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 repack_limit
;
57 struct ctdb_context
*ctdb
;
58 struct ctdb_db_context
*ctdb_db
;
59 struct tdb_context
*dest_db
;
60 trbt_tree_t
*delete_list
;
61 uint32_t delete_count
;
62 struct ctdb_marshall_buffer
**vacuum_fetch_list
;
68 uint32_t fast_added_to_vacuum_fetch_list
;
69 uint32_t fast_added_to_delete_list
;
70 uint32_t fast_deleted
;
71 uint32_t fast_skipped
;
74 uint32_t full_scheduled
;
75 uint32_t full_skipped
;
79 uint32_t delete_remote_error
;
80 uint32_t delete_local_error
;
81 uint32_t delete_deleted
;
82 uint32_t delete_skipped
;
85 /* this structure contains the information for one record to be deleted */
86 struct delete_record_data
{
87 struct ctdb_context
*ctdb
;
88 struct ctdb_db_context
*ctdb_db
;
89 struct ctdb_ltdb_header hdr
;
94 struct delete_records_list
{
95 struct ctdb_marshall_buffer
*records
;
96 struct vacuum_data
*vdata
;
99 static int insert_record_into_delete_queue(struct ctdb_db_context
*ctdb_db
,
100 const struct ctdb_ltdb_header
*hdr
,
104 * Store key and header in a tree, indexed by the key hash.
106 static int insert_delete_record_data_into_tree(struct ctdb_context
*ctdb
,
107 struct ctdb_db_context
*ctdb_db
,
109 const struct ctdb_ltdb_header
*hdr
,
112 struct delete_record_data
*dd
;
116 len
= offsetof(struct delete_record_data
, keydata
) + key
.dsize
;
118 dd
= (struct delete_record_data
*)talloc_size(tree
, len
);
120 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
123 talloc_set_name_const(dd
, "struct delete_record_data");
126 dd
->ctdb_db
= ctdb_db
;
127 dd
->key
.dsize
= key
.dsize
;
128 dd
->key
.dptr
= dd
->keydata
;
129 memcpy(dd
->keydata
, key
.dptr
, key
.dsize
);
133 hash
= ctdb_hash(&key
);
135 trbt_insert32(tree
, hash
, dd
);
140 static int add_record_to_delete_list(struct vacuum_data
*vdata
, TDB_DATA key
,
141 struct ctdb_ltdb_header
*hdr
)
143 struct ctdb_context
*ctdb
= vdata
->ctdb
;
144 struct ctdb_db_context
*ctdb_db
= vdata
->ctdb_db
;
148 hash
= ctdb_hash(&key
);
150 if (trbt_lookup32(vdata
->delete_list
, hash
)) {
151 DEBUG(DEBUG_INFO
, (__location__
" Hash collision when vacuuming, skipping this record.\n"));
155 ret
= insert_delete_record_data_into_tree(ctdb
, ctdb_db
,
162 vdata
->delete_count
++;
168 * Add a record to the list of records to be sent
169 * to their lmaster with VACUUM_FETCH.
171 static int add_record_to_vacuum_fetch_list(struct vacuum_data
*vdata
,
174 struct ctdb_context
*ctdb
= vdata
->ctdb
;
175 struct ctdb_rec_data
*rec
;
178 struct ctdb_marshall_buffer
*vfl
;
180 lmaster
= ctdb_lmaster(ctdb
, &key
);
182 vfl
= vdata
->vacuum_fetch_list
[lmaster
];
184 rec
= ctdb_marshall_record(vfl
, ctdb
->pnn
, key
, NULL
, tdb_null
);
186 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
187 vdata
->traverse_error
= true;
191 old_size
= talloc_get_size(vfl
);
192 vfl
= talloc_realloc_size(NULL
, vfl
, old_size
+ rec
->length
);
194 DEBUG(DEBUG_ERR
,(__location__
" Failed to expand\n"));
195 vdata
->traverse_error
= true;
198 vdata
->vacuum_fetch_list
[lmaster
] = vfl
;
201 memcpy(old_size
+(uint8_t *)vfl
, rec
, rec
->length
);
208 static void ctdb_vacuum_event(struct event_context
*ev
, struct timed_event
*te
,
209 struct timeval t
, void *private_data
);
211 static int vacuum_record_parser(TDB_DATA key
, TDB_DATA data
, void *private_data
)
213 struct ctdb_ltdb_header
*header
=
214 (struct ctdb_ltdb_header
*)private_data
;
216 if (data
.dsize
!= sizeof(struct ctdb_ltdb_header
)) {
220 *header
= *(struct ctdb_ltdb_header
*)data
.dptr
;
226 * traverse function for gathering the records that can be deleted
228 static int vacuum_traverse(struct tdb_context
*tdb
, TDB_DATA key
, TDB_DATA data
,
231 struct vacuum_data
*vdata
= talloc_get_type(private_data
,
233 struct ctdb_context
*ctdb
= vdata
->ctdb
;
234 struct ctdb_db_context
*ctdb_db
= vdata
->ctdb_db
;
236 struct ctdb_ltdb_header
*hdr
;
241 lmaster
= ctdb_lmaster(ctdb
, &key
);
242 if (lmaster
>= ctdb
->num_nodes
) {
244 DEBUG(DEBUG_CRIT
, (__location__
245 " lmaster[%u] >= ctdb->num_nodes[%u] for key"
248 (unsigned)ctdb
->num_nodes
,
249 (unsigned)ctdb_hash(&key
)));
253 if (data
.dsize
!= sizeof(struct ctdb_ltdb_header
)) {
254 /* it is not a deleted record */
255 vdata
->full_skipped
++;
259 hdr
= (struct ctdb_ltdb_header
*)data
.dptr
;
261 if (hdr
->dmaster
!= ctdb
->pnn
) {
262 vdata
->full_skipped
++;
267 * Add the record to this process's delete_queue for processing
268 * in the subsequent traverse in the fast vacuum run.
270 res
= insert_record_into_delete_queue(ctdb_db
, hdr
, key
);
274 vdata
->full_scheduled
++;
281 * traverse the tree of records to delete and marshall them into
284 static int delete_marshall_traverse(void *param
, void *data
)
286 struct delete_record_data
*dd
= talloc_get_type(data
, struct delete_record_data
);
287 struct delete_records_list
*recs
= talloc_get_type(param
, struct delete_records_list
);
288 struct ctdb_rec_data
*rec
;
291 rec
= ctdb_marshall_record(dd
, recs
->records
->db_id
, dd
->key
, &dd
->hdr
, tdb_null
);
293 DEBUG(DEBUG_ERR
, (__location__
" failed to marshall record\n"));
297 old_size
= talloc_get_size(recs
->records
);
298 recs
->records
= talloc_realloc_size(NULL
, recs
->records
, old_size
+ rec
->length
);
299 if (recs
->records
== NULL
) {
300 DEBUG(DEBUG_ERR
,(__location__
" Failed to expand\n"));
303 recs
->records
->count
++;
304 memcpy(old_size
+(uint8_t *)(recs
->records
), rec
, rec
->length
);
309 * Variant of delete_marshall_traverse() that bumps the
310 * RSN of each traversed record in the database.
312 * This is needed to ensure that when rolling out our
313 * empty record copy before remote deletion, we as the
314 * record's dmaster keep a higher RSN than the non-dmaster
315 * nodes. This is needed to prevent old copies from
316 * resurrection in recoveries.
318 static int delete_marshall_traverse_first(void *param
, void *data
)
320 struct delete_record_data
*dd
= talloc_get_type(data
, struct delete_record_data
);
321 struct delete_records_list
*recs
= talloc_get_type(param
, struct delete_records_list
);
322 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
323 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
324 struct ctdb_ltdb_header header
;
326 uint32_t hash
= ctdb_hash(&(dd
->key
));
329 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
332 (__location__
" Error getting chainlock on record with "
333 "key hash [0x%08x] on database db[%s].\n",
334 hash
, ctdb_db
->db_name
));
335 recs
->vdata
->delete_skipped
++;
341 * Verify that the record is still empty, its RSN has not
342 * changed and that we are still its lmaster and dmaster.
345 res
= tdb_parse_record(ctdb_db
->ltdb
->tdb
, dd
->key
,
346 vacuum_record_parser
, &header
);
351 if (header
.flags
& CTDB_REC_RO_FLAGS
) {
352 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
353 "on database db[%s] has read-only flags. "
355 hash
, ctdb_db
->db_name
));
359 if (header
.dmaster
!= ctdb
->pnn
) {
360 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
361 "on database db[%s] has been migrated away. "
363 hash
, ctdb_db
->db_name
));
367 if (header
.rsn
!= dd
->hdr
.rsn
) {
368 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
369 "on database db[%s] seems to have been "
370 "migrated away and back again (with empty "
371 "data). skipping.\n",
372 hash
, ctdb_db
->db_name
));
376 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
378 if (lmaster
!= ctdb
->pnn
) {
379 DEBUG(DEBUG_INFO
, (__location__
": not lmaster for record in "
380 "delete list (key hash [0x%08x], db[%s]). "
381 "Strange! skipping.\n",
382 hash
, ctdb_db
->db_name
));
387 * Increment the record's RSN to ensure the dmaster (i.e. the current
388 * node) has the highest RSN of the record in the cluster.
389 * This is to prevent old record copies from resurrecting in recoveries
390 * if something should fail during the deletion process.
391 * Note that ctdb_ltdb_store_server() increments the RSN if called
392 * on the record's dmaster.
395 res
= ctdb_ltdb_store(ctdb_db
, dd
->key
, &header
, tdb_null
);
397 DEBUG(DEBUG_ERR
, (__location__
": Failed to store record with "
398 "key hash [0x%08x] on database db[%s].\n",
399 hash
, ctdb_db
->db_name
));
403 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
408 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
410 recs
->vdata
->delete_skipped
++;
419 return delete_marshall_traverse(param
, data
);
423 * traverse function for the traversal of the delete_queue,
424 * the fast-path vacuuming list.
426 * - If the record has been migrated off the node
427 * or has been revived (filled with data) on the node,
428 * then skip the record.
430 * - If the current node is the record's lmaster and it is
431 * a record that has never been migrated with data, then
432 * delete the record from the local tdb.
434 * - If the current node is the record's lmaster and it has
435 * been migrated with data, then schedule it for the normal
436 * vacuuming procedure (i.e. add it to the delete_list).
438 * - If the current node is NOT the record's lmaster then
439 * add it to the list of records that are to be sent to
440 * the lmaster with the VACUUM_FETCH message.
442 static int delete_queue_traverse(void *param
, void *data
)
444 struct delete_record_data
*dd
=
445 talloc_get_type(data
, struct delete_record_data
);
446 struct vacuum_data
*vdata
= talloc_get_type(param
, struct vacuum_data
);
447 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
448 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
; /* or dd->ctdb ??? */
450 struct ctdb_ltdb_header header
;
452 uint32_t hash
= ctdb_hash(&(dd
->key
));
456 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
459 (__location__
" Error getting chainlock on record with "
460 "key hash [0x%08x] on database db[%s].\n",
461 hash
, ctdb_db
->db_name
));
466 res
= tdb_parse_record(ctdb_db
->ltdb
->tdb
, dd
->key
,
467 vacuum_record_parser
, &header
);
472 if (header
.dmaster
!= ctdb
->pnn
) {
473 /* The record has been migrated off the node. Skip. */
477 if (header
.rsn
!= dd
->hdr
.rsn
) {
479 * The record has been migrated off the node and back again.
480 * But not requeued for deletion. Skip it.
486 * We are dmaster, and the record has no data, and it has
487 * not been migrated after it has been queued for deletion.
489 * At this stage, the record could still have been revived locally
490 * and last been written with empty data. This can only be
491 * fixed with the addition of an active or delete flag. (TODO)
494 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
496 if (lmaster
!= ctdb
->pnn
) {
497 res
= add_record_to_vacuum_fetch_list(vdata
, dd
->key
);
501 (__location__
" Error adding record to list "
502 "of records to send to lmaster.\n"));
505 vdata
->fast_added_to_vacuum_fetch_list
++;
510 /* use header->flags or dd->hdr.flags ?? */
511 if (dd
->hdr
.flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
) {
512 res
= add_record_to_delete_list(vdata
, dd
->key
, &dd
->hdr
);
516 (__location__
" Error adding record to list "
517 "of records for deletion on lmaster.\n"));
520 vdata
->fast_added_to_delete_list
++;
523 res
= tdb_delete(ctdb_db
->ltdb
->tdb
, dd
->key
);
527 (__location__
" Error deleting record with key "
528 "hash [0x%08x] from local data base db[%s].\n",
529 hash
, ctdb_db
->db_name
));
535 (__location__
" Deleted record with key hash "
536 "[0x%08x] from local data base db[%s].\n",
537 hash
, ctdb_db
->db_name
));
538 vdata
->fast_deleted
++;
544 vdata
->fast_skipped
++;
547 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
553 * Delete the records that we are lmaster and dmaster for and
554 * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
557 static int delete_record_traverse(void *param
, void *data
)
559 struct delete_record_data
*dd
=
560 talloc_get_type(data
, struct delete_record_data
);
561 struct vacuum_data
*vdata
= talloc_get_type(param
, struct vacuum_data
);
562 struct ctdb_db_context
*ctdb_db
= dd
->ctdb_db
;
563 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
565 struct ctdb_ltdb_header header
;
567 uint32_t hash
= ctdb_hash(&(dd
->key
));
569 res
= tdb_chainlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
572 (__location__
" Error getting chainlock on record with "
573 "key hash [0x%08x] on database db[%s].\n",
574 hash
, ctdb_db
->db_name
));
575 vdata
->delete_local_error
++;
576 vdata
->delete_left
--;
582 * Verify that the record is still empty, its RSN has not
583 * changed and that we are still its lmaster and dmaster.
586 res
= tdb_parse_record(ctdb_db
->ltdb
->tdb
, dd
->key
,
587 vacuum_record_parser
, &header
);
592 if (header
.flags
& CTDB_REC_RO_FLAGS
) {
593 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
594 "on database db[%s] has read-only flags. "
596 hash
, ctdb_db
->db_name
));
600 if (header
.dmaster
!= ctdb
->pnn
) {
601 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
602 "on database db[%s] has been migrated away. "
604 hash
, ctdb_db
->db_name
));
608 if (header
.rsn
!= dd
->hdr
.rsn
+ 1) {
610 * The record has been migrated off the node and back again.
611 * But not requeued for deletion. Skip it.
612 * (Note that the first marshall traverse has bumped the RSN
615 DEBUG(DEBUG_INFO
, (__location__
": record with hash [0x%08x] "
616 "on database db[%s] seems to have been "
617 "migrated away and back again (with empty "
618 "data). skipping.\n",
619 hash
, ctdb_db
->db_name
));
623 lmaster
= ctdb_lmaster(ctdb_db
->ctdb
, &dd
->key
);
625 if (lmaster
!= ctdb
->pnn
) {
626 DEBUG(DEBUG_INFO
, (__location__
": not lmaster for record in "
627 "delete list (key hash [0x%08x], db[%s]). "
628 "Strange! skipping.\n",
629 hash
, ctdb_db
->db_name
));
633 res
= tdb_delete(ctdb_db
->ltdb
->tdb
, dd
->key
);
637 (__location__
" Error deleting record with key hash "
638 "[0x%08x] from local data base db[%s].\n",
639 hash
, ctdb_db
->db_name
));
640 vdata
->delete_local_error
++;
645 (__location__
" Deleted record with key hash [0x%08x] from "
646 "local data base db[%s].\n", hash
, ctdb_db
->db_name
));
648 vdata
->delete_deleted
++;
652 vdata
->delete_skipped
++;
655 tdb_chainunlock(ctdb_db
->ltdb
->tdb
, dd
->key
);
658 vdata
->delete_left
--;
664 * Traverse the delete_queue.
665 * Records are either deleted directly or filled
666 * into the delete list or the vacuum fetch lists
667 * for further processing.
669 static void ctdb_process_delete_queue(struct ctdb_db_context
*ctdb_db
,
670 struct vacuum_data
*vdata
)
675 ret
= trbt_traversearray32(ctdb_db
->delete_queue
, 1,
676 delete_queue_traverse
, vdata
);
679 DEBUG(DEBUG_ERR
, (__location__
" Error traversing "
680 "the delete queue.\n"));
683 sum
= vdata
->fast_deleted
684 + vdata
->fast_skipped
686 + vdata
->fast_added_to_delete_list
687 + vdata
->fast_added_to_vacuum_fetch_list
;
689 if (vdata
->fast_total
!= sum
) {
690 DEBUG(DEBUG_ERR
, (__location__
" Inconsistency in fast vacuum "
691 "counts for db[%s]: total[%u] != sum[%u]\n",
692 ctdb_db
->db_name
, (unsigned)vdata
->fast_total
,
696 if (vdata
->fast_total
> 0) {
699 " fast vacuuming delete_queue traverse statistics: "
708 (unsigned)vdata
->fast_total
,
709 (unsigned)vdata
->fast_deleted
,
710 (unsigned)vdata
->fast_skipped
,
711 (unsigned)vdata
->fast_error
,
712 (unsigned)vdata
->fast_added_to_delete_list
,
713 (unsigned)vdata
->fast_added_to_vacuum_fetch_list
));
720 * read-only traverse of the database, looking for records that
721 * might be able to be vacuumed.
723 * This is not done each time but only every tunable
724 * VacuumFastPathCount times.
726 static void ctdb_vacuum_traverse_db(struct ctdb_db_context
*ctdb_db
,
727 struct vacuum_data
*vdata
)
731 ret
= tdb_traverse_read(ctdb_db
->ltdb
->tdb
, vacuum_traverse
, vdata
);
732 if (ret
== -1 || vdata
->traverse_error
) {
733 DEBUG(DEBUG_ERR
, (__location__
" Traverse error in vacuuming "
734 "'%s'\n", ctdb_db
->db_name
));
737 if (vdata
->full_total
> 0) {
740 " full vacuuming db traverse statistics: "
747 (unsigned)vdata
->full_total
,
748 (unsigned)vdata
->full_skipped
,
749 (unsigned)vdata
->full_error
,
750 (unsigned)vdata
->full_scheduled
));
757 * Process the vacuum fetch lists:
758 * For records for which we are not the lmaster, tell the lmaster to
761 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context
*ctdb_db
,
762 struct vacuum_data
*vdata
)
765 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
767 for (i
= 0; i
< ctdb
->num_nodes
; i
++) {
769 struct ctdb_marshall_buffer
*vfl
= vdata
->vacuum_fetch_list
[i
];
771 if (ctdb
->nodes
[i
]->pnn
== ctdb
->pnn
) {
775 if (vfl
->count
== 0) {
779 DEBUG(DEBUG_INFO
, ("Found %u records for lmaster %u in '%s'\n",
780 vfl
->count
, ctdb
->nodes
[i
]->pnn
,
783 data
.dsize
= talloc_get_size(vfl
);
784 data
.dptr
= (void *)vfl
;
785 if (ctdb_client_send_message(ctdb
, ctdb
->nodes
[i
]->pnn
,
786 CTDB_SRVID_VACUUM_FETCH
,
789 DEBUG(DEBUG_ERR
, (__location__
" Failed to send vacuum "
790 "fetch message to %u\n",
791 ctdb
->nodes
[i
]->pnn
));
799 * Process the delete list:
801 * This is the last step of vacuuming that consistently deletes
802 * those records that have been migrated with data and can hence
803 * not be deleted when leaving a node.
805 * In this step, the lmaster does the final deletion of those empty
806 * records that it is also dmaster for. It has ususally received
807 * at least some of these records previously from the former dmasters
808 * with the vacuum fetch message.
810 * This last step is implemented as a 3-phase process to protect from
811 * races leading to data corruption:
813 * 1) Send the lmaster's copy to all other active nodes with the
814 * RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
815 * 2) Send the records that could successfully be stored remotely
816 * in step #1 to all active nodes with the TRY_DELETE_RECORDS
817 * control. The remote notes delete their local copy.
818 * 3) The lmaster locally deletes its copies of all records that
819 * could successfully be deleted remotely in step #2.
821 static void ctdb_process_delete_list(struct ctdb_db_context
*ctdb_db
,
822 struct vacuum_data
*vdata
)
825 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
826 struct delete_records_list
*recs
;
828 struct ctdb_node_map
*nodemap
;
829 uint32_t *active_nodes
;
830 int num_active_nodes
;
834 if (vdata
->delete_count
== 0) {
838 tmp_ctx
= talloc_new(vdata
);
839 if (tmp_ctx
== NULL
) {
840 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
844 vdata
->delete_left
= vdata
->delete_count
;
847 * get the list of currently active nodes
850 ret
= ctdb_ctrl_getnodemap(ctdb
, TIMELIMIT(),
855 DEBUG(DEBUG_ERR
,(__location__
" unable to get node map\n"));
859 active_nodes
= list_of_active_nodes(ctdb
, nodemap
,
860 nodemap
, /* talloc context */
861 false /* include self */);
863 num_active_nodes
= talloc_get_size(active_nodes
)/sizeof(*active_nodes
);
866 * Now delete the records all active nodes in a three-phase process:
867 * 1) send all active remote nodes the current empty copy with this
869 * 2) if all nodes could store the new copy,
870 * tell all the active remote nodes to delete all their copy
871 * 3) if all remote nodes deleted their record copy, delete it locally
876 * Send currently empty record copy to all active nodes for storing.
879 recs
= talloc_zero(tmp_ctx
, struct delete_records_list
);
881 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
884 recs
->records
= (struct ctdb_marshall_buffer
*)
885 talloc_zero_size(recs
,
886 offsetof(struct ctdb_marshall_buffer
, data
));
887 if (recs
->records
== NULL
) {
888 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
891 recs
->records
->db_id
= ctdb_db
->db_id
;
895 * traverse the tree of all records we want to delete and
896 * create a blob we can send to the other nodes.
898 * We call delete_marshall_traverse_first() to bump the
899 * records' RSNs in the database, to ensure we (as dmaster)
900 * keep the highest RSN of the records in the cluster.
902 ret
= trbt_traversearray32(vdata
->delete_list
, 1,
903 delete_marshall_traverse_first
, recs
);
905 DEBUG(DEBUG_ERR
, (__location__
" Error traversing the "
906 "delete list for first marshalling.\n"));
909 indata
.dsize
= talloc_get_size(recs
->records
);
910 indata
.dptr
= (void *)recs
->records
;
912 for (i
= 0; i
< num_active_nodes
; i
++) {
913 struct ctdb_marshall_buffer
*records
;
914 struct ctdb_rec_data
*rec
;
918 ret
= ctdb_control(ctdb
, active_nodes
[i
], 0,
919 CTDB_CONTROL_RECEIVE_RECORDS
, 0,
920 indata
, recs
, &outdata
, &res
,
922 if (ret
!= 0 || res
!= 0) {
923 DEBUG(DEBUG_ERR
, ("Error storing record copies on "
924 "node %u: ret[%d] res[%d]\n",
925 active_nodes
[i
], ret
, res
));
930 * outdata contains the list of records coming back
931 * from the node: These are the records that the
932 * remote node could not store. We remove these from
933 * the list to process further.
935 records
= (struct ctdb_marshall_buffer
*)outdata
.dptr
;
936 rec
= (struct ctdb_rec_data
*)&records
->data
[0];
937 while (records
->count
-- > 1) {
938 TDB_DATA reckey
, recdata
;
939 struct ctdb_ltdb_header
*rechdr
;
940 struct delete_record_data
*dd
;
942 reckey
.dptr
= &rec
->data
[0];
943 reckey
.dsize
= rec
->keylen
;
944 recdata
.dptr
= &rec
->data
[reckey
.dsize
];
945 recdata
.dsize
= rec
->datalen
;
947 if (recdata
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
948 DEBUG(DEBUG_CRIT
,(__location__
" bad ltdb record\n"));
951 rechdr
= (struct ctdb_ltdb_header
*)recdata
.dptr
;
952 recdata
.dptr
+= sizeof(*rechdr
);
953 recdata
.dsize
-= sizeof(*rechdr
);
955 dd
= (struct delete_record_data
*)trbt_lookup32(
960 * The other node could not store the record
961 * copy and it is the first node that failed.
962 * So we should remove it from the tree and
966 vdata
->delete_remote_error
++;
967 vdata
->delete_left
--;
970 rec
= (struct ctdb_rec_data
*)(rec
->length
+ (uint8_t *)rec
);
974 if (vdata
->delete_left
== 0) {
980 * Send the remaining records to all active nodes for deletion.
982 * The lmaster's (i.e. our) copies of these records have been stored
983 * successfully on the other nodes.
987 * Create a marshall blob from the remaining list of records to delete.
990 talloc_free(recs
->records
);
992 recs
->records
= (struct ctdb_marshall_buffer
*)
993 talloc_zero_size(recs
,
994 offsetof(struct ctdb_marshall_buffer
, data
));
995 if (recs
->records
== NULL
) {
996 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
999 recs
->records
->db_id
= ctdb_db
->db_id
;
1001 ret
= trbt_traversearray32(vdata
->delete_list
, 1,
1002 delete_marshall_traverse
, recs
);
1004 DEBUG(DEBUG_ERR
, (__location__
" Error traversing the "
1005 "delete list for second marshalling.\n"));
1008 indata
.dsize
= talloc_get_size(recs
->records
);
1009 indata
.dptr
= (void *)recs
->records
;
1011 for (i
= 0; i
< num_active_nodes
; i
++) {
1012 struct ctdb_marshall_buffer
*records
;
1013 struct ctdb_rec_data
*rec
;
1017 ret
= ctdb_control(ctdb
, active_nodes
[i
], 0,
1018 CTDB_CONTROL_TRY_DELETE_RECORDS
, 0,
1019 indata
, recs
, &outdata
, &res
,
1021 if (ret
!= 0 || res
!= 0) {
1022 DEBUG(DEBUG_ERR
, ("Failed to delete records on "
1023 "node %u: ret[%d] res[%d]\n",
1024 active_nodes
[i
], ret
, res
));
1029 * outdata contains the list of records coming back
1030 * from the node: These are the records that the
1031 * remote node could not delete. We remove these from
1032 * the list to delete locally.
1034 records
= (struct ctdb_marshall_buffer
*)outdata
.dptr
;
1035 rec
= (struct ctdb_rec_data
*)&records
->data
[0];
1036 while (records
->count
-- > 1) {
1037 TDB_DATA reckey
, recdata
;
1038 struct ctdb_ltdb_header
*rechdr
;
1039 struct delete_record_data
*dd
;
1041 reckey
.dptr
= &rec
->data
[0];
1042 reckey
.dsize
= rec
->keylen
;
1043 recdata
.dptr
= &rec
->data
[reckey
.dsize
];
1044 recdata
.dsize
= rec
->datalen
;
1046 if (recdata
.dsize
< sizeof(struct ctdb_ltdb_header
)) {
1047 DEBUG(DEBUG_CRIT
,(__location__
" bad ltdb record\n"));
1050 rechdr
= (struct ctdb_ltdb_header
*)recdata
.dptr
;
1051 recdata
.dptr
+= sizeof(*rechdr
);
1052 recdata
.dsize
-= sizeof(*rechdr
);
1054 dd
= (struct delete_record_data
*)trbt_lookup32(
1056 ctdb_hash(&reckey
));
1059 * The other node could not delete the
1060 * record and it is the first node that
1061 * failed. So we should remove it from
1062 * the tree and update statistics.
1065 vdata
->delete_remote_error
++;
1066 vdata
->delete_left
--;
1069 rec
= (struct ctdb_rec_data
*)(rec
->length
+ (uint8_t *)rec
);
1073 if (vdata
->delete_left
== 0) {
1079 * Delete the remaining records locally.
1081 * These records have successfully been deleted on all
1082 * active remote nodes.
1085 ret
= trbt_traversearray32(vdata
->delete_list
, 1,
1086 delete_record_traverse
, vdata
);
1088 DEBUG(DEBUG_ERR
, (__location__
" Error traversing the "
1089 "delete list for deletion.\n"));
1094 if (vdata
->delete_left
!= 0) {
1095 DEBUG(DEBUG_ERR
, (__location__
" Vaccum db[%s] error: "
1096 "there are %u records left for deletion after "
1097 "processing delete list\n",
1099 (unsigned)vdata
->delete_left
));
1102 sum
= vdata
->delete_deleted
1103 + vdata
->delete_skipped
1104 + vdata
->delete_remote_error
1105 + vdata
->delete_local_error
1106 + vdata
->delete_left
;
1108 if (vdata
->delete_count
!= sum
) {
1109 DEBUG(DEBUG_ERR
, (__location__
" Inconsistency in vacuum "
1110 "delete list counts for db[%s]: total[%u] != sum[%u]\n",
1111 ctdb_db
->db_name
, (unsigned)vdata
->delete_count
,
1115 if (vdata
->delete_count
> 0) {
1118 " vacuum delete list statistics: "
1127 (unsigned)vdata
->delete_count
,
1128 (unsigned)vdata
->delete_deleted
,
1129 (unsigned)vdata
->delete_skipped
,
1130 (unsigned)vdata
->delete_remote_error
,
1131 (unsigned)vdata
->delete_local_error
,
1132 (unsigned)vdata
->delete_left
));
1136 talloc_free(tmp_ctx
);
1142 * initialize the vacuum_data
1144 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context
*ctdb_db
,
1145 struct vacuum_data
*vdata
)
1148 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1150 vdata
->fast_added_to_delete_list
= 0;
1151 vdata
->fast_added_to_vacuum_fetch_list
= 0;
1152 vdata
->fast_deleted
= 0;
1153 vdata
->fast_skipped
= 0;
1154 vdata
->fast_error
= 0;
1155 vdata
->fast_total
= 0;
1156 vdata
->full_scheduled
= 0;
1157 vdata
->full_skipped
= 0;
1158 vdata
->full_error
= 0;
1159 vdata
->full_total
= 0;
1160 vdata
->delete_count
= 0;
1161 vdata
->delete_left
= 0;
1162 vdata
->delete_remote_error
= 0;
1163 vdata
->delete_local_error
= 0;
1164 vdata
->delete_skipped
= 0;
1165 vdata
->delete_deleted
= 0;
1167 /* the list needs to be of length num_nodes */
1168 vdata
->vacuum_fetch_list
= talloc_zero_array(vdata
,
1169 struct ctdb_marshall_buffer
*,
1171 if (vdata
->vacuum_fetch_list
== NULL
) {
1172 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1175 for (i
= 0; i
< ctdb
->num_nodes
; i
++) {
1176 vdata
->vacuum_fetch_list
[i
] = (struct ctdb_marshall_buffer
*)
1177 talloc_zero_size(vdata
->vacuum_fetch_list
,
1178 offsetof(struct ctdb_marshall_buffer
, data
));
1179 if (vdata
->vacuum_fetch_list
[i
] == NULL
) {
1180 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1183 vdata
->vacuum_fetch_list
[i
]->db_id
= ctdb_db
->db_id
;
1191 * - Always do the fast vacuuming run, which traverses
1192 * the in-memory delete queue: these records have been
1193 * scheduled for deletion.
1194 * - Only if explicitly requested, the database is traversed
1195 * in order to use the traditional heuristics on empty records
1196 * to trigger deletion.
1197 * This is done only every VacuumFastPathCount'th vacuuming run.
1199 * The traverse runs fill two lists:
1201 * - The delete_list:
1202 * This is the list of empty records the current
1203 * node is lmaster and dmaster for. These records are later
1204 * deleted first on other nodes and then locally.
1206 * The fast vacuuming run has a short cut for those records
1207 * that have never been migrated with data: these records
1208 * are immediately deleted locally, since they have left
1209 * no trace on other nodes.
1211 * - The vacuum_fetch lists
1212 * (one for each other lmaster node):
1213 * The records in this list are sent for deletion to
1214 * their lmaster in a bulk VACUUM_FETCH message.
1216 * The lmaster then migrates all these records to itelf
1217 * so that they can be vacuumed there.
1219 * This executes in the child context.
1221 static int ctdb_vacuum_db(struct ctdb_db_context
*ctdb_db
,
1222 struct vacuum_data
*vdata
,
1223 bool full_vacuum_run
)
1225 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1228 DEBUG(DEBUG_INFO
, (__location__
" Entering %s vacuum run for db "
1229 "%s db_id[0x%08x]\n",
1230 full_vacuum_run
? "full" : "fast",
1231 ctdb_db
->db_name
, ctdb_db
->db_id
));
1233 ret
= ctdb_ctrl_getvnnmap(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
, ctdb
, &ctdb
->vnn_map
);
1235 DEBUG(DEBUG_ERR
, ("Unable to get vnnmap from local node\n"));
1239 pnn
= ctdb_ctrl_getpnn(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
);
1241 DEBUG(DEBUG_ERR
, ("Unable to get pnn from local node\n"));
1247 ret
= ctdb_vacuum_init_vacuum_data(ctdb_db
, vdata
);
1252 if (full_vacuum_run
) {
1253 ctdb_vacuum_traverse_db(ctdb_db
, vdata
);
1256 ctdb_process_delete_queue(ctdb_db
, vdata
);
1258 ctdb_process_vacuum_fetch_lists(ctdb_db
, vdata
);
1260 ctdb_process_delete_list(ctdb_db
, vdata
);
1262 /* this ensures we run our event queue */
1263 ctdb_ctrl_getpnn(ctdb
, TIMELIMIT(), CTDB_CURRENT_NODE
);
1270 * traverse function for repacking
1272 static int repack_traverse(struct tdb_context
*tdb
, TDB_DATA key
, TDB_DATA data
,
1275 struct vacuum_data
*vdata
= (struct vacuum_data
*)private_data
;
1277 if (vdata
->vacuum
) {
1278 uint32_t hash
= ctdb_hash(&key
);
1279 struct delete_record_data
*kd
;
1281 * check if we can ignore this record because it's in the delete_list
1283 kd
= (struct delete_record_data
*)trbt_lookup32(vdata
->delete_list
, hash
);
1285 * there might be hash collisions so we have to compare the keys here to be sure
1287 if (kd
&& kd
->key
.dsize
== key
.dsize
&& memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) == 0) {
1288 struct ctdb_ltdb_header
*hdr
= (struct ctdb_ltdb_header
*)data
.dptr
;
1290 * we have to check if the record hasn't changed in the meantime in order to
1291 * savely remove it from the database
1293 if (data
.dsize
== sizeof(struct ctdb_ltdb_header
) &&
1294 hdr
->dmaster
== kd
->ctdb
->pnn
&&
1295 ctdb_lmaster(kd
->ctdb
, &(kd
->key
)) == kd
->ctdb
->pnn
&&
1296 kd
->hdr
.rsn
== hdr
->rsn
) {
1302 if (tdb_store(vdata
->dest_db
, key
, data
, TDB_INSERT
) != 0) {
1303 vdata
->traverse_error
= true;
1313 static int ctdb_repack_tdb(struct tdb_context
*tdb
, TALLOC_CTX
*mem_ctx
, struct vacuum_data
*vdata
)
1315 struct tdb_context
*tmp_db
;
1317 if (tdb_transaction_start(tdb
) != 0) {
1318 DEBUG(DEBUG_ERR
,(__location__
" Failed to start transaction\n"));
1322 tmp_db
= tdb_open("tmpdb", tdb_hash_size(tdb
),
1323 TDB_INTERNAL
|TDB_DISALLOW_NESTING
,
1325 if (tmp_db
== NULL
) {
1326 DEBUG(DEBUG_ERR
,(__location__
" Failed to create tmp_db\n"));
1327 tdb_transaction_cancel(tdb
);
1331 vdata
->traverse_error
= false;
1332 vdata
->dest_db
= tmp_db
;
1333 vdata
->vacuum
= true;
1334 vdata
->vacuumed
= 0;
1338 * repack and vacuum on-the-fly by not writing the records that are
1341 if (tdb_traverse_read(tdb
, repack_traverse
, vdata
) == -1) {
1342 DEBUG(DEBUG_ERR
,(__location__
" Failed to traverse copying out\n"));
1343 tdb_transaction_cancel(tdb
);
1348 DEBUG(DEBUG_INFO
,(__location__
" %u records vacuumed\n", vdata
->vacuumed
));
1350 if (vdata
->traverse_error
) {
1351 DEBUG(DEBUG_ERR
,(__location__
" Error during traversal\n"));
1352 tdb_transaction_cancel(tdb
);
1357 if (tdb_wipe_all(tdb
) != 0) {
1358 DEBUG(DEBUG_ERR
,(__location__
" Failed to wipe database\n"));
1359 tdb_transaction_cancel(tdb
);
1364 vdata
->traverse_error
= false;
1365 vdata
->dest_db
= tdb
;
1366 vdata
->vacuum
= false;
1369 if (tdb_traverse_read(tmp_db
, repack_traverse
, vdata
) == -1) {
1370 DEBUG(DEBUG_ERR
,(__location__
" Failed to traverse copying back\n"));
1371 tdb_transaction_cancel(tdb
);
1376 if (vdata
->traverse_error
) {
1377 DEBUG(DEBUG_ERR
,(__location__
" Error during second traversal\n"));
1378 tdb_transaction_cancel(tdb
);
1386 if (tdb_transaction_commit(tdb
) != 0) {
1387 DEBUG(DEBUG_ERR
,(__location__
" Failed to commit\n"));
1390 DEBUG(DEBUG_INFO
,(__location__
" %u records copied\n", vdata
->copied
));
1396 * repack and vaccum a db
1397 * called from the child context
1399 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context
*ctdb_db
,
1400 TALLOC_CTX
*mem_ctx
,
1401 bool full_vacuum_run
)
1403 uint32_t repack_limit
= ctdb_db
->ctdb
->tunable
.repack_limit
;
1404 const char *name
= ctdb_db
->db_name
;
1405 int freelist_size
= 0;
1406 struct vacuum_data
*vdata
;
1408 vdata
= talloc_zero(mem_ctx
, struct vacuum_data
);
1409 if (vdata
== NULL
) {
1410 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1414 vdata
->ctdb
= ctdb_db
->ctdb
;
1415 vdata
->repack_limit
= repack_limit
;
1416 vdata
->delete_list
= trbt_create(vdata
, 0);
1417 vdata
->ctdb_db
= ctdb_db
;
1418 if (vdata
->delete_list
== NULL
) {
1419 DEBUG(DEBUG_ERR
,(__location__
" Out of memory\n"));
1424 vdata
->start
= timeval_current();
1427 * gather all records that can be deleted in vdata
1429 if (ctdb_vacuum_db(ctdb_db
, vdata
, full_vacuum_run
) != 0) {
1430 DEBUG(DEBUG_ERR
,(__location__
" Failed to vacuum '%s'\n", name
));
1433 if (repack_limit
!= 0) {
1434 freelist_size
= tdb_freelist_size(ctdb_db
->ltdb
->tdb
);
1435 if (freelist_size
== -1) {
1436 DEBUG(DEBUG_ERR
,(__location__
" Failed to get freelist size for '%s'\n", name
));
1443 * decide if a repack is necessary
1445 if ((repack_limit
== 0 || (uint32_t)freelist_size
< repack_limit
))
1451 DEBUG(DEBUG_INFO
,("Repacking %s with %u freelist entries and %u records to delete\n",
1452 name
, freelist_size
, vdata
->delete_left
));
1455 * repack and implicitely get rid of the records we can delete
1457 if (ctdb_repack_tdb(ctdb_db
->ltdb
->tdb
, mem_ctx
, vdata
) != 0) {
1458 DEBUG(DEBUG_ERR
,(__location__
" Failed to repack '%s'\n", name
));
1467 static uint32_t get_vacuum_interval(struct ctdb_db_context
*ctdb_db
)
1469 uint32_t interval
= ctdb_db
->ctdb
->tunable
.vacuum_interval
;
1474 static int vacuum_child_destructor(struct ctdb_vacuum_child_context
*child_ctx
)
1476 double l
= timeval_elapsed(&child_ctx
->start_time
);
1477 struct ctdb_db_context
*ctdb_db
= child_ctx
->vacuum_handle
->ctdb_db
;
1478 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1480 DEBUG(DEBUG_INFO
,("Vacuuming took %.3f seconds for database %s\n", l
, ctdb_db
->db_name
));
1482 if (child_ctx
->child_pid
!= -1) {
1483 ctdb_kill(ctdb
, child_ctx
->child_pid
, SIGKILL
);
1485 /* Bump the number of successful fast-path runs. */
1486 child_ctx
->vacuum_handle
->fast_path_count
++;
1489 DLIST_REMOVE(ctdb
->vacuumers
, child_ctx
);
1491 event_add_timed(ctdb
->ev
, child_ctx
->vacuum_handle
,
1492 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1493 ctdb_vacuum_event
, child_ctx
->vacuum_handle
);
1499 * this event is generated when a vacuum child process times out
1501 static void vacuum_child_timeout(struct event_context
*ev
, struct timed_event
*te
,
1502 struct timeval t
, void *private_data
)
1504 struct ctdb_vacuum_child_context
*child_ctx
= talloc_get_type(private_data
, struct ctdb_vacuum_child_context
);
1506 DEBUG(DEBUG_ERR
,("Vacuuming child process timed out for db %s\n", child_ctx
->vacuum_handle
->ctdb_db
->db_name
));
1508 child_ctx
->status
= VACUUM_TIMEOUT
;
1510 talloc_free(child_ctx
);
1515 * this event is generated when a vacuum child process has completed
1517 static void vacuum_child_handler(struct event_context
*ev
, struct fd_event
*fde
,
1518 uint16_t flags
, void *private_data
)
1520 struct ctdb_vacuum_child_context
*child_ctx
= talloc_get_type(private_data
, struct ctdb_vacuum_child_context
);
1524 DEBUG(DEBUG_INFO
,("Vacuuming child process %d finished for db %s\n", child_ctx
->child_pid
, child_ctx
->vacuum_handle
->ctdb_db
->db_name
));
1525 child_ctx
->child_pid
= -1;
1527 ret
= read(child_ctx
->fd
[0], &c
, 1);
1528 if (ret
!= 1 || c
!= 0) {
1529 child_ctx
->status
= VACUUM_ERROR
;
1530 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
));
1532 child_ctx
->status
= VACUUM_OK
;
1535 talloc_free(child_ctx
);
1539 * this event is called every time we need to start a new vacuum process
1542 ctdb_vacuum_event(struct event_context
*ev
, struct timed_event
*te
,
1543 struct timeval t
, void *private_data
)
1545 struct ctdb_vacuum_handle
*vacuum_handle
= talloc_get_type(private_data
, struct ctdb_vacuum_handle
);
1546 struct ctdb_db_context
*ctdb_db
= vacuum_handle
->ctdb_db
;
1547 struct ctdb_context
*ctdb
= ctdb_db
->ctdb
;
1548 struct ctdb_vacuum_child_context
*child_ctx
;
1549 struct tevent_fd
*fde
;
1552 /* we dont vacuum if we are in recovery mode, or db frozen */
1553 if (ctdb
->recovery_mode
== CTDB_RECOVERY_ACTIVE
||
1554 ctdb
->freeze_mode
[ctdb_db
->priority
] != CTDB_FREEZE_NONE
) {
1555 DEBUG(DEBUG_INFO
, ("Not vacuuming %s (%s)\n", ctdb_db
->db_name
,
1556 ctdb
->recovery_mode
== CTDB_RECOVERY_ACTIVE
? "in recovery"
1557 : ctdb
->freeze_mode
[ctdb_db
->priority
] == CTDB_FREEZE_PENDING
1560 event_add_timed(ctdb
->ev
, vacuum_handle
,
1561 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1562 ctdb_vacuum_event
, vacuum_handle
);
1566 child_ctx
= talloc(vacuum_handle
, struct ctdb_vacuum_child_context
);
1567 if (child_ctx
== NULL
) {
1568 DEBUG(DEBUG_CRIT
, (__location__
" Failed to allocate child context for vacuuming of %s\n", ctdb_db
->db_name
));
1569 ctdb_fatal(ctdb
, "Out of memory when crating vacuum child context. Shutting down\n");
1573 ret
= pipe(child_ctx
->fd
);
1575 talloc_free(child_ctx
);
1576 DEBUG(DEBUG_ERR
, ("Failed to create pipe for vacuum child process.\n"));
1577 event_add_timed(ctdb
->ev
, vacuum_handle
,
1578 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1579 ctdb_vacuum_event
, vacuum_handle
);
1583 if (vacuum_handle
->fast_path_count
> ctdb
->tunable
.vacuum_fast_path_count
) {
1584 vacuum_handle
->fast_path_count
= 0;
1587 child_ctx
->child_pid
= ctdb_fork(ctdb
);
1588 if (child_ctx
->child_pid
== (pid_t
)-1) {
1589 close(child_ctx
->fd
[0]);
1590 close(child_ctx
->fd
[1]);
1591 talloc_free(child_ctx
);
1592 DEBUG(DEBUG_ERR
, ("Failed to fork vacuum child process.\n"));
1593 event_add_timed(ctdb
->ev
, vacuum_handle
,
1594 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1595 ctdb_vacuum_event
, vacuum_handle
);
1600 if (child_ctx
->child_pid
== 0) {
1602 bool full_vacuum_run
= false;
1603 close(child_ctx
->fd
[0]);
1605 DEBUG(DEBUG_INFO
,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db
->db_name
));
1606 ctdb_set_process_name("ctdb_vacuum");
1607 if (switch_from_server_to_client(ctdb
, "vacuum-%s", ctdb_db
->db_name
) != 0) {
1608 DEBUG(DEBUG_CRIT
, (__location__
"ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1615 if ((ctdb
->tunable
.vacuum_fast_path_count
> 0) &&
1616 (vacuum_handle
->fast_path_count
== 0))
1618 full_vacuum_run
= true;
1620 cc
= ctdb_vacuum_and_repack_db(ctdb_db
, child_ctx
,
1623 write(child_ctx
->fd
[1], &cc
, 1);
1627 set_close_on_exec(child_ctx
->fd
[0]);
1628 close(child_ctx
->fd
[1]);
1630 child_ctx
->status
= VACUUM_RUNNING
;
1631 child_ctx
->start_time
= timeval_current();
1633 DLIST_ADD(ctdb
->vacuumers
, child_ctx
);
1634 talloc_set_destructor(child_ctx
, vacuum_child_destructor
);
1637 * Clear the fastpath vacuuming list in the parent.
1639 talloc_free(ctdb_db
->delete_queue
);
1640 ctdb_db
->delete_queue
= trbt_create(ctdb_db
, 0);
1641 if (ctdb_db
->delete_queue
== NULL
) {
1642 /* fatal here? ... */
1643 ctdb_fatal(ctdb
, "Out of memory when re-creating vacuum tree "
1644 "in parent context. Shutting down\n");
1647 event_add_timed(ctdb
->ev
, child_ctx
,
1648 timeval_current_ofs(ctdb
->tunable
.vacuum_max_run_time
, 0),
1649 vacuum_child_timeout
, child_ctx
);
1651 DEBUG(DEBUG_DEBUG
, (__location__
" Created PIPE FD:%d to child vacuum process\n", child_ctx
->fd
[0]));
1653 fde
= event_add_fd(ctdb
->ev
, child_ctx
, child_ctx
->fd
[0],
1654 EVENT_FD_READ
, vacuum_child_handler
, child_ctx
);
1655 tevent_fd_set_auto_close(fde
);
1657 vacuum_handle
->child_ctx
= child_ctx
;
1658 child_ctx
->vacuum_handle
= vacuum_handle
;
1661 void ctdb_stop_vacuuming(struct ctdb_context
*ctdb
)
1663 /* Simply free them all. */
1664 while (ctdb
->vacuumers
) {
1665 DEBUG(DEBUG_INFO
, ("Aborting vacuuming for %s (%i)\n",
1666 ctdb
->vacuumers
->vacuum_handle
->ctdb_db
->db_name
,
1667 (int)ctdb
->vacuumers
->child_pid
));
1668 /* vacuum_child_destructor kills it, removes from list */
1669 talloc_free(ctdb
->vacuumers
);
1673 /* this function initializes the vacuuming context for a database
1674 * starts the vacuuming events
1676 int ctdb_vacuum_init(struct ctdb_db_context
*ctdb_db
)
1678 if (ctdb_db
->persistent
!= 0) {
1679 DEBUG(DEBUG_ERR
,("Vacuuming is disabled for persistent database %s\n", ctdb_db
->db_name
));
1683 ctdb_db
->vacuum_handle
= talloc(ctdb_db
, struct ctdb_vacuum_handle
);
1684 CTDB_NO_MEMORY(ctdb_db
->ctdb
, ctdb_db
->vacuum_handle
);
1686 ctdb_db
->vacuum_handle
->ctdb_db
= ctdb_db
;
1687 ctdb_db
->vacuum_handle
->fast_path_count
= 0;
1689 event_add_timed(ctdb_db
->ctdb
->ev
, ctdb_db
->vacuum_handle
,
1690 timeval_current_ofs(get_vacuum_interval(ctdb_db
), 0),
1691 ctdb_vacuum_event
, ctdb_db
->vacuum_handle
);
1696 static void remove_record_from_delete_queue(struct ctdb_db_context
*ctdb_db
,
1697 const struct ctdb_ltdb_header
*hdr
,
1700 struct delete_record_data
*kd
;
1703 hash
= (uint32_t)ctdb_hash(&key
);
1705 DEBUG(DEBUG_DEBUG
, (__location__
1706 " remove_record_from_delete_queue: "
1711 "migrated_with_data[%s]\n",
1712 ctdb_db
->db_name
, ctdb_db
->db_id
,
1714 ctdb_lmaster(ctdb_db
->ctdb
, &key
),
1715 hdr
->flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
? "yes" : "no"));
1717 kd
= (struct delete_record_data
*)trbt_lookup32(ctdb_db
->delete_queue
, hash
);
1719 DEBUG(DEBUG_DEBUG
, (__location__
1720 " remove_record_from_delete_queue: "
1721 "record not in queue (hash[0x%08x])\n.",
1726 if ((kd
->key
.dsize
!= key
.dsize
) ||
1727 (memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) != 0))
1729 DEBUG(DEBUG_DEBUG
, (__location__
1730 " remove_record_from_delete_queue: "
1731 "hash collision for key with hash[0x%08x] "
1732 "in db[%s] - skipping\n",
1733 hash
, ctdb_db
->db_name
));
1737 DEBUG(DEBUG_DEBUG
, (__location__
1738 " remove_record_from_delete_queue: "
1739 "removing key with hash[0x%08x]\n",
1748 * Insert a record into the ctdb_db context's delete queue,
1749 * handling hash collisions.
1751 static int insert_record_into_delete_queue(struct ctdb_db_context
*ctdb_db
,
1752 const struct ctdb_ltdb_header
*hdr
,
1755 struct delete_record_data
*kd
;
1759 hash
= (uint32_t)ctdb_hash(&key
);
1761 DEBUG(DEBUG_INFO
, (__location__
" schedule for deletion: db[%s] "
1765 "migrated_with_data[%s]\n",
1766 ctdb_db
->db_name
, ctdb_db
->db_id
,
1768 ctdb_lmaster(ctdb_db
->ctdb
, &key
),
1769 hdr
->flags
& CTDB_REC_FLAG_MIGRATED_WITH_DATA
? "yes" : "no"));
1771 kd
= (struct delete_record_data
*)trbt_lookup32(ctdb_db
->delete_queue
, hash
);
1773 if ((kd
->key
.dsize
!= key
.dsize
) ||
1774 (memcmp(kd
->key
.dptr
, key
.dptr
, key
.dsize
) != 0))
1777 (__location__
" schedule for deletion: "
1778 "hash collision for key hash [0x%08x]. "
1779 "Skipping the record.\n", hash
));
1783 (__location__
" schedule for deletion: "
1784 "updating entry for key with hash [0x%08x].\n",
1789 ret
= insert_delete_record_data_into_tree(ctdb_db
->ctdb
, ctdb_db
,
1790 ctdb_db
->delete_queue
,
1794 (__location__
" schedule for deletion: error "
1795 "inserting key with hash [0x%08x] into delete queue\n",
1804 * Schedule a record for deletetion.
1805 * Called from the parent context.
1807 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context
*ctdb
,
1810 struct ctdb_control_schedule_for_deletion
*dd
;
1811 struct ctdb_db_context
*ctdb_db
;
1815 dd
= (struct ctdb_control_schedule_for_deletion
*)indata
.dptr
;
1817 ctdb_db
= find_ctdb_db(ctdb
, dd
->db_id
);
1818 if (ctdb_db
== NULL
) {
1819 DEBUG(DEBUG_ERR
, (__location__
" Unknown db id 0x%08x\n",
1824 key
.dsize
= dd
->keylen
;
1827 ret
= insert_record_into_delete_queue(ctdb_db
, &dd
->hdr
, key
);
1832 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context
*ctdb_db
,
1833 const struct ctdb_ltdb_header
*hdr
,
1837 struct ctdb_control_schedule_for_deletion
*dd
;
1841 if (ctdb_db
->ctdb
->ctdbd_pid
== getpid()) {
1842 /* main daemon - directly queue */
1843 ret
= insert_record_into_delete_queue(ctdb_db
, hdr
, key
);
1848 /* if we dont have a connection to the daemon we can not send
1849 a control. For example sometimes from update_record control child
1852 if (!ctdb_db
->ctdb
->can_send_controls
) {
1857 /* child process: send the main daemon a control */
1858 indata
.dsize
= offsetof(struct ctdb_control_schedule_for_deletion
, key
) + key
.dsize
;
1859 indata
.dptr
= talloc_zero_array(ctdb_db
, uint8_t, indata
.dsize
);
1860 if (indata
.dptr
== NULL
) {
1861 DEBUG(DEBUG_ERR
, (__location__
" out of memory\n"));
1864 dd
= (struct ctdb_control_schedule_for_deletion
*)(void *)indata
.dptr
;
1865 dd
->db_id
= ctdb_db
->db_id
;
1867 dd
->keylen
= key
.dsize
;
1868 memcpy(dd
->key
, key
.dptr
, key
.dsize
);
1870 ret
= ctdb_control(ctdb_db
->ctdb
,
1873 CTDB_CONTROL_SCHEDULE_FOR_DELETION
,
1874 CTDB_CTRL_FLAG_NOREPLY
, /* flags */
1879 NULL
, /* timeout : NULL == wait forever */
1880 NULL
); /* error message */
1882 talloc_free(indata
.dptr
);
1884 if (ret
!= 0 || status
!= 0) {
1885 DEBUG(DEBUG_ERR
, (__location__
" Error sending "
1886 "SCHEDULE_FOR_DELETION "
1896 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context
*ctdb_db
,
1897 const struct ctdb_ltdb_header
*hdr
,
1900 if (ctdb_db
->ctdb
->ctdbd_pid
!= getpid()) {
1902 * Only remove the record from the delete queue if called
1903 * in the main daemon.
1908 remove_record_from_delete_queue(ctdb_db
, hdr
, key
);