2 Unix SMB/CIFS implementation.
4 trivial database library
6 Copyright (C) Andrew Tridgell 2005
8 ** NOTE! The following LGPL license applies to the tdb
9 ** library. This does NOT imply that all of Samba is released
12 This library is free software; you can redistribute it and/or
13 modify it under the terms of the GNU Lesser General Public
14 License as published by the Free Software Foundation; either
15 version 3 of the License, or (at your option) any later version.
17 This library is distributed in the hope that it will be useful,
18 but WITHOUT ANY WARRANTY; without even the implied warranty of
19 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 Lesser General Public License for more details.
22 You should have received a copy of the GNU Lesser General Public
23 License along with this library; if not, see <http://www.gnu.org/licenses/>.
26 #include "tdb_private.h"
31 - only allow a single transaction at a time per database. This makes
32 using the transaction API simpler, as otherwise the caller would
33 have to cope with temporary failures in transactions that conflict
34 with other current transactions
36 - keep the transaction recovery information in the same file as the
37 database, using a special 'transaction recovery' record pointed at
38 by the header. This removes the need for extra journal files as
39 used by some other databases
41 - dynamically allocated the transaction recover record, re-using it
42 for subsequent transactions. If a larger record is needed then
43 tdb_free() the old record to place it on the normal tdb freelist
44 before allocating the new record
46 - during transactions, keep a linked list of writes all that have
47 been performed by intercepting all tdb_write() calls. The hooked
48 transaction versions of tdb_read() and tdb_write() check this
49 linked list and try to use the elements of the list in preference
52 - don't allow any locks to be held when a transaction starts,
53 otherwise we can end up with deadlock (plus lack of lock nesting
54 in posix locks would mean the lock is lost)
56 - if the caller gains a lock during the transaction but doesn't
57 release it then fail the commit
59 - allow for nested calls to tdb_transaction_start(), re-using the
60 existing transaction record. If the inner transaction is cancelled
61 then a subsequent commit will fail
63 - keep a mirrored copy of the tdb hash chain heads to allow for the
64 fast hash heads scan on traverse, updating the mirrored copy in
65 the transaction version of tdb_write
67 - allow callers to mix transaction and non-transaction use of tdb,
68 although once a transaction is started then an exclusive lock is
69 gained until the transaction is committed or cancelled
71 - the commit stategy involves first saving away all modified data
72 into a linearised buffer in the transaction recovery area, then
73 marking the transaction recovery area with a magic value to
74 indicate a valid recovery record. In total 4 fsync/msync calls are
75 needed per commit to prevent race conditions. It might be possible
76 to reduce this to 3 or even 2 with some more work.
78 - check for a valid recovery record on open of the tdb, while the
79 open lock is held. Automatically recover from the transaction
80 recovery area if needed, then continue with the open as
81 usual. This allows for smooth crash recovery with no administrator
84 - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85 still available, but no transaction recovery area is used and no
86 fsync/msync calls are made.
88 - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89 tdb_add_flags() transaction nesting is enabled.
90 It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
91 The default is that transaction nesting is allowed.
92 Note: this default may change in future versions of tdb.
94 Beware. when transactions are nested a transaction successfully
95 completed with tdb_transaction_commit() can be silently unrolled later.
97 - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
98 tdb_add_flags() transaction nesting is disabled.
99 It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
100 An attempt create a nested transaction will fail with TDB_ERR_NESTING.
101 The default is that transaction nesting is allowed.
102 Note: this default may change in future versions of tdb.
107 hold the context of any current transaction
109 struct tdb_transaction {
110 /* we keep a mirrored copy of the tdb hash heads here so
111 tdb_next_hash_chain() can operate efficiently */
112 uint32_t *hash_heads;
114 /* the original io methods - used to do IOs to the real db */
115 const struct tdb_methods *io_methods;
117 /* the list of transaction blocks. When a block is first
118 written to, it gets created in this list */
121 uint32_t block_size; /* bytes in each block */
122 uint32_t last_block_size; /* number of valid bytes in the last block */
124 /* non-zero when an internal transaction error has
125 occurred. All write operations will then fail until the
126 transaction is ended */
127 int transaction_error;
129 /* when inside a transaction we need to keep track of any
130 nested tdb_transaction_start() calls, as these are allowed,
131 but don't create a new transaction */
134 /* set when a prepare has already occurred */
136 tdb_off_t magic_offset;
138 /* old file size before transaction */
139 tdb_len_t old_map_size;
141 /* did we expand in this transaction */
147 read while in a transaction. We need to check first if the data is in our list
148 of transaction elements, then if not do a real read
150 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
151 tdb_len_t len, int cv)
155 /* break it down into block sized ops */
156 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
157 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
158 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
163 buf = (void *)(len2 + (char *)buf);
170 blk = off / tdb->transaction->block_size;
172 /* see if we have it in the block list */
173 if (tdb->transaction->num_blocks <= blk ||
174 tdb->transaction->blocks[blk] == NULL) {
175 /* nope, do a real read */
176 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
182 /* it is in the block list. Now check for the last block */
183 if (blk == tdb->transaction->num_blocks-1) {
184 if (len > tdb->transaction->last_block_size) {
189 /* now copy it out of this block */
190 memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
192 tdb_convert(buf, len);
197 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
198 tdb->ecode = TDB_ERR_IO;
199 tdb->transaction->transaction_error = 1;
205 write while in a transaction
207 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
208 const void *buf, tdb_len_t len)
212 /* Only a commit is allowed on a prepared transaction */
213 if (tdb->transaction->prepared) {
214 tdb->ecode = TDB_ERR_EINVAL;
215 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
216 tdb->transaction->transaction_error = 1;
220 /* if the write is to a hash head, then update the transaction
222 if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
223 off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
224 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
225 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
228 /* break it up into block sized chunks */
229 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
230 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
231 if (transaction_write(tdb, off, buf, len2) != 0) {
237 buf = (const void *)(len2 + (const char *)buf);
245 blk = off / tdb->transaction->block_size;
246 off = off % tdb->transaction->block_size;
248 if (tdb->transaction->num_blocks <= blk) {
249 uint8_t **new_blocks;
250 /* expand the blocks array */
251 if (tdb->transaction->blocks == NULL) {
252 new_blocks = (uint8_t **)malloc(
253 (blk+1)*sizeof(uint8_t *));
255 new_blocks = (uint8_t **)realloc(
256 tdb->transaction->blocks,
257 (blk+1)*sizeof(uint8_t *));
259 if (new_blocks == NULL) {
260 tdb->ecode = TDB_ERR_OOM;
263 memset(&new_blocks[tdb->transaction->num_blocks], 0,
264 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
265 tdb->transaction->blocks = new_blocks;
266 tdb->transaction->num_blocks = blk+1;
267 tdb->transaction->last_block_size = 0;
270 /* allocate and fill a block? */
271 if (tdb->transaction->blocks[blk] == NULL) {
272 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
273 if (tdb->transaction->blocks[blk] == NULL) {
274 tdb->ecode = TDB_ERR_OOM;
275 tdb->transaction->transaction_error = 1;
278 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
279 tdb_len_t len2 = tdb->transaction->block_size;
280 if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
281 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
283 if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
284 tdb->transaction->blocks[blk],
286 SAFE_FREE(tdb->transaction->blocks[blk]);
287 tdb->ecode = TDB_ERR_IO;
290 if (blk == tdb->transaction->num_blocks-1) {
291 tdb->transaction->last_block_size = len2;
296 /* overwrite part of an existing block */
298 memset(tdb->transaction->blocks[blk] + off, 0, len);
300 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
302 if (blk == tdb->transaction->num_blocks-1) {
303 if (len + off > tdb->transaction->last_block_size) {
304 tdb->transaction->last_block_size = len + off;
311 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
312 (blk*tdb->transaction->block_size) + off, len));
313 tdb->transaction->transaction_error = 1;
319 write while in a transaction - this varient never expands the transaction blocks, it only
320 updates existing blocks. This means it cannot change the recovery size
322 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
323 const void *buf, tdb_len_t len)
327 /* break it up into block sized chunks */
328 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
329 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
330 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
336 buf = (const void *)(len2 + (const char *)buf);
344 blk = off / tdb->transaction->block_size;
345 off = off % tdb->transaction->block_size;
347 if (tdb->transaction->num_blocks <= blk ||
348 tdb->transaction->blocks[blk] == NULL) {
352 if (blk == tdb->transaction->num_blocks-1 &&
353 off + len > tdb->transaction->last_block_size) {
354 if (off >= tdb->transaction->last_block_size) {
357 len = tdb->transaction->last_block_size - off;
360 /* overwrite part of an existing block */
361 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
368 accelerated hash chain head search, using the cached hash heads
370 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
373 for (;h < tdb->header.hash_size;h++) {
374 /* the +1 takes account of the freelist */
375 if (0 != tdb->transaction->hash_heads[h+1]) {
383 out of bounds check during a transaction
385 static int transaction_oob(struct tdb_context *tdb, tdb_off_t off,
386 tdb_len_t len, int probe)
388 if (off + len >= off && off + len <= tdb->map_size) {
391 tdb->ecode = TDB_ERR_IO;
396 transaction version of tdb_expand().
398 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
401 /* add a write to the transaction elements, so subsequent
402 reads see the zero data */
403 if (transaction_write(tdb, size, NULL, addition) != 0) {
407 tdb->transaction->expanded = true;
412 static const struct tdb_methods transaction_methods = {
415 transaction_next_hash_chain,
417 transaction_expand_file,
422 start a tdb transaction. No token is returned, as only a single
423 transaction is allowed to be pending per tdb_context
425 static int _tdb_transaction_start(struct tdb_context *tdb,
426 enum tdb_lock_flags lockflags)
428 /* some sanity checks */
429 if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
430 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
431 tdb->ecode = TDB_ERR_EINVAL;
435 /* cope with nested tdb_transaction_start() calls */
436 if (tdb->transaction != NULL) {
437 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
438 tdb->ecode = TDB_ERR_NESTING;
441 tdb->transaction->nesting++;
442 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
443 tdb->transaction->nesting));
447 if (tdb_have_extra_locks(tdb)) {
448 /* the caller must not have any locks when starting a
449 transaction as otherwise we'll be screwed by lack
450 of nested locks in posix */
451 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
452 tdb->ecode = TDB_ERR_LOCK;
456 if (tdb->travlocks.next != NULL) {
457 /* you cannot use transactions inside a traverse (although you can use
458 traverse inside a transaction) as otherwise you can end up with
460 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
461 tdb->ecode = TDB_ERR_LOCK;
465 tdb->transaction = (struct tdb_transaction *)
466 calloc(sizeof(struct tdb_transaction), 1);
467 if (tdb->transaction == NULL) {
468 tdb->ecode = TDB_ERR_OOM;
472 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
473 tdb->transaction->block_size = tdb->page_size;
475 /* get the transaction write lock. This is a blocking lock. As
476 discussed with Volker, there are a number of ways we could
477 make this async, which we will probably do in the future */
478 if (tdb_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
479 SAFE_FREE(tdb->transaction->blocks);
480 SAFE_FREE(tdb->transaction);
481 if ((lockflags & TDB_LOCK_WAIT) == 0) {
482 tdb->ecode = TDB_ERR_NOLOCK;
487 /* get a read lock from the freelist to the end of file. This
488 is upgraded to a write lock during the commit */
489 if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
490 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
491 goto fail_allrecord_lock;
494 /* setup a copy of the hash table heads so the hash scan in
495 traverse can be fast */
496 tdb->transaction->hash_heads = (uint32_t *)
497 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
498 if (tdb->transaction->hash_heads == NULL) {
499 tdb->ecode = TDB_ERR_OOM;
502 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
503 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
504 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
505 tdb->ecode = TDB_ERR_IO;
509 /* make sure we know about any file expansions already done by
511 tdb->methods->tdb_oob(tdb, tdb->map_size, 1, 1);
512 tdb->transaction->old_map_size = tdb->map_size;
514 /* finally hook the io methods, replacing them with
515 transaction specific methods */
516 tdb->transaction->io_methods = tdb->methods;
517 tdb->methods = &transaction_methods;
519 /* Trace at the end, so we get sequence number correct. */
520 tdb_trace(tdb, "tdb_transaction_start");
524 tdb_allrecord_unlock(tdb, F_RDLCK, false);
526 tdb_transaction_unlock(tdb, F_WRLCK);
527 SAFE_FREE(tdb->transaction->blocks);
528 SAFE_FREE(tdb->transaction->hash_heads);
529 SAFE_FREE(tdb->transaction);
533 _PUBLIC_ int tdb_transaction_start(struct tdb_context *tdb)
535 return _tdb_transaction_start(tdb, TDB_LOCK_WAIT);
538 _PUBLIC_ int tdb_transaction_start_nonblock(struct tdb_context *tdb)
540 return _tdb_transaction_start(tdb, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
546 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
548 if (tdb->flags & TDB_NOSYNC) {
552 #ifdef HAVE_FDATASYNC
553 if (fdatasync(tdb->fd) != 0) {
555 if (fsync(tdb->fd) != 0) {
557 tdb->ecode = TDB_ERR_IO;
558 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
563 tdb_off_t moffset = offset & ~(tdb->page_size-1);
564 if (msync(moffset + (char *)tdb->map_ptr,
565 length + (offset - moffset), MS_SYNC) != 0) {
566 tdb->ecode = TDB_ERR_IO;
567 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
577 static int _tdb_transaction_cancel(struct tdb_context *tdb)
581 if (tdb->transaction == NULL) {
582 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
586 if (tdb->transaction->nesting != 0) {
587 tdb->transaction->transaction_error = 1;
588 tdb->transaction->nesting--;
592 tdb->map_size = tdb->transaction->old_map_size;
594 /* free all the transaction blocks */
595 for (i=0;i<tdb->transaction->num_blocks;i++) {
596 if (tdb->transaction->blocks[i] != NULL) {
597 free(tdb->transaction->blocks[i]);
600 SAFE_FREE(tdb->transaction->blocks);
602 if (tdb->transaction->magic_offset) {
603 const struct tdb_methods *methods = tdb->transaction->io_methods;
604 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
606 /* remove the recovery marker */
607 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
608 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
609 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
614 /* This also removes the OPEN_LOCK, if we have it. */
615 tdb_release_transaction_locks(tdb);
617 /* restore the normal io methods */
618 tdb->methods = tdb->transaction->io_methods;
620 SAFE_FREE(tdb->transaction->hash_heads);
621 SAFE_FREE(tdb->transaction);
627 cancel the current transaction
629 _PUBLIC_ int tdb_transaction_cancel(struct tdb_context *tdb)
631 tdb_trace(tdb, "tdb_transaction_cancel");
632 return _tdb_transaction_cancel(tdb);
636 work out how much space the linearised recovery data will consume
638 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
640 tdb_len_t recovery_size = 0;
643 recovery_size = sizeof(uint32_t);
644 for (i=0;i<tdb->transaction->num_blocks;i++) {
645 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
648 if (tdb->transaction->blocks[i] == NULL) {
651 recovery_size += 2*sizeof(tdb_off_t);
652 if (i == tdb->transaction->num_blocks-1) {
653 recovery_size += tdb->transaction->last_block_size;
655 recovery_size += tdb->transaction->block_size;
659 return recovery_size;
662 int tdb_recovery_area(struct tdb_context *tdb,
663 const struct tdb_methods *methods,
664 tdb_off_t *recovery_offset,
665 struct tdb_record *rec)
667 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, recovery_offset) == -1) {
671 if (*recovery_offset == 0) {
676 if (methods->tdb_read(tdb, *recovery_offset, rec, sizeof(*rec),
681 /* ignore invalid recovery regions: can happen in crash */
682 if (rec->magic != TDB_RECOVERY_MAGIC &&
683 rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
684 *recovery_offset = 0;
691 allocate the recovery area, or use an existing recovery area if it is
694 static int tdb_recovery_allocate(struct tdb_context *tdb,
695 tdb_len_t *recovery_size,
696 tdb_off_t *recovery_offset,
697 tdb_len_t *recovery_max_size)
699 struct tdb_record rec;
700 const struct tdb_methods *methods = tdb->transaction->io_methods;
701 tdb_off_t recovery_head;
703 if (tdb_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
704 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
708 *recovery_size = tdb_recovery_size(tdb);
710 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
711 /* it fits in the existing area */
712 *recovery_max_size = rec.rec_len;
713 *recovery_offset = recovery_head;
717 /* we need to free up the old recovery area, then allocate a
718 new one at the end of the file. Note that we cannot use
719 tdb_allocate() to allocate the new one as that might return
720 us an area that is being currently used (as of the start of
722 if (recovery_head != 0) {
723 if (tdb_free(tdb, recovery_head, &rec) == -1) {
724 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
729 /* the tdb_free() call might have increased the recovery size */
730 *recovery_size = tdb_recovery_size(tdb);
732 /* round up to a multiple of page size */
733 *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
734 *recovery_offset = tdb->map_size;
735 recovery_head = *recovery_offset;
737 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
738 (tdb->map_size - tdb->transaction->old_map_size) +
739 sizeof(rec) + *recovery_max_size) == -1) {
740 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
744 /* remap the file (if using mmap) */
745 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
747 /* we have to reset the old map size so that we don't try to expand the file
748 again in the transaction commit, which would destroy the recovery area */
749 tdb->transaction->old_map_size = tdb->map_size;
751 /* write the recovery header offset and sync - we can sync without a race here
752 as the magic ptr in the recovery record has not been set */
753 CONVERT(recovery_head);
754 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
755 &recovery_head, sizeof(tdb_off_t)) == -1) {
756 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
759 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
760 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
769 setup the recovery data that will be used on a crash during commit
771 static int transaction_setup_recovery(struct tdb_context *tdb,
772 tdb_off_t *magic_offset)
774 tdb_len_t recovery_size;
775 unsigned char *data, *p;
776 const struct tdb_methods *methods = tdb->transaction->io_methods;
777 struct tdb_record *rec;
778 tdb_off_t recovery_offset, recovery_max_size;
779 tdb_off_t old_map_size = tdb->transaction->old_map_size;
780 uint32_t magic, tailer;
784 check that the recovery area has enough space
786 if (tdb_recovery_allocate(tdb, &recovery_size,
787 &recovery_offset, &recovery_max_size) == -1) {
791 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
793 tdb->ecode = TDB_ERR_OOM;
797 rec = (struct tdb_record *)data;
798 memset(rec, 0, sizeof(*rec));
800 rec->magic = TDB_RECOVERY_INVALID_MAGIC;
801 rec->data_len = recovery_size;
802 rec->rec_len = recovery_max_size;
803 rec->key_len = old_map_size;
806 /* build the recovery data into a single blob to allow us to do a single
807 large write, which should be more efficient */
808 p = data + sizeof(*rec);
809 for (i=0;i<tdb->transaction->num_blocks;i++) {
813 if (tdb->transaction->blocks[i] == NULL) {
817 offset = i * tdb->transaction->block_size;
818 length = tdb->transaction->block_size;
819 if (i == tdb->transaction->num_blocks-1) {
820 length = tdb->transaction->last_block_size;
823 if (offset >= old_map_size) {
826 if (offset + length > tdb->transaction->old_map_size) {
827 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
829 tdb->ecode = TDB_ERR_CORRUPT;
832 memcpy(p, &offset, 4);
833 memcpy(p+4, &length, 4);
837 /* the recovery area contains the old data, not the
838 new data, so we have to call the original tdb_read
840 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
842 tdb->ecode = TDB_ERR_IO;
849 tailer = sizeof(*rec) + recovery_max_size;
850 memcpy(p, &tailer, 4);
855 /* write the recovery data to the recovery area */
856 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
857 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
859 tdb->ecode = TDB_ERR_IO;
862 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
863 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
865 tdb->ecode = TDB_ERR_IO;
869 /* as we don't have ordered writes, we have to sync the recovery
870 data before we update the magic to indicate that the recovery
872 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
879 magic = TDB_RECOVERY_MAGIC;
882 *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
884 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
885 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
886 tdb->ecode = TDB_ERR_IO;
889 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
890 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
891 tdb->ecode = TDB_ERR_IO;
895 /* ensure the recovery magic marker is on disk */
896 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
903 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
905 const struct tdb_methods *methods;
907 if (tdb->transaction == NULL) {
908 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
912 if (tdb->transaction->prepared) {
913 tdb->ecode = TDB_ERR_EINVAL;
914 _tdb_transaction_cancel(tdb);
915 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
919 if (tdb->transaction->transaction_error) {
920 tdb->ecode = TDB_ERR_IO;
921 _tdb_transaction_cancel(tdb);
922 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
927 if (tdb->transaction->nesting != 0) {
931 /* check for a null transaction */
932 if (tdb->transaction->blocks == NULL) {
936 methods = tdb->transaction->io_methods;
938 /* if there are any locks pending then the caller has not
939 nested their locks properly, so fail the transaction */
940 if (tdb_have_extra_locks(tdb)) {
941 tdb->ecode = TDB_ERR_LOCK;
942 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
943 _tdb_transaction_cancel(tdb);
947 /* upgrade the main transaction lock region to a write lock */
948 if (tdb_allrecord_upgrade(tdb) == -1) {
949 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
950 _tdb_transaction_cancel(tdb);
954 /* get the open lock - this prevents new users attaching to the database
956 if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
957 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
958 _tdb_transaction_cancel(tdb);
962 if (!(tdb->flags & TDB_NOSYNC)) {
963 /* write the recovery data to the end of the file */
964 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
965 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
966 _tdb_transaction_cancel(tdb);
971 tdb->transaction->prepared = true;
973 /* expand the file to the new size if needed */
974 if (tdb->map_size != tdb->transaction->old_map_size) {
975 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
977 tdb->transaction->old_map_size) == -1) {
978 tdb->ecode = TDB_ERR_IO;
979 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
980 _tdb_transaction_cancel(tdb);
983 tdb->map_size = tdb->transaction->old_map_size;
984 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
987 /* Keep the open lock until the actual commit */
993 prepare to commit the current transaction
995 _PUBLIC_ int tdb_transaction_prepare_commit(struct tdb_context *tdb)
997 tdb_trace(tdb, "tdb_transaction_prepare_commit");
998 return _tdb_transaction_prepare_commit(tdb);
1001 /* A repack is worthwhile if the largest is less than half total free. */
1002 static bool repack_worthwhile(struct tdb_context *tdb)
1005 struct tdb_record rec;
1006 tdb_len_t total = 0, largest = 0;
1008 if (tdb_ofs_read(tdb, FREELIST_TOP, &ptr) == -1) {
1012 while (ptr != 0 && tdb_rec_free_read(tdb, ptr, &rec) == 0) {
1013 total += rec.rec_len;
1014 if (rec.rec_len > largest) {
1015 largest = rec.rec_len;
1020 return total > largest * 2;
1024 commit the current transaction
1026 _PUBLIC_ int tdb_transaction_commit(struct tdb_context *tdb)
1028 const struct tdb_methods *methods;
1030 bool need_repack = false;
1032 if (tdb->transaction == NULL) {
1033 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1037 tdb_trace(tdb, "tdb_transaction_commit");
1039 if (tdb->transaction->transaction_error) {
1040 tdb->ecode = TDB_ERR_IO;
1041 _tdb_transaction_cancel(tdb);
1042 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1047 if (tdb->transaction->nesting != 0) {
1048 tdb->transaction->nesting--;
1052 /* check for a null transaction */
1053 if (tdb->transaction->blocks == NULL) {
1054 _tdb_transaction_cancel(tdb);
1058 if (!tdb->transaction->prepared) {
1059 int ret = _tdb_transaction_prepare_commit(tdb);
1064 methods = tdb->transaction->io_methods;
1066 /* perform all the writes */
1067 for (i=0;i<tdb->transaction->num_blocks;i++) {
1071 if (tdb->transaction->blocks[i] == NULL) {
1075 offset = i * tdb->transaction->block_size;
1076 length = tdb->transaction->block_size;
1077 if (i == tdb->transaction->num_blocks-1) {
1078 length = tdb->transaction->last_block_size;
1081 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1082 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1084 /* we've overwritten part of the data and
1085 possibly expanded the file, so we need to
1086 run the crash recovery code */
1087 tdb->methods = methods;
1088 tdb_transaction_recover(tdb);
1090 _tdb_transaction_cancel(tdb);
1092 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1095 SAFE_FREE(tdb->transaction->blocks[i]);
1098 /* Do this before we drop lock or blocks. */
1099 if (tdb->transaction->expanded) {
1100 need_repack = repack_worthwhile(tdb);
1103 SAFE_FREE(tdb->transaction->blocks);
1104 tdb->transaction->num_blocks = 0;
1106 /* ensure the new data is on disk */
1107 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1112 TODO: maybe write to some dummy hdr field, or write to magic
1113 offset without mmap, before the last sync, instead of the
1117 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1118 don't change the mtime of the file, this means the file may
1119 not be backed up (as tdb rounding to block sizes means that
1120 file size changes are quite rare too). The following forces
1121 mtime changes when a transaction completes */
1123 utime(tdb->name, NULL);
1126 /* use a transaction cancel to free memory and remove the
1127 transaction locks */
1128 _tdb_transaction_cancel(tdb);
1131 return tdb_repack(tdb);
1139 recover from an aborted transaction. Must be called with exclusive
1140 database write access already established (including the open
1141 lock to prevent new processes attaching)
1143 int tdb_transaction_recover(struct tdb_context *tdb)
1145 tdb_off_t recovery_head, recovery_eof;
1146 unsigned char *data, *p;
1148 struct tdb_record rec;
1150 /* find the recovery area */
1151 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1152 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1153 tdb->ecode = TDB_ERR_IO;
1157 if (recovery_head == 0) {
1158 /* we have never allocated a recovery record */
1162 /* read the recovery record */
1163 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1164 sizeof(rec), DOCONV()) == -1) {
1165 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1166 tdb->ecode = TDB_ERR_IO;
1170 if (rec.magic != TDB_RECOVERY_MAGIC) {
1171 /* there is no valid recovery data */
1175 if (tdb->read_only) {
1176 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1177 tdb->ecode = TDB_ERR_CORRUPT;
1181 recovery_eof = rec.key_len;
1183 data = (unsigned char *)malloc(rec.data_len);
1185 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1186 tdb->ecode = TDB_ERR_OOM;
1190 /* read the full recovery data */
1191 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1192 rec.data_len, 0) == -1) {
1193 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1194 tdb->ecode = TDB_ERR_IO;
1198 /* recover the file data */
1200 while (p+8 < data + rec.data_len) {
1206 memcpy(&len, p+4, 4);
1208 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1210 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1211 tdb->ecode = TDB_ERR_IO;
1219 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1220 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1221 tdb->ecode = TDB_ERR_IO;
1225 /* if the recovery area is after the recovered eof then remove it */
1226 if (recovery_eof <= recovery_head) {
1227 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1228 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1229 tdb->ecode = TDB_ERR_IO;
1234 /* remove the recovery magic */
1235 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1237 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1238 tdb->ecode = TDB_ERR_IO;
1242 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1243 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1244 tdb->ecode = TDB_ERR_IO;
1248 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",
1255 /* Any I/O failures we say "needs recovery". */
1256 bool tdb_needs_recovery(struct tdb_context *tdb)
1258 tdb_off_t recovery_head;
1259 struct tdb_record rec;
1261 /* find the recovery area */
1262 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1266 if (recovery_head == 0) {
1267 /* we have never allocated a recovery record */
1271 /* read the recovery record */
1272 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1273 sizeof(rec), DOCONV()) == -1) {
1277 return (rec.magic == TDB_RECOVERY_MAGIC);