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);
458 if (tdb_has_open_lock(tdb))
459 tdb_unlock_open(tdb);
461 SAFE_FREE(tdb->transaction);
465 start a tdb transaction. No token is returned, as only a single
466 transaction is allowed to be pending per tdb_context
468 int tdb_transaction_start(struct tdb_context *tdb)
470 /* some sanity checks */
471 if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
472 tdb->ecode = TDB_ERR_EINVAL;
473 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
474 "tdb_transaction_start: cannot start a transaction"
475 " on a read-only or internal db\n");
479 /* cope with nested tdb_transaction_start() calls */
480 if (tdb->transaction != NULL) {
481 tdb->ecode = TDB_ERR_NESTING;
485 if (tdb_has_hash_locks(tdb)) {
486 /* the caller must not have any locks when starting a
487 transaction as otherwise we'll be screwed by lack
488 of nested locks in posix */
489 tdb->ecode = TDB_ERR_LOCK;
490 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
491 "tdb_transaction_start: cannot start a transaction"
492 " with locks held\n");
496 tdb->transaction = (struct tdb_transaction *)
497 calloc(sizeof(struct tdb_transaction), 1);
498 if (tdb->transaction == NULL) {
499 tdb->ecode = TDB_ERR_OOM;
503 /* get the transaction write lock. This is a blocking lock. As
504 discussed with Volker, there are a number of ways we could
505 make this async, which we will probably do in the future */
506 if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
507 SAFE_FREE(tdb->transaction->blocks);
508 SAFE_FREE(tdb->transaction);
512 /* get a read lock over entire file. This is upgraded to a write
513 lock during the commit */
514 if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
515 goto fail_allrecord_lock;
518 /* make sure we know about any file expansions already done by
520 tdb->methods->oob(tdb, tdb->map_size + 1, true);
521 tdb->transaction->old_map_size = tdb->map_size;
523 /* finally hook the io methods, replacing them with
524 transaction specific methods */
525 tdb->transaction->io_methods = tdb->methods;
526 tdb->methods = &transaction_methods;
530 tdb_transaction_unlock(tdb, F_WRLCK);
531 SAFE_FREE(tdb->transaction->blocks);
532 SAFE_FREE(tdb->transaction);
538 cancel the current transaction
540 void tdb_transaction_cancel(struct tdb_context *tdb)
542 _tdb_transaction_cancel(tdb);
546 work out how much space the linearised recovery data will consume
548 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
550 tdb_len_t recovery_size = 0;
553 recovery_size = sizeof(tdb_len_t);
554 for (i=0;i<tdb->transaction->num_blocks;i++) {
555 if (i * getpagesize() >= tdb->transaction->old_map_size) {
558 if (tdb->transaction->blocks[i] == NULL) {
561 recovery_size += 2*sizeof(tdb_off_t);
562 if (i == tdb->transaction->num_blocks-1) {
563 recovery_size += tdb->transaction->last_block_size;
565 recovery_size += getpagesize();
569 return recovery_size;
573 allocate the recovery area, or use an existing recovery area if it is
576 static int tdb_recovery_allocate(struct tdb_context *tdb,
577 tdb_len_t *recovery_size,
578 tdb_off_t *recovery_offset,
579 tdb_len_t *recovery_max_size)
581 struct tdb_recovery_record rec;
582 const struct tdb_methods *methods = tdb->transaction->io_methods;
583 tdb_off_t recovery_head;
586 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
587 if (recovery_head == TDB_OFF_ERR) {
588 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
589 "tdb_recovery_allocate:"
590 " failed to read recovery head\n");
594 if (recovery_head != 0) {
595 if (methods->read(tdb, recovery_head, &rec, sizeof(rec))) {
596 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
597 "tdb_recovery_allocate:"
598 " failed to read recovery record\n");
601 tdb_convert(tdb, &rec, sizeof(rec));
602 /* ignore invalid recovery regions: can happen in crash */
603 if (rec.magic != TDB_RECOVERY_MAGIC &&
604 rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
609 *recovery_size = tdb_recovery_size(tdb);
611 if (recovery_head != 0 && *recovery_size <= rec.max_len) {
612 /* it fits in the existing area */
613 *recovery_max_size = rec.max_len;
614 *recovery_offset = recovery_head;
618 /* we need to free up the old recovery area, then allocate a
619 new one at the end of the file. Note that we cannot use
620 normal allocation to allocate the new one as that might return
621 us an area that is being currently used (as of the start of
623 if (recovery_head != 0) {
624 if (add_free_record(tdb, recovery_head,
625 sizeof(rec) + rec.max_len) != 0) {
626 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
627 "tdb_recovery_allocate:"
628 " failed to free previous recovery area\n");
633 /* the tdb_free() call might have increased the recovery size */
634 *recovery_size = tdb_recovery_size(tdb);
636 /* round up to a multiple of page size */
638 = (((sizeof(rec) + *recovery_size) + getpagesize()-1)
639 & ~(getpagesize()-1))
641 *recovery_offset = tdb->map_size;
642 recovery_head = *recovery_offset;
644 /* Restore ->map_size before calling underlying expand_file.
645 Also so that we don't try to expand the file again in the
646 transaction commit, which would destroy the recovery
648 addition = (tdb->map_size - tdb->transaction->old_map_size) +
649 sizeof(rec) + *recovery_max_size;
650 tdb->map_size = tdb->transaction->old_map_size;
651 if (methods->expand_file(tdb, addition) == -1) {
652 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
653 "tdb_recovery_allocate:"
654 " failed to create recovery area\n");
658 /* we have to reset the old map size so that we don't try to
659 expand the file again in the transaction commit, which
660 would destroy the recovery area */
661 tdb->transaction->old_map_size = tdb->map_size;
663 /* write the recovery header offset and sync - we can sync without a race here
664 as the magic ptr in the recovery record has not been set */
665 tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
666 if (methods->write(tdb, offsetof(struct tdb_header, recovery),
667 &recovery_head, sizeof(tdb_off_t)) == -1) {
668 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
669 "tdb_recovery_allocate:"
670 " failed to write recovery head\n");
673 transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
679 /* Set up header for the recovery record. */
680 static void set_recovery_header(struct tdb_recovery_record *rec,
682 uint64_t datalen, uint64_t actuallen,
686 rec->max_len = actuallen;
692 setup the recovery data that will be used on a crash during commit
694 static int transaction_setup_recovery(struct tdb_context *tdb,
695 tdb_off_t *magic_offset)
697 tdb_len_t recovery_size;
698 unsigned char *data, *p;
699 const struct tdb_methods *methods = tdb->transaction->io_methods;
700 struct tdb_recovery_record *rec;
701 tdb_off_t recovery_offset, recovery_max_size;
702 tdb_off_t old_map_size = tdb->transaction->old_map_size;
703 uint64_t magic, tailer;
707 check that the recovery area has enough space
709 if (tdb_recovery_allocate(tdb, &recovery_size,
710 &recovery_offset, &recovery_max_size) == -1) {
714 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
716 tdb->ecode = TDB_ERR_OOM;
720 rec = (struct tdb_recovery_record *)data;
721 set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
722 recovery_size, recovery_max_size, old_map_size);
723 tdb_convert(tdb, rec, sizeof(*rec));
725 /* build the recovery data into a single blob to allow us to do a single
726 large write, which should be more efficient */
727 p = data + sizeof(*rec);
728 for (i=0;i<tdb->transaction->num_blocks;i++) {
732 if (tdb->transaction->blocks[i] == NULL) {
736 offset = i * getpagesize();
737 length = getpagesize();
738 if (i == tdb->transaction->num_blocks-1) {
739 length = tdb->transaction->last_block_size;
742 if (offset >= old_map_size) {
745 if (offset + length > tdb->map_size) {
746 tdb->ecode = TDB_ERR_CORRUPT;
747 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
748 "tdb_transaction_setup_recovery:"
749 " transaction data over new region boundary\n");
753 memcpy(p, &offset, sizeof(offset));
754 memcpy(p + sizeof(offset), &length, sizeof(length));
755 tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
757 /* the recovery area contains the old data, not the
758 new data, so we have to call the original tdb_read
760 if (methods->read(tdb, offset,
761 p + sizeof(offset) + sizeof(length),
766 p += sizeof(offset) + sizeof(length) + length;
770 tailer = sizeof(*rec) + recovery_max_size;
771 memcpy(p, &tailer, sizeof(tailer));
772 tdb_convert(tdb, p, sizeof(tailer));
774 /* write the recovery data to the recovery area */
775 if (methods->write(tdb, recovery_offset, data,
776 sizeof(*rec) + recovery_size) == -1) {
777 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
778 "tdb_transaction_setup_recovery:"
779 " failed to write recovery data\n");
783 transaction_write_existing(tdb, recovery_offset, data,
784 sizeof(*rec) + recovery_size);
786 /* as we don't have ordered writes, we have to sync the recovery
787 data before we update the magic to indicate that the recovery
789 if (transaction_sync(tdb, recovery_offset,
790 sizeof(*rec) + recovery_size) == -1) {
797 magic = TDB_RECOVERY_MAGIC;
798 tdb_convert(tdb, &magic, sizeof(magic));
800 *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
803 if (methods->write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
804 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
805 "tdb_transaction_setup_recovery:"
806 " failed to write recovery magic\n");
809 transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
811 /* ensure the recovery magic marker is on disk */
812 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
819 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
821 const struct tdb_methods *methods;
823 if (tdb->transaction == NULL) {
824 tdb->ecode = TDB_ERR_EINVAL;
825 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
826 "tdb_transaction_prepare_commit: no transaction\n");
830 if (tdb->transaction->prepared) {
831 tdb->ecode = TDB_ERR_EINVAL;
832 _tdb_transaction_cancel(tdb);
833 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
834 "tdb_transaction_prepare_commit:"
835 " transaction already prepared\n");
839 if (tdb->transaction->transaction_error) {
840 tdb->ecode = TDB_ERR_IO;
841 _tdb_transaction_cancel(tdb);
842 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
843 "tdb_transaction_prepare_commit:"
844 " transaction error pending\n");
849 if (tdb->transaction->nesting != 0) {
850 tdb->transaction->nesting--;
854 /* check for a null transaction */
855 if (tdb->transaction->blocks == NULL) {
859 methods = tdb->transaction->io_methods;
861 /* upgrade the main transaction lock region to a write lock */
862 if (tdb_allrecord_upgrade(tdb) == -1) {
863 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
864 "tdb_transaction_prepare_commit:"
865 " failed to upgrade hash locks\n");
866 _tdb_transaction_cancel(tdb);
870 /* get the open lock - this prevents new users attaching to the database
872 if (tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK) == -1) {
873 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
874 "tdb_transaction_prepare_commit:"
875 " failed to get open lock\n");
876 _tdb_transaction_cancel(tdb);
880 /* Since we have whole db locked, we don't need the expansion lock. */
881 if (!(tdb->flags & TDB_NOSYNC)) {
882 /* write the recovery data to the end of the file */
883 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
884 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
885 "tdb_transaction_prepare_commit:"
886 " failed to setup recovery data\n");
887 _tdb_transaction_cancel(tdb);
892 tdb->transaction->prepared = true;
894 /* expand the file to the new size if needed */
895 if (tdb->map_size != tdb->transaction->old_map_size) {
896 tdb_len_t add = tdb->map_size - tdb->transaction->old_map_size;
897 /* Restore original map size for tdb_expand_file */
898 tdb->map_size = tdb->transaction->old_map_size;
899 if (methods->expand_file(tdb, add) == -1) {
900 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
901 "tdb_transaction_prepare_commit:"
902 " expansion failed\n");
903 _tdb_transaction_cancel(tdb);
908 /* Keep the open lock until the actual commit */
914 prepare to commit the current transaction
916 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
918 return _tdb_transaction_prepare_commit(tdb);
922 commit the current transaction
924 int tdb_transaction_commit(struct tdb_context *tdb)
926 const struct tdb_methods *methods;
929 if (tdb->transaction == NULL) {
930 tdb->ecode = TDB_ERR_EINVAL;
931 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
932 "tdb_transaction_commit: no transaction\n");
936 tdb_trace(tdb, "tdb_transaction_commit");
938 if (tdb->transaction->transaction_error) {
939 tdb->ecode = TDB_ERR_IO;
940 tdb_transaction_cancel(tdb);
941 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
942 "tdb_transaction_commit: transaction error pending\n");
947 if (tdb->transaction->nesting != 0) {
948 tdb->transaction->nesting--;
952 /* check for a null transaction */
953 if (tdb->transaction->blocks == NULL) {
954 _tdb_transaction_cancel(tdb);
958 if (!tdb->transaction->prepared) {
959 int ret = _tdb_transaction_prepare_commit(tdb);
964 methods = tdb->transaction->io_methods;
966 /* perform all the writes */
967 for (i=0;i<tdb->transaction->num_blocks;i++) {
971 if (tdb->transaction->blocks[i] == NULL) {
975 offset = i * getpagesize();
976 length = getpagesize();
977 if (i == tdb->transaction->num_blocks-1) {
978 length = tdb->transaction->last_block_size;
981 if (methods->write(tdb, offset, tdb->transaction->blocks[i],
983 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
984 "tdb_transaction_commit:"
985 " write failed during commit\n");
987 /* we've overwritten part of the data and
988 possibly expanded the file, so we need to
989 run the crash recovery code */
990 tdb->methods = methods;
991 tdb_transaction_recover(tdb);
993 _tdb_transaction_cancel(tdb);
997 SAFE_FREE(tdb->transaction->blocks[i]);
1000 SAFE_FREE(tdb->transaction->blocks);
1001 tdb->transaction->num_blocks = 0;
1003 /* ensure the new data is on disk */
1004 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1009 TODO: maybe write to some dummy hdr field, or write to magic
1010 offset without mmap, before the last sync, instead of the
1014 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1015 don't change the mtime of the file, this means the file may
1016 not be backed up (as tdb rounding to block sizes means that
1017 file size changes are quite rare too). The following forces
1018 mtime changes when a transaction completes */
1020 utime(tdb->name, NULL);
1023 /* use a transaction cancel to free memory and remove the
1024 transaction locks */
1025 _tdb_transaction_cancel(tdb);
1032 recover from an aborted transaction. Must be called with exclusive
1033 database write access already established (including the open
1034 lock to prevent new processes attaching)
1036 int tdb_transaction_recover(struct tdb_context *tdb)
1038 tdb_off_t recovery_head, recovery_eof;
1039 unsigned char *data, *p;
1040 struct tdb_recovery_record rec;
1042 /* find the recovery area */
1043 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1044 if (recovery_head == TDB_OFF_ERR) {
1045 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1046 "tdb_transaction_recover:"
1047 " failed to read recovery head\n");
1051 if (recovery_head == 0) {
1052 /* we have never allocated a recovery record */
1056 /* read the recovery record */
1057 if (tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec)) == -1) {
1058 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1059 "tdb_transaction_recover:"
1060 " failed to read recovery record\n");
1064 if (rec.magic != TDB_RECOVERY_MAGIC) {
1065 /* there is no valid recovery data */
1069 if (tdb->read_only) {
1070 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1071 "tdb_transaction_recover:"
1072 " attempt to recover read only database\n");
1073 tdb->ecode = TDB_ERR_CORRUPT;
1077 recovery_eof = rec.eof;
1079 data = (unsigned char *)malloc(rec.len);
1081 tdb->ecode = TDB_ERR_OOM;
1082 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1083 "tdb_transaction_recover:"
1084 " failed to allocate recovery data\n");
1088 /* read the full recovery data */
1089 if (tdb->methods->read(tdb, recovery_head + sizeof(rec), data,
1091 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1092 "tdb_transaction_recover:"
1093 " failed to read recovery data\n");
1097 /* recover the file data */
1099 while (p+sizeof(tdb_off_t)+sizeof(tdb_len_t) < data + rec.len) {
1102 tdb_convert(tdb, p, sizeof(ofs) + sizeof(len));
1103 memcpy(&ofs, p, sizeof(ofs));
1104 memcpy(&len, p + sizeof(ofs), sizeof(len));
1105 p += sizeof(ofs) + sizeof(len);
1107 if (tdb->methods->write(tdb, ofs, p, len) == -1) {
1109 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1110 "tdb_transaction_recover:"
1111 " failed to recover %zu bytes at offset %zu\n",
1112 (size_t)len, (size_t)ofs);
1120 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1121 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1122 "tdb_transaction_recover: failed to sync recovery\n");
1126 /* if the recovery area is after the recovered eof then remove it */
1127 if (recovery_eof <= recovery_head) {
1128 if (tdb_write_off(tdb, offsetof(struct tdb_header,recovery), 0)
1130 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1131 "tdb_transaction_recover:"
1132 " failed to remove recovery head\n");
1137 /* remove the recovery magic */
1138 if (tdb_write_off(tdb,
1140 + offsetof(struct tdb_recovery_record, magic),
1141 TDB_RECOVERY_INVALID_MAGIC) == -1) {
1142 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1143 "tdb_transaction_recover:"
1144 " failed to remove recovery magic\n");
1148 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1149 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
1150 "tdb_transaction_recover: failed to sync2 recovery\n");
1154 tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
1155 "tdb_transaction_recover: recovered %zu byte database\n",
1156 (size_t)recovery_eof);
1162 /* Any I/O failures we say "needs recovery". */
1163 bool tdb_needs_recovery(struct tdb_context *tdb)
1165 tdb_off_t recovery_head;
1166 struct tdb_recovery_record rec;
1168 /* find the recovery area */
1169 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1170 if (recovery_head == TDB_OFF_ERR) {
1174 if (recovery_head == 0) {
1175 /* we have never allocated a recovery record */
1179 /* read the recovery record */
1180 if (tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec)) == -1) {
1184 return (rec.magic == TDB_RECOVERY_MAGIC);