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 /* we should re-pack on commit */
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 len, int probe)
387 if (len <= tdb->map_size) {
390 tdb->ecode = TDB_ERR_IO;
395 transaction version of tdb_expand().
397 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
400 /* add a write to the transaction elements, so subsequent
401 reads see the zero data */
402 if (transaction_write(tdb, size, NULL, addition) != 0) {
406 tdb->transaction->need_repack = true;
411 static const struct tdb_methods transaction_methods = {
414 transaction_next_hash_chain,
416 transaction_expand_file,
422 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
424 if (tdb->flags & TDB_NOSYNC) {
428 if (fsync(tdb->fd) != 0) {
429 tdb->ecode = TDB_ERR_IO;
430 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
435 tdb_off_t moffset = offset & ~(tdb->page_size-1);
436 if (msync(moffset + (char *)tdb->map_ptr,
437 length + (offset - moffset), MS_SYNC) != 0) {
438 tdb->ecode = TDB_ERR_IO;
439 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
449 static int _tdb_transaction_cancel(struct tdb_context *tdb)
453 if (tdb->transaction == NULL) {
454 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
458 if (tdb->transaction->nesting != 0) {
459 tdb->transaction->transaction_error = 1;
460 tdb->transaction->nesting--;
464 tdb->map_size = tdb->transaction->old_map_size;
466 /* free all the transaction blocks */
467 for (i=0;i<tdb->transaction->num_blocks;i++) {
468 if (tdb->transaction->blocks[i] != NULL) {
469 free(tdb->transaction->blocks[i]);
472 SAFE_FREE(tdb->transaction->blocks);
474 if (tdb->transaction->magic_offset) {
475 const struct tdb_methods *methods = tdb->transaction->io_methods;
476 uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
478 /* remove the recovery marker */
479 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
480 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
481 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
486 /* This also removes the OPEN_LOCK, if we have it. */
487 tdb_release_extra_locks(tdb);
489 /* restore the normal io methods */
490 tdb->methods = tdb->transaction->io_methods;
492 tdb_transaction_unlock(tdb, F_WRLCK);
493 SAFE_FREE(tdb->transaction->hash_heads);
494 SAFE_FREE(tdb->transaction);
500 start a tdb transaction. No token is returned, as only a single
501 transaction is allowed to be pending per tdb_context
503 int tdb_transaction_start(struct tdb_context *tdb)
505 /* some sanity checks */
506 if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
507 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
508 tdb->ecode = TDB_ERR_EINVAL;
512 /* cope with nested tdb_transaction_start() calls */
513 if (tdb->transaction != NULL) {
514 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
515 tdb->ecode = TDB_ERR_NESTING;
518 tdb_trace(tdb, "tdb_transaction_start");
519 tdb->transaction->nesting++;
520 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
521 tdb->transaction->nesting));
525 if (tdb_have_extra_locks(tdb)) {
526 /* the caller must not have any locks when starting a
527 transaction as otherwise we'll be screwed by lack
528 of nested locks in posix */
529 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
530 tdb->ecode = TDB_ERR_LOCK;
534 if (tdb->travlocks.next != NULL) {
535 /* you cannot use transactions inside a traverse (although you can use
536 traverse inside a transaction) as otherwise you can end up with
538 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
539 tdb->ecode = TDB_ERR_LOCK;
543 tdb->transaction = (struct tdb_transaction *)
544 calloc(sizeof(struct tdb_transaction), 1);
545 if (tdb->transaction == NULL) {
546 tdb->ecode = TDB_ERR_OOM;
550 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
551 tdb->transaction->block_size = tdb->page_size;
553 /* get the transaction write lock. This is a blocking lock. As
554 discussed with Volker, there are a number of ways we could
555 make this async, which we will probably do in the future */
556 if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
557 SAFE_FREE(tdb->transaction->blocks);
558 SAFE_FREE(tdb->transaction);
562 /* get a read lock from the freelist to the end of file. This
563 is upgraded to a write lock during the commit */
564 if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
565 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
566 goto fail_allrecord_lock;
569 /* setup a copy of the hash table heads so the hash scan in
570 traverse can be fast */
571 tdb->transaction->hash_heads = (uint32_t *)
572 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
573 if (tdb->transaction->hash_heads == NULL) {
574 tdb->ecode = TDB_ERR_OOM;
577 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
578 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
579 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
580 tdb->ecode = TDB_ERR_IO;
584 /* make sure we know about any file expansions already done by
586 tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
587 tdb->transaction->old_map_size = tdb->map_size;
589 /* finally hook the io methods, replacing them with
590 transaction specific methods */
591 tdb->transaction->io_methods = tdb->methods;
592 tdb->methods = &transaction_methods;
594 /* Trace at the end, so we get sequence number correct. */
595 tdb_trace(tdb, "tdb_transaction_start");
599 tdb_allrecord_unlock(tdb, F_RDLCK, false);
601 tdb_transaction_unlock(tdb, F_WRLCK);
602 SAFE_FREE(tdb->transaction->blocks);
603 SAFE_FREE(tdb->transaction->hash_heads);
604 SAFE_FREE(tdb->transaction);
610 cancel the current transaction
612 int tdb_transaction_cancel(struct tdb_context *tdb)
614 tdb_trace(tdb, "tdb_transaction_cancel");
615 return _tdb_transaction_cancel(tdb);
619 work out how much space the linearised recovery data will consume
621 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
623 tdb_len_t recovery_size = 0;
626 recovery_size = sizeof(uint32_t);
627 for (i=0;i<tdb->transaction->num_blocks;i++) {
628 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
631 if (tdb->transaction->blocks[i] == NULL) {
634 recovery_size += 2*sizeof(tdb_off_t);
635 if (i == tdb->transaction->num_blocks-1) {
636 recovery_size += tdb->transaction->last_block_size;
638 recovery_size += tdb->transaction->block_size;
642 return recovery_size;
646 allocate the recovery area, or use an existing recovery area if it is
649 static int tdb_recovery_allocate(struct tdb_context *tdb,
650 tdb_len_t *recovery_size,
651 tdb_off_t *recovery_offset,
652 tdb_len_t *recovery_max_size)
654 struct tdb_record rec;
655 const struct tdb_methods *methods = tdb->transaction->io_methods;
656 tdb_off_t recovery_head, new_end;
658 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
659 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
665 if (recovery_head != 0) {
666 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
667 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
670 /* ignore invalid recovery regions: can happen in crash */
671 if (rec.magic != TDB_RECOVERY_MAGIC &&
672 rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
677 *recovery_size = tdb_recovery_size(tdb);
679 /* Existing recovery area? */
680 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
681 /* it fits in the existing area */
682 *recovery_max_size = rec.rec_len;
683 *recovery_offset = recovery_head;
687 /* If recovery area in middle of file, we need a new one. */
688 if (recovery_head == 0
689 || recovery_head + sizeof(rec) + rec.rec_len != tdb->map_size) {
690 /* we need to free up the old recovery area, then allocate a
691 new one at the end of the file. Note that we cannot use
692 tdb_allocate() to allocate the new one as that might return
693 us an area that is being currently used (as of the start of
696 if (tdb_free(tdb, recovery_head, &rec) == -1) {
697 TDB_LOG((tdb, TDB_DEBUG_FATAL,
698 "tdb_recovery_allocate: failed to"
699 " free previous recovery area\n"));
703 /* the tdb_free() call might have increased
704 * the recovery size */
705 *recovery_size = tdb_recovery_size(tdb);
708 /* New head will be at end of file. */
709 recovery_head = tdb->map_size;
712 /* Now we know where it will be. */
713 *recovery_offset = recovery_head;
715 /* Expand by more than we need, so we don't do it often. */
716 *recovery_max_size = tdb_expand_adjust(tdb->map_size,
721 new_end = recovery_head + sizeof(rec) + *recovery_max_size;
723 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
724 new_end - tdb->transaction->old_map_size)
726 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
730 /* remap the file (if using mmap) */
731 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
733 /* we have to reset the old map size so that we don't try to expand the file
734 again in the transaction commit, which would destroy the recovery area */
735 tdb->transaction->old_map_size = tdb->map_size;
737 /* write the recovery header offset and sync - we can sync without a race here
738 as the magic ptr in the recovery record has not been set */
739 CONVERT(recovery_head);
740 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
741 &recovery_head, sizeof(tdb_off_t)) == -1) {
742 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
745 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
746 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
755 setup the recovery data that will be used on a crash during commit
757 static int transaction_setup_recovery(struct tdb_context *tdb,
758 tdb_off_t *magic_offset)
760 tdb_len_t recovery_size;
761 unsigned char *data, *p;
762 const struct tdb_methods *methods = tdb->transaction->io_methods;
763 struct tdb_record *rec;
764 tdb_off_t recovery_offset, recovery_max_size;
765 tdb_off_t old_map_size = tdb->transaction->old_map_size;
766 uint32_t magic, tailer;
770 check that the recovery area has enough space
772 if (tdb_recovery_allocate(tdb, &recovery_size,
773 &recovery_offset, &recovery_max_size) == -1) {
777 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
779 tdb->ecode = TDB_ERR_OOM;
783 rec = (struct tdb_record *)data;
784 memset(rec, 0, sizeof(*rec));
786 rec->magic = TDB_RECOVERY_INVALID_MAGIC;
787 rec->data_len = recovery_size;
788 rec->rec_len = recovery_max_size;
789 rec->key_len = old_map_size;
792 /* build the recovery data into a single blob to allow us to do a single
793 large write, which should be more efficient */
794 p = data + sizeof(*rec);
795 for (i=0;i<tdb->transaction->num_blocks;i++) {
799 if (tdb->transaction->blocks[i] == NULL) {
803 offset = i * tdb->transaction->block_size;
804 length = tdb->transaction->block_size;
805 if (i == tdb->transaction->num_blocks-1) {
806 length = tdb->transaction->last_block_size;
809 if (offset >= old_map_size) {
812 if (offset + length > tdb->transaction->old_map_size) {
813 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
815 tdb->ecode = TDB_ERR_CORRUPT;
818 memcpy(p, &offset, 4);
819 memcpy(p+4, &length, 4);
823 /* the recovery area contains the old data, not the
824 new data, so we have to call the original tdb_read
826 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
828 tdb->ecode = TDB_ERR_IO;
835 tailer = sizeof(*rec) + recovery_max_size;
836 memcpy(p, &tailer, 4);
839 /* write the recovery data to the recovery area */
840 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
841 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
843 tdb->ecode = TDB_ERR_IO;
846 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
847 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
849 tdb->ecode = TDB_ERR_IO;
853 /* as we don't have ordered writes, we have to sync the recovery
854 data before we update the magic to indicate that the recovery
856 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
863 magic = TDB_RECOVERY_MAGIC;
866 *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
868 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
869 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
870 tdb->ecode = TDB_ERR_IO;
873 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
874 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
875 tdb->ecode = TDB_ERR_IO;
879 /* ensure the recovery magic marker is on disk */
880 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
887 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
889 const struct tdb_methods *methods;
891 if (tdb->transaction == NULL) {
892 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
896 if (tdb->transaction->prepared) {
897 tdb->ecode = TDB_ERR_EINVAL;
898 _tdb_transaction_cancel(tdb);
899 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
903 if (tdb->transaction->transaction_error) {
904 tdb->ecode = TDB_ERR_IO;
905 _tdb_transaction_cancel(tdb);
906 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
911 if (tdb->transaction->nesting != 0) {
912 tdb->transaction->nesting--;
916 /* check for a null transaction */
917 if (tdb->transaction->blocks == NULL) {
921 methods = tdb->transaction->io_methods;
923 /* if there are any locks pending then the caller has not
924 nested their locks properly, so fail the transaction */
925 if (tdb_have_extra_locks(tdb)) {
926 tdb->ecode = TDB_ERR_LOCK;
927 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
928 _tdb_transaction_cancel(tdb);
932 /* upgrade the main transaction lock region to a write lock */
933 if (tdb_allrecord_upgrade(tdb) == -1) {
934 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
935 _tdb_transaction_cancel(tdb);
939 /* get the open lock - this prevents new users attaching to the database
941 if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
942 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
943 _tdb_transaction_cancel(tdb);
947 if (!(tdb->flags & TDB_NOSYNC)) {
948 /* write the recovery data to the end of the file */
949 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
950 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
951 _tdb_transaction_cancel(tdb);
956 tdb->transaction->prepared = true;
958 /* expand the file to the new size if needed */
959 if (tdb->map_size != tdb->transaction->old_map_size) {
960 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
962 tdb->transaction->old_map_size) == -1) {
963 tdb->ecode = TDB_ERR_IO;
964 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
965 _tdb_transaction_cancel(tdb);
968 tdb->map_size = tdb->transaction->old_map_size;
969 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
972 /* Keep the open lock until the actual commit */
978 prepare to commit the current transaction
980 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
982 tdb_trace(tdb, "tdb_transaction_prepare_commit");
983 return _tdb_transaction_prepare_commit(tdb);
987 commit the current transaction
989 int tdb_transaction_commit(struct tdb_context *tdb)
991 const struct tdb_methods *methods;
995 if (tdb->transaction == NULL) {
996 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1000 tdb_trace(tdb, "tdb_transaction_commit");
1002 if (tdb->transaction->transaction_error) {
1003 tdb->ecode = TDB_ERR_IO;
1004 tdb_transaction_cancel(tdb);
1005 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1010 if (tdb->transaction->nesting != 0) {
1011 tdb->transaction->nesting--;
1015 /* check for a null transaction */
1016 if (tdb->transaction->blocks == NULL) {
1017 _tdb_transaction_cancel(tdb);
1021 if (!tdb->transaction->prepared) {
1022 int ret = _tdb_transaction_prepare_commit(tdb);
1027 methods = tdb->transaction->io_methods;
1029 /* perform all the writes */
1030 for (i=0;i<tdb->transaction->num_blocks;i++) {
1034 if (tdb->transaction->blocks[i] == NULL) {
1038 offset = i * tdb->transaction->block_size;
1039 length = tdb->transaction->block_size;
1040 if (i == tdb->transaction->num_blocks-1) {
1041 length = tdb->transaction->last_block_size;
1044 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1045 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1047 /* we've overwritten part of the data and
1048 possibly expanded the file, so we need to
1049 run the crash recovery code */
1050 tdb->methods = methods;
1051 tdb_transaction_recover(tdb);
1053 _tdb_transaction_cancel(tdb);
1055 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1058 SAFE_FREE(tdb->transaction->blocks[i]);
1061 SAFE_FREE(tdb->transaction->blocks);
1062 tdb->transaction->num_blocks = 0;
1064 /* ensure the new data is on disk */
1065 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1070 TODO: maybe write to some dummy hdr field, or write to magic
1071 offset without mmap, before the last sync, instead of the
1075 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1076 don't change the mtime of the file, this means the file may
1077 not be backed up (as tdb rounding to block sizes means that
1078 file size changes are quite rare too). The following forces
1079 mtime changes when a transaction completes */
1081 utime(tdb->name, NULL);
1084 need_repack = tdb->transaction->need_repack;
1086 /* use a transaction cancel to free memory and remove the
1087 transaction locks */
1088 _tdb_transaction_cancel(tdb);
1091 return tdb_repack(tdb);
1099 recover from an aborted transaction. Must be called with exclusive
1100 database write access already established (including the open
1101 lock to prevent new processes attaching)
1103 int tdb_transaction_recover(struct tdb_context *tdb)
1105 tdb_off_t recovery_head, recovery_eof;
1106 unsigned char *data, *p;
1108 struct tdb_record rec;
1110 /* find the recovery area */
1111 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1112 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1113 tdb->ecode = TDB_ERR_IO;
1117 if (recovery_head == 0) {
1118 /* we have never allocated a recovery record */
1122 /* read the recovery record */
1123 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1124 sizeof(rec), DOCONV()) == -1) {
1125 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1126 tdb->ecode = TDB_ERR_IO;
1130 if (rec.magic != TDB_RECOVERY_MAGIC) {
1131 /* there is no valid recovery data */
1135 if (tdb->read_only) {
1136 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1137 tdb->ecode = TDB_ERR_CORRUPT;
1141 recovery_eof = rec.key_len;
1143 data = (unsigned char *)malloc(rec.data_len);
1145 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1146 tdb->ecode = TDB_ERR_OOM;
1150 /* read the full recovery data */
1151 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1152 rec.data_len, 0) == -1) {
1153 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1154 tdb->ecode = TDB_ERR_IO;
1158 /* recover the file data */
1160 while (p+8 < data + rec.data_len) {
1166 memcpy(&len, p+4, 4);
1168 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1170 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1171 tdb->ecode = TDB_ERR_IO;
1179 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1180 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1181 tdb->ecode = TDB_ERR_IO;
1185 /* if the recovery area is after the recovered eof then remove it */
1186 if (recovery_eof <= recovery_head) {
1187 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1188 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1189 tdb->ecode = TDB_ERR_IO;
1194 /* remove the recovery magic */
1195 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1197 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1198 tdb->ecode = TDB_ERR_IO;
1202 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1203 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1204 tdb->ecode = TDB_ERR_IO;
1208 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",
1215 /* Any I/O failures we say "needs recovery". */
1216 bool tdb_needs_recovery(struct tdb_context *tdb)
1218 tdb_off_t recovery_head;
1219 struct tdb_record rec;
1221 /* find the recovery area */
1222 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1226 if (recovery_head == 0) {
1227 /* we have never allocated a recovery record */
1231 /* read the recovery record */
1232 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1233 sizeof(rec), DOCONV()) == -1) {
1237 return (rec.magic == TDB_RECOVERY_MAGIC);