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 "tdb1_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 tdb1_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 tdb1_write() calls. The hooked
48 transaction versions of tdb1_read() and tdb1_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 tdb1_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 tdb1_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 TDB1_NOSYNC is passed to flags in tdb1_open then transactions are
85 still available, but no transaction recovery area is used and no
86 fsync/msync calls are made.
88 - if TDB1_ALLOW_NESTING is passed to flags in tdb open, or added using
89 tdb1_add_flags() transaction nesting is enabled.
90 It resets the TDB1_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 tdb1_transaction_commit() can be silently unrolled later.
97 - if TDB1_DISALLOW_NESTING is passed to flags in tdb open, or added using
98 tdb1_add_flags() transaction nesting is disabled.
99 It resets the TDB1_ALLOW_NESTING flag, as both cannot be used together.
100 An attempt create a nested transaction will fail with TDB1_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 tdb1_transaction {
110 /* we keep a mirrored copy of the tdb hash heads here so
111 tdb1_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 tdb1_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 tdb1_transaction_start() calls, as these are allowed,
131 but don't create a new transaction */
134 /* set when a prepare has already occurred */
136 tdb1_off_t magic_offset;
138 /* old file size before transaction */
139 tdb1_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 transaction1_read(struct tdb1_context *tdb, tdb1_off_t off, void *buf,
151 tdb1_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 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
158 if (transaction1_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->tdb1_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 tdb1_convert(buf, len);
197 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
198 tdb->ecode = TDB1_ERR_IO;
199 tdb->transaction->transaction_error = 1;
205 write while in a transaction
207 static int transaction1_write(struct tdb1_context *tdb, tdb1_off_t off,
208 const void *buf, tdb1_len_t len)
212 /* Only a commit is allowed on a prepared transaction */
213 if (tdb->transaction->prepared) {
214 tdb->ecode = TDB1_ERR_EINVAL;
215 TDB1_LOG((tdb, TDB1_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(tdb1_off_t) && off >= TDB1_FREELIST_TOP &&
223 off < TDB1_FREELIST_TOP+TDB1_HASHTABLE_SIZE(tdb)) {
224 uint32_t chain = (off-TDB1_FREELIST_TOP) / sizeof(tdb1_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 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
231 if (transaction1_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 = TDB1_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 = TDB1_ERR_OOM;
275 tdb->transaction->transaction_error = 1;
278 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
279 tdb1_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->tdb1_read(tdb, blk * tdb->transaction->block_size,
284 tdb->transaction->blocks[blk],
286 SAFE_FREE(tdb->transaction->blocks[blk]);
287 tdb->ecode = TDB1_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 TDB1_LOG((tdb, TDB1_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 transaction1_write_existing(struct tdb1_context *tdb, tdb1_off_t off,
323 const void *buf, tdb1_len_t len)
327 /* break it up into block sized chunks */
328 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
329 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
330 if (transaction1_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 transaction1_next_hash_chain(struct tdb1_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 transaction1_oob(struct tdb1_context *tdb, tdb1_off_t len, int probe)
387 if (len <= tdb->map_size) {
390 tdb->ecode = TDB1_ERR_IO;
395 transaction version of tdb1_expand().
397 static int transaction1_expand_file(struct tdb1_context *tdb, tdb1_off_t size,
400 /* add a write to the transaction elements, so subsequent
401 reads see the zero data */
402 if (transaction1_write(tdb, size, NULL, addition) != 0) {
406 tdb->transaction->expanded = true;
411 static const struct tdb1_methods transaction1_methods = {
414 transaction1_next_hash_chain,
416 transaction1_expand_file,
421 start a tdb transaction. No token is returned, as only a single
422 transaction is allowed to be pending per tdb1_context
424 static int _tdb1_transaction_start(struct tdb1_context *tdb)
426 /* some sanity checks */
427 if (tdb->read_only || (tdb->flags & TDB1_INTERNAL) || tdb->traverse_read) {
428 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction on a read-only or internal db\n"));
429 tdb->ecode = TDB1_ERR_EINVAL;
433 /* cope with nested tdb1_transaction_start() calls */
434 if (tdb->transaction != NULL) {
435 if (!(tdb->flags & TDB1_ALLOW_NESTING)) {
436 tdb->ecode = TDB1_ERR_NESTING;
439 tdb->transaction->nesting++;
440 TDB1_LOG((tdb, TDB1_DEBUG_TRACE, "tdb1_transaction_start: nesting %d\n",
441 tdb->transaction->nesting));
445 if (tdb1_have_extra_locks(tdb)) {
446 /* the caller must not have any locks when starting a
447 transaction as otherwise we'll be screwed by lack
448 of nested locks in posix */
449 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction with locks held\n"));
450 tdb->ecode = TDB1_ERR_LOCK;
454 if (tdb->travlocks.next != NULL) {
455 /* you cannot use transactions inside a traverse (although you can use
456 traverse inside a transaction) as otherwise you can end up with
458 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction within a traverse\n"));
459 tdb->ecode = TDB1_ERR_LOCK;
463 tdb->transaction = (struct tdb1_transaction *)
464 calloc(sizeof(struct tdb1_transaction), 1);
465 if (tdb->transaction == NULL) {
466 tdb->ecode = TDB1_ERR_OOM;
470 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
471 tdb->transaction->block_size = tdb->page_size;
473 /* get the transaction write lock. This is a blocking lock. As
474 discussed with Volker, there are a number of ways we could
475 make this async, which we will probably do in the future */
476 if (tdb1_transaction_lock(tdb, F_WRLCK, TDB1_LOCK_WAIT) == -1) {
477 SAFE_FREE(tdb->transaction->blocks);
478 SAFE_FREE(tdb->transaction);
482 /* get a read lock from the freelist to the end of file. This
483 is upgraded to a write lock during the commit */
484 if (tdb1_allrecord_lock(tdb, F_RDLCK, TDB1_LOCK_WAIT, true) == -1) {
485 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: failed to get hash locks\n"));
486 goto fail_allrecord_lock;
489 /* setup a copy of the hash table heads so the hash scan in
490 traverse can be fast */
491 tdb->transaction->hash_heads = (uint32_t *)
492 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
493 if (tdb->transaction->hash_heads == NULL) {
494 tdb->ecode = TDB1_ERR_OOM;
497 if (tdb->methods->tdb1_read(tdb, TDB1_FREELIST_TOP, tdb->transaction->hash_heads,
498 TDB1_HASHTABLE_SIZE(tdb), 0) != 0) {
499 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_start: failed to read hash heads\n"));
500 tdb->ecode = TDB1_ERR_IO;
504 /* make sure we know about any file expansions already done by
506 tdb->methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
507 tdb->transaction->old_map_size = tdb->map_size;
509 /* finally hook the io methods, replacing them with
510 transaction specific methods */
511 tdb->transaction->io_methods = tdb->methods;
512 tdb->methods = &transaction1_methods;
517 tdb1_allrecord_unlock(tdb, F_RDLCK);
519 tdb1_transaction_unlock(tdb, F_WRLCK);
520 SAFE_FREE(tdb->transaction->blocks);
521 SAFE_FREE(tdb->transaction->hash_heads);
522 SAFE_FREE(tdb->transaction);
526 int tdb1_transaction_start(struct tdb1_context *tdb)
528 return _tdb1_transaction_start(tdb);
534 static int transaction1_sync(struct tdb1_context *tdb, tdb1_off_t offset, tdb1_len_t length)
536 if (tdb->flags & TDB1_NOSYNC) {
541 if (fdatasync(tdb->fd) != 0) {
543 if (fsync(tdb->fd) != 0) {
545 tdb->ecode = TDB1_ERR_IO;
546 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction: fsync failed\n"));
551 tdb1_off_t moffset = offset & ~(tdb->page_size-1);
552 if (msync(moffset + (char *)tdb->map_ptr,
553 length + (offset - moffset), MS_SYNC) != 0) {
554 tdb->ecode = TDB1_ERR_IO;
555 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction: msync failed - %s\n",
565 static int _tdb1_transaction_cancel(struct tdb1_context *tdb)
569 if (tdb->transaction == NULL) {
570 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_cancel: no transaction\n"));
574 if (tdb->transaction->nesting != 0) {
575 tdb->transaction->transaction_error = 1;
576 tdb->transaction->nesting--;
580 tdb->map_size = tdb->transaction->old_map_size;
582 /* free all the transaction blocks */
583 for (i=0;i<tdb->transaction->num_blocks;i++) {
584 if (tdb->transaction->blocks[i] != NULL) {
585 free(tdb->transaction->blocks[i]);
588 SAFE_FREE(tdb->transaction->blocks);
590 if (tdb->transaction->magic_offset) {
591 const struct tdb1_methods *methods = tdb->transaction->io_methods;
592 const uint32_t invalid = TDB1_RECOVERY_INVALID_MAGIC;
594 /* remove the recovery marker */
595 if (methods->tdb1_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
596 transaction1_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
597 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_cancel: failed to remove recovery magic\n"));
602 /* This also removes the OPEN_LOCK, if we have it. */
603 tdb1_release_transaction_locks(tdb);
605 /* restore the normal io methods */
606 tdb->methods = tdb->transaction->io_methods;
608 SAFE_FREE(tdb->transaction->hash_heads);
609 SAFE_FREE(tdb->transaction);
615 cancel the current transaction
617 int tdb1_transaction_cancel(struct tdb1_context *tdb)
619 return _tdb1_transaction_cancel(tdb);
623 work out how much space the linearised recovery data will consume
625 static tdb1_len_t tdb1_recovery_size(struct tdb1_context *tdb)
627 tdb1_len_t recovery_size = 0;
630 recovery_size = sizeof(uint32_t);
631 for (i=0;i<tdb->transaction->num_blocks;i++) {
632 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
635 if (tdb->transaction->blocks[i] == NULL) {
638 recovery_size += 2*sizeof(tdb1_off_t);
639 if (i == tdb->transaction->num_blocks-1) {
640 recovery_size += tdb->transaction->last_block_size;
642 recovery_size += tdb->transaction->block_size;
646 return recovery_size;
649 int tdb1_recovery_area(struct tdb1_context *tdb,
650 const struct tdb1_methods *methods,
651 tdb1_off_t *recovery_offset,
652 struct tdb1_record *rec)
654 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, recovery_offset) == -1) {
658 if (*recovery_offset == 0) {
663 if (methods->tdb1_read(tdb, *recovery_offset, rec, sizeof(*rec),
664 TDB1_DOCONV()) == -1) {
668 /* ignore invalid recovery regions: can happen in crash */
669 if (rec->magic != TDB1_RECOVERY_MAGIC &&
670 rec->magic != TDB1_RECOVERY_INVALID_MAGIC) {
671 *recovery_offset = 0;
678 allocate the recovery area, or use an existing recovery area if it is
681 static int tdb1_recovery_allocate(struct tdb1_context *tdb,
682 tdb1_len_t *recovery_size,
683 tdb1_off_t *recovery_offset,
684 tdb1_len_t *recovery_max_size)
686 struct tdb1_record rec;
687 const struct tdb1_methods *methods = tdb->transaction->io_methods;
688 tdb1_off_t recovery_head;
690 if (tdb1_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
691 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to read recovery head\n"));
695 *recovery_size = tdb1_recovery_size(tdb);
697 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
698 /* it fits in the existing area */
699 *recovery_max_size = rec.rec_len;
700 *recovery_offset = recovery_head;
704 /* we need to free up the old recovery area, then allocate a
705 new one at the end of the file. Note that we cannot use
706 tdb1_allocate() to allocate the new one as that might return
707 us an area that is being currently used (as of the start of
709 if (recovery_head != 0) {
710 if (tdb1_free(tdb, recovery_head, &rec) == -1) {
711 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to free previous recovery area\n"));
716 /* the tdb1_free() call might have increased the recovery size */
717 *recovery_size = tdb1_recovery_size(tdb);
719 /* round up to a multiple of page size */
720 *recovery_max_size = TDB1_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
721 *recovery_offset = tdb->map_size;
722 recovery_head = *recovery_offset;
724 if (methods->tdb1_expand_file(tdb, tdb->transaction->old_map_size,
725 (tdb->map_size - tdb->transaction->old_map_size) +
726 sizeof(rec) + *recovery_max_size) == -1) {
727 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to create recovery area\n"));
731 /* remap the file (if using mmap) */
732 methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
734 /* we have to reset the old map size so that we don't try to expand the file
735 again in the transaction commit, which would destroy the recovery area */
736 tdb->transaction->old_map_size = tdb->map_size;
738 /* write the recovery header offset and sync - we can sync without a race here
739 as the magic ptr in the recovery record has not been set */
740 TDB1_CONV(recovery_head);
741 if (methods->tdb1_write(tdb, TDB1_RECOVERY_HEAD,
742 &recovery_head, sizeof(tdb1_off_t)) == -1) {
743 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to write recovery head\n"));
746 if (transaction1_write_existing(tdb, TDB1_RECOVERY_HEAD, &recovery_head, sizeof(tdb1_off_t)) == -1) {
747 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to write recovery head\n"));
756 setup the recovery data that will be used on a crash during commit
758 static int transaction1_setup_recovery(struct tdb1_context *tdb,
759 tdb1_off_t *magic_offset)
761 tdb1_len_t recovery_size;
762 unsigned char *data, *p;
763 const struct tdb1_methods *methods = tdb->transaction->io_methods;
764 struct tdb1_record *rec;
765 tdb1_off_t recovery_offset, recovery_max_size;
766 tdb1_off_t old_map_size = tdb->transaction->old_map_size;
767 uint32_t magic, tailer;
771 check that the recovery area has enough space
773 if (tdb1_recovery_allocate(tdb, &recovery_size,
774 &recovery_offset, &recovery_max_size) == -1) {
778 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
780 tdb->ecode = TDB1_ERR_OOM;
784 rec = (struct tdb1_record *)data;
785 memset(rec, 0, sizeof(*rec));
787 rec->magic = TDB1_RECOVERY_INVALID_MAGIC;
788 rec->data_len = recovery_size;
789 rec->rec_len = recovery_max_size;
790 rec->key_len = old_map_size;
793 /* build the recovery data into a single blob to allow us to do a single
794 large write, which should be more efficient */
795 p = data + sizeof(*rec);
796 for (i=0;i<tdb->transaction->num_blocks;i++) {
800 if (tdb->transaction->blocks[i] == NULL) {
804 offset = i * tdb->transaction->block_size;
805 length = tdb->transaction->block_size;
806 if (i == tdb->transaction->num_blocks-1) {
807 length = tdb->transaction->last_block_size;
810 if (offset >= old_map_size) {
813 if (offset + length > tdb->transaction->old_map_size) {
814 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: transaction data over new region boundary\n"));
816 tdb->ecode = TDB1_ERR_CORRUPT;
819 memcpy(p, &offset, 4);
820 memcpy(p+4, &length, 4);
824 /* the recovery area contains the old data, not the
825 new data, so we have to call the original tdb1_read
827 if (methods->tdb1_read(tdb, offset, p + 8, length, 0) != 0) {
829 tdb->ecode = TDB1_ERR_IO;
836 tailer = sizeof(*rec) + recovery_max_size;
837 memcpy(p, &tailer, 4);
842 /* write the recovery data to the recovery area */
843 if (methods->tdb1_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
844 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write recovery data\n"));
846 tdb->ecode = TDB1_ERR_IO;
849 if (transaction1_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
850 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write secondary recovery data\n"));
852 tdb->ecode = TDB1_ERR_IO;
856 /* as we don't have ordered writes, we have to sync the recovery
857 data before we update the magic to indicate that the recovery
859 if (transaction1_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
866 magic = TDB1_RECOVERY_MAGIC;
869 *magic_offset = recovery_offset + offsetof(struct tdb1_record, magic);
871 if (methods->tdb1_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
872 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write recovery magic\n"));
873 tdb->ecode = TDB1_ERR_IO;
876 if (transaction1_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
877 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write secondary recovery magic\n"));
878 tdb->ecode = TDB1_ERR_IO;
882 /* ensure the recovery magic marker is on disk */
883 if (transaction1_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
890 static int _tdb1_transaction_prepare_commit(struct tdb1_context *tdb)
892 const struct tdb1_methods *methods;
894 if (tdb->transaction == NULL) {
895 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: no transaction\n"));
899 if (tdb->transaction->prepared) {
900 tdb->ecode = TDB1_ERR_EINVAL;
901 _tdb1_transaction_cancel(tdb);
902 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: transaction already prepared\n"));
906 if (tdb->transaction->transaction_error) {
907 tdb->ecode = TDB1_ERR_IO;
908 _tdb1_transaction_cancel(tdb);
909 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: transaction error pending\n"));
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 /* if there are any locks pending then the caller has not
926 nested their locks properly, so fail the transaction */
927 if (tdb1_have_extra_locks(tdb)) {
928 tdb->ecode = TDB1_ERR_LOCK;
929 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: locks pending on commit\n"));
930 _tdb1_transaction_cancel(tdb);
934 /* upgrade the main transaction lock region to a write lock */
935 if (tdb1_allrecord_upgrade(tdb) == -1) {
936 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: failed to upgrade hash locks\n"));
937 _tdb1_transaction_cancel(tdb);
941 /* get the open lock - this prevents new users attaching to the database
943 if (tdb1_nest_lock(tdb, TDB1_OPEN_LOCK, F_WRLCK, TDB1_LOCK_WAIT) == -1) {
944 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: failed to get open lock\n"));
945 _tdb1_transaction_cancel(tdb);
949 if (!(tdb->flags & TDB1_NOSYNC)) {
950 /* write the recovery data to the end of the file */
951 if (transaction1_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
952 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_prepare_commit: failed to setup recovery data\n"));
953 _tdb1_transaction_cancel(tdb);
958 tdb->transaction->prepared = true;
960 /* expand the file to the new size if needed */
961 if (tdb->map_size != tdb->transaction->old_map_size) {
962 if (methods->tdb1_expand_file(tdb, tdb->transaction->old_map_size,
964 tdb->transaction->old_map_size) == -1) {
965 tdb->ecode = TDB1_ERR_IO;
966 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_prepare_commit: expansion failed\n"));
967 _tdb1_transaction_cancel(tdb);
970 tdb->map_size = tdb->transaction->old_map_size;
971 methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
974 /* Keep the open lock until the actual commit */
980 prepare to commit the current transaction
982 int tdb1_transaction_prepare_commit(struct tdb1_context *tdb)
984 return _tdb1_transaction_prepare_commit(tdb);
987 /* A repack is worthwhile if the largest is less than half total free. */
988 static bool repack_worthwhile(struct tdb1_context *tdb)
991 struct tdb1_record rec;
992 tdb1_len_t total = 0, largest = 0;
994 if (tdb1_ofs_read(tdb, TDB1_FREELIST_TOP, &ptr) == -1) {
998 while (ptr != 0 && tdb1_rec_free_read(tdb, ptr, &rec) == 0) {
999 total += rec.rec_len;
1000 if (rec.rec_len > largest) {
1001 largest = rec.rec_len;
1006 return total > largest * 2;
1010 commit the current transaction
1012 int tdb1_transaction_commit(struct tdb1_context *tdb)
1014 const struct tdb1_methods *methods;
1016 bool need_repack = false;
1018 if (tdb->transaction == NULL) {
1019 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_commit: no transaction\n"));
1023 if (tdb->transaction->transaction_error) {
1024 tdb->ecode = TDB1_ERR_IO;
1025 _tdb1_transaction_cancel(tdb);
1026 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_commit: transaction error pending\n"));
1031 if (tdb->transaction->nesting != 0) {
1032 tdb->transaction->nesting--;
1036 /* check for a null transaction */
1037 if (tdb->transaction->blocks == NULL) {
1038 _tdb1_transaction_cancel(tdb);
1042 if (!tdb->transaction->prepared) {
1043 int ret = _tdb1_transaction_prepare_commit(tdb);
1048 methods = tdb->transaction->io_methods;
1050 /* perform all the writes */
1051 for (i=0;i<tdb->transaction->num_blocks;i++) {
1055 if (tdb->transaction->blocks[i] == NULL) {
1059 offset = i * tdb->transaction->block_size;
1060 length = tdb->transaction->block_size;
1061 if (i == tdb->transaction->num_blocks-1) {
1062 length = tdb->transaction->last_block_size;
1065 if (methods->tdb1_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1066 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_commit: write failed during commit\n"));
1068 /* we've overwritten part of the data and
1069 possibly expanded the file, so we need to
1070 run the crash recovery code */
1071 tdb->methods = methods;
1072 tdb1_transaction_recover(tdb);
1074 _tdb1_transaction_cancel(tdb);
1076 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_commit: write failed\n"));
1079 SAFE_FREE(tdb->transaction->blocks[i]);
1082 /* Do this before we drop lock or blocks. */
1083 if (tdb->transaction->expanded) {
1084 need_repack = repack_worthwhile(tdb);
1087 SAFE_FREE(tdb->transaction->blocks);
1088 tdb->transaction->num_blocks = 0;
1090 /* ensure the new data is on disk */
1091 if (transaction1_sync(tdb, 0, tdb->map_size) == -1) {
1096 TODO: maybe write to some dummy hdr field, or write to magic
1097 offset without mmap, before the last sync, instead of the
1101 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1102 don't change the mtime of the file, this means the file may
1103 not be backed up (as tdb rounding to block sizes means that
1104 file size changes are quite rare too). The following forces
1105 mtime changes when a transaction completes */
1107 utime(tdb->name, NULL);
1110 /* use a transaction cancel to free memory and remove the
1111 transaction locks */
1112 _tdb1_transaction_cancel(tdb);
1115 return tdb1_repack(tdb);
1123 recover from an aborted transaction. Must be called with exclusive
1124 database write access already established (including the open
1125 lock to prevent new processes attaching)
1127 int tdb1_transaction_recover(struct tdb1_context *tdb)
1129 tdb1_off_t recovery_head, recovery_eof;
1130 unsigned char *data, *p;
1132 struct tdb1_record rec;
1134 /* find the recovery area */
1135 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1136 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery head\n"));
1137 tdb->ecode = TDB1_ERR_IO;
1141 if (recovery_head == 0) {
1142 /* we have never allocated a recovery record */
1146 /* read the recovery record */
1147 if (tdb->methods->tdb1_read(tdb, recovery_head, &rec,
1148 sizeof(rec), TDB1_DOCONV()) == -1) {
1149 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery record\n"));
1150 tdb->ecode = TDB1_ERR_IO;
1154 if (rec.magic != TDB1_RECOVERY_MAGIC) {
1155 /* there is no valid recovery data */
1159 if (tdb->read_only) {
1160 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: attempt to recover read only database\n"));
1161 tdb->ecode = TDB1_ERR_CORRUPT;
1165 recovery_eof = rec.key_len;
1167 data = (unsigned char *)malloc(rec.data_len);
1169 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to allocate recovery data\n"));
1170 tdb->ecode = TDB1_ERR_OOM;
1174 /* read the full recovery data */
1175 if (tdb->methods->tdb1_read(tdb, recovery_head + sizeof(rec), data,
1176 rec.data_len, 0) == -1) {
1177 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery data\n"));
1178 tdb->ecode = TDB1_ERR_IO;
1182 /* recover the file data */
1184 while (p+8 < data + rec.data_len) {
1186 if (TDB1_DOCONV()) {
1190 memcpy(&len, p+4, 4);
1192 if (tdb->methods->tdb1_write(tdb, ofs, p+8, len) == -1) {
1194 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1195 tdb->ecode = TDB1_ERR_IO;
1203 if (transaction1_sync(tdb, 0, tdb->map_size) == -1) {
1204 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to sync recovery\n"));
1205 tdb->ecode = TDB1_ERR_IO;
1209 /* if the recovery area is after the recovered eof then remove it */
1210 if (recovery_eof <= recovery_head) {
1211 if (tdb1_ofs_write(tdb, TDB1_RECOVERY_HEAD, &zero) == -1) {
1212 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to remove recovery head\n"));
1213 tdb->ecode = TDB1_ERR_IO;
1218 /* remove the recovery magic */
1219 if (tdb1_ofs_write(tdb, recovery_head + offsetof(struct tdb1_record, magic),
1221 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to remove recovery magic\n"));
1222 tdb->ecode = TDB1_ERR_IO;
1226 if (transaction1_sync(tdb, 0, recovery_eof) == -1) {
1227 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to sync2 recovery\n"));
1228 tdb->ecode = TDB1_ERR_IO;
1232 TDB1_LOG((tdb, TDB1_DEBUG_TRACE, "tdb1_transaction_recover: recovered %d byte database\n",
1239 /* Any I/O failures we say "needs recovery". */
1240 bool tdb1_needs_recovery(struct tdb1_context *tdb)
1242 tdb1_off_t recovery_head;
1243 struct tdb1_record rec;
1245 /* find the recovery area */
1246 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1250 if (recovery_head == 0) {
1251 /* we have never allocated a recovery record */
1255 /* read the recovery record */
1256 if (tdb->methods->tdb1_read(tdb, recovery_head, &rec,
1257 sizeof(rec), TDB1_DOCONV()) == -1) {
1261 return (rec.magic == TDB1_RECOVERY_MAGIC);