data
*vdata
,
173 struct ctdb_context
*ctdb
= vdata
->ctdb
;
174 struct ctdb_rec_data
*rec
;
177 struct ctdb_marshall_buffer
*vfl
;
179 lmaster
= ctdb_lmaster(ctdb
, &key
);
181 vfl
= vdata
->vacuum_fetch_list
[lmaster
];
183 rec
= ctdb_marshall_record(vfl
, ctdb
->pnn
, key
, NULL
, tdb_null
);
185 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
186 vdata
->traverse_error
= true;
190 old_size
= talloc_get_size(vfl
);
191 vfl
= talloc_realloc_size(NULL
, vfl
, old_size
+ rec
->length
);
193 DEBUG(DEBUG_ERR
,(__location__
" Failed to expand\n"));
194 vdata
->traverse_error
= true;
197 vdata
->vacuum_fetch_list
[lmaster
] = vfl
;
200 memcpy(old_size
+(uint8_t *)vfl
, rec
, rec
->length
);
209 static void ctdb_vacuum_event(struct event_context
*ev
, struct timed_event
*te
,
210 struct timeval t
, void *private_data
);
214 * traverse function for gathering the records that can be deleted
216 static int vacuum_traverse(struct tdb_context
*tdb
, TDB_DATA key
, TDB_DATA data
, void *private)
218 struct vacuum_data
*vdata
= talloc_get_type(private, struct vacuum_data
);
219 struct ctdb_context
*ctdb
= vdata
->ctdb
;
221 struct ctdb_ltdb_header
*hdr
;
226 lmaster
= ctdb_lmaster(ctdb
, &key
);
227 if (lmaster
>= ctdb
->num_nodes
) {
229 DEBUG(DEBUG_CRIT
, (__location__
230 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
233 (unsigned)ctdb
->num_nodes
,
234 (unsigned)ctdb_hash(&key
)));
238 if (data
.dsize
!= sizeof(struct ctdb_ltdb_header
)) {
239 /* it is not a deleted record */
240 vdata
->full_skipped
++;
244 hdr
= (struct ctdb_ltdb_header
*)data
.dptr
;
246 if (hdr
->dmaster
!= ctdb
->pnn
) {
247 vdata
->full_skipped
++;
251 if (lmaster
== ctdb
->pnn
) {
253 * We are both lmaster and dmaster, and the record is empty.
254 * So we should be able to delete it.
256 res
= add_record_to_delete_list(vdata
, key
, hdr
);
260 vdata
->full_added_to_delete_list
++;
264 * We are not lmaster.
265 * Add the record to the blob ready to send to the nodes.
267 res
= add_record_to_vacuum_fetch_list(vdata
, key
);
271 vdata
->full_added_to_vacuum_fetch_list
++;
279 * traverse the tree of records to delete and marshall them into
282 static int delete_marshall_traverse(void *param
, void *data
)
284 struct delete_record_data
*dd
= talloc_get_type(data
, struct delete_record_data
);
285 struct delete_records_list
*recs
= talloc_get_type(param
, struct delete_records_list
);
286 struct ctdb_rec_data
*rec
;
289 rec
= ctdb_marshall_record(dd
, recs
->records
->db_id
, dd
->key
, &dd
->hdr
, tdb_null
);
291 DEBUG(DEBUG_ERR
, (__location__
" failed to marshall record\n"));
295 old_size
= talloc_get_size(recs
->records
);
296 recs
->records
= talloc_realloc_size(NULL
, recs
->records
, old_size
+ rec
->length
);
297 if (recs
->records
== NULL
) {
298 DEBUG(DEBUG_ERR
,(__location__
" Failed to expand\n"));
301 recs
->records
->count
++;
302 memcpy(old_size
+(uint8_t *)(recs
->records
), rec
, rec
->length
);
307 * Variant of delete_marshall_traverse() that bumps the
308 * RSN of each traversed record in the database.
310 * This is needed to ensure that when rolling out our
311 * empty record copy before remote deletion, we as the
312 * record's dmaster keep a higher RSN than the non-dmaster
313 * nodes. This is needed to prevent old copies from
314 * resurrection in recoveries.
316 static int delete_marshall_traverse_first(void *param
, void *data
)
318 struct delete_record_data
*dd
= talloc_get_type(data
, struct delete_record_data
);
319 struct delete_records_list
*recs
= talloc_get_type(param
, struct delete_records_list
);
320 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
321 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
322 struct ctdb_ltdb_header
*header
;
323 TDB_DATA tdb_data
, ctdb_data
;
325 uint32_t hash
= ctdb_hash(&(dd
->key
));
328 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
331 (__location__
" Error getting chainlock on record with "
332 "key hash [0x%08x] on database db[%s].\n",
333 hash
, ctdb_db
->db_name
));
334 recs
->vdata
->delete_skipped
++;
340 * Verify that the record is still empty, its RSN has not
341 * changed and that we are still its lmaster and dmaster.
344 tdb_data
= tdb_fetch(ctdb_db
->ltdb
->tdb
, dd
->key
);
345 if (tdb_data
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
346 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
347 "on database db[%s] does not exist or is not"
348 " a ctdb-record. skipping.\n",
349 hash
, ctdb_db
->db_name
));
353 if (tdb_data
.dsize
> sizeof(struct ctdb_ltdb_header
)) {
354 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
355 "on database db[%s] has been recycled. "
357 hash
, ctdb_db
->db_name
));
361 header
= (struct ctdb_ltdb_header
*)tdb_data
.dptr
;
363 if (header
->flags
& CTDB_REC_RO_FLAGS
) {
364 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
365 "on database db[%s] has read-only flags. "
367 hash
, ctdb_db
->db_name
));
371 if (header
->dmaster
!= ctdb
->pnn
) {
372 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
373 "on database db[%s] has been migrated away. "
375 hash
, ctdb_db
->db_name
));
379 if (header
->rsn
!= dd
->hdr
.rsn
) {
380 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
381 "on database db[%s] seems to have been "
382 "migrated away and back again (with empty "
383 "data). skipping.\n",
384 hash
, ctdb_db
->db_name
));
388 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
390 if (lmaster
!= ctdb
->pnn
) {
391 DEBUG(DEBUG_INFO
, (__location__
": not lmaster for record in "
392 "delete list (key hash [0x%08x], db[%s]). "
393 "Strange! skipping.\n",
394 hash
, ctdb_db
->db_name
));
399 * Increment the record's RSN to ensure the dmaster (i.e. the current
400 * node) has the highest RSN of the record in the cluster.
401 * This is to prevent old record copies from resurrecting in recoveries
402 * if something should fail during the deletion process.
403 * Note that ctdb_ltdb_store_server() increments the RSN if called
404 * on the record's dmaster.
407 ctdb_data
.dptr
= tdb_data
.dptr
+ sizeof(struct ctdb_ltdb_header
);
408 ctdb_data
.dsize
= tdb_data
.dsize
- sizeof(struct ctdb_ltdb_header
);
410 res
= ctdb_ltdb_store(ctdb_db
, dd
->key
, header
, ctdb_data
);
412 DEBUG(DEBUG_ERR
, (__location__
": Failed to store record with "
413 "key hash [0x%08x] on database db[%s].\n",
414 hash
, ctdb_db
->db_name
));
418 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
423 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
425 recs
->vdata
->delete_skipped
++;
430 if (tdb_data
.dptr
!= NULL
) {
438 return delete_marshall_traverse(param
, data
);
442 * traverse function for the traversal of the delete_queue,
443 * the fast-path vacuuming list.
445 * - If the record has been migrated off the node
446 * or has been revived (filled with data) on the node,
447 * then skip the record.
449 * - If the current node is the record's lmaster and it is
450 * a record that has never been migrated with data, then
451 * delete the record from the local tdb.
453 * - If the current node is the record's lmaster and it has
454 * been migrated with data, then schedule it for the normal
455 * vacuuming procedure (i.e. add it to the delete_list).
457 * - If the current node is NOT the record's lmaster then
458 * add it to the list of records that are to be sent to
459 * the lmaster with the VACUUM_FETCH message.
461 static int delete_queue_traverse(void *param
, void *data
)
463 struct delete_record_data
*dd
=
464 talloc_get_type(data
, struct delete_record_data
);
465 struct vacuum_data
*vdata
= talloc_get_type(param
, struct vacuum_data
);
466 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
467 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
; /* or dd->ctdb ??? */
469 struct ctdb_ltdb_header
*header
;
472 uint32_t hash
= ctdb_hash(&(dd
->key
));
476 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
479 (__location__
" Error getting chainlock on record with "
480 "key hash [0x%08x] on database db[%s].\n",
481 hash
, ctdb_db
->db_name
));
486 tdb_data
= tdb_fetch(ctdb_db
->ltdb
->tdb
, dd
->key
);
487 if (tdb_data
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
488 /* Does not exist or not a ctdb record. Skip. */
492 if (tdb_data
.dsize
> sizeof(struct ctdb_ltdb_header
)) {
493 /* The record has been recycled (filled with data). Skip. */
497 header
= (struct ctdb_ltdb_header
*)tdb_data
.dptr
;
499 if (header
->dmaster
!= ctdb
->pnn
) {
500 /* The record has been migrated off the node. Skip. */
504 if (header
->rsn
!= dd
->hdr
.rsn
) {
506 * The record has been migrated off the node and back again.
507 * But not requeued for deletion. Skip it.
513 * We are dmaster, and the record has no data, and it has
514 * not been migrated after it has been queued for deletion.
516 * At this stage, the record could still have been revived locally
517 * and last been written with empty data. This can only be
518 * fixed with the addition of an active or delete flag. (TODO)
521 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
523 if (lmaster
!= ctdb
->pnn
) {
524 res
= add_record_to_vacuum_fetch_list(vdata
, dd
->key
);
528 (__location__
" Error adding record to list "
529 "of records to send to lmaster.\n"));
532 vdata
->fast_added_to_vacuum_fetch_list
++;
537 /* use header->flags or dd->hdr.flags ?? */
538 if (dd
->hdr
.flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
) {
539 res
= add_record_to_delete_list(vdata
, dd
->key
, &dd
->hdr
);
543 (__location__
" Error adding record to list "
544 "of records for deletion on lmaster.\n"));
547 vdata
->fast_added_to_delete_list
++;
550 res
= tdb_delete(ctdb_db
->ltdb
->tdb
, dd
->key
);
554 (__location__
" Error deleting record with key "
555 "hash [0x%08x] from local data base db[%s].\n",
556 hash
, ctdb_db
->db_name
));
562 (__location__
" Deleted record with key hash "
563 "[0x%08x] from local data base db[%s].\n",
564 hash
, ctdb_db
->db_name
));
565 vdata
->fast_deleted
++;
571 vdata
->fast_skipped
++;
576 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
582 * Delete the records that we are lmaster and dmaster for and
583 * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
586 static int delete_record_traverse(void *param
, void *data
)
588 struct delete_record_data
*dd
=
589 talloc_get_type(data
, struct delete_record_data
);
590 struct vacuum_data
*vdata
= talloc_get_type(param
, struct vacuum_data
);
591 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
592 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
594 struct ctdb_ltdb_header
*header
;
597 uint32_t hash
= ctdb_hash(&(dd
->key
));
599 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
602 (__location__
" Error getting chainlock on record with "
603 "key hash [0x%08x] on database db[%s].\n",
604 hash
, ctdb_db
->db_name
));
605 vdata
->delete_local_error
++;
606 vdata
->delete_left
--;
612 * Verify that the record is still empty, its RSN has not
613 * changed and that we are still its lmaster and dmaster.
616 tdb_data
= tdb_fetch(ctdb_db
->ltdb
->tdb
, dd
->key
);
617 if (tdb_data
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
618 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
619 "on database db[%s] does not exist or is not"
620 " a ctdb-record. skipping.\n",
621 hash
, ctdb_db
->db_name
));
625 if (tdb_data
.dsize
> sizeof(struct ctdb_ltdb_header
)) {
626 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
627 "on database db[%s] has been recycled. "
629 hash
, ctdb_db
->db_name
));
633 header
= (struct ctdb_ltdb_header
*)tdb_data
.dptr
;
635 if (header
->flags
& CTDB_REC_RO_FLAGS
) {
636 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
637 "on database db[%s] has read-only flags. "
639 hash
, ctdb_db
->db_name
));
643 if (header
->dmaster
!= ctdb
->pnn
) {
644 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
645 "on database db[%s] has been migrated away. "
647 hash
, ctdb_db
->db_name
));
651 if (header
->rsn
!= dd
->hdr
.rsn
+ 1) {
653 * The record has been migrated off the node and back again.
654 * But not requeued for deletion. Skip it.
655 * (Note that the first marshall traverse has bumped the RSN
658 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
659 "on database db[%s] seems to have been "
660 "migrated away and back again (with empty "
661 "data). skipping.\n",
662 hash
, ctdb_db
->db_name
));
666 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
668 if (lmaster
!= ctdb
->pnn
) {
669 DEBUG(DEBUG_INFO
, (__location__
": not lmaster for record in "
670 "delete list (key hash [0x%08x], db[%s]). "
671 "Strange! skipping.\n",
672 hash
, ctdb_db
->db_name
));
676 res
= tdb_delete(ctdb_db
->ltdb
->tdb
, dd
->key
);
680 (__location__
" Error deleting record with key hash "
681 "[0x%08x] from local data base db[%s].\n",
682 hash
, ctdb_db
->db_name
));
683 vdata
->delete_local_error
++;
688 (__location__
" Deleted record with key hash [0x%08x] from "
689 "local data base db[%s].\n", hash
, ctdb_db
->db_name
));
691 vdata
->delete_deleted
++;
695 vdata
->delete_skipped
++;
700 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
703 vdata
->delete_left
--;
709 * Fast vacuuming run:
710 * Traverse the delete_queue.
711 * This fills the same lists as the database traverse.
713 static void ctdb_vacuum_db_fast(struct ctdb_db_context
*ctdb_db
,
714 struct vacuum_data
*vdata
)
716 trbt_traversearray32(ctdb_db
->delete_queue
, 1, delete_queue_traverse
, vdata
);
718 if (vdata
->fast_total
> 0) {
721 " fast vacuuming delete_queue traverse statistics: "
730 (unsigned)vdata
->fast_total
,
731 (unsigned)vdata
->fast_deleted
,
732 (unsigned)vdata
->fast_skipped
,
733 (unsigned)vdata
->fast_error
,
734 (unsigned)vdata
->fast_added_to_delete_list
,
735 (unsigned)vdata
->fast_added_to_vacuum_fetch_list
));
743 * read-only traverse of the database, looking for records that
744 * might be able to be vacuumed.
746 * This is not done each time but only every tunable
747 * VacuumFastPathCount times.
749 static int ctdb_vacuum_db_full(struct ctdb_db_context
*ctdb_db
,
750 struct vacuum_data
*vdata
,
751 bool full_vacuum_run
)
755 if (!full_vacuum_run
) {
759 ret
= tdb_traverse_read(ctdb_db
->ltdb
->tdb
, vacuum_traverse
, vdata
);
760 if (ret
== -1 || vdata
->traverse_error
) {
761 DEBUG(DEBUG_ERR
, (__location__
" Traverse error in vacuuming "
762 "'%s'\n", ctdb_db
->db_name
));
766 if (vdata
->full_total
> 0) {
769 " full vacuuming db traverse statistics: "
777 (unsigned)vdata
->full_total
,
778 (unsigned)vdata
->full_skipped
,
779 (unsigned)vdata
->full_error
,
780 (unsigned)vdata
->full_added_to_delete_list
,
781 (unsigned)vdata
->full_added_to_vacuum_fetch_list
));
788 * Process the vacuum fetch lists:
789 * For records for which we are not the lmaster, tell the lmaster to
792 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context
*ctdb_db
,
793 struct vacuum_data
*vdata
)
796 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
798 for (i
= 0; i
< ctdb
->num_nodes
; i
++) {
800 struct ctdb_marshall_buffer
*vfl
= vdata
->vacuum_fetch_list
[i
];
802 if (ctdb
->nodes
[i
]->pnn
== ctdb
->pnn
) {
806 if (vfl
->count
== 0) {
810 DEBUG(DEBUG_INFO
, ("Found %u records for lmaster %u in '%s'\n",
811 vfl
->count
, ctdb
->nodes
[i
]->pnn
,
814 data
.dsize
= talloc_get_size(vfl
);
815 data
.dptr
= (void *)vfl
;
816 if (ctdb_client_send_message(ctdb
, ctdb
->nodes
[i
]->pnn
,
817 CTDB_SRVID_VACUUM_FETCH
,
820 DEBUG(DEBUG_ERR
, (__location__
" Failed to send vacuum "
821 "fetch message to %u\n",
822 ctdb
->nodes
[i
]->pnn
));
831 * Process the delete list:
833 * This is the last step of vacuuming that consistently deletes
834 * those records that have been migrated with data and can hence
835 * not be deleted when leaving a node.
837 * In this step, the lmaster does the final deletion of those empty
838 * records that it is also dmaster for. It has ususally received
839 * at least some of these records previously from the former dmasters
840 * with the vacuum fetch message.
842 * This last step is implemented as a 3-phase process to protect from
843 * races leading to data corruption:
845 * 1) Send the lmaster's copy to all other active nodes with the
846 * RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
847 * 2) Send the records that could successfully be stored remotely
848 * in step #1 to all active nodes with the TRY_DELETE_RECORDS
849 * control. The remote notes delete their local copy.
850 * 3) The lmaster locally deletes its copies of all records that
851 * could successfully be deleted remotely in step #2.
853 static int ctdb_process_delete_list(struct ctdb_db_context
*ctdb_db
,
854 struct vacuum_data
*vdata
)
857 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
858 struct delete_records_list
*recs
;
860 struct ctdb_node_map
*nodemap
;
861 uint32_t *active_nodes
;
862 int num_active_nodes
;
865 if (vdata
->delete_count
== 0) {
869 tmp_ctx
= talloc_new(vdata
);
870 if (tmp_ctx
== NULL
) {
871 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
875 vdata
->delete_left
= vdata
->delete_count
;
878 * get the list of currently active nodes
881 ret
= ctdb_ctrl_getnodemap(ctdb
, TIMELIMIT(),
886 DEBUG(DEBUG_ERR
,(__location__
" unable to get node map\n"));
891 active_nodes
= list_of_active_nodes(ctdb
, nodemap
,
892 nodemap
, /* talloc context */
893 false /* include self */);
895 num_active_nodes
= talloc_get_size(active_nodes
)/sizeof(*active_nodes
);
898 * Now delete the records all active nodes in a three-phase process:
899 * 1) send all active remote nodes the current empty copy with this
901 * 2) if all nodes could store the new copy,
902 * tell all the active remote nodes to delete all their copy
903 * 3) if all remote nodes deleted their record copy, delete it locally
908 * Send currently empty record copy to all active nodes for storing.
911 recs
= talloc_zero(tmp_ctx
, struct delete_records_list
);
913 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
917 recs
->records
= (struct ctdb_marshall_buffer
*)
918 talloc_zero_size(recs
,
919 offsetof(struct ctdb_marshall_buffer
, data
));
920 if (recs
->records
== NULL
) {
921 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
925 recs
->records
->db_id
= ctdb_db
->db_id
;
929 * traverse the tree of all records we want to delete and
930 * create a blob we can send to the other nodes.
932 * We call delete_marshall_traverse_first() to bump the
933 * records' RSNs in the database, to ensure we (as dmaster)
934 * keep the highest RSN of the records in the cluster.
936 trbt_traversearray32(vdata
->delete_list
, 1,
937 delete_marshall_traverse_first
, recs
);
939 indata
.dsize
= talloc_get_size(recs
->records
);
940 indata
.dptr
= (void *)recs
->records
;
942 for (i
= 0; i
< num_active_nodes
; i
++) {
943 struct ctdb_marshall_buffer
*records
;
944 struct ctdb_rec_data
*rec
;
948 ret
= ctdb_control(ctdb
, active_nodes
[i
], 0,
949 CTDB_CONTROL_RECEIVE_RECORDS
, 0,
950 indata
, recs
, &outdata
, &res
,
952 if (ret
!= 0 || res
!= 0) {
953 DEBUG(DEBUG_ERR
, ("Error storing record copies on "
954 "node %u: ret[%d] res[%d]\n",
955 active_nodes
[i
], ret
, res
));
961 * outdata contains the list of records coming back
962 * from the node: These are the records that the
963 * remote node could not store. We remove these from
964 * the list to process further.
966 records
= (struct ctdb_marshall_buffer
*)outdata
.dptr
;
967 rec
= (struct ctdb_rec_data
*)&records
->data
[0];
968 while (records
->count
-- > 1) {
969 TDB_DATA reckey
, recdata
;
970 struct ctdb_ltdb_header
*rechdr
;
971 struct delete_record_data
*dd
;
973 reckey
.dptr
= &rec
->data
[0];
974 reckey
.dsize
= rec
->keylen
;
975 recdata
.dptr
= &rec
->data
[reckey
.dsize
];
976 recdata
.dsize
= rec
->datalen
;
978 if (recdata
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
979 DEBUG(DEBUG_CRIT
,(__location__
" bad ltdb record\n"));
983 rechdr
= (struct ctdb_ltdb_header
*)recdata
.dptr
;
984 recdata
.dptr
+= sizeof(*rechdr
);
985 recdata
.dsize
-= sizeof(*rechdr
);
987 dd
= (struct delete_record_data
*)trbt_lookup32(
992 * The other node could not store the record
993 * copy and it is the first node that failed.
994 * So we should remove it from the tree and
998 vdata
->delete_remote_error
++;
999 vdata
->delete_left
--;
1002 rec
= (struct ctdb_rec_data
*)(rec
->length
+ (uint8_t *)rec
);
1006 if (vdata
->delete_left
== 0) {
1012 * Send the remaining records to all active nodes for deletion.
1014 * The lmaster's (i.e. our) copies of these records have been stored
1015 * successfully on the other nodes.
1019 * Create a marshall blob from the remaining list of records to delete.
1022 talloc_free(recs
->records
);
1024 recs
->records
= (struct ctdb_marshall_buffer
*)
1025 talloc_zero_size(recs
,
1026 offsetof(struct ctdb_marshall_buffer
, data
));
1027 if (recs
->records
== NULL
) {
1028 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1032 recs
->records
->db_id
= ctdb_db
->db_id
;
1034 trbt_traversearray32(vdata
->delete_list
, 1,
1035 delete_marshall_traverse
, recs
);
1037 indata
.dsize
= talloc_get_size(recs
->records
);
1038 indata
.dptr
= (void *)recs
->records
;
1040 for (i
= 0; i
< num_active_nodes
; i
++) {
1041 struct ctdb_marshall_buffer
*records
;
1042 struct ctdb_rec_data
*rec
;
1046 ret
= ctdb_control(ctdb
, active_nodes
[i
], 0,
1047 CTDB_CONTROL_TRY_DELETE_RECORDS
, 0,
1048 indata
, recs
, &outdata
, &res
,
1050 if (ret
!= 0 || res
!= 0) {
1051 DEBUG(DEBUG_ERR
, ("Failed to delete records on "
1052 "node %u: ret[%d] res[%d]\n",
1053 active_nodes
[i
], ret
, res
));
1059 * outdata contains the list of records coming back
1060 * from the node: These are the records that the
1061 * remote node could not delete. We remove these from
1062 * the list to delete locally.
1064 records
= (struct ctdb_marshall_buffer
*)outdata
.dptr
;
1065 rec
= (struct ctdb_rec_data
*)&records
->data
[0];
1066 while (records
->count
-- > 1) {
1067 TDB_DATA reckey
, recdata
;
1068 struct ctdb_ltdb_header
*rechdr
;
1069 struct delete_record_data
*dd
;
1071 reckey
.dptr
= &rec
->data
[0];
1072 reckey
.dsize
= rec
->keylen
;
1073 recdata
.dptr
= &rec
->data
[reckey
.dsize
];
1074 recdata
.dsize
= rec
->datalen
;
1076 if (recdata
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
1077 DEBUG(DEBUG_CRIT
,(__location__
" bad ltdb record\n"));
1081 rechdr
= (struct ctdb_ltdb_header
*)recdata
.dptr
;
1082 recdata
.dptr
+= sizeof(*rechdr
);
1083 recdata
.dsize
-= sizeof(*rechdr
);
1085 dd
= (struct delete_record_data
*)trbt_lookup32(
1087 ctdb_hash(&reckey
));
1090 * The other node could not delete the
1091 * record and it is the first node that
1092 * failed. So we should remove it from
1093 * the tree and update statistics.
1096 vdata
->delete_remote_error
++;
1097 vdata
->delete_left
--;
1100 rec
= (struct ctdb_rec_data
*)(rec
->length
+ (uint8_t *)rec
);
1104 if (vdata
->delete_left
== 0) {
1110 * Delete the remaining records locally.
1112 * These records have successfully been deleted on all
1113 * active remote nodes.
1116 trbt_traversearray32(vdata
->delete_list
, 1,
1117 delete_record_traverse
, vdata
);
1121 if (vdata
->delete_count
> 0) {
1124 " vacuum delete list statistics: "
1133 (unsigned)vdata
->delete_count
,
1134 (unsigned)vdata
->delete_deleted
,
1135 (unsigned)vdata
->delete_skipped
,
1136 (unsigned)vdata
->delete_remote_error
,
1137 (unsigned)vdata
->delete_local_error
,
1138 (unsigned)vdata
->delete_left
));
1144 talloc_free(tmp_ctx
);
1150 * initialize the vacuum_data
1152 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context
*ctdb_db
,
1153 struct vacuum_data
*vdata
)
1156 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1158 vdata
->fast_added_to_delete_list
= 0;
1159 vdata
->fast_added_to_vacuum_fetch_list
= 0;
1160 vdata
->fast_deleted
= 0;
1161 vdata
->fast_skipped
= 0;
1162 vdata
->fast_error
= 0;
1163 vdata
->fast_total
= 0;
1164 vdata
->full_added_to_delete_list
= 0;
1165 vdata
->full_added_to_vacuum_fetch_list
= 0;
1166 vdata
->full_skipped
= 0;
1167 vdata
->full_error
= 0;
1168 vdata
->full_total
= 0;
1169 vdata
->delete_count
= 0;
1170 vdata
->delete_left
= 0;
1171 vdata
->delete_remote_error
= 0;
1172 vdata
->delete_local_error
= 0;
1173 vdata
->delete_skipped
= 0;
1174 vdata
->delete_deleted
= 0;
1176 /* the list needs to be of length num_nodes */
1177 vdata
->vacuum_fetch_list
= talloc_zero_array(vdata
,
1178 struct ctdb_marshall_buffer
*,
1180 if (vdata
->vacuum_fetch_list
== NULL
) {
1181 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1184 for (i
= 0; i
< ctdb
->num_nodes
; i
++) {
1185 vdata
->vacuum_fetch_list
[i
] = (struct ctdb_marshall_buffer
*)
1186 talloc_zero_size(vdata
->vacuum_fetch_list
,
1187 offsetof(struct ctdb_marshall_buffer
, data
));
1188 if (vdata
->vacuum_fetch_list
[i
] == NULL
) {
1189 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1192 vdata
->vacuum_fetch_list
[i
]->db_id
= ctdb_db
->db_id
;
1200 * - Always do the fast vacuuming run, which traverses
1201 * the in-memory delete queue: these records have been
1202 * scheduled for deletion.
1203 * - Only if explicitly requested, the database is traversed
1204 * in order to use the traditional heuristics on empty records
1205 * to trigger deletion.
1206 * This is done only every VacuumFastPathCount'th vacuuming run.
1208 * The traverse runs fill two lists:
1210 * - The delete_list:
1211 * This is the list of empty records the current
1212 * node is lmaster and dmaster for. These records are later
1213 * deleted first on other nodes and then locally.
1215 * The fast vacuuming run has a short cut for those records
1216 * that have never been migrated with data: these records
1217 * are immediately deleted locally, since they have left
1218 * no trace on other nodes.
1220 * - The vacuum_fetch lists
1221 * (one for each other lmaster node):
1222 * The records in this list are sent for deletion to
1223 * their lmaster in a bulk VACUUM_FETCH message.
1225 * The lmaster then migrates all these records to itelf
1226 * so that they can be vacuumed there.
1228 * This executes in the child context.
1230 static int ctdb_vacuum_db(struct ctdb_db_context
*ctdb_db
,
1231 struct vacuum_data
*vdata
,
1232 bool full_vacuum_run
)
1234 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1237 DEBUG(DEBUG_INFO
, (__location__
" Entering %s vacuum run for db "
1238 "%s db_id[0x%08x]\n",
1239 full_vacuum_run
? "full" : "fast",
1240 ctdb_db
->db_name
, ctdb_db
->db_id
));
1242 ret
= ctdb_ctrl_getvnnmap(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
, ctdb
, &ctdb
->vnn_map
);
1244 DEBUG(DEBUG_ERR
, ("Unable to get vnnmap from local node\n"));
1248 pnn
= ctdb_ctrl_getpnn(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
);
1250 DEBUG(DEBUG_ERR
, ("Unable to get pnn from local node\n"));
1256 ret
= ctdb_vacuum_init_vacuum_data(ctdb_db
, vdata
);
1261 ctdb_vacuum_db_fast(ctdb_db
, vdata
);
1263 ret
= ctdb_vacuum_db_full(ctdb_db
, vdata
, full_vacuum_run
);
1268 ret
= ctdb_process_vacuum_fetch_lists(ctdb_db
, vdata
);
1273 ret
= ctdb_process_delete_list(ctdb_db
, vdata
);
1278 /* this ensures we run our event queue */
1279 ctdb_ctrl_getpnn(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
);
1286 * traverse function for repacking
1288 static int repack_traverse(struct tdb_context
*tdb
, TDB_DATA key
, TDB_DATA data
, void *private)
1290 struct vacuum_data
*vdata
= (struct vacuum_data
*)private;
1292 if (vdata
->vacuum
) {
1293 uint32_t hash
= ctdb_hash(&key
);
1294 struct delete_record_data
*kd
;
1296 * check if we can ignore this record because it's in the delete_list
1298 kd
= (struct delete_record_data
*)trbt_lookup32(vdata
->delete_list
, hash
);
1300 * there might be hash collisions so we have to compare the keys here to be sure
1302 if (kd
&& kd
->key
.dsize
== key
.dsize
&& memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) == 0) {
1303 struct ctdb_ltdb_header
*hdr
= (struct ctdb_ltdb_header
*)data
.dptr
;
1305 * we have to check if the record hasn't changed in the meantime in order to
1306 * savely remove it from the database
1308 if (data
.dsize
== sizeof(struct ctdb_ltdb_header
) &&
1309 hdr
->dmaster
== kd
->ctdb
->pnn
&&
1310 ctdb_lmaster(kd
->ctdb
, &(kd
->key
)) == kd
->ctdb
->pnn
&&
1311 kd
->hdr
.rsn
== hdr
->rsn
) {
1317 if (tdb_store(vdata
->dest_db
, key
, data
, TDB_INSERT
) != 0) {
1318 vdata
->traverse_error
= true;
1328 static int ctdb_repack_tdb(struct tdb_context
*tdb
, TALLOC_CTX
*mem_ctx
, struct vacuum_data
*vdata
)
1330 struct tdb_context
*tmp_db
;
1332 if (tdb_transaction_start(tdb
) != 0) {
1333 DEBUG(DEBUG_ERR
,(__location__
" Failed to start transaction\n"));
1337 tmp_db
= tdb_open("tmpdb", tdb_hash_size(tdb
),
1338 TDB_INTERNAL
|TDB_DISALLOW_NESTING
,
1340 if (tmp_db
== NULL
) {
1341 DEBUG(DEBUG_ERR
,(__location__
" Failed to create tmp_db\n"));
1342 tdb_transaction_cancel(tdb
);
1346 vdata
->traverse_error
= false;
1347 vdata
->dest_db
= tmp_db
;
1348 vdata
->vacuum
= true;
1349 vdata
->vacuumed
= 0;
1353 * repack and vacuum on-the-fly by not writing the records that are
1356 if (tdb_traverse_read(tdb
, repack_traverse
, vdata
) == -1) {
1357 DEBUG(DEBUG_ERR
,(__location__
" Failed to traverse copying out\n"));
1358 tdb_transaction_cancel(tdb
);
1363 DEBUG(DEBUG_INFO
,(__location__
" %u records vacuumed\n", vdata
->vacuumed
));
1365 if (vdata
->traverse_error
) {
1366 DEBUG(DEBUG_ERR
,(__location__
" Error during traversal\n"));
1367 tdb_transaction_cancel(tdb
);
1372 if (tdb_wipe_all(tdb
) != 0) {
1373 DEBUG(DEBUG_ERR
,(__location__
" Failed to wipe database\n"));
1374 tdb_transaction_cancel(tdb
);
1379 vdata
->traverse_error
= false;
1380 vdata
->dest_db
= tdb
;
1381 vdata
->vacuum
= false;
1384 if (tdb_traverse_read(tmp_db
, repack_traverse
, vdata
) == -1) {
1385 DEBUG(DEBUG_ERR
,(__location__
" Failed to traverse copying back\n"));
1386 tdb_transaction_cancel(tdb
);
1391 if (vdata
->traverse_error
) {
1392 DEBUG(DEBUG_ERR
,(__location__
" Error during second traversal\n"));
1393 tdb_transaction_cancel(tdb
);
1401 if (tdb_transaction_commit(tdb
) != 0) {
1402 DEBUG(DEBUG_ERR
,(__location__
" Failed to commit\n"));
1405 DEBUG(DEBUG_INFO
,(__location__
" %u records copied\n", vdata
->copied
));
1411 * repack and vaccum a db
1412 * called from the child context
1414 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context
*ctdb_db
,
1415 TALLOC_CTX
*mem_ctx
,
1416 bool full_vacuum_run
)
1418 uint32_t repack_limit
= ctdb_db
->ctdb
->tunable
.repack_limit
;
1419 uint32_t vacuum_limit
= ctdb_db
->ctdb
->tunable
.vacuum_limit
;
1420 const char *name
= ctdb_db
->db_name
;
1421 int freelist_size
= 0;
1422 struct vacuum_data
*vdata
;
1424 vdata
= talloc_zero(mem_ctx
, struct vacuum_data
);
1425 if (vdata
== NULL
) {
1426 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1430 vdata
->ctdb
= ctdb_db
->ctdb
;
1431 vdata
->vacuum_limit
= vacuum_limit
;
1432 vdata
->repack_limit
= repack_limit
;
1433 vdata
->delete_list
= trbt_create(vdata
, 0);
1434 vdata
->ctdb_db
= ctdb_db
;
1435 if (vdata
->delete_list
== NULL
) {
1436 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1441 vdata
->start
= timeval_current();
1444 * gather all records that can be deleted in vdata
1446 if (ctdb_vacuum_db(ctdb_db
, vdata
, full_vacuum_run
) != 0) {
1447 DEBUG(DEBUG_ERR
,(__location__
" Failed to vacuum '%s'\n", name
));
1450 if (repack_limit
!= 0) {
1451 freelist_size
= tdb_freelist_size(ctdb_db
->ltdb
->tdb
);
1452 if (freelist_size
== -1) {
1453 DEBUG(DEBUG_ERR
,(__location__
" Failed to get freelist size for '%s'\n", name
));
1460 * decide if a repack is necessary
1462 if ((repack_limit
== 0 || (uint32_t)freelist_size
< repack_limit
) &&
1463 vdata
->delete_left
< vacuum_limit
)
1469 DEBUG(DEBUG_INFO
,("Repacking %s with %u freelist entries and %u records to delete\n",
1470 name
, freelist_size
, vdata
->delete_left
));
1473 * repack and implicitely get rid of the records we can delete
1475 if (ctdb_repack_tdb(ctdb_db
->ltdb
->tdb
, mem_ctx
, vdata
) != 0) {
1476 DEBUG(DEBUG_ERR
,(__location__
" Failed to repack '%s'\n", name
));
1485 static uint32_t get_vacuum_interval(struct ctdb_db_context
*ctdb_db
)
1487 uint32_t interval
= ctdb_db
->ctdb
->tunable
.vacuum_interval
;
1492 static int vacuum_child_destructor(struct ctdb_vacuum_child_context
*child_ctx
)
1494 double l
= timeval_elapsed(&child_ctx
->start_time
);
1495 struct ctdb_db_context
*ctdb_db
= child_ctx
->vacuum_handle
->ctdb_db
;
1496 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1498 DEBUG(DEBUG_INFO
,("Vacuuming took %.3f seconds for database %s\n", l
, ctdb_db
->db_name
));
1500 if (child_ctx
->child_pid
!= -1) {
1501 ctdb_kill(ctdb
, child_ctx
->child_pid
, SIGKILL
);
1503 /* Bump the number of successful fast-path runs. */
1504 child_ctx
->vacuum_handle
->fast_path_count
++;
1507 DLIST_REMOVE(ctdb
->vacuumers
, child_ctx
);
1509 event_add_timed(ctdb
->ev
, child_ctx
->vacuum_handle
,
1510 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1511 ctdb_vacuum_event
, child_ctx
->vacuum_handle
);
1517 * this event is generated when a vacuum child process times out
1519 static void vacuum_child_timeout(struct event_context
*ev
, struct timed_event
*te
,
1520 struct timeval t
, void *private_data
)
1522 struct ctdb_vacuum_child_context
*child_ctx
= talloc_get_type(private_data
, struct ctdb_vacuum_child_context
);
1524 DEBUG(DEBUG_ERR
,("Vacuuming child process timed out for db %s\n", child_ctx
->vacuum_handle
->ctdb_db
->db_name
));
1526 child_ctx
->status
= VACUUM_TIMEOUT
;
1528 talloc_free(child_ctx
);
1533 * this event is generated when a vacuum child process has completed
1535 static void vacuum_child_handler(struct event_context
*ev
, struct fd_event
*fde
,
1536 uint16_t flags
, void *private_data
)
1538 struct ctdb_vacuum_child_context
*child_ctx
= talloc_get_type(private_data
, struct ctdb_vacuum_child_context
);
1542 DEBUG(DEBUG_INFO
,("Vacuuming child process %d finished for db %s\n", child_ctx
->child_pid
, child_ctx
->vacuum_handle
->ctdb_db
->db_name
));
1543 child_ctx
->child_pid
= -1;
1545 ret
= read(child_ctx
->fd
[0], &c
, 1);
1546 if (ret
!= 1 || c
!= 0) {
1547 child_ctx
->status
= VACUUM_ERROR
;
1548 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
));
1550 child_ctx
->status
= VACUUM_OK
;
1553 talloc_free(child_ctx
);
1557 * this event is called every time we need to start a new vacuum process
1560 ctdb_vacuum_event(struct event_context
*ev
, struct timed_event
*te
,
1561 struct timeval t
, void *private_data
)
1563 struct ctdb_vacuum_handle
*vacuum_handle
= talloc_get_type(private_data
, struct ctdb_vacuum_handle
);
1564 struct ctdb_db_context
*ctdb_db
= vacuum_handle
->ctdb_db
;
1565 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1566 struct ctdb_vacuum_child_context
*child_ctx
;
1567 struct tevent_fd
*fde
;
1570 /* we dont vacuum if we are in recovery mode, or db frozen */
1571 if (ctdb
->recovery_mode
== CTDB_RECOVERY_ACTIVE
||
1572 ctdb
->freeze_mode
[ctdb_db
->priority
] != CTDB_FREEZE_NONE
) {
1573 DEBUG(DEBUG_INFO
, ("Not vacuuming %s (%s)\n", ctdb_db
->db_name
,
1574 ctdb
->recovery_mode
== CTDB_RECOVERY_ACTIVE
? "in recovery"
1575 : ctdb
->freeze_mode
[ctdb_db
->priority
] == CTDB_FREEZE_PENDING
1578 event_add_timed(ctdb
->ev
, vacuum_handle
,
1579 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1580 ctdb_vacuum_event
, vacuum_handle
);
1584 child_ctx
= talloc(vacuum_handle
, struct ctdb_vacuum_child_context
);
1585 if (child_ctx
== NULL
) {
1586 DEBUG(DEBUG_CRIT
, (__location__
" Failed to allocate child context for vacuuming of %s\n", ctdb_db
->db_name
));
1587 ctdb_fatal(ctdb
, "Out of memory when crating vacuum child context. Shutting down\n");
1591 ret
= pipe(child_ctx
->fd
);
1593 talloc_free(child_ctx
);
1594 DEBUG(DEBUG_ERR
, ("Failed to create pipe for vacuum child process.\n"));
1595 event_add_timed(ctdb
->ev
, vacuum_handle
,
1596 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1597 ctdb_vacuum_event
, vacuum_handle
);
1601 if (vacuum_handle
->fast_path_count
> ctdb
->tunable
.vacuum_fast_path_count
) {
1602 vacuum_handle
->fast_path_count
= 0;
1605 child_ctx
->child_pid
= ctdb_fork(ctdb
);
1606 if (child_ctx
->child_pid
== (pid_t
)-1) {
1607 close(child_ctx
->fd
[0]);
1608 close(child_ctx
->fd
[1]);
1609 talloc_free(child_ctx
);
1610 DEBUG(DEBUG_ERR
, ("Failed to fork vacuum child process.\n"));
1611 event_add_timed(ctdb
->ev
, vacuum_handle
,
1612 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1613 ctdb_vacuum_event
, vacuum_handle
);
1618 if (child_ctx
->child_pid
== 0) {
1620 bool full_vacuum_run
= false;
1621 close(child_ctx
->fd
[0]);
1623 DEBUG(DEBUG_INFO
,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db
->db_name
));
1624 ctdb_set_process_name("ctdb_vacuum");
1625 if (switch_from_server_to_client(ctdb
, "vacuum-%s", ctdb_db
->db_name
) != 0) {
1626 DEBUG(DEBUG_CRIT
, (__location__
"ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1633 if ((ctdb
->tunable
.vacuum_fast_path_count
> 0) &&
1634 (vacuum_handle
->fast_path_count
== 0))
1636 full_vacuum_run
= true;
1638 cc
= ctdb_vacuum_and_repack_db(ctdb_db
, child_ctx
,
1641 write(child_ctx
->fd
[1], &cc
, 1);
1645 set_close_on_exec(child_ctx
->fd
[0]);
1646 close(child_ctx
->fd
[1]);
1648 child_ctx
->status
= VACUUM_RUNNING
;
1649 child_ctx
->start_time
= timeval_current();
1651 DLIST_ADD(ctdb
->vacuumers
, child_ctx
);
1652 talloc_set_destructor(child_ctx
, vacuum_child_destructor
);
1655 * Clear the fastpath vacuuming list in the parent.
1657 talloc_free(ctdb_db
->delete_queue
);
1658 ctdb_db
->delete_queue
= trbt_create(ctdb_db
, 0);
1659 if (ctdb_db
->delete_queue
== NULL
) {
1660 /* fatal here? ... */
1661 ctdb_fatal(ctdb
, "Out of memory when re-creating vacuum tree "
1662 "in parent context. Shutting down\n");
1665 event_add_timed(ctdb
->ev
, child_ctx
,
1666 timeval_current_ofs(ctdb
->tunable
.vacuum_max_run_time
, 0),
1667 vacuum_child_timeout
, child_ctx
);
1669 DEBUG(DEBUG_DEBUG
, (__location__
" Created PIPE FD:%d to child vacuum process\n", child_ctx
->fd
[0]));
1671 fde
= event_add_fd(ctdb
->ev
, child_ctx
, child_ctx
->fd
[0],
1672 EVENT_FD_READ
, vacuum_child_handler
, child_ctx
);
1673 tevent_fd_set_auto_close(fde
);
1675 vacuum_handle
->child_ctx
= child_ctx
;
1676 child_ctx
->vacuum_handle
= vacuum_handle
;
1679 void ctdb_stop_vacuuming(struct ctdb_context
*ctdb
)
1681 /* Simply free them all. */
1682 while (ctdb
->vacuumers
) {
1683 DEBUG(DEBUG_INFO
, ("Aborting vacuuming for %s (%i)\n",
1684 ctdb
->vacuumers
->vacuum_handle
->ctdb_db
->db_name
,
1685 (int)ctdb
->vacuumers
->child_pid
));
1686 /* vacuum_child_destructor kills it, removes from list */
1687 talloc_free(ctdb
->vacuumers
);
1691 /* this function initializes the vacuuming context for a database
1692 * starts the vacuuming events
1694 int ctdb_vacuum_init(struct ctdb_db_context
*ctdb_db
)
1696 if (ctdb_db
->persistent
!= 0) {
1697 DEBUG(DEBUG_ERR
,("Vacuuming is disabled for persistent database %s\n", ctdb_db
->db_name
));
1701 ctdb_db
->vacuum_handle
= talloc(ctdb_db
, struct ctdb_vacuum_handle
);
1702 CTDB_NO_MEMORY(ctdb_db
->ctdb
, ctdb_db
->vacuum_handle
);
1704 ctdb_db
->vacuum_handle
->ctdb_db
= ctdb_db
;
1705 ctdb_db
->vacuum_handle
->fast_path_count
= 0;
1707 event_add_timed(ctdb_db
->ctdb
->ev
, ctdb_db
->vacuum_handle
,
1708 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1709 ctdb_vacuum_event
, ctdb_db
->vacuum_handle
);
1714 static void remove_record_from_delete_queue(struct ctdb_db_context
*ctdb_db
,
1715 const struct ctdb_ltdb_header
*hdr
,
1718 struct delete_record_data
*kd
;
1721 hash
= (uint32_t)ctdb_hash(&key
);
1723 DEBUG(DEBUG_DEBUG
, (__location__
1724 " remove_record_from_delete_queue: "
1729 "migrated_with_data[%s]\n",
1730 ctdb_db
->db_name
, ctdb_db
->db_id
,
1732 ctdb_lmaster(ctdb_db
->ctdb
, &key
),
1733 hdr
->flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
? "yes" : "no"));
1735 kd
= (struct delete_record_data
*)trbt_lookup32(ctdb_db
->delete_queue
, hash
);
1737 DEBUG(DEBUG_DEBUG
, (__location__
1738 " remove_record_from_delete_queue: "
1739 "record not in queue (hash[0x%08x])\n.",
1744 if ((kd
->key
.dsize
!= key
.dsize
) ||
1745 (memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) != 0))
1747 DEBUG(DEBUG_DEBUG
, (__location__
1748 " remove_record_from_delete_queue: "
1749 "hash collision for key with hash[0x%08x] "
1750 "in db[%s] - skipping\n",
1751 hash
, ctdb_db
->db_name
));
1755 DEBUG(DEBUG_DEBUG
, (__location__
1756 " remove_record_from_delete_queue: "
1757 "removing key with hash[0x%08x]\n",
1766 * Insert a record into the ctdb_db context's delete queue,
1767 * handling hash collisions.
1769 static int insert_record_into_delete_queue(struct ctdb_db_context
*ctdb_db
,
1770 const struct ctdb_ltdb_header
*hdr
,
1773 struct delete_record_data
*kd
;
1777 hash
= (uint32_t)ctdb_hash(&key
);
1779 DEBUG(DEBUG_INFO
, (__location__
" schedule for deletion: db[%s] "
1783 "migrated_with_data[%s]\n",
1784 ctdb_db
->db_name
, ctdb_db
->db_id
,
1786 ctdb_lmaster(ctdb_db
->ctdb
, &key
),
1787 hdr
->flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
? "yes" : "no"));
1789 kd
= (struct delete_record_data
*)trbt_lookup32(ctdb_db
->delete_queue
, hash
);
1791 if ((kd
->key
.dsize
!= key
.dsize
) ||
1792 (memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) != 0))
1795 (__location__
" schedule for deletion: "
1796 "hash collision for key hash [0x%08x]. "
1797 "Skipping the record.\n", hash
));
1801 (__location__
" schedule for deletion: "
1802 "updating entry for key with hash [0x%08x].\n",
1807 ret
= insert_delete_record_data_into_tree(ctdb_db
->ctdb
, ctdb_db
,
1808 ctdb_db
->delete_queue
,
1812 (__location__
" schedule for deletion: error "
1813 "inserting key with hash [0x%08x] into delete queue\n",
1822 * Schedule a record for deletetion.
1823 * Called from the parent context.
1825 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context
*ctdb
,
1828 struct ctdb_control_schedule_for_deletion
*dd
;
1829 struct ctdb_db_context
*ctdb_db
;
1833 dd
= (struct ctdb_control_schedule_for_deletion
*)indata
.dptr
;
1835 ctdb_db
= find_ctdb_db(ctdb
, dd
->db_id
);
1836 if (ctdb_db
== NULL
) {
1837 DEBUG(DEBUG_ERR
, (__location__
" Unknown db id 0x%08x\n",
1842 key
.dsize
= dd
->keylen
;
1845 ret
= insert_record_into_delete_queue(ctdb_db
, &dd
->hdr
, key
);
1850 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context
*ctdb_db
,
1851 const struct ctdb_ltdb_header
*hdr
,
1855 struct ctdb_control_schedule_for_deletion
*dd
;
1859 if (ctdb_db
->ctdb
->ctdbd_pid
== getpid()) {
1860 /* main daemon - directly queue */
1861 ret
= insert_record_into_delete_queue(ctdb_db
, hdr
, key
);
1866 /* if we dont have a connection to the daemon we can not send
1867 a control. For example sometimes from update_record control child
1870 if (!ctdb_db
->ctdb
->can_send_controls
) {
1875 /* child process: send the main daemon a control */
1876 indata
.dsize
= offsetof(struct ctdb_control_schedule_for_deletion
, key
) + key
.dsize
;
1877 indata
.dptr
= talloc_zero_array(ctdb_db
, uint8_t, indata
.dsize
);
1878 if (indata
.dptr
== NULL
) {
1879 DEBUG(DEBUG_ERR
, (__location__
" out of memory\n"));
1882 dd
= (struct ctdb_control_schedule_for_deletion
*)(void *)indata
.dptr
;
1883 dd
->db_id
= ctdb_db
->db_id
;
1885 dd
->keylen
= key
.dsize
;
1886 memcpy(dd
->key
, key
.dptr
, key
.dsize
);
1888 ret
= ctdb_control(ctdb_db
->ctdb
,
1891 CTDB_CONTROL_SCHEDULE_FOR_DELETION
,
1892 CTDB_CTRL_FLAG_NOREPLY
, /* flags */
1897 NULL
, /* timeout : NULL == wait forever */
1898 NULL
); /* error message */
1900 talloc_free(indata
.dptr
);
1902 if (ret
!= 0 || status
!= 0) {
1903 DEBUG(DEBUG_ERR
, (__location__
" Error sending "
1904 "SCHEDULE_FOR_DELETION "
1914 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context
*ctdb_db
,
1915 const struct ctdb_ltdb_header
*hdr
,
1918 if (ctdb_db
->ctdb
->ctdbd_pid
!= getpid()) {
1920 * Only remove the record from the delete queue if called
1921 * in the main daemon.
1926 remove_record_from_delete_queue(ctdb_db
, hdr
, key
);