2 Unix SMB/CIFS implementation.
4 trivial database library
6 Copyright (C) Andrew Tridgell 2005
7 Copyright (C) Rusty Russell 2010
9 ** NOTE! The following LGPL license applies to the tdb
10 ** library. This does NOT imply that all of Samba is released
13 This library is free software; you can redistribute it and/or
14 modify it under the terms of the GNU Lesser General Public
15 License as published by the Free Software Foundation; either
16 version 3 of the License, or (at your option) any later version.
18 This library is distributed in the hope that it will be useful,
19 but WITHOUT ANY WARRANTY; without even the implied warranty of
20 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21 Lesser General Public License for more details.
23 You should have received a copy of the GNU Lesser General Public
24 License along with this library; if not, see <http://www.gnu.org/licenses/>.
28 #define SAFE_FREE(x) do { if ((x) != NULL) {free(x); (x)=NULL;} } while(0)
33 - only allow a single transaction at a time per database. This makes
34 using the transaction API simpler, as otherwise the caller would
35 have to cope with temporary failures in transactions that conflict
36 with other current transactions
38 - keep the transaction recovery information in the same file as the
39 database, using a special 'transaction recovery' record pointed at
40 by the header. This removes the need for extra journal files as
41 used by some other databases
43 - dynamically allocated the transaction recover record, re-using it
44 for subsequent transactions. If a larger record is needed then
45 tdb_free() the old record to place it on the normal tdb freelist
46 before allocating the new record
48 - during transactions, keep a linked list of writes all that have
49 been performed by intercepting all tdb_write() calls. The hooked
50 transaction versions of tdb_read() and tdb_write() check this
51 linked list and try to use the elements of the list in preference
54 - don't allow any locks to be held when a transaction starts,
55 otherwise we can end up with deadlock (plus lack of lock nesting
56 in posix locks would mean the lock is lost)
58 - if the caller gains a lock during the transaction but doesn't
59 release it then fail the commit
61 - allow for nested calls to tdb_transaction_start(), re-using the
62 existing transaction record. If the inner transaction is cancelled
63 then a subsequent commit will fail
65 - keep a mirrored copy of the tdb hash chain heads to allow for the
66 fast hash heads scan on traverse, updating the mirrored copy in
67 the transaction version of tdb_write
69 - allow callers to mix transaction and non-transaction use of tdb,
70 although once a transaction is started then an exclusive lock is
71 gained until the transaction is committed or cancelled
73 - the commit stategy involves first saving away all modified data
74 into a linearised buffer in the transaction recovery area, then
75 marking the transaction recovery area with a magic value to
76 indicate a valid recovery record. In total 4 fsync/msync calls are
77 needed per commit to prevent race conditions. It might be possible
78 to reduce this to 3 or even 2 with some more work.
80 - check for a valid recovery record on open of the tdb, while the
81 open lock is held. Automatically recover from the transaction
82 recovery area if needed, then continue with the open as
83 usual. This allows for smooth crash recovery with no administrator
86 - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
87 still available, but no transaction recovery area is used and no
88 fsync/msync calls are made.
93 hold the context of any current transaction
95 struct tdb_transaction {
96 /* the original io methods - used to do IOs to the real db */
97 const struct tdb_methods *io_methods;
99 /* the list of transaction blocks. When a block is first
100 written to, it gets created in this list */
103 size_t last_block_size; /* number of valid bytes in the last block */
105 /* non-zero when an internal transaction error has
106 occurred. All write operations will then fail until the
107 transaction is ended */
108 int transaction_error;
110 /* when inside a transaction we need to keep track of any
111 nested tdb_transaction_start() calls, as these are allowed,
112 but don't create a new transaction */
115 /* set when a prepare has already occurred */
117 tdb_off_t magic_offset;
119 /* old file size before transaction */
120 tdb_len_t old_map_size;
125 read while in a transaction. We need to check first if the data is in our list
126 of transaction elements, then if not do a real read
128 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
133 /* break it down into block sized ops */
134 while (len + (off % getpagesize()) > getpagesize()) {
135 tdb_len_t len2 = getpagesize() - (off % getpagesize());
136 if (transaction_read(tdb, off, buf, len2) != 0) {
141 buf = (void *)(len2 + (char *)buf);
148 blk = off / getpagesize();
150 /* see if we have it in the block list */
151 if (tdb->transaction->num_blocks <= blk ||
152 tdb->transaction->blocks[blk] == NULL) {
153 /* nope, do a real read */
154 if (tdb->transaction->io_methods->read(tdb, off, buf, len) != 0) {
160 /* it is in the block list. Now check for the last block */
161 if (blk == tdb->transaction->num_blocks-1) {
162 if (len > tdb->transaction->last_block_size) {
167 /* now copy it out of this block */
168 memcpy(buf, tdb->transaction->blocks[blk] + (off % getpagesize()), len);
172 tdb->ecode = TDB_ERR_IO;
173 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
174 "transaction_read: failed at off=%llu len=%llu\n",
175 (long long)off, (long long)len);
176 tdb->transaction->transaction_error = 1;
182 write while in a transaction
184 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
185 const void *buf, tdb_len_t len)
189 /* Only a commit is allowed on a prepared transaction */
190 if (tdb->transaction->prepared) {
191 tdb->ecode = TDB_ERR_EINVAL;
192 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
193 "transaction_write: transaction already prepared,"
194 " write not allowed\n");
195 tdb->transaction->transaction_error = 1;
199 /* break it up into block sized chunks */
200 while (len + (off % getpagesize()) > getpagesize()) {
201 tdb_len_t len2 = getpagesize() - (off % getpagesize());
202 if (transaction_write(tdb, off, buf, len2) != 0) {
208 buf = (const void *)(len2 + (const char *)buf);
216 blk = off / getpagesize();
217 off = off % getpagesize();
219 if (tdb->transaction->num_blocks <= blk) {
220 uint8_t **new_blocks;
221 /* expand the blocks array */
222 if (tdb->transaction->blocks == NULL) {
223 new_blocks = (uint8_t **)malloc(
224 (blk+1)*sizeof(uint8_t *));
226 new_blocks = (uint8_t **)realloc(
227 tdb->transaction->blocks,
228 (blk+1)*sizeof(uint8_t *));
230 if (new_blocks == NULL) {
231 tdb->ecode = TDB_ERR_OOM;
234 memset(&new_blocks[tdb->transaction->num_blocks], 0,
235 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
236 tdb->transaction->blocks = new_blocks;
237 tdb->transaction->num_blocks = blk+1;
238 tdb->transaction->last_block_size = 0;
241 /* allocate and fill a block? */
242 if (tdb->transaction->blocks[blk] == NULL) {
243 tdb->transaction->blocks[blk] = (uint8_t *)calloc(getpagesize(), 1);
244 if (tdb->transaction->blocks[blk] == NULL) {
245 tdb->ecode = TDB_ERR_OOM;
246 tdb->transaction->transaction_error = 1;
249 if (tdb->transaction->old_map_size > blk * getpagesize()) {
250 tdb_len_t len2 = getpagesize();
251 if (len2 + (blk * getpagesize()) > tdb->transaction->old_map_size) {
252 len2 = tdb->transaction->old_map_size - (blk * getpagesize());
254 if (tdb->transaction->io_methods->read(tdb, blk * getpagesize(),
255 tdb->transaction->blocks[blk],
257 SAFE_FREE(tdb->transaction->blocks[blk]);
260 if (blk == tdb->transaction->num_blocks-1) {
261 tdb->transaction->last_block_size = len2;
266 /* overwrite part of an existing block */
268 memset(tdb->transaction->blocks[blk] + off, 0, len);
270 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
272 if (blk == tdb->transaction->num_blocks-1) {
273 if (len + off > tdb->transaction->last_block_size) {
274 tdb->transaction->last_block_size = len + off;
281 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
282 "transaction_write: failed at off=%llu len=%llu\n",
283 (long long)((blk*getpagesize()) + off),
285 tdb->transaction->transaction_error = 1;
291 write while in a transaction - this varient never expands the transaction blocks, it only
292 updates existing blocks. This means it cannot change the recovery size
294 static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
295 const void *buf, tdb_len_t len)
299 /* break it up into block sized chunks */
300 while (len + (off % getpagesize()) > getpagesize()) {
301 tdb_len_t len2 = getpagesize() - (off % getpagesize());
302 transaction_write_existing(tdb, off, buf, len2);
306 buf = (const void *)(len2 + (const char *)buf);
314 blk = off / getpagesize();
315 off = off % getpagesize();
317 if (tdb->transaction->num_blocks <= blk ||
318 tdb->transaction->blocks[blk] == NULL) {
322 if (blk == tdb->transaction->num_blocks-1 &&
323 off + len > tdb->transaction->last_block_size) {
324 if (off >= tdb->transaction->last_block_size) {
327 len = tdb->transaction->last_block_size - off;
330 /* overwrite part of an existing block */
331 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
336 out of bounds check during a transaction
338 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, bool probe)
340 if (len <= tdb->map_size) {
343 tdb->ecode = TDB_ERR_IO;
348 transaction version of tdb_expand().
350 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t addition)
352 /* add a write to the transaction elements, so subsequent
353 reads see the zero data */
354 if (transaction_write(tdb, tdb->map_size, NULL, addition) != 0) {
357 tdb->map_size += addition;
361 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
368 static const struct tdb_methods transaction_methods = {
372 transaction_expand_file,
379 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
381 if (tdb->flags & TDB_NOSYNC) {
385 if (fsync(tdb->fd) != 0) {
386 tdb->ecode = TDB_ERR_IO;
387 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
388 "tdb_transaction: fsync failed\n");
393 tdb_off_t moffset = offset & ~(getpagesize()-1);
394 if (msync(moffset + (char *)tdb->map_ptr,
395 length + (offset - moffset), MS_SYNC) != 0) {
396 tdb->ecode = TDB_ERR_IO;
397 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
398 "tdb_transaction: msync failed - %s\n",
408 static void _tdb_transaction_cancel(struct tdb_context *tdb)
412 if (tdb->transaction == NULL) {
413 tdb->ecode = TDB_ERR_EINVAL;
414 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
415 "tdb_transaction_cancel: no transaction\n");
419 if (tdb->transaction->nesting != 0) {
420 tdb->transaction->transaction_error = 1;
421 tdb->transaction->nesting--;
425 tdb->map_size = tdb->transaction->old_map_size;
427 /* free all the transaction blocks */
428 for (i=0;i<tdb->transaction->num_blocks;i++) {
429 if (tdb->transaction->blocks[i] != NULL) {
430 free(tdb->transaction->blocks[i]);
433 SAFE_FREE(tdb->transaction->blocks);
435 if (tdb->transaction->magic_offset) {
436 const struct tdb_methods *methods = tdb->transaction->io_methods;
437 uint64_t invalid = TDB_RECOVERY_INVALID_MAGIC;
439 /* remove the recovery marker */
440 if (methods->write(tdb, tdb->transaction->magic_offset,
441 &invalid, sizeof(invalid)) == -1 ||
442 transaction_sync(tdb, tdb->transaction->magic_offset,
443 sizeof(invalid)) == -1) {
444 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
445 "tdb_transaction_cancel: failed to remove"
446 " recovery magic\n");
450 if (tdb->allrecord_lock.count)
451 tdb_allrecord_unlock(tdb, tdb->allrecord_lock.ltype);
453 /* restore the normal io methods */
454 tdb->methods = tdb->transaction->io_methods;
456 tdb_transaction_unlock(tdb, F_WRLCK);
457 tdb_unlock_expand(tdb, F_WRLCK);
459 if (tdb_has_open_lock(tdb))
460 tdb_unlock_open(tdb);
462 SAFE_FREE(tdb->transaction);
466 start a tdb transaction. No token is returned, as only a single
467 transaction is allowed to be pending per tdb_context
469 int tdb_transaction_start(struct tdb_context *tdb)
471 /* some sanity checks */
472 if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
473 tdb->ecode = TDB_ERR_EINVAL;
474 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
475 "tdb_transaction_start: cannot start a transaction"
476 " on a read-only or internal db\n");
480 /* cope with nested tdb_transaction_start() calls */
481 if (tdb->transaction != NULL) {
482 tdb->ecode = TDB_ERR_NESTING;
486 if (tdb_has_hash_locks(tdb)) {
487 /* the caller must not have any locks when starting a
488 transaction as otherwise we'll be screwed by lack
489 of nested locks in posix */
490 tdb->ecode = TDB_ERR_LOCK;
491 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
492 "tdb_transaction_start: cannot start a transaction"
493 " with locks held\n");
497 tdb->transaction = (struct tdb_transaction *)
498 calloc(sizeof(struct tdb_transaction), 1);
499 if (tdb->transaction == NULL) {
500 tdb->ecode = TDB_ERR_OOM;
504 /* get the transaction write lock. This is a blocking lock. As
505 discussed with Volker, there are a number of ways we could
506 make this async, which we will probably do in the future */
507 if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
508 SAFE_FREE(tdb->transaction->blocks);
509 SAFE_FREE(tdb->transaction);
513 /* get a read lock over entire file. This is upgraded to a write
514 lock during the commit */
515 if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
516 goto fail_allrecord_lock;
519 if (tdb_lock_expand(tdb, F_WRLCK) != 0) {
520 goto fail_expand_lock;
523 /* make sure we know about any file expansions already done by
525 tdb->methods->oob(tdb, tdb->map_size + 1, true);
526 tdb->transaction->old_map_size = tdb->map_size;
528 /* finally hook the io methods, replacing them with
529 transaction specific methods */
530 tdb->transaction->io_methods = tdb->methods;
531 tdb->methods = &transaction_methods;
535 tdb_allrecord_unlock(tdb, F_RDLCK);
537 tdb_transaction_unlock(tdb, F_WRLCK);
538 SAFE_FREE(tdb->transaction->blocks);
539 SAFE_FREE(tdb->transaction);
545 cancel the current transaction
547 void tdb_transaction_cancel(struct tdb_context *tdb)
549 _tdb_transaction_cancel(tdb);
553 work out how much space the linearised recovery data will consume
555 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
557 tdb_len_t recovery_size = 0;
560 recovery_size = sizeof(tdb_len_t);
561 for (i=0;i<tdb->transaction->num_blocks;i++) {
562 if (i * getpagesize() >= tdb->transaction->old_map_size) {
565 if (tdb->transaction->blocks[i] == NULL) {
568 recovery_size += 2*sizeof(tdb_off_t);
569 if (i == tdb->transaction->num_blocks-1) {
570 recovery_size += tdb->transaction->last_block_size;
572 recovery_size += getpagesize();
576 return recovery_size;
580 allocate the recovery area, or use an existing recovery area if it is
583 static int tdb_recovery_allocate(struct tdb_context *tdb,
584 tdb_len_t *recovery_size,
585 tdb_off_t *recovery_offset,
586 tdb_len_t *recovery_max_size)
588 struct tdb_recovery_record rec;
589 const struct tdb_methods *methods = tdb->transaction->io_methods;
590 tdb_off_t recovery_head;
593 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
594 if (recovery_head == TDB_OFF_ERR) {
595 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
596 "tdb_recovery_allocate:"
597 " failed to read recovery head\n");
601 if (recovery_head != 0) {
602 if (methods->read(tdb, recovery_head, &rec, sizeof(rec))) {
603 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
604 "tdb_recovery_allocate:"
605 " failed to read recovery record\n");
608 tdb_convert(tdb, &rec, sizeof(rec));
609 /* ignore invalid recovery regions: can happen in crash */
610 if (rec.magic != TDB_RECOVERY_MAGIC &&
611 rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
616 *recovery_size = tdb_recovery_size(tdb);
618 if (recovery_head != 0 && *recovery_size <= rec.max_len) {
619 /* it fits in the existing area */
620 *recovery_max_size = rec.max_len;
621 *recovery_offset = recovery_head;
625 /* we need to free up the old recovery area, then allocate a
626 new one at the end of the file. Note that we cannot use
627 normal allocation to allocate the new one as that might return
628 us an area that is being currently used (as of the start of
630 if (recovery_head != 0) {
631 if (add_free_record(tdb, recovery_head,
632 sizeof(rec) + rec.max_len) != 0) {
633 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
634 "tdb_recovery_allocate:"
635 " failed to free previous recovery area\n");
640 /* the tdb_free() call might have increased the recovery size */
641 *recovery_size = tdb_recovery_size(tdb);
643 /* round up to a multiple of page size */
645 = (((sizeof(rec) + *recovery_size) + getpagesize()-1)
646 & ~(getpagesize()-1))
648 *recovery_offset = tdb->map_size;
649 recovery_head = *recovery_offset;
651 /* Restore ->map_size before calling underlying expand_file.
652 Also so that we don't try to expand the file again in the
653 transaction commit, which would destroy the recovery
655 addition = (tdb->map_size - tdb->transaction->old_map_size) +
656 sizeof(rec) + *recovery_max_size;
657 tdb->map_size = tdb->transaction->old_map_size;
658 if (methods->expand_file(tdb, addition) == -1) {
659 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
660 "tdb_recovery_allocate:"
661 " failed to create recovery area\n");
665 /* we have to reset the old map size so that we don't try to
666 expand the file again in the transaction commit, which
667 would destroy the recovery area */
668 tdb->transaction->old_map_size = tdb->map_size;
670 /* write the recovery header offset and sync - we can sync without a race here
671 as the magic ptr in the recovery record has not been set */
672 tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
673 if (methods->write(tdb, offsetof(struct tdb_header, recovery),
674 &recovery_head, sizeof(tdb_off_t)) == -1) {
675 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
676 "tdb_recovery_allocate:"
677 " failed to write recovery head\n");
680 transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
686 /* Set up header for the recovery record. */
687 static void set_recovery_header(struct tdb_recovery_record *rec,
689 uint64_t datalen, uint64_t actuallen,
693 rec->max_len = actuallen;
699 setup the recovery data that will be used on a crash during commit
701 static int transaction_setup_recovery(struct tdb_context *tdb,
702 tdb_off_t *magic_offset)
704 tdb_len_t recovery_size;
705 unsigned char *data, *p;
706 const struct tdb_methods *methods = tdb->transaction->io_methods;
707 struct tdb_recovery_record *rec;
708 tdb_off_t recovery_offset, recovery_max_size;
709 tdb_off_t old_map_size = tdb->transaction->old_map_size;
710 uint64_t magic, tailer;
714 check that the recovery area has enough space
716 if (tdb_recovery_allocate(tdb, &recovery_size,
717 &recovery_offset, &recovery_max_size) == -1) {
721 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
723 tdb->ecode = TDB_ERR_OOM;
727 rec = (struct tdb_recovery_record *)data;
728 set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
729 recovery_size, recovery_max_size, old_map_size);
730 tdb_convert(tdb, rec, sizeof(*rec));
732 /* build the recovery data into a single blob to allow us to do a single
733 large write, which should be more efficient */
734 p = data + sizeof(*rec);
735 for (i=0;i<tdb->transaction->num_blocks;i++) {
739 if (tdb->transaction->blocks[i] == NULL) {
743 offset = i * getpagesize();
744 length = getpagesize();
745 if (i == tdb->transaction->num_blocks-1) {
746 length = tdb->transaction->last_block_size;
749 if (offset >= old_map_size) {
752 if (offset + length > tdb->map_size) {
753 tdb->ecode = TDB_ERR_CORRUPT;
754 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
755 "tdb_transaction_setup_recovery:"
756 " transaction data over new region boundary\n");
760 memcpy(p, &offset, sizeof(offset));
761 memcpy(p + sizeof(offset), &length, sizeof(length));
762 tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
764 /* the recovery area contains the old data, not the
765 new data, so we have to call the original tdb_read
767 if (methods->read(tdb, offset,
768 p + sizeof(offset) + sizeof(length),
773 p += sizeof(offset) + sizeof(length) + length;
777 tailer = sizeof(*rec) + recovery_max_size;
778 memcpy(p, &tailer, sizeof(tailer));
779 tdb_convert(tdb, p, sizeof(tailer));
781 /* write the recovery data to the recovery area */
782 if (methods->write(tdb, recovery_offset, data,
783 sizeof(*rec) + recovery_size) == -1) {
784 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
785 "tdb_transaction_setup_recovery:"
786 " failed to write recovery data\n");
790 transaction_write_existing(tdb, recovery_offset, data,
791 sizeof(*rec) + recovery_size);
793 /* as we don't have ordered writes, we have to sync the recovery
794 data before we update the magic to indicate that the recovery
796 if (transaction_sync(tdb, recovery_offset,
797 sizeof(*rec) + recovery_size) == -1) {
804 magic = TDB_RECOVERY_MAGIC;
805 tdb_convert(tdb, &magic, sizeof(magic));
807 *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
810 if (methods->write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
811 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
812 "tdb_transaction_setup_recovery:"
813 " failed to write recovery magic\n");
816 transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
818 /* ensure the recovery magic marker is on disk */
819 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
826 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
828 const struct tdb_methods *methods;
830 if (tdb->transaction == NULL) {
831 tdb->ecode = TDB_ERR_EINVAL;
832 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
833 "tdb_transaction_prepare_commit: no transaction\n");
837 if (tdb->transaction->prepared) {
838 tdb->ecode = TDB_ERR_EINVAL;
839 _tdb_transaction_cancel(tdb);
840 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
841 "tdb_transaction_prepare_commit:"
842 " transaction already prepared\n");
846 if (tdb->transaction->transaction_error) {
847 tdb->ecode = TDB_ERR_IO;
848 _tdb_transaction_cancel(tdb);
849 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
850 "tdb_transaction_prepare_commit:"
851 " transaction error pending\n");
856 if (tdb->transaction->nesting != 0) {
857 tdb->transaction->nesting--;
861 /* check for a null transaction */
862 if (tdb->transaction->blocks == NULL) {
866 methods = tdb->transaction->io_methods;
868 /* upgrade the main transaction lock region to a write lock */
869 if (tdb_allrecord_upgrade(tdb) == -1) {
870 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
871 "tdb_transaction_prepare_commit:"
872 " failed to upgrade hash locks\n");
873 _tdb_transaction_cancel(tdb);
877 /* get the open lock - this prevents new users attaching to the database
879 if (tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK) == -1) {
880 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
881 "tdb_transaction_prepare_commit:"
882 " failed to get open lock\n");
883 _tdb_transaction_cancel(tdb);
887 if (!(tdb->flags & TDB_NOSYNC)) {
888 /* write the recovery data to the end of the file */
889 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
890 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
891 "tdb_transaction_prepare_commit:"
892 " failed to setup recovery data\n");
893 _tdb_transaction_cancel(tdb);
898 tdb->transaction->prepared = true;
900 /* expand the file to the new size if needed */
901 if (tdb->map_size != tdb->transaction->old_map_size) {
902 tdb_len_t add = tdb->map_size - tdb->transaction->old_map_size;
903 /* Restore original map size for tdb_expand_file */
904 tdb->map_size = tdb->transaction->old_map_size;
905 if (methods->expand_file(tdb, add) == -1) {
906 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
907 "tdb_transaction_prepare_commit:"
908 " expansion failed\n");
909 _tdb_transaction_cancel(tdb);
914 /* Keep the open lock until the actual commit */
920 prepare to commit the current transaction
922 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
924 return _tdb_transaction_prepare_commit(tdb);
928 commit the current transaction
930 int tdb_transaction_commit(struct tdb_context *tdb)
932 const struct tdb_methods *methods;
935 if (tdb->transaction == NULL) {
936 tdb->ecode = TDB_ERR_EINVAL;
937 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
938 "tdb_transaction_commit: no transaction\n");
942 tdb_trace(tdb, "tdb_transaction_commit");
944 if (tdb->transaction->transaction_error) {
945 tdb->ecode = TDB_ERR_IO;
946 tdb_transaction_cancel(tdb);
947 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
948 "tdb_transaction_commit: transaction error pending\n");
953 if (tdb->transaction->nesting != 0) {
954 tdb->transaction->nesting--;
958 /* check for a null transaction */
959 if (tdb->transaction->blocks == NULL) {
960 _tdb_transaction_cancel(tdb);
964 if (!tdb->transaction->prepared) {
965 int ret = _tdb_transaction_prepare_commit(tdb);
970 methods = tdb->transaction->io_methods;
972 /* perform all the writes */
973 for (i=0;i<tdb->transaction->num_blocks;i++) {
977 if (tdb->transaction->blocks[i] == NULL) {
981 offset = i * getpagesize();
982 length = getpagesize();
983 if (i == tdb->transaction->num_blocks-1) {
984 length = tdb->transaction->last_block_size;
987 if (methods->write(tdb, offset, tdb->transaction->blocks[i],
989 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
990 "tdb_transaction_commit:"
991 " write failed during commit\n");
993 /* we've overwritten part of the data and
994 possibly expanded the file, so we need to
995 run the crash recovery code */
996 tdb->methods = methods;
997 tdb_transaction_recover(tdb);
999 _tdb_transaction_cancel(tdb);
1003 SAFE_FREE(tdb->transaction->blocks[i]);
1006 SAFE_FREE(tdb->transaction->blocks);
1007 tdb->transaction->num_blocks = 0;
1009 /* ensure the new data is on disk */
1010 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1015 TODO: maybe write to some dummy hdr field, or write to magic
1016 offset without mmap, before the last sync, instead of the
1020 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1021 don't change the mtime of the file, this means the file may
1022 not be backed up (as tdb rounding to block sizes means that
1023 file size changes are quite rare too). The following forces
1024 mtime changes when a transaction completes */
1026 utime(tdb->name, NULL);
1029 /* use a transaction cancel to free memory and remove the
1030 transaction locks */
1031 _tdb_transaction_cancel(tdb);
1038 recover from an aborted transaction. Must be called with exclusive
1039 database write access already established (including the open
1040 lock to prevent new processes attaching)
1042 int tdb_transaction_recover(struct tdb_context *tdb)
1044 tdb_off_t recovery_head, recovery_eof;
1045 unsigned char *data, *p;
1046 struct tdb_recovery_record rec;
1048 /* find the recovery area */
1049 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1050 if (recovery_head == TDB_OFF_ERR) {
1051 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1052 "tdb_transaction_recover:"
1053 " failed to read recovery head\n");
1057 if (recovery_head == 0) {
1058 /* we have never allocated a recovery record */
1062 /* read the recovery record */
1063 if (tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec)) == -1) {
1064 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1065 "tdb_transaction_recover:"
1066 " failed to read recovery record\n");
1070 if (rec.magic != TDB_RECOVERY_MAGIC) {
1071 /* there is no valid recovery data */
1075 if (tdb->read_only) {
1076 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1077 "tdb_transaction_recover:"
1078 " attempt to recover read only database\n");
1079 tdb->ecode = TDB_ERR_CORRUPT;
1083 recovery_eof = rec.eof;
1085 data = (unsigned char *)malloc(rec.len);
1087 tdb->ecode = TDB_ERR_OOM;
1088 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1089 "tdb_transaction_recover:"
1090 " failed to allocate recovery data\n");
1094 /* read the full recovery data */
1095 if (tdb->methods->read(tdb, recovery_head + sizeof(rec), data,
1097 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1098 "tdb_transaction_recover:"
1099 " failed to read recovery data\n");
1103 /* recover the file data */
1105 while (p+sizeof(tdb_off_t)+sizeof(tdb_len_t) < data + rec.len) {
1108 tdb_convert(tdb, p, sizeof(ofs) + sizeof(len));
1109 memcpy(&ofs, p, sizeof(ofs));
1110 memcpy(&len, p + sizeof(ofs), sizeof(len));
1111 p += sizeof(ofs) + sizeof(len);
1113 if (tdb->methods->write(tdb, ofs, p, len) == -1) {
1115 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1116 "tdb_transaction_recover:"
1117 " failed to recover %zu bytes at offset %zu\n",
1118 (size_t)len, (size_t)ofs);
1126 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1127 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1128 "tdb_transaction_recover: failed to sync recovery\n");
1132 /* if the recovery area is after the recovered eof then remove it */
1133 if (recovery_eof <= recovery_head) {
1134 if (tdb_write_off(tdb, offsetof(struct tdb_header,recovery), 0)
1136 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1137 "tdb_transaction_recover:"
1138 " failed to remove recovery head\n");
1143 /* remove the recovery magic */
1144 if (tdb_write_off(tdb,
1146 + offsetof(struct tdb_recovery_record, magic),
1147 TDB_RECOVERY_INVALID_MAGIC) == -1) {
1148 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1149 "tdb_transaction_recover:"
1150 " failed to remove recovery magic\n");
1154 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1155 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1156 "tdb_transaction_recover: failed to sync2 recovery\n");
1160 tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
1161 "tdb_transaction_recover: recovered %zu byte database\n",
1162 (size_t)recovery_eof);
1168 /* Any I/O failures we say "needs recovery". */
1169 bool tdb_needs_recovery(struct tdb_context *tdb)
1171 tdb_off_t recovery_head;
1172 struct tdb_recovery_record rec;
1174 /* find the recovery area */
1175 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1176 if (recovery_head == TDB_OFF_ERR) {
1180 if (recovery_head == 0) {
1181 /* we have never allocated a recovery record */
1185 /* read the recovery record */
1186 if (tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec)) == -1) {
1190 return (rec.magic == TDB_RECOVERY_MAGIC);