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 */
113 unsigned int nesting;
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 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
535 return tdb->last_error
536 = tdb_logerr(tdb, TDB_ERR_IO,
538 "tdb_transaction_start:"
539 " already inside transaction");
541 tdb->transaction->nesting++;
545 if (tdb_has_hash_locks(tdb)) {
546 /* the caller must not have any locks when starting a
547 transaction as otherwise we'll be screwed by lack
548 of nested locks in POSIX */
549 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
551 "tdb_transaction_start:"
553 " transaction with locks"
557 tdb->transaction = (struct tdb_transaction *)
558 calloc(sizeof(struct tdb_transaction), 1);
559 if (tdb->transaction == NULL) {
560 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_OOM,
562 "tdb_transaction_start:"
566 /* get the transaction write lock. This is a blocking lock. As
567 discussed with Volker, there are a number of ways we could
568 make this async, which we will probably do in the future */
569 ecode = tdb_transaction_lock(tdb, F_WRLCK);
570 if (ecode != TDB_SUCCESS) {
571 SAFE_FREE(tdb->transaction->blocks);
572 SAFE_FREE(tdb->transaction);
573 return tdb->last_error = ecode;
576 /* get a read lock over entire file. This is upgraded to a write
577 lock during the commit */
578 ecode = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true);
579 if (ecode != TDB_SUCCESS) {
580 goto fail_allrecord_lock;
583 /* make sure we know about any file expansions already done by
585 tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
586 tdb->transaction->old_map_size = tdb->file->map_size;
588 /* finally hook the io methods, replacing them with
589 transaction specific methods */
590 tdb->transaction->io_methods = tdb->methods;
591 tdb->methods = &transaction_methods;
592 return tdb->last_error = TDB_SUCCESS;
595 tdb_transaction_unlock(tdb, F_WRLCK);
596 SAFE_FREE(tdb->transaction->blocks);
597 SAFE_FREE(tdb->transaction);
598 return tdb->last_error = ecode;
603 cancel the current transaction
605 void tdb_transaction_cancel(struct tdb_context *tdb)
607 _tdb_transaction_cancel(tdb);
611 work out how much space the linearised recovery data will consume
613 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
615 tdb_len_t recovery_size = 0;
618 recovery_size = sizeof(tdb_len_t);
619 for (i=0;i<tdb->transaction->num_blocks;i++) {
620 if (i * PAGESIZE >= tdb->transaction->old_map_size) {
623 if (tdb->transaction->blocks[i] == NULL) {
626 recovery_size += 2*sizeof(tdb_off_t);
627 if (i == tdb->transaction->num_blocks-1) {
628 recovery_size += tdb->transaction->last_block_size;
630 recovery_size += PAGESIZE;
634 return recovery_size;
638 allocate the recovery area, or use an existing recovery area if it is
641 static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
642 tdb_len_t *recovery_size,
643 tdb_off_t *recovery_offset,
644 tdb_len_t *recovery_max_size)
646 struct tdb_recovery_record rec;
647 const struct tdb_methods *methods = tdb->transaction->io_methods;
648 tdb_off_t recovery_head;
650 enum TDB_ERROR ecode;
652 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
653 if (TDB_OFF_IS_ERR(recovery_head)) {
654 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
655 "tdb_recovery_allocate:"
656 " failed to read recovery head");
659 if (recovery_head != 0) {
660 ecode = methods->tread(tdb, recovery_head, &rec, sizeof(rec));
661 if (ecode != TDB_SUCCESS) {
662 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
663 "tdb_recovery_allocate:"
664 " failed to read recovery record");
666 tdb_convert(tdb, &rec, sizeof(rec));
667 /* ignore invalid recovery regions: can happen in crash */
668 if (rec.magic != TDB_RECOVERY_MAGIC &&
669 rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
674 *recovery_size = tdb_recovery_size(tdb);
676 if (recovery_head != 0 && *recovery_size <= rec.max_len) {
677 /* it fits in the existing area */
678 *recovery_max_size = rec.max_len;
679 *recovery_offset = recovery_head;
683 /* we need to free up the old recovery area, then allocate a
684 new one at the end of the file. Note that we cannot use
685 normal allocation to allocate the new one as that might return
686 us an area that is being currently used (as of the start of
688 if (recovery_head != 0) {
690 ecode = add_free_record(tdb, recovery_head,
691 sizeof(rec) + rec.max_len,
693 if (ecode != TDB_SUCCESS) {
694 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
695 "tdb_recovery_allocate:"
696 " failed to free previous"
701 /* the tdb_free() call might have increased the recovery size */
702 *recovery_size = tdb_recovery_size(tdb);
704 /* round up to a multiple of page size */
706 = (((sizeof(rec) + *recovery_size) + PAGESIZE-1)
709 *recovery_offset = tdb->file->map_size;
710 recovery_head = *recovery_offset;
712 /* Restore ->map_size before calling underlying expand_file.
713 Also so that we don't try to expand the file again in the
714 transaction commit, which would destroy the recovery
716 addition = (tdb->file->map_size - tdb->transaction->old_map_size) +
717 sizeof(rec) + *recovery_max_size;
718 tdb->file->map_size = tdb->transaction->old_map_size;
719 ecode = methods->expand_file(tdb, addition);
720 if (ecode != TDB_SUCCESS) {
721 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
722 "tdb_recovery_allocate:"
723 " failed to create recovery area");
726 /* we have to reset the old map size so that we don't try to
727 expand the file again in the transaction commit, which
728 would destroy the recovery area */
729 tdb->transaction->old_map_size = tdb->file->map_size;
731 /* write the recovery header offset and sync - we can sync without a race here
732 as the magic ptr in the recovery record has not been set */
733 tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
734 ecode = methods->twrite(tdb, offsetof(struct tdb_header, recovery),
735 &recovery_head, sizeof(tdb_off_t));
736 if (ecode != TDB_SUCCESS) {
737 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
738 "tdb_recovery_allocate:"
739 " failed to write recovery head");
741 transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
747 /* Set up header for the recovery record. */
748 static void set_recovery_header(struct tdb_recovery_record *rec,
750 uint64_t datalen, uint64_t actuallen,
754 rec->max_len = actuallen;
760 setup the recovery data that will be used on a crash during commit
762 static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
763 tdb_off_t *magic_offset)
765 /* Initialized for GCC's 4.4.5 overzealous uninitialized warnings. */
766 tdb_len_t recovery_size = 0;
767 tdb_off_t recovery_offset = 0, recovery_max_size = 0;
768 unsigned char *data, *p;
769 const struct tdb_methods *methods = tdb->transaction->io_methods;
770 struct tdb_recovery_record *rec;
771 tdb_off_t old_map_size = tdb->transaction->old_map_size;
772 uint64_t magic, tailer;
774 enum TDB_ERROR ecode;
777 check that the recovery area has enough space
779 ecode = tdb_recovery_allocate(tdb, &recovery_size,
780 &recovery_offset, &recovery_max_size);
781 if (ecode != TDB_SUCCESS) {
785 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
787 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
788 "transaction_setup_recovery:"
792 rec = (struct tdb_recovery_record *)data;
793 set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
794 recovery_size, recovery_max_size, old_map_size);
795 tdb_convert(tdb, rec, sizeof(*rec));
797 /* build the recovery data into a single blob to allow us to do a single
798 large write, which should be more efficient */
799 p = data + sizeof(*rec);
800 for (i=0;i<tdb->transaction->num_blocks;i++) {
804 if (tdb->transaction->blocks[i] == NULL) {
808 offset = i * PAGESIZE;
810 if (i == tdb->transaction->num_blocks-1) {
811 length = tdb->transaction->last_block_size;
814 if (offset >= old_map_size) {
817 if (offset + length > tdb->file->map_size) {
819 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
820 "tdb_transaction_setup_recovery:"
821 " transaction data over new region"
824 memcpy(p, &offset, sizeof(offset));
825 memcpy(p + sizeof(offset), &length, sizeof(length));
826 tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
828 /* the recovery area contains the old data, not the
829 new data, so we have to call the original tdb_read
831 ecode = methods->tread(tdb, offset,
832 p + sizeof(offset) + sizeof(length),
834 if (ecode != TDB_SUCCESS) {
838 p += sizeof(offset) + sizeof(length) + length;
842 tailer = sizeof(*rec) + recovery_max_size;
843 memcpy(p, &tailer, sizeof(tailer));
844 tdb_convert(tdb, p, sizeof(tailer));
846 /* write the recovery data to the recovery area */
847 ecode = methods->twrite(tdb, recovery_offset, data,
848 sizeof(*rec) + recovery_size);
849 if (ecode != TDB_SUCCESS) {
851 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
852 "tdb_transaction_setup_recovery:"
853 " failed to write recovery data");
855 transaction_write_existing(tdb, recovery_offset, data,
856 sizeof(*rec) + recovery_size);
858 /* as we don't have ordered writes, we have to sync the recovery
859 data before we update the magic to indicate that the recovery
861 ecode = transaction_sync(tdb, recovery_offset,
862 sizeof(*rec) + recovery_size);
863 if (ecode != TDB_SUCCESS) {
870 magic = TDB_RECOVERY_MAGIC;
871 tdb_convert(tdb, &magic, sizeof(magic));
873 *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
876 ecode = methods->twrite(tdb, *magic_offset, &magic, sizeof(magic));
877 if (ecode != TDB_SUCCESS) {
878 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
879 "tdb_transaction_setup_recovery:"
880 " failed to write recovery magic");
882 transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
884 /* ensure the recovery magic marker is on disk */
885 return transaction_sync(tdb, *magic_offset, sizeof(magic));
888 static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
890 const struct tdb_methods *methods;
891 enum TDB_ERROR ecode;
893 if (tdb->transaction == NULL) {
894 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
895 "tdb_transaction_prepare_commit:"
899 if (tdb->transaction->prepared) {
900 _tdb_transaction_cancel(tdb);
901 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
902 "tdb_transaction_prepare_commit:"
903 " transaction already prepared");
906 if (tdb->transaction->transaction_error) {
907 _tdb_transaction_cancel(tdb);
908 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
909 "tdb_transaction_prepare_commit:"
910 " transaction error pending");
914 if (tdb->transaction->nesting != 0) {
918 /* check for a null transaction */
919 if (tdb->transaction->blocks == NULL) {
923 methods = tdb->transaction->io_methods;
925 /* upgrade the main transaction lock region to a write lock */
926 ecode = tdb_allrecord_upgrade(tdb);
927 if (ecode != TDB_SUCCESS) {
931 /* get the open lock - this prevents new users attaching to the database
933 ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
934 if (ecode != TDB_SUCCESS) {
938 /* Since we have whole db locked, we don't need the expansion lock. */
939 if (!(tdb->flags & TDB_NOSYNC)) {
940 /* write the recovery data to the end of the file */
941 ecode = transaction_setup_recovery(tdb,
944 if (ecode != TDB_SUCCESS) {
949 tdb->transaction->prepared = true;
951 /* expand the file to the new size if needed */
952 if (tdb->file->map_size != tdb->transaction->old_map_size) {
955 add = tdb->file->map_size - tdb->transaction->old_map_size;
956 /* Restore original map size for tdb_expand_file */
957 tdb->file->map_size = tdb->transaction->old_map_size;
958 ecode = methods->expand_file(tdb, add);
959 if (ecode != TDB_SUCCESS) {
964 /* Keep the open lock until the actual commit */
969 prepare to commit the current transaction
971 enum TDB_ERROR tdb_transaction_prepare_commit(struct tdb_context *tdb)
973 return _tdb_transaction_prepare_commit(tdb);
977 commit the current transaction
979 enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
981 const struct tdb_methods *methods;
983 enum TDB_ERROR ecode;
985 if (tdb->transaction == NULL) {
986 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
988 "tdb_transaction_commit:"
992 tdb_trace(tdb, "tdb_transaction_commit");
994 if (tdb->transaction->nesting != 0) {
995 tdb->transaction->nesting--;
996 return tdb->last_error = TDB_SUCCESS;
999 /* check for a null transaction */
1000 if (tdb->transaction->blocks == NULL) {
1001 _tdb_transaction_cancel(tdb);
1002 return tdb->last_error = TDB_SUCCESS;
1005 if (!tdb->transaction->prepared) {
1006 ecode = _tdb_transaction_prepare_commit(tdb);
1007 if (ecode != TDB_SUCCESS) {
1008 _tdb_transaction_cancel(tdb);
1009 return tdb->last_error = ecode;
1013 methods = tdb->transaction->io_methods;
1015 /* perform all the writes */
1016 for (i=0;i<tdb->transaction->num_blocks;i++) {
1020 if (tdb->transaction->blocks[i] == NULL) {
1024 offset = i * PAGESIZE;
1026 if (i == tdb->transaction->num_blocks-1) {
1027 length = tdb->transaction->last_block_size;
1030 ecode = methods->twrite(tdb, offset,
1031 tdb->transaction->blocks[i], length);
1032 if (ecode != TDB_SUCCESS) {
1033 /* we've overwritten part of the data and
1034 possibly expanded the file, so we need to
1035 run the crash recovery code */
1036 tdb->methods = methods;
1037 tdb_transaction_recover(tdb);
1039 _tdb_transaction_cancel(tdb);
1041 return tdb->last_error = ecode;
1043 SAFE_FREE(tdb->transaction->blocks[i]);
1046 SAFE_FREE(tdb->transaction->blocks);
1047 tdb->transaction->num_blocks = 0;
1049 /* ensure the new data is on disk */
1050 ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1051 if (ecode != TDB_SUCCESS) {
1052 return tdb->last_error = ecode;
1056 TODO: maybe write to some dummy hdr field, or write to magic
1057 offset without mmap, before the last sync, instead of the
1061 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1062 don't change the mtime of the file, this means the file may
1063 not be backed up (as tdb rounding to block sizes means that
1064 file size changes are quite rare too). The following forces
1065 mtime changes when a transaction completes */
1067 utime(tdb->name, NULL);
1070 /* use a transaction cancel to free memory and remove the
1071 transaction locks: it "restores" map_size, too. */
1072 tdb->transaction->old_map_size = tdb->file->map_size;
1073 _tdb_transaction_cancel(tdb);
1075 return tdb->last_error = TDB_SUCCESS;
1080 recover from an aborted transaction. Must be called with exclusive
1081 database write access already established (including the open
1082 lock to prevent new processes attaching)
1084 enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb)
1086 tdb_off_t recovery_head, recovery_eof;
1087 unsigned char *data, *p;
1088 struct tdb_recovery_record rec;
1089 enum TDB_ERROR ecode;
1091 /* find the recovery area */
1092 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1093 if (TDB_OFF_IS_ERR(recovery_head)) {
1094 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
1095 "tdb_transaction_recover:"
1096 " failed to read recovery head");
1099 if (recovery_head == 0) {
1100 /* we have never allocated a recovery record */
1104 /* read the recovery record */
1105 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1106 if (ecode != TDB_SUCCESS) {
1107 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1108 "tdb_transaction_recover:"
1109 " failed to read recovery record");
1112 if (rec.magic != TDB_RECOVERY_MAGIC) {
1113 /* there is no valid recovery data */
1117 if (tdb->read_only) {
1118 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
1119 "tdb_transaction_recover:"
1120 " attempt to recover read only database");
1123 recovery_eof = rec.eof;
1125 data = (unsigned char *)malloc(rec.len);
1127 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
1128 "tdb_transaction_recover:"
1129 " failed to allocate recovery data");
1132 /* read the full recovery data */
1133 ecode = tdb->methods->tread(tdb, recovery_head + sizeof(rec), data,
1135 if (ecode != TDB_SUCCESS) {
1136 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1137 "tdb_transaction_recover:"
1138 " failed to read recovery data");
1141 /* recover the file data */
1143 while (p+sizeof(tdb_off_t)+sizeof(tdb_len_t) < data + rec.len) {
1146 tdb_convert(tdb, p, sizeof(ofs) + sizeof(len));
1147 memcpy(&ofs, p, sizeof(ofs));
1148 memcpy(&len, p + sizeof(ofs), sizeof(len));
1149 p += sizeof(ofs) + sizeof(len);
1151 ecode = tdb->methods->twrite(tdb, ofs, p, len);
1152 if (ecode != TDB_SUCCESS) {
1154 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1155 "tdb_transaction_recover:"
1156 " failed to recover %zu bytes"
1158 (size_t)len, (size_t)ofs);
1165 ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1166 if (ecode != TDB_SUCCESS) {
1167 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1168 "tdb_transaction_recover:"
1169 " failed to sync recovery");
1172 /* if the recovery area is after the recovered eof then remove it */
1173 if (recovery_eof <= recovery_head) {
1174 ecode = tdb_write_off(tdb, offsetof(struct tdb_header,
1177 if (ecode != TDB_SUCCESS) {
1178 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1179 "tdb_transaction_recover:"
1180 " failed to remove recovery head");
1184 /* remove the recovery magic */
1185 ecode = tdb_write_off(tdb,
1187 + offsetof(struct tdb_recovery_record, magic),
1188 TDB_RECOVERY_INVALID_MAGIC);
1189 if (ecode != TDB_SUCCESS) {
1190 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1191 "tdb_transaction_recover:"
1192 " failed to remove recovery magic");
1195 ecode = transaction_sync(tdb, 0, recovery_eof);
1196 if (ecode != TDB_SUCCESS) {
1197 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1198 "tdb_transaction_recover:"
1199 " failed to sync2 recovery");
1202 tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
1203 "tdb_transaction_recover: recovered %zu byte database",
1204 (size_t)recovery_eof);
1210 tdb_bool_err tdb_needs_recovery(struct tdb_context *tdb)
1212 tdb_off_t recovery_head;
1213 struct tdb_recovery_record rec;
1214 enum TDB_ERROR ecode;
1216 /* find the recovery area */
1217 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1218 if (TDB_OFF_IS_ERR(recovery_head)) {
1219 return recovery_head;
1222 if (recovery_head == 0) {
1223 /* we have never allocated a recovery record */
1227 /* read the recovery record */
1228 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1229 if (ecode != TDB_SUCCESS) {
1233 return (rec.magic == TDB_RECOVERY_MAGIC);