2 Unix SMB/CIFS implementation.
4 trivial database library
6 Copyright (C) Andrew Tridgell 2005
7 Copyright (C) Rusty Russell 2010
9 ** NOTE! The following LGPL license applies to the tdb
10 ** library. This does NOT imply that all of Samba is released
13 This library is free software; you can redistribute it and/or
14 modify it under the terms of the GNU Lesser General Public
15 License as published by the Free Software Foundation; either
16 version 3 of the License, or (at your option) any later version.
18 This library is distributed in the hope that it will be useful,
19 but WITHOUT ANY WARRANTY; without even the implied warranty of
20 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21 Lesser General Public License for more details.
23 You should have received a copy of the GNU Lesser General Public
24 License along with this library; if not, see <http://www.gnu.org/licenses/>.
28 #define SAFE_FREE(x) do { if ((x) != NULL) {free(x); (x)=NULL;} } while(0)
33 - only allow a single transaction at a time per database. This makes
34 using the transaction API simpler, as otherwise the caller would
35 have to cope with temporary failures in transactions that conflict
36 with other current transactions
38 - keep the transaction recovery information in the same file as the
39 database, using a special 'transaction recovery' record pointed at
40 by the header. This removes the need for extra journal files as
41 used by some other databases
43 - dynamically allocated the transaction recover record, re-using it
44 for subsequent transactions. If a larger record is needed then
45 tdb_free() the old record to place it on the normal tdb freelist
46 before allocating the new record
48 - during transactions, keep a linked list of writes all that have
49 been performed by intercepting all tdb_write() calls. The hooked
50 transaction versions of tdb_read() and tdb_write() check this
51 linked list and try to use the elements of the list in preference
54 - don't allow any locks to be held when a transaction starts,
55 otherwise we can end up with deadlock (plus lack of lock nesting
56 in POSIX locks would mean the lock is lost)
58 - if the caller gains a lock during the transaction but doesn't
59 release it then fail the commit
61 - allow for nested calls to tdb_transaction_start(), re-using the
62 existing transaction record. If the inner transaction is canceled
63 then a subsequent commit will fail
65 - keep a mirrored copy of the tdb hash chain heads to allow for the
66 fast hash heads scan on traverse, updating the mirrored copy in
67 the transaction version of tdb_write
69 - allow callers to mix transaction and non-transaction use of tdb,
70 although once a transaction is started then an exclusive lock is
71 gained until the transaction is committed or canceled
73 - the commit stategy involves first saving away all modified data
74 into a linearised buffer in the transaction recovery area, then
75 marking the transaction recovery area with a magic value to
76 indicate a valid recovery record. In total 4 fsync/msync calls are
77 needed per commit to prevent race conditions. It might be possible
78 to reduce this to 3 or even 2 with some more work.
80 - check for a valid recovery record on open of the tdb, while the
81 open lock is held. Automatically recover from the transaction
82 recovery area if needed, then continue with the open as
83 usual. This allows for smooth crash recovery with no administrator
86 - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
87 still available, but no transaction recovery area is used and no
88 fsync/msync calls are made.
93 hold the context of any current transaction
95 struct tdb_transaction {
96 /* the original io methods - used to do IOs to the real db */
97 const struct tdb_methods *io_methods;
99 /* the list of transaction blocks. When a block is first
100 written to, it gets created in this list */
103 size_t last_block_size; /* number of valid bytes in the last block */
105 /* non-zero when an internal transaction error has
106 occurred. All write operations will then fail until the
107 transaction is ended */
108 int transaction_error;
110 /* when inside a transaction we need to keep track of any
111 nested tdb_transaction_start() calls, as these are allowed,
112 but don't create a new transaction */
115 /* set when a prepare has already occurred */
117 tdb_off_t magic_offset;
119 /* old file size before transaction */
120 tdb_len_t old_map_size;
123 /* This doesn't really need to be pagesize, but we use it for similar reasons. */
124 #define PAGESIZE 4096
127 read while in a transaction. We need to check first if the data is in our list
128 of transaction elements, then if not do a real read
130 static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
131 void *buf, tdb_len_t len)
134 enum TDB_ERROR ecode;
136 /* break it down into block sized ops */
137 while (len + (off % PAGESIZE) > PAGESIZE) {
138 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
139 ecode = transaction_read(tdb, off, buf, len2);
140 if (ecode != TDB_SUCCESS) {
145 buf = (void *)(len2 + (char *)buf);
152 blk = off / PAGESIZE;
154 /* see if we have it in the block list */
155 if (tdb->transaction->num_blocks <= blk ||
156 tdb->transaction->blocks[blk] == NULL) {
157 /* nope, do a real read */
158 ecode = tdb->transaction->io_methods->tread(tdb, off, buf, len);
159 if (ecode != TDB_SUCCESS) {
165 /* it is in the block list. Now check for the last block */
166 if (blk == tdb->transaction->num_blocks-1) {
167 if (len > tdb->transaction->last_block_size) {
173 /* now copy it out of this block */
174 memcpy(buf, tdb->transaction->blocks[blk] + (off % PAGESIZE), len);
178 tdb->transaction->transaction_error = 1;
179 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
180 "transaction_read: failed at off=%zu len=%zu",
181 (size_t)off, (size_t)len);
186 write while in a transaction
188 static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
189 const void *buf, tdb_len_t len)
192 enum TDB_ERROR ecode;
194 /* Only a commit is allowed on a prepared transaction */
195 if (tdb->transaction->prepared) {
196 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
197 "transaction_write: transaction already"
198 " prepared, write not allowed");
202 /* break it up into block sized chunks */
203 while (len + (off % PAGESIZE) > PAGESIZE) {
204 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
205 ecode = transaction_write(tdb, off, buf, len2);
206 if (ecode != TDB_SUCCESS) {
212 buf = (const void *)(len2 + (const char *)buf);
220 blk = off / PAGESIZE;
221 off = off % PAGESIZE;
223 if (tdb->transaction->num_blocks <= blk) {
224 uint8_t **new_blocks;
225 /* expand the blocks array */
226 if (tdb->transaction->blocks == NULL) {
227 new_blocks = (uint8_t **)malloc(
228 (blk+1)*sizeof(uint8_t *));
230 new_blocks = (uint8_t **)realloc(
231 tdb->transaction->blocks,
232 (blk+1)*sizeof(uint8_t *));
234 if (new_blocks == NULL) {
235 ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
237 " failed to allocate");
240 memset(&new_blocks[tdb->transaction->num_blocks], 0,
241 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
242 tdb->transaction->blocks = new_blocks;
243 tdb->transaction->num_blocks = blk+1;
244 tdb->transaction->last_block_size = 0;
247 /* allocate and fill a block? */
248 if (tdb->transaction->blocks[blk] == NULL) {
249 tdb->transaction->blocks[blk] = (uint8_t *)calloc(PAGESIZE, 1);
250 if (tdb->transaction->blocks[blk] == NULL) {
251 ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
253 " failed to allocate");
256 if (tdb->transaction->old_map_size > blk * PAGESIZE) {
257 tdb_len_t len2 = PAGESIZE;
258 if (len2 + (blk * PAGESIZE) > tdb->transaction->old_map_size) {
259 len2 = tdb->transaction->old_map_size - (blk * PAGESIZE);
261 ecode = tdb->transaction->io_methods->tread(tdb,
263 tdb->transaction->blocks[blk],
265 if (ecode != TDB_SUCCESS) {
266 ecode = tdb_logerr(tdb, ecode,
270 " read old block: %s",
272 SAFE_FREE(tdb->transaction->blocks[blk]);
275 if (blk == tdb->transaction->num_blocks-1) {
276 tdb->transaction->last_block_size = len2;
281 /* overwrite part of an existing block */
283 memset(tdb->transaction->blocks[blk] + off, 0, len);
285 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
287 if (blk == tdb->transaction->num_blocks-1) {
288 if (len + off > tdb->transaction->last_block_size) {
289 tdb->transaction->last_block_size = len + off;
296 tdb->transaction->transaction_error = 1;
302 write while in a transaction - this variant never expands the transaction blocks, it only
303 updates existing blocks. This means it cannot change the recovery size
305 static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
306 const void *buf, tdb_len_t len)
310 /* break it up into block sized chunks */
311 while (len + (off % PAGESIZE) > PAGESIZE) {
312 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
313 transaction_write_existing(tdb, off, buf, len2);
317 buf = (const void *)(len2 + (const char *)buf);
325 blk = off / PAGESIZE;
326 off = off % PAGESIZE;
328 if (tdb->transaction->num_blocks <= blk ||
329 tdb->transaction->blocks[blk] == NULL) {
333 if (blk == tdb->transaction->num_blocks-1 &&
334 off + len > tdb->transaction->last_block_size) {
335 if (off >= tdb->transaction->last_block_size) {
338 len = tdb->transaction->last_block_size - off;
341 /* overwrite part of an existing block */
342 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
347 out of bounds check during a transaction
349 static enum TDB_ERROR transaction_oob(struct tdb_context *tdb, tdb_off_t len,
352 if (len <= tdb->file->map_size) {
356 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
357 "tdb_oob len %lld beyond transaction size %lld",
359 (long long)tdb->file->map_size);
365 transaction version of tdb_expand().
367 static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
370 enum TDB_ERROR ecode;
372 /* add a write to the transaction elements, so subsequent
373 reads see the zero data */
374 ecode = transaction_write(tdb, tdb->file->map_size, NULL, addition);
375 if (ecode == TDB_SUCCESS) {
376 tdb->file->map_size += addition;
381 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
382 size_t len, bool write_mode)
384 size_t blk = off / PAGESIZE, end_blk;
386 /* This is wrong for zero-length blocks, but will fail gracefully */
387 end_blk = (off + len - 1) / PAGESIZE;
389 /* Can only do direct if in single block and we've already copied. */
393 if (blk >= tdb->transaction->num_blocks)
395 if (tdb->transaction->blocks[blk] == NULL)
397 return tdb->transaction->blocks[blk] + off % PAGESIZE;
400 /* Single which we have copied? */
402 && blk < tdb->transaction->num_blocks
403 && tdb->transaction->blocks[blk])
404 return tdb->transaction->blocks[blk] + off % PAGESIZE;
406 /* Otherwise must be all not copied. */
407 while (blk <= end_blk) {
408 if (blk >= tdb->transaction->num_blocks)
410 if (tdb->transaction->blocks[blk])
414 return tdb->transaction->io_methods->direct(tdb, off, len, false);
417 static const struct tdb_methods transaction_methods = {
421 transaction_expand_file,
428 static enum TDB_ERROR transaction_sync(struct tdb_context *tdb,
429 tdb_off_t offset, tdb_len_t length)
431 if (tdb->flags & TDB_NOSYNC) {
435 if (fsync(tdb->file->fd) != 0) {
436 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
437 "tdb_transaction: fsync failed: %s",
441 if (tdb->file->map_ptr) {
442 tdb_off_t moffset = offset & ~(PAGESIZE-1);
443 if (msync(moffset + (char *)tdb->file->map_ptr,
444 length + (offset - moffset), MS_SYNC) != 0) {
445 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
446 "tdb_transaction: msync failed: %s",
455 static void _tdb_transaction_cancel(struct tdb_context *tdb)
458 enum TDB_ERROR ecode;
460 if (tdb->transaction == NULL) {
461 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
462 "tdb_transaction_cancel: no transaction");
466 if (tdb->transaction->nesting != 0) {
467 tdb->transaction->transaction_error = 1;
468 tdb->transaction->nesting--;
472 tdb->file->map_size = tdb->transaction->old_map_size;
474 /* free all the transaction blocks */
475 for (i=0;i<tdb->transaction->num_blocks;i++) {
476 if (tdb->transaction->blocks[i] != NULL) {
477 free(tdb->transaction->blocks[i]);
480 SAFE_FREE(tdb->transaction->blocks);
482 if (tdb->transaction->magic_offset) {
483 const struct tdb_methods *methods = tdb->transaction->io_methods;
484 uint64_t invalid = TDB_RECOVERY_INVALID_MAGIC;
486 /* remove the recovery marker */
487 ecode = methods->twrite(tdb, tdb->transaction->magic_offset,
488 &invalid, sizeof(invalid));
489 if (ecode == TDB_SUCCESS)
490 ecode = transaction_sync(tdb,
491 tdb->transaction->magic_offset,
493 if (ecode != TDB_SUCCESS) {
494 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
495 "tdb_transaction_cancel: failed to remove"
500 if (tdb->file->allrecord_lock.count)
501 tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
503 /* restore the normal io methods */
504 tdb->methods = tdb->transaction->io_methods;
506 tdb_transaction_unlock(tdb, F_WRLCK);
508 if (tdb_has_open_lock(tdb))
509 tdb_unlock_open(tdb);
511 SAFE_FREE(tdb->transaction);
515 start a tdb transaction. No token is returned, as only a single
516 transaction is allowed to be pending per tdb_context
518 enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
520 enum TDB_ERROR ecode;
522 /* some sanity checks */
523 if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
524 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
526 "tdb_transaction_start:"
529 "read-only or internal db");
532 /* cope with nested tdb_transaction_start() calls */
533 if (tdb->transaction != NULL) {
534 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_IO,
536 "tdb_transaction_start:"
541 if (tdb_has_hash_locks(tdb)) {
542 /* the caller must not have any locks when starting a
543 transaction as otherwise we'll be screwed by lack
544 of nested locks in POSIX */
545 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
547 "tdb_transaction_start:"
549 " transaction with locks"
553 tdb->transaction = (struct tdb_transaction *)
554 calloc(sizeof(struct tdb_transaction), 1);
555 if (tdb->transaction == NULL) {
556 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_OOM,
558 "tdb_transaction_start:"
562 /* get the transaction write lock. This is a blocking lock. As
563 discussed with Volker, there are a number of ways we could
564 make this async, which we will probably do in the future */
565 ecode = tdb_transaction_lock(tdb, F_WRLCK);
566 if (ecode != TDB_SUCCESS) {
567 SAFE_FREE(tdb->transaction->blocks);
568 SAFE_FREE(tdb->transaction);
569 return tdb->last_error = ecode;
572 /* get a read lock over entire file. This is upgraded to a write
573 lock during the commit */
574 ecode = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true);
575 if (ecode != TDB_SUCCESS) {
576 goto fail_allrecord_lock;
579 /* make sure we know about any file expansions already done by
581 tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
582 tdb->transaction->old_map_size = tdb->file->map_size;
584 /* finally hook the io methods, replacing them with
585 transaction specific methods */
586 tdb->transaction->io_methods = tdb->methods;
587 tdb->methods = &transaction_methods;
588 return tdb->last_error = TDB_SUCCESS;
591 tdb_transaction_unlock(tdb, F_WRLCK);
592 SAFE_FREE(tdb->transaction->blocks);
593 SAFE_FREE(tdb->transaction);
594 return tdb->last_error = ecode;
599 cancel the current transaction
601 void tdb_transaction_cancel(struct tdb_context *tdb)
603 _tdb_transaction_cancel(tdb);
607 work out how much space the linearised recovery data will consume
609 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
611 tdb_len_t recovery_size = 0;
614 recovery_size = sizeof(tdb_len_t);
615 for (i=0;i<tdb->transaction->num_blocks;i++) {
616 if (i * PAGESIZE >= tdb->transaction->old_map_size) {
619 if (tdb->transaction->blocks[i] == NULL) {
622 recovery_size += 2*sizeof(tdb_off_t);
623 if (i == tdb->transaction->num_blocks-1) {
624 recovery_size += tdb->transaction->last_block_size;
626 recovery_size += PAGESIZE;
630 return recovery_size;
634 allocate the recovery area, or use an existing recovery area if it is
637 static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
638 tdb_len_t *recovery_size,
639 tdb_off_t *recovery_offset,
640 tdb_len_t *recovery_max_size)
642 struct tdb_recovery_record rec;
643 const struct tdb_methods *methods = tdb->transaction->io_methods;
644 tdb_off_t recovery_head;
646 enum TDB_ERROR ecode;
648 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
649 if (TDB_OFF_IS_ERR(recovery_head)) {
650 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
651 "tdb_recovery_allocate:"
652 " failed to read recovery head");
655 if (recovery_head != 0) {
656 ecode = methods->tread(tdb, recovery_head, &rec, sizeof(rec));
657 if (ecode != TDB_SUCCESS) {
658 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
659 "tdb_recovery_allocate:"
660 " failed to read recovery record");
662 tdb_convert(tdb, &rec, sizeof(rec));
663 /* ignore invalid recovery regions: can happen in crash */
664 if (rec.magic != TDB_RECOVERY_MAGIC &&
665 rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
670 *recovery_size = tdb_recovery_size(tdb);
672 if (recovery_head != 0 && *recovery_size <= rec.max_len) {
673 /* it fits in the existing area */
674 *recovery_max_size = rec.max_len;
675 *recovery_offset = recovery_head;
679 /* we need to free up the old recovery area, then allocate a
680 new one at the end of the file. Note that we cannot use
681 normal allocation to allocate the new one as that might return
682 us an area that is being currently used (as of the start of
684 if (recovery_head != 0) {
686 ecode = add_free_record(tdb, recovery_head,
687 sizeof(rec) + rec.max_len);
688 if (ecode != TDB_SUCCESS) {
689 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
690 "tdb_recovery_allocate:"
691 " failed to free previous"
696 /* the tdb_free() call might have increased the recovery size */
697 *recovery_size = tdb_recovery_size(tdb);
699 /* round up to a multiple of page size */
701 = (((sizeof(rec) + *recovery_size) + PAGESIZE-1)
704 *recovery_offset = tdb->file->map_size;
705 recovery_head = *recovery_offset;
707 /* Restore ->map_size before calling underlying expand_file.
708 Also so that we don't try to expand the file again in the
709 transaction commit, which would destroy the recovery
711 addition = (tdb->file->map_size - tdb->transaction->old_map_size) +
712 sizeof(rec) + *recovery_max_size;
713 tdb->file->map_size = tdb->transaction->old_map_size;
714 ecode = methods->expand_file(tdb, addition);
715 if (ecode != TDB_SUCCESS) {
716 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
717 "tdb_recovery_allocate:"
718 " failed to create recovery area");
721 /* we have to reset the old map size so that we don't try to
722 expand the file again in the transaction commit, which
723 would destroy the recovery area */
724 tdb->transaction->old_map_size = tdb->file->map_size;
726 /* write the recovery header offset and sync - we can sync without a race here
727 as the magic ptr in the recovery record has not been set */
728 tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
729 ecode = methods->twrite(tdb, offsetof(struct tdb_header, recovery),
730 &recovery_head, sizeof(tdb_off_t));
731 if (ecode != TDB_SUCCESS) {
732 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
733 "tdb_recovery_allocate:"
734 " failed to write recovery head");
736 transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
742 /* Set up header for the recovery record. */
743 static void set_recovery_header(struct tdb_recovery_record *rec,
745 uint64_t datalen, uint64_t actuallen,
749 rec->max_len = actuallen;
755 setup the recovery data that will be used on a crash during commit
757 static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
758 tdb_off_t *magic_offset)
760 /* Initialized for GCC's 4.4.5 overzealous uninitialized warnings. */
761 tdb_len_t recovery_size = 0;
762 tdb_off_t recovery_offset = 0, recovery_max_size = 0;
763 unsigned char *data, *p;
764 const struct tdb_methods *methods = tdb->transaction->io_methods;
765 struct tdb_recovery_record *rec;
766 tdb_off_t old_map_size = tdb->transaction->old_map_size;
767 uint64_t magic, tailer;
769 enum TDB_ERROR ecode;
772 check that the recovery area has enough space
774 ecode = tdb_recovery_allocate(tdb, &recovery_size,
775 &recovery_offset, &recovery_max_size);
776 if (ecode != TDB_SUCCESS) {
780 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
782 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
783 "transaction_setup_recovery:"
787 rec = (struct tdb_recovery_record *)data;
788 set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
789 recovery_size, recovery_max_size, old_map_size);
790 tdb_convert(tdb, rec, sizeof(*rec));
792 /* build the recovery data into a single blob to allow us to do a single
793 large write, which should be more efficient */
794 p = data + sizeof(*rec);
795 for (i=0;i<tdb->transaction->num_blocks;i++) {
799 if (tdb->transaction->blocks[i] == NULL) {
803 offset = i * PAGESIZE;
805 if (i == tdb->transaction->num_blocks-1) {
806 length = tdb->transaction->last_block_size;
809 if (offset >= old_map_size) {
812 if (offset + length > tdb->file->map_size) {
814 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
815 "tdb_transaction_setup_recovery:"
816 " transaction data over new region"
819 memcpy(p, &offset, sizeof(offset));
820 memcpy(p + sizeof(offset), &length, sizeof(length));
821 tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
823 /* the recovery area contains the old data, not the
824 new data, so we have to call the original tdb_read
826 ecode = methods->tread(tdb, offset,
827 p + sizeof(offset) + sizeof(length),
829 if (ecode != TDB_SUCCESS) {
833 p += sizeof(offset) + sizeof(length) + length;
837 tailer = sizeof(*rec) + recovery_max_size;
838 memcpy(p, &tailer, sizeof(tailer));
839 tdb_convert(tdb, p, sizeof(tailer));
841 /* write the recovery data to the recovery area */
842 ecode = methods->twrite(tdb, recovery_offset, data,
843 sizeof(*rec) + recovery_size);
844 if (ecode != TDB_SUCCESS) {
846 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
847 "tdb_transaction_setup_recovery:"
848 " failed to write recovery data");
850 transaction_write_existing(tdb, recovery_offset, data,
851 sizeof(*rec) + recovery_size);
853 /* as we don't have ordered writes, we have to sync the recovery
854 data before we update the magic to indicate that the recovery
856 ecode = transaction_sync(tdb, recovery_offset,
857 sizeof(*rec) + recovery_size);
858 if (ecode != TDB_SUCCESS) {
865 magic = TDB_RECOVERY_MAGIC;
866 tdb_convert(tdb, &magic, sizeof(magic));
868 *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
871 ecode = methods->twrite(tdb, *magic_offset, &magic, sizeof(magic));
872 if (ecode != TDB_SUCCESS) {
873 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
874 "tdb_transaction_setup_recovery:"
875 " failed to write recovery magic");
877 transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
879 /* ensure the recovery magic marker is on disk */
880 return transaction_sync(tdb, *magic_offset, sizeof(magic));
883 static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
885 const struct tdb_methods *methods;
886 enum TDB_ERROR ecode;
888 if (tdb->transaction == NULL) {
889 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
890 "tdb_transaction_prepare_commit:"
894 if (tdb->transaction->prepared) {
895 _tdb_transaction_cancel(tdb);
896 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
897 "tdb_transaction_prepare_commit:"
898 " transaction already prepared");
901 if (tdb->transaction->transaction_error) {
902 _tdb_transaction_cancel(tdb);
903 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
904 "tdb_transaction_prepare_commit:"
905 " transaction error pending");
909 if (tdb->transaction->nesting != 0) {
910 tdb->transaction->nesting--;
914 /* check for a null transaction */
915 if (tdb->transaction->blocks == NULL) {
919 methods = tdb->transaction->io_methods;
921 /* upgrade the main transaction lock region to a write lock */
922 ecode = tdb_allrecord_upgrade(tdb);
923 if (ecode != TDB_SUCCESS) {
927 /* get the open lock - this prevents new users attaching to the database
929 ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
930 if (ecode != TDB_SUCCESS) {
934 /* Since we have whole db locked, we don't need the expansion lock. */
935 if (!(tdb->flags & TDB_NOSYNC)) {
936 /* write the recovery data to the end of the file */
937 ecode = transaction_setup_recovery(tdb,
940 if (ecode != TDB_SUCCESS) {
945 tdb->transaction->prepared = true;
947 /* expand the file to the new size if needed */
948 if (tdb->file->map_size != tdb->transaction->old_map_size) {
951 add = tdb->file->map_size - tdb->transaction->old_map_size;
952 /* Restore original map size for tdb_expand_file */
953 tdb->file->map_size = tdb->transaction->old_map_size;
954 ecode = methods->expand_file(tdb, add);
955 if (ecode != TDB_SUCCESS) {
960 /* Keep the open lock until the actual commit */
965 prepare to commit the current transaction
967 enum TDB_ERROR tdb_transaction_prepare_commit(struct tdb_context *tdb)
969 return _tdb_transaction_prepare_commit(tdb);
973 commit the current transaction
975 enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
977 const struct tdb_methods *methods;
979 enum TDB_ERROR ecode;
981 if (tdb->transaction == NULL) {
982 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
984 "tdb_transaction_commit:"
988 tdb_trace(tdb, "tdb_transaction_commit");
990 if (tdb->transaction->nesting != 0) {
991 tdb->transaction->nesting--;
992 return tdb->last_error = TDB_SUCCESS;
995 /* check for a null transaction */
996 if (tdb->transaction->blocks == NULL) {
997 _tdb_transaction_cancel(tdb);
998 return tdb->last_error = TDB_SUCCESS;
1001 if (!tdb->transaction->prepared) {
1002 ecode = _tdb_transaction_prepare_commit(tdb);
1003 if (ecode != TDB_SUCCESS) {
1004 _tdb_transaction_cancel(tdb);
1005 return tdb->last_error = ecode;
1009 methods = tdb->transaction->io_methods;
1011 /* perform all the writes */
1012 for (i=0;i<tdb->transaction->num_blocks;i++) {
1016 if (tdb->transaction->blocks[i] == NULL) {
1020 offset = i * PAGESIZE;
1022 if (i == tdb->transaction->num_blocks-1) {
1023 length = tdb->transaction->last_block_size;
1026 ecode = methods->twrite(tdb, offset,
1027 tdb->transaction->blocks[i], length);
1028 if (ecode != TDB_SUCCESS) {
1029 /* we've overwritten part of the data and
1030 possibly expanded the file, so we need to
1031 run the crash recovery code */
1032 tdb->methods = methods;
1033 tdb_transaction_recover(tdb);
1035 _tdb_transaction_cancel(tdb);
1037 return tdb->last_error = ecode;
1039 SAFE_FREE(tdb->transaction->blocks[i]);
1042 SAFE_FREE(tdb->transaction->blocks);
1043 tdb->transaction->num_blocks = 0;
1045 /* ensure the new data is on disk */
1046 ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1047 if (ecode != TDB_SUCCESS) {
1048 return tdb->last_error = ecode;
1052 TODO: maybe write to some dummy hdr field, or write to magic
1053 offset without mmap, before the last sync, instead of the
1057 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1058 don't change the mtime of the file, this means the file may
1059 not be backed up (as tdb rounding to block sizes means that
1060 file size changes are quite rare too). The following forces
1061 mtime changes when a transaction completes */
1063 utime(tdb->name, NULL);
1066 /* use a transaction cancel to free memory and remove the
1067 transaction locks: it "restores" map_size, too. */
1068 tdb->transaction->old_map_size = tdb->file->map_size;
1069 _tdb_transaction_cancel(tdb);
1071 return tdb->last_error = TDB_SUCCESS;
1076 recover from an aborted transaction. Must be called with exclusive
1077 database write access already established (including the open
1078 lock to prevent new processes attaching)
1080 enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb)
1082 tdb_off_t recovery_head, recovery_eof;
1083 unsigned char *data, *p;
1084 struct tdb_recovery_record rec;
1085 enum TDB_ERROR ecode;
1087 /* find the recovery area */
1088 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1089 if (TDB_OFF_IS_ERR(recovery_head)) {
1090 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
1091 "tdb_transaction_recover:"
1092 " failed to read recovery head");
1095 if (recovery_head == 0) {
1096 /* we have never allocated a recovery record */
1100 /* read the recovery record */
1101 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1102 if (ecode != TDB_SUCCESS) {
1103 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1104 "tdb_transaction_recover:"
1105 " failed to read recovery record");
1108 if (rec.magic != TDB_RECOVERY_MAGIC) {
1109 /* there is no valid recovery data */
1113 if (tdb->read_only) {
1114 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
1115 "tdb_transaction_recover:"
1116 " attempt to recover read only database");
1119 recovery_eof = rec.eof;
1121 data = (unsigned char *)malloc(rec.len);
1123 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
1124 "tdb_transaction_recover:"
1125 " failed to allocate recovery data");
1128 /* read the full recovery data */
1129 ecode = tdb->methods->tread(tdb, recovery_head + sizeof(rec), data,
1131 if (ecode != TDB_SUCCESS) {
1132 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1133 "tdb_transaction_recover:"
1134 " failed to read recovery data");
1137 /* recover the file data */
1139 while (p+sizeof(tdb_off_t)+sizeof(tdb_len_t) < data + rec.len) {
1142 tdb_convert(tdb, p, sizeof(ofs) + sizeof(len));
1143 memcpy(&ofs, p, sizeof(ofs));
1144 memcpy(&len, p + sizeof(ofs), sizeof(len));
1145 p += sizeof(ofs) + sizeof(len);
1147 ecode = tdb->methods->twrite(tdb, ofs, p, len);
1148 if (ecode != TDB_SUCCESS) {
1150 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1151 "tdb_transaction_recover:"
1152 " failed to recover %zu bytes"
1154 (size_t)len, (size_t)ofs);
1161 ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1162 if (ecode != TDB_SUCCESS) {
1163 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1164 "tdb_transaction_recover:"
1165 " failed to sync recovery");
1168 /* if the recovery area is after the recovered eof then remove it */
1169 if (recovery_eof <= recovery_head) {
1170 ecode = tdb_write_off(tdb, offsetof(struct tdb_header,
1173 if (ecode != TDB_SUCCESS) {
1174 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1175 "tdb_transaction_recover:"
1176 " failed to remove recovery head");
1180 /* remove the recovery magic */
1181 ecode = tdb_write_off(tdb,
1183 + offsetof(struct tdb_recovery_record, magic),
1184 TDB_RECOVERY_INVALID_MAGIC);
1185 if (ecode != TDB_SUCCESS) {
1186 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1187 "tdb_transaction_recover:"
1188 " failed to remove recovery magic");
1191 ecode = transaction_sync(tdb, 0, recovery_eof);
1192 if (ecode != TDB_SUCCESS) {
1193 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1194 "tdb_transaction_recover:"
1195 " failed to sync2 recovery");
1198 tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
1199 "tdb_transaction_recover: recovered %zu byte database",
1200 (size_t)recovery_eof);
1206 tdb_bool_err tdb_needs_recovery(struct tdb_context *tdb)
1208 tdb_off_t recovery_head;
1209 struct tdb_recovery_record rec;
1210 enum TDB_ERROR ecode;
1212 /* find the recovery area */
1213 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1214 if (TDB_OFF_IS_ERR(recovery_head)) {
1215 return recovery_head;
1218 if (recovery_head == 0) {
1219 /* we have never allocated a recovery record */
1223 /* read the recovery record */
1224 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1225 if (ecode != TDB_SUCCESS) {
1229 return (rec.magic == TDB_RECOVERY_MAGIC);