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 fsync/msync calls are made. This means we
86 are still proof against a process dying during transaction commit,
87 but not against machine reboot.
89 - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
90 tdb_add_flags() transaction nesting is enabled.
91 It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
92 The default is that transaction nesting is allowed.
93 Note: this default may change in future versions of tdb.
95 Beware. when transactions are nested a transaction successfully
96 completed with tdb_transaction_commit() can be silently unrolled later.
98 - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
99 tdb_add_flags() transaction nesting is disabled.
100 It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
101 An attempt create a nested transaction will fail with TDB_ERR_NESTING.
102 The default is that transaction nesting is allowed.
103 Note: this default may change in future versions of tdb.
108 hold the context of any current transaction
110 struct tdb_transaction {
111 /* we keep a mirrored copy of the tdb hash heads here so
112 tdb_next_hash_chain() can operate efficiently */
113 uint32_t *hash_heads;
115 /* the original io methods - used to do IOs to the real db */
116 const struct tdb_methods *io_methods;
118 /* the list of transaction blocks. When a block is first
119 written to, it gets created in this list */
122 uint32_t block_size; /* bytes in each block */
123 uint32_t last_block_size; /* number of valid bytes in the last block */
125 /* non-zero when an internal transaction error has
126 occurred. All write operations will then fail until the
127 transaction is ended */
128 int transaction_error;
130 /* when inside a transaction we need to keep track of any
131 nested tdb_transaction_start() calls, as these are allowed,
132 but don't create a new transaction */
135 /* set when a prepare has already occurred */
137 tdb_off_t magic_offset;
139 /* old file size before transaction */
140 tdb_len_t old_map_size;
142 /* did we expand in this transaction */
148 read while in a transaction. We need to check first if the data is in our list
149 of transaction elements, then if not do a real read
151 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
152 tdb_len_t len, int cv)
156 /* break it down into block sized ops */
157 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
158 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
159 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
164 buf = (void *)(len2 + (char *)buf);
171 blk = off / tdb->transaction->block_size;
173 /* see if we have it in the block list */
174 if (tdb->transaction->num_blocks <= blk ||
175 tdb->transaction->blocks[blk] == NULL) {
176 /* nope, do a real read */
177 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
183 /* it is in the block list. Now check for the last block */
184 if (blk == tdb->transaction->num_blocks-1) {
185 if (len > tdb->transaction->last_block_size) {
190 /* now copy it out of this block */
191 memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
193 tdb_convert(buf, len);
198 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%u len=%u\n", off, len));
199 tdb->ecode = TDB_ERR_IO;
200 tdb->transaction->transaction_error = 1;
206 write while in a transaction
208 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
209 const void *buf, tdb_len_t len)
213 /* Only a commit is allowed on a prepared transaction */
214 if (tdb->transaction->prepared) {
215 tdb->ecode = TDB_ERR_EINVAL;
216 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
217 tdb->transaction->transaction_error = 1;
221 /* if the write is to a hash head, then update the transaction
223 if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
224 off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
225 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
226 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
229 /* break it up into block sized chunks */
230 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
231 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
232 if (transaction_write(tdb, off, buf, len2) != 0) {
238 buf = (const void *)(len2 + (const char *)buf);
246 blk = off / tdb->transaction->block_size;
247 off = off % tdb->transaction->block_size;
249 if (tdb->transaction->num_blocks <= blk) {
250 uint8_t **new_blocks;
251 /* expand the blocks array */
252 new_blocks = (uint8_t **)realloc(tdb->transaction->blocks,
253 (blk+1)*sizeof(uint8_t *));
254 if (new_blocks == NULL) {
255 tdb->ecode = TDB_ERR_OOM;
258 memset(&new_blocks[tdb->transaction->num_blocks], 0,
259 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
260 tdb->transaction->blocks = new_blocks;
261 tdb->transaction->num_blocks = blk+1;
262 tdb->transaction->last_block_size = 0;
265 /* allocate and fill a block? */
266 if (tdb->transaction->blocks[blk] == NULL) {
267 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
268 if (tdb->transaction->blocks[blk] == NULL) {
269 tdb->ecode = TDB_ERR_OOM;
270 tdb->transaction->transaction_error = 1;
273 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
274 tdb_len_t len2 = tdb->transaction->block_size;
275 if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
276 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
278 if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
279 tdb->transaction->blocks[blk],
281 SAFE_FREE(tdb->transaction->blocks[blk]);
282 tdb->ecode = TDB_ERR_IO;
285 if (blk == tdb->transaction->num_blocks-1) {
286 tdb->transaction->last_block_size = len2;
291 /* overwrite part of an existing block */
293 memset(tdb->transaction->blocks[blk] + off, 0, len);
295 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
297 if (blk == tdb->transaction->num_blocks-1) {
298 if (len + off > tdb->transaction->last_block_size) {
299 tdb->transaction->last_block_size = len + off;
306 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%u len=%u\n",
307 (blk*tdb->transaction->block_size) + off, len));
308 tdb->transaction->transaction_error = 1;
314 write while in a transaction - this variant never expands the transaction blocks, it only
315 updates existing blocks. This means it cannot change the recovery size
317 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
318 const void *buf, tdb_len_t len)
322 /* break it up into block sized chunks */
323 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
324 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
325 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
331 buf = (const void *)(len2 + (const char *)buf);
339 blk = off / tdb->transaction->block_size;
340 off = off % tdb->transaction->block_size;
342 if (tdb->transaction->num_blocks <= blk ||
343 tdb->transaction->blocks[blk] == NULL) {
347 if (blk == tdb->transaction->num_blocks-1 &&
348 off + len > tdb->transaction->last_block_size) {
349 if (off >= tdb->transaction->last_block_size) {
352 len = tdb->transaction->last_block_size - off;
355 /* overwrite part of an existing block */
356 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
363 accelerated hash chain head search, using the cached hash heads
365 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
368 for (;h < tdb->hash_size;h++) {
369 /* the +1 takes account of the freelist */
370 if (0 != tdb->transaction->hash_heads[h+1]) {
378 out of bounds check during a transaction
380 static int transaction_oob(struct tdb_context *tdb, tdb_off_t off,
381 tdb_len_t len, int probe)
383 if (off + len >= off && off + len <= tdb->map_size) {
386 tdb->ecode = TDB_ERR_IO;
391 transaction version of tdb_expand().
393 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
396 const char buf_zero[8192] = {0};
397 size_t buf_len = sizeof(buf_zero);
399 while (addition > 0) {
400 size_t n = MIN(addition, buf_len);
403 ret = transaction_write(tdb, size, buf_zero, n);
412 tdb->transaction->expanded = true;
417 static const struct tdb_methods transaction_methods = {
420 transaction_next_hash_chain,
422 transaction_expand_file,
426 * Is a transaction currently active on this context?
429 _PUBLIC_ bool tdb_transaction_active(struct tdb_context *tdb)
431 return (tdb->transaction != NULL);
435 start a tdb transaction. No token is returned, as only a single
436 transaction is allowed to be pending per tdb_context
438 static int _tdb_transaction_start(struct tdb_context *tdb,
439 enum tdb_lock_flags lockflags)
441 /* some sanity checks */
442 if (tdb->read_only || (tdb->flags & TDB_INTERNAL)
443 || tdb->traverse_read) {
444 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
445 tdb->ecode = TDB_ERR_EINVAL;
449 /* cope with nested tdb_transaction_start() calls */
450 if (tdb->transaction != NULL) {
451 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
452 tdb->ecode = TDB_ERR_NESTING;
455 tdb->transaction->nesting++;
456 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
457 tdb->transaction->nesting));
461 if (tdb_have_extra_locks(tdb)) {
462 /* the caller must not have any locks when starting a
463 transaction as otherwise we'll be screwed by lack
464 of nested locks in posix */
465 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
466 tdb->ecode = TDB_ERR_LOCK;
470 if (tdb->travlocks.next != NULL) {
471 /* you cannot use transactions inside a traverse (although you can use
472 traverse inside a transaction) as otherwise you can end up with
474 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
475 tdb->ecode = TDB_ERR_LOCK;
479 tdb->transaction = (struct tdb_transaction *)
480 calloc(sizeof(struct tdb_transaction), 1);
481 if (tdb->transaction == NULL) {
482 tdb->ecode = TDB_ERR_OOM;
486 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
487 tdb->transaction->block_size = tdb->page_size;
489 /* get the transaction write lock. This is a blocking lock. As
490 discussed with Volker, there are a number of ways we could
491 make this async, which we will probably do in the future */
492 if (tdb_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
493 SAFE_FREE(tdb->transaction->blocks);
494 SAFE_FREE(tdb->transaction);
495 if ((lockflags & TDB_LOCK_WAIT) == 0) {
496 tdb->ecode = TDB_ERR_NOLOCK;
498 TDB_LOG((tdb, TDB_DEBUG_ERROR,
499 "tdb_transaction_start: "
500 "failed to get transaction lock\n"));
505 /* get a read lock from the freelist to the end of file. This
506 is upgraded to a write lock during the commit */
507 if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
508 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
509 goto fail_allrecord_lock;
512 /* setup a copy of the hash table heads so the hash scan in
513 traverse can be fast */
514 tdb->transaction->hash_heads = (uint32_t *)
515 calloc(tdb->hash_size+1, sizeof(uint32_t));
516 if (tdb->transaction->hash_heads == NULL) {
517 tdb->ecode = TDB_ERR_OOM;
520 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
521 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
522 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
523 tdb->ecode = TDB_ERR_IO;
527 /* make sure we know about any file expansions already done by
529 tdb->methods->tdb_oob(tdb, tdb->map_size, 1, 1);
530 tdb->transaction->old_map_size = tdb->map_size;
532 /* finally hook the io methods, replacing them with
533 transaction specific methods */
534 tdb->transaction->io_methods = tdb->methods;
535 tdb->methods = &transaction_methods;
537 /* Trace at the end, so we get sequence number correct. */
538 tdb_trace(tdb, "tdb_transaction_start");
542 tdb_allrecord_unlock(tdb, F_RDLCK, false);
544 tdb_transaction_unlock(tdb, F_WRLCK);
545 SAFE_FREE(tdb->transaction->blocks);
546 SAFE_FREE(tdb->transaction->hash_heads);
547 SAFE_FREE(tdb->transaction);
551 _PUBLIC_ int tdb_transaction_start(struct tdb_context *tdb)
553 return _tdb_transaction_start(tdb, TDB_LOCK_WAIT);
556 _PUBLIC_ int tdb_transaction_start_nonblock(struct tdb_context *tdb)
558 return _tdb_transaction_start(tdb, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
564 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
566 if (tdb->flags & TDB_NOSYNC) {
570 #ifdef HAVE_FDATASYNC
571 if (fdatasync(tdb->fd) != 0) {
573 if (fsync(tdb->fd) != 0) {
575 tdb->ecode = TDB_ERR_IO;
576 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
581 tdb_off_t moffset = offset & ~(tdb->page_size-1);
582 if (msync(moffset + (char *)tdb->map_ptr,
583 length + (offset - moffset), MS_SYNC) != 0) {
584 tdb->ecode = TDB_ERR_IO;
585 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
595 static int _tdb_transaction_cancel(struct tdb_context *tdb)
599 if (tdb->transaction == NULL) {
600 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
604 if (tdb->transaction->nesting != 0) {
605 tdb->transaction->transaction_error = 1;
606 tdb->transaction->nesting--;
610 tdb->map_size = tdb->transaction->old_map_size;
612 /* free all the transaction blocks */
613 for (i=0;i<tdb->transaction->num_blocks;i++) {
614 if (tdb->transaction->blocks[i] != NULL) {
615 free(tdb->transaction->blocks[i]);
618 SAFE_FREE(tdb->transaction->blocks);
620 if (tdb->transaction->magic_offset) {
621 const struct tdb_methods *methods = tdb->transaction->io_methods;
622 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
624 /* remove the recovery marker */
625 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
626 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
627 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
632 /* This also removes the OPEN_LOCK, if we have it. */
633 tdb_release_transaction_locks(tdb);
635 /* restore the normal io methods */
636 tdb->methods = tdb->transaction->io_methods;
638 SAFE_FREE(tdb->transaction->hash_heads);
639 SAFE_FREE(tdb->transaction);
645 cancel the current transaction
647 _PUBLIC_ int tdb_transaction_cancel(struct tdb_context *tdb)
649 tdb_trace(tdb, "tdb_transaction_cancel");
650 return _tdb_transaction_cancel(tdb);
654 work out how much space the linearised recovery data will consume
656 static bool tdb_recovery_size(struct tdb_context *tdb, tdb_len_t *result)
658 tdb_len_t recovery_size = 0;
661 recovery_size = sizeof(uint32_t);
662 for (i=0;i<tdb->transaction->num_blocks;i++) {
663 tdb_len_t block_size;
664 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
667 if (tdb->transaction->blocks[i] == NULL) {
670 if (!tdb_add_len_t(recovery_size, 2*sizeof(tdb_off_t),
674 if (i == tdb->transaction->num_blocks-1) {
675 block_size = tdb->transaction->last_block_size;
677 block_size = tdb->transaction->block_size;
679 if (!tdb_add_len_t(recovery_size, block_size,
685 *result = recovery_size;
689 int tdb_recovery_area(struct tdb_context *tdb,
690 const struct tdb_methods *methods,
691 tdb_off_t *recovery_offset,
692 struct tdb_record *rec)
694 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, recovery_offset) == -1) {
698 if (*recovery_offset == 0) {
703 if (methods->tdb_read(tdb, *recovery_offset, rec, sizeof(*rec),
708 /* ignore invalid recovery regions: can happen in crash */
709 if (rec->magic != TDB_RECOVERY_MAGIC &&
710 rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
711 *recovery_offset = 0;
718 allocate the recovery area, or use an existing recovery area if it is
721 static int tdb_recovery_allocate(struct tdb_context *tdb,
722 tdb_len_t *recovery_size,
723 tdb_off_t *recovery_offset,
724 tdb_len_t *recovery_max_size)
726 struct tdb_record rec;
727 const struct tdb_methods *methods = tdb->transaction->io_methods;
728 tdb_off_t recovery_head, new_end;
730 if (tdb_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
731 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
735 if (!tdb_recovery_size(tdb, recovery_size)) {
736 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
737 "overflow recovery size\n"));
741 /* Existing recovery area? */
742 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
743 /* it fits in the existing area */
744 *recovery_max_size = rec.rec_len;
745 *recovery_offset = recovery_head;
749 /* If recovery area in middle of file, we need a new one. */
750 if (recovery_head == 0
751 || recovery_head + sizeof(rec) + rec.rec_len != tdb->map_size) {
752 /* we need to free up the old recovery area, then allocate a
753 new one at the end of the file. Note that we cannot use
754 tdb_allocate() to allocate the new one as that might return
755 us an area that is being currently used (as of the start of
758 if (tdb_free(tdb, recovery_head, &rec) == -1) {
759 TDB_LOG((tdb, TDB_DEBUG_FATAL,
760 "tdb_recovery_allocate: failed to"
761 " free previous recovery area\n"));
765 /* the tdb_free() call might have increased
766 * the recovery size */
767 if (!tdb_recovery_size(tdb, recovery_size)) {
768 TDB_LOG((tdb, TDB_DEBUG_FATAL,
769 "tdb_recovery_allocate: "
770 "overflow recovery size\n"));
775 /* New head will be at end of file. */
776 recovery_head = tdb->map_size;
779 /* Now we know where it will be. */
780 *recovery_offset = recovery_head;
782 /* Expand by more than we need, so we don't do it often. */
783 *recovery_max_size = tdb_expand_adjust(tdb->map_size,
788 if (!tdb_add_off_t(recovery_head, sizeof(rec), &new_end) ||
789 !tdb_add_off_t(new_end, *recovery_max_size, &new_end)) {
790 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
791 "overflow recovery area\n"));
795 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
796 new_end - tdb->transaction->old_map_size)
798 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
802 /* remap the file (if using mmap) */
803 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
805 /* we have to reset the old map size so that we don't try to expand the file
806 again in the transaction commit, which would destroy the recovery area */
807 tdb->transaction->old_map_size = tdb->map_size;
809 /* write the recovery header offset and sync - we can sync without a race here
810 as the magic ptr in the recovery record has not been set */
811 CONVERT(recovery_head);
812 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
813 &recovery_head, sizeof(tdb_off_t)) == -1) {
814 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
817 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
818 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
827 setup the recovery data that will be used on a crash during commit
829 static int transaction_setup_recovery(struct tdb_context *tdb,
830 tdb_off_t *magic_offset)
832 tdb_len_t recovery_size;
833 unsigned char *data, *p;
834 const struct tdb_methods *methods = tdb->transaction->io_methods;
835 struct tdb_record *rec;
836 tdb_off_t recovery_offset, recovery_max_size;
837 tdb_off_t old_map_size = tdb->transaction->old_map_size;
838 uint32_t magic, tailer;
842 check that the recovery area has enough space
844 if (tdb_recovery_allocate(tdb, &recovery_size,
845 &recovery_offset, &recovery_max_size) == -1) {
849 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
851 tdb->ecode = TDB_ERR_OOM;
855 rec = (struct tdb_record *)data;
856 memset(rec, 0, sizeof(*rec));
858 rec->magic = TDB_RECOVERY_INVALID_MAGIC;
859 rec->data_len = recovery_size;
860 rec->rec_len = recovery_max_size;
861 rec->key_len = old_map_size;
864 /* build the recovery data into a single blob to allow us to do a single
865 large write, which should be more efficient */
866 p = data + sizeof(*rec);
867 for (i=0;i<tdb->transaction->num_blocks;i++) {
871 if (tdb->transaction->blocks[i] == NULL) {
875 offset = i * tdb->transaction->block_size;
876 length = tdb->transaction->block_size;
877 if (i == tdb->transaction->num_blocks-1) {
878 length = tdb->transaction->last_block_size;
881 if (offset >= old_map_size) {
884 if (offset + length > tdb->transaction->old_map_size) {
885 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
887 tdb->ecode = TDB_ERR_CORRUPT;
890 memcpy(p, &offset, 4);
891 memcpy(p+4, &length, 4);
895 /* the recovery area contains the old data, not the
896 new data, so we have to call the original tdb_read
898 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
900 tdb->ecode = TDB_ERR_IO;
907 tailer = sizeof(*rec) + recovery_max_size;
908 memcpy(p, &tailer, 4);
913 /* write the recovery data to the recovery area */
914 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
915 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
917 tdb->ecode = TDB_ERR_IO;
920 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
921 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
923 tdb->ecode = TDB_ERR_IO;
927 /* as we don't have ordered writes, we have to sync the recovery
928 data before we update the magic to indicate that the recovery
930 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
937 magic = TDB_RECOVERY_MAGIC;
940 *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
942 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
943 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
944 tdb->ecode = TDB_ERR_IO;
947 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
948 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
949 tdb->ecode = TDB_ERR_IO;
953 /* ensure the recovery magic marker is on disk */
954 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
961 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
963 const struct tdb_methods *methods;
965 if (tdb->transaction == NULL) {
966 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
970 if (tdb->transaction->prepared) {
971 tdb->ecode = TDB_ERR_EINVAL;
972 _tdb_transaction_cancel(tdb);
973 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
977 if (tdb->transaction->transaction_error) {
978 tdb->ecode = TDB_ERR_IO;
979 _tdb_transaction_cancel(tdb);
980 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
985 if (tdb->transaction->nesting != 0) {
989 /* check for a null transaction */
990 if (tdb->transaction->blocks == NULL) {
994 methods = tdb->transaction->io_methods;
996 /* if there are any locks pending then the caller has not
997 nested their locks properly, so fail the transaction */
998 if (tdb_have_extra_locks(tdb)) {
999 tdb->ecode = TDB_ERR_LOCK;
1000 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
1001 _tdb_transaction_cancel(tdb);
1005 /* upgrade the main transaction lock region to a write lock */
1006 if (tdb_allrecord_upgrade(tdb) == -1) {
1007 if (tdb->ecode == TDB_ERR_RDONLY && tdb->read_only) {
1008 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1009 "tdb_transaction_prepare_commit: "
1010 "failed to upgrade hash locks: "
1011 "database is read only\n"));
1012 } else if (tdb->ecode == TDB_ERR_RDONLY
1013 && tdb->traverse_read) {
1014 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1015 "tdb_transaction_prepare_commit: "
1016 "failed to upgrade hash locks: "
1017 "a database traverse is in progress\n"));
1019 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1020 "tdb_transaction_prepare_commit: "
1021 "failed to upgrade hash locks: %s\n",
1022 tdb_errorstr(tdb)));
1024 _tdb_transaction_cancel(tdb);
1028 /* get the open lock - this prevents new users attaching to the database
1029 during the commit */
1030 if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
1031 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
1032 _tdb_transaction_cancel(tdb);
1036 /* write the recovery data to the end of the file */
1037 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
1038 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
1039 _tdb_transaction_cancel(tdb);
1043 tdb->transaction->prepared = true;
1045 /* expand the file to the new size if needed */
1046 if (tdb->map_size != tdb->transaction->old_map_size) {
1047 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
1049 tdb->transaction->old_map_size) == -1) {
1050 tdb->ecode = TDB_ERR_IO;
1051 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
1052 _tdb_transaction_cancel(tdb);
1055 tdb->map_size = tdb->transaction->old_map_size;
1056 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
1059 /* Keep the open lock until the actual commit */
1065 prepare to commit the current transaction
1067 _PUBLIC_ int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1069 tdb_trace(tdb, "tdb_transaction_prepare_commit");
1070 return _tdb_transaction_prepare_commit(tdb);
1073 /* A repack is worthwhile if the largest is less than half total free. */
1074 static bool repack_worthwhile(struct tdb_context *tdb)
1077 struct tdb_record rec;
1078 tdb_len_t total = 0, largest = 0;
1080 if (tdb_ofs_read(tdb, FREELIST_TOP, &ptr) == -1) {
1084 while (ptr != 0 && tdb_rec_free_read(tdb, ptr, &rec) == 0) {
1085 total += rec.rec_len;
1086 if (rec.rec_len > largest) {
1087 largest = rec.rec_len;
1092 return total > largest * 2;
1096 commit the current transaction
1098 _PUBLIC_ int tdb_transaction_commit(struct tdb_context *tdb)
1100 const struct tdb_methods *methods;
1102 bool need_repack = false;
1104 if (tdb->transaction == NULL) {
1105 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1109 tdb_trace(tdb, "tdb_transaction_commit");
1111 if (tdb->transaction->transaction_error) {
1112 tdb->ecode = TDB_ERR_IO;
1113 _tdb_transaction_cancel(tdb);
1114 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1119 if (tdb->transaction->nesting != 0) {
1120 tdb->transaction->nesting--;
1124 /* check for a null transaction */
1125 if (tdb->transaction->blocks == NULL) {
1126 _tdb_transaction_cancel(tdb);
1130 if (!tdb->transaction->prepared) {
1131 int ret = _tdb_transaction_prepare_commit(tdb);
1136 methods = tdb->transaction->io_methods;
1138 /* perform all the writes */
1139 for (i=0;i<tdb->transaction->num_blocks;i++) {
1143 if (tdb->transaction->blocks[i] == NULL) {
1147 offset = i * tdb->transaction->block_size;
1148 length = tdb->transaction->block_size;
1149 if (i == tdb->transaction->num_blocks-1) {
1150 length = tdb->transaction->last_block_size;
1153 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1154 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1156 /* we've overwritten part of the data and
1157 possibly expanded the file, so we need to
1158 run the crash recovery code */
1159 tdb->methods = methods;
1160 tdb_transaction_recover(tdb);
1162 _tdb_transaction_cancel(tdb);
1164 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1167 SAFE_FREE(tdb->transaction->blocks[i]);
1170 /* Do this before we drop lock or blocks. */
1171 if (tdb->transaction->expanded) {
1172 need_repack = repack_worthwhile(tdb);
1175 SAFE_FREE(tdb->transaction->blocks);
1176 tdb->transaction->num_blocks = 0;
1178 /* ensure the new data is on disk */
1179 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1184 TODO: maybe write to some dummy hdr field, or write to magic
1185 offset without mmap, before the last sync, instead of the
1189 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1190 don't change the mtime of the file, this means the file may
1191 not be backed up (as tdb rounding to block sizes means that
1192 file size changes are quite rare too). The following forces
1193 mtime changes when a transaction completes */
1195 utime(tdb->name, NULL);
1198 /* use a transaction cancel to free memory and remove the
1199 transaction locks */
1200 _tdb_transaction_cancel(tdb);
1203 return tdb_repack(tdb);
1211 recover from an aborted transaction. Must be called with exclusive
1212 database write access already established (including the open
1213 lock to prevent new processes attaching)
1215 int tdb_transaction_recover(struct tdb_context *tdb)
1217 tdb_off_t recovery_head, recovery_eof;
1218 unsigned char *data, *p;
1220 struct tdb_record rec;
1222 /* find the recovery area */
1223 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1224 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1225 tdb->ecode = TDB_ERR_IO;
1229 if (recovery_head == 0) {
1230 /* we have never allocated a recovery record */
1234 /* read the recovery record */
1235 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1236 sizeof(rec), DOCONV()) == -1) {
1237 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1238 tdb->ecode = TDB_ERR_IO;
1242 if (rec.magic != TDB_RECOVERY_MAGIC) {
1243 /* there is no valid recovery data */
1247 if (tdb->read_only) {
1248 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1249 tdb->ecode = TDB_ERR_CORRUPT;
1253 recovery_eof = rec.key_len;
1255 data = (unsigned char *)malloc(rec.data_len);
1257 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1258 tdb->ecode = TDB_ERR_OOM;
1262 /* read the full recovery data */
1263 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1264 rec.data_len, 0) == -1) {
1265 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1266 tdb->ecode = TDB_ERR_IO;
1270 /* recover the file data */
1272 while (p+8 < data + rec.data_len) {
1278 memcpy(&len, p+4, 4);
1280 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1282 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %u bytes at offset %u\n", len, ofs));
1283 tdb->ecode = TDB_ERR_IO;
1291 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1292 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1293 tdb->ecode = TDB_ERR_IO;
1297 /* if the recovery area is after the recovered eof then remove it */
1298 if (recovery_eof <= recovery_head) {
1299 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1300 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1301 tdb->ecode = TDB_ERR_IO;
1306 /* remove the recovery magic */
1307 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1309 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1310 tdb->ecode = TDB_ERR_IO;
1314 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1315 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1316 tdb->ecode = TDB_ERR_IO;
1320 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %u byte database\n",
1327 /* Any I/O failures we say "needs recovery". */
1328 bool tdb_needs_recovery(struct tdb_context *tdb)
1330 tdb_off_t recovery_head;
1331 struct tdb_record rec;
1333 /* find the recovery area */
1334 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1338 if (recovery_head == 0) {
1339 /* we have never allocated a recovery record */
1343 /* read the recovery record */
1344 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1345 sizeof(rec), DOCONV()) == -1) {
1349 return (rec.magic == TDB_RECOVERY_MAGIC);