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 enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
129 void *buf, tdb_len_t len)
132 enum TDB_ERROR ecode;
134 /* break it down into block sized ops */
135 while (len + (off % getpagesize()) > getpagesize()) {
136 tdb_len_t len2 = getpagesize() - (off % getpagesize());
137 ecode = transaction_read(tdb, off, buf, len2);
138 if (ecode != TDB_SUCCESS) {
143 buf = (void *)(len2 + (char *)buf);
150 blk = off / getpagesize();
152 /* see if we have it in the block list */
153 if (tdb->transaction->num_blocks <= blk ||
154 tdb->transaction->blocks[blk] == NULL) {
155 /* nope, do a real read */
156 ecode = tdb->transaction->io_methods->tread(tdb, off, buf, len);
157 if (ecode != TDB_SUCCESS) {
163 /* it is in the block list. Now check for the last block */
164 if (blk == tdb->transaction->num_blocks-1) {
165 if (len > tdb->transaction->last_block_size) {
171 /* now copy it out of this block */
172 memcpy(buf, tdb->transaction->blocks[blk] + (off % getpagesize()), len);
176 tdb->transaction->transaction_error = 1;
177 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
178 "transaction_read: failed at off=%zu len=%zu",
179 (size_t)off, (size_t)len);
184 write while in a transaction
186 static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
187 const void *buf, tdb_len_t len)
190 enum TDB_ERROR ecode;
192 /* Only a commit is allowed on a prepared transaction */
193 if (tdb->transaction->prepared) {
194 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
195 "transaction_write: transaction already"
196 " prepared, write not allowed");
200 /* break it up into block sized chunks */
201 while (len + (off % getpagesize()) > getpagesize()) {
202 tdb_len_t len2 = getpagesize() - (off % getpagesize());
203 ecode = transaction_write(tdb, off, buf, len2);
204 if (ecode != TDB_SUCCESS) {
210 buf = (const void *)(len2 + (const char *)buf);
218 blk = off / getpagesize();
219 off = off % getpagesize();
221 if (tdb->transaction->num_blocks <= blk) {
222 uint8_t **new_blocks;
223 /* expand the blocks array */
224 if (tdb->transaction->blocks == NULL) {
225 new_blocks = (uint8_t **)malloc(
226 (blk+1)*sizeof(uint8_t *));
228 new_blocks = (uint8_t **)realloc(
229 tdb->transaction->blocks,
230 (blk+1)*sizeof(uint8_t *));
232 if (new_blocks == NULL) {
233 ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
235 " failed to allocate");
238 memset(&new_blocks[tdb->transaction->num_blocks], 0,
239 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
240 tdb->transaction->blocks = new_blocks;
241 tdb->transaction->num_blocks = blk+1;
242 tdb->transaction->last_block_size = 0;
245 /* allocate and fill a block? */
246 if (tdb->transaction->blocks[blk] == NULL) {
247 tdb->transaction->blocks[blk] = (uint8_t *)calloc(getpagesize(), 1);
248 if (tdb->transaction->blocks[blk] == NULL) {
249 ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
251 " failed to allocate");
254 if (tdb->transaction->old_map_size > blk * getpagesize()) {
255 tdb_len_t len2 = getpagesize();
256 if (len2 + (blk * getpagesize()) > tdb->transaction->old_map_size) {
257 len2 = tdb->transaction->old_map_size - (blk * getpagesize());
259 ecode = tdb->transaction->io_methods->tread(tdb,
261 tdb->transaction->blocks[blk],
263 if (ecode != TDB_SUCCESS) {
264 ecode = tdb_logerr(tdb, ecode,
268 " read old block: %s",
270 SAFE_FREE(tdb->transaction->blocks[blk]);
273 if (blk == tdb->transaction->num_blocks-1) {
274 tdb->transaction->last_block_size = len2;
279 /* overwrite part of an existing block */
281 memset(tdb->transaction->blocks[blk] + off, 0, len);
283 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
285 if (blk == tdb->transaction->num_blocks-1) {
286 if (len + off > tdb->transaction->last_block_size) {
287 tdb->transaction->last_block_size = len + off;
294 tdb->transaction->transaction_error = 1;
300 write while in a transaction - this varient never expands the transaction blocks, it only
301 updates existing blocks. This means it cannot change the recovery size
303 static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
304 const void *buf, tdb_len_t len)
308 /* break it up into block sized chunks */
309 while (len + (off % getpagesize()) > getpagesize()) {
310 tdb_len_t len2 = getpagesize() - (off % getpagesize());
311 transaction_write_existing(tdb, off, buf, len2);
315 buf = (const void *)(len2 + (const char *)buf);
323 blk = off / getpagesize();
324 off = off % getpagesize();
326 if (tdb->transaction->num_blocks <= blk ||
327 tdb->transaction->blocks[blk] == NULL) {
331 if (blk == tdb->transaction->num_blocks-1 &&
332 off + len > tdb->transaction->last_block_size) {
333 if (off >= tdb->transaction->last_block_size) {
336 len = tdb->transaction->last_block_size - off;
339 /* overwrite part of an existing block */
340 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
345 out of bounds check during a transaction
347 static enum TDB_ERROR transaction_oob(struct tdb_context *tdb, tdb_off_t len,
350 if (len <= tdb->map_size) {
354 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
355 "tdb_oob len %lld beyond transaction size %lld",
357 (long long)tdb->map_size);
363 transaction version of tdb_expand().
365 static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
368 enum TDB_ERROR ecode;
370 /* add a write to the transaction elements, so subsequent
371 reads see the zero data */
372 ecode = transaction_write(tdb, tdb->map_size, NULL, addition);
373 if (ecode == TDB_SUCCESS) {
374 tdb->map_size += addition;
379 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
380 size_t len, bool write_mode)
382 size_t blk = off / getpagesize(), end_blk;
384 /* This is wrong for zero-length blocks, but will fail gracefully */
385 end_blk = (off + len - 1) / getpagesize();
387 /* Can only do direct if in single block and we've already copied. */
391 if (blk >= tdb->transaction->num_blocks)
393 if (tdb->transaction->blocks[blk] == NULL)
395 return tdb->transaction->blocks[blk] + off % getpagesize();
398 /* Single which we have copied? */
400 && blk < tdb->transaction->num_blocks
401 && tdb->transaction->blocks[blk])
402 return tdb->transaction->blocks[blk] + off % getpagesize();
404 /* Otherwise must be all not copied. */
405 while (blk < end_blk) {
406 if (blk >= tdb->transaction->num_blocks)
408 if (tdb->transaction->blocks[blk])
412 return tdb->transaction->io_methods->direct(tdb, off, len, false);
415 static const struct tdb_methods transaction_methods = {
419 transaction_expand_file,
426 static enum TDB_ERROR transaction_sync(struct tdb_context *tdb,
427 tdb_off_t offset, tdb_len_t length)
429 if (tdb->flags & TDB_NOSYNC) {
433 if (fsync(tdb->fd) != 0) {
434 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
435 "tdb_transaction: fsync failed: %s",
440 tdb_off_t moffset = offset & ~(getpagesize()-1);
441 if (msync(moffset + (char *)tdb->map_ptr,
442 length + (offset - moffset), MS_SYNC) != 0) {
443 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
444 "tdb_transaction: msync failed: %s",
453 static void _tdb_transaction_cancel(struct tdb_context *tdb)
456 enum TDB_ERROR ecode;
458 if (tdb->transaction == NULL) {
459 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
460 "tdb_transaction_cancel: no transaction");
464 if (tdb->transaction->nesting != 0) {
465 tdb->transaction->transaction_error = 1;
466 tdb->transaction->nesting--;
470 tdb->map_size = tdb->transaction->old_map_size;
472 /* free all the transaction blocks */
473 for (i=0;i<tdb->transaction->num_blocks;i++) {
474 if (tdb->transaction->blocks[i] != NULL) {
475 free(tdb->transaction->blocks[i]);
478 SAFE_FREE(tdb->transaction->blocks);
480 if (tdb->transaction->magic_offset) {
481 const struct tdb_methods *methods = tdb->transaction->io_methods;
482 uint64_t invalid = TDB_RECOVERY_INVALID_MAGIC;
484 /* remove the recovery marker */
485 ecode = methods->twrite(tdb, tdb->transaction->magic_offset,
486 &invalid, sizeof(invalid));
487 if (ecode == TDB_SUCCESS)
488 ecode = transaction_sync(tdb,
489 tdb->transaction->magic_offset,
491 if (ecode != TDB_SUCCESS) {
492 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
493 "tdb_transaction_cancel: failed to remove"
498 if (tdb->allrecord_lock.count)
499 tdb_allrecord_unlock(tdb, tdb->allrecord_lock.ltype);
501 /* restore the normal io methods */
502 tdb->methods = tdb->transaction->io_methods;
504 tdb_transaction_unlock(tdb, F_WRLCK);
506 if (tdb_has_open_lock(tdb))
507 tdb_unlock_open(tdb);
509 SAFE_FREE(tdb->transaction);
513 start a tdb transaction. No token is returned, as only a single
514 transaction is allowed to be pending per tdb_context
516 int tdb_transaction_start(struct tdb_context *tdb)
518 enum TDB_ERROR ecode;
520 /* some sanity checks */
521 if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
522 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
523 "tdb_transaction_start: cannot start a transaction"
524 " on a read-only or internal db");
528 /* cope with nested tdb_transaction_start() calls */
529 if (tdb->transaction != NULL) {
530 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_USE_ERROR,
531 "tdb_transaction_start:"
532 " already inside transaction");
536 if (tdb_has_hash_locks(tdb)) {
537 /* the caller must not have any locks when starting a
538 transaction as otherwise we'll be screwed by lack
539 of nested locks in posix */
540 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
541 "tdb_transaction_start: cannot start a transaction"
546 tdb->transaction = (struct tdb_transaction *)
547 calloc(sizeof(struct tdb_transaction), 1);
548 if (tdb->transaction == NULL) {
549 tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
550 "tdb_transaction_start: cannot allocate");
554 /* get the transaction write lock. This is a blocking lock. As
555 discussed with Volker, there are a number of ways we could
556 make this async, which we will probably do in the future */
557 ecode = tdb_transaction_lock(tdb, F_WRLCK);
558 if (ecode != TDB_SUCCESS) {
560 SAFE_FREE(tdb->transaction->blocks);
561 SAFE_FREE(tdb->transaction);
565 /* get a read lock over entire file. This is upgraded to a write
566 lock during the commit */
567 ecode = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true);
568 if (ecode != TDB_SUCCESS) {
570 goto fail_allrecord_lock;
573 /* make sure we know about any file expansions already done by
575 tdb->methods->oob(tdb, tdb->map_size + 1, true);
576 tdb->transaction->old_map_size = tdb->map_size;
578 /* finally hook the io methods, replacing them with
579 transaction specific methods */
580 tdb->transaction->io_methods = tdb->methods;
581 tdb->methods = &transaction_methods;
585 tdb_transaction_unlock(tdb, F_WRLCK);
586 SAFE_FREE(tdb->transaction->blocks);
587 SAFE_FREE(tdb->transaction);
593 cancel the current transaction
595 void tdb_transaction_cancel(struct tdb_context *tdb)
597 _tdb_transaction_cancel(tdb);
601 work out how much space the linearised recovery data will consume
603 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
605 tdb_len_t recovery_size = 0;
608 recovery_size = sizeof(tdb_len_t);
609 for (i=0;i<tdb->transaction->num_blocks;i++) {
610 if (i * getpagesize() >= tdb->transaction->old_map_size) {
613 if (tdb->transaction->blocks[i] == NULL) {
616 recovery_size += 2*sizeof(tdb_off_t);
617 if (i == tdb->transaction->num_blocks-1) {
618 recovery_size += tdb->transaction->last_block_size;
620 recovery_size += getpagesize();
624 return recovery_size;
628 allocate the recovery area, or use an existing recovery area if it is
631 static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
632 tdb_len_t *recovery_size,
633 tdb_off_t *recovery_offset,
634 tdb_len_t *recovery_max_size)
636 struct tdb_recovery_record rec;
637 const struct tdb_methods *methods = tdb->transaction->io_methods;
638 tdb_off_t recovery_head;
640 enum TDB_ERROR ecode;
642 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
643 if (TDB_OFF_IS_ERR(recovery_head)) {
644 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
645 "tdb_recovery_allocate:"
646 " failed to read recovery head");
649 if (recovery_head != 0) {
650 ecode = methods->tread(tdb, recovery_head, &rec, sizeof(rec));
651 if (ecode != TDB_SUCCESS) {
652 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
653 "tdb_recovery_allocate:"
654 " failed to read recovery record");
656 tdb_convert(tdb, &rec, sizeof(rec));
657 /* ignore invalid recovery regions: can happen in crash */
658 if (rec.magic != TDB_RECOVERY_MAGIC &&
659 rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
664 *recovery_size = tdb_recovery_size(tdb);
666 if (recovery_head != 0 && *recovery_size <= rec.max_len) {
667 /* it fits in the existing area */
668 *recovery_max_size = rec.max_len;
669 *recovery_offset = recovery_head;
673 /* we need to free up the old recovery area, then allocate a
674 new one at the end of the file. Note that we cannot use
675 normal allocation to allocate the new one as that might return
676 us an area that is being currently used (as of the start of
678 if (recovery_head != 0) {
679 add_stat(tdb, frees, 1);
680 ecode = add_free_record(tdb, recovery_head,
681 sizeof(rec) + rec.max_len);
682 if (ecode != TDB_SUCCESS) {
683 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
684 "tdb_recovery_allocate:"
685 " failed to free previous"
690 /* the tdb_free() call might have increased the recovery size */
691 *recovery_size = tdb_recovery_size(tdb);
693 /* round up to a multiple of page size */
695 = (((sizeof(rec) + *recovery_size) + getpagesize()-1)
696 & ~(getpagesize()-1))
698 *recovery_offset = tdb->map_size;
699 recovery_head = *recovery_offset;
701 /* Restore ->map_size before calling underlying expand_file.
702 Also so that we don't try to expand the file again in the
703 transaction commit, which would destroy the recovery
705 addition = (tdb->map_size - tdb->transaction->old_map_size) +
706 sizeof(rec) + *recovery_max_size;
707 tdb->map_size = tdb->transaction->old_map_size;
708 ecode = methods->expand_file(tdb, addition);
709 if (ecode != TDB_SUCCESS) {
710 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
711 "tdb_recovery_allocate:"
712 " failed to create recovery area");
715 /* we have to reset the old map size so that we don't try to
716 expand the file again in the transaction commit, which
717 would destroy the recovery area */
718 tdb->transaction->old_map_size = tdb->map_size;
720 /* write the recovery header offset and sync - we can sync without a race here
721 as the magic ptr in the recovery record has not been set */
722 tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
723 ecode = methods->twrite(tdb, offsetof(struct tdb_header, recovery),
724 &recovery_head, sizeof(tdb_off_t));
725 if (ecode != TDB_SUCCESS) {
726 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
727 "tdb_recovery_allocate:"
728 " failed to write recovery head");
730 transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
736 /* Set up header for the recovery record. */
737 static void set_recovery_header(struct tdb_recovery_record *rec,
739 uint64_t datalen, uint64_t actuallen,
743 rec->max_len = actuallen;
749 setup the recovery data that will be used on a crash during commit
751 static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
752 tdb_off_t *magic_offset)
754 tdb_len_t recovery_size;
755 unsigned char *data, *p;
756 const struct tdb_methods *methods = tdb->transaction->io_methods;
757 struct tdb_recovery_record *rec;
758 tdb_off_t recovery_offset, recovery_max_size;
759 tdb_off_t old_map_size = tdb->transaction->old_map_size;
760 uint64_t magic, tailer;
762 enum TDB_ERROR ecode;
765 check that the recovery area has enough space
767 ecode = tdb_recovery_allocate(tdb, &recovery_size,
768 &recovery_offset, &recovery_max_size);
769 if (ecode != TDB_SUCCESS) {
773 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
775 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
776 "transaction_setup_recovery:"
780 rec = (struct tdb_recovery_record *)data;
781 set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
782 recovery_size, recovery_max_size, old_map_size);
783 tdb_convert(tdb, rec, sizeof(*rec));
785 /* build the recovery data into a single blob to allow us to do a single
786 large write, which should be more efficient */
787 p = data + sizeof(*rec);
788 for (i=0;i<tdb->transaction->num_blocks;i++) {
792 if (tdb->transaction->blocks[i] == NULL) {
796 offset = i * getpagesize();
797 length = getpagesize();
798 if (i == tdb->transaction->num_blocks-1) {
799 length = tdb->transaction->last_block_size;
802 if (offset >= old_map_size) {
805 if (offset + length > tdb->map_size) {
807 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
808 "tdb_transaction_setup_recovery:"
809 " transaction data over new region"
812 memcpy(p, &offset, sizeof(offset));
813 memcpy(p + sizeof(offset), &length, sizeof(length));
814 tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
816 /* the recovery area contains the old data, not the
817 new data, so we have to call the original tdb_read
819 ecode = methods->tread(tdb, offset,
820 p + sizeof(offset) + sizeof(length),
822 if (ecode != TDB_SUCCESS) {
826 p += sizeof(offset) + sizeof(length) + length;
830 tailer = sizeof(*rec) + recovery_max_size;
831 memcpy(p, &tailer, sizeof(tailer));
832 tdb_convert(tdb, p, sizeof(tailer));
834 /* write the recovery data to the recovery area */
835 ecode = methods->twrite(tdb, recovery_offset, data,
836 sizeof(*rec) + recovery_size);
837 if (ecode != TDB_SUCCESS) {
839 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
840 "tdb_transaction_setup_recovery:"
841 " failed to write recovery data");
843 transaction_write_existing(tdb, recovery_offset, data,
844 sizeof(*rec) + recovery_size);
846 /* as we don't have ordered writes, we have to sync the recovery
847 data before we update the magic to indicate that the recovery
849 ecode = transaction_sync(tdb, recovery_offset,
850 sizeof(*rec) + recovery_size);
851 if (ecode != TDB_SUCCESS) {
858 magic = TDB_RECOVERY_MAGIC;
859 tdb_convert(tdb, &magic, sizeof(magic));
861 *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
864 ecode = methods->twrite(tdb, *magic_offset, &magic, sizeof(magic));
865 if (ecode != TDB_SUCCESS) {
866 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
867 "tdb_transaction_setup_recovery:"
868 " failed to write recovery magic");
870 transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
872 /* ensure the recovery magic marker is on disk */
873 return transaction_sync(tdb, *magic_offset, sizeof(magic));
876 static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
878 const struct tdb_methods *methods;
879 enum TDB_ERROR ecode;
881 if (tdb->transaction == NULL) {
882 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
883 "tdb_transaction_prepare_commit:"
887 if (tdb->transaction->prepared) {
888 _tdb_transaction_cancel(tdb);
889 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
890 "tdb_transaction_prepare_commit:"
891 " transaction already prepared");
894 if (tdb->transaction->transaction_error) {
895 _tdb_transaction_cancel(tdb);
896 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
897 "tdb_transaction_prepare_commit:"
898 " transaction error pending");
902 if (tdb->transaction->nesting != 0) {
903 tdb->transaction->nesting--;
907 /* check for a null transaction */
908 if (tdb->transaction->blocks == NULL) {
912 methods = tdb->transaction->io_methods;
914 /* upgrade the main transaction lock region to a write lock */
915 ecode = tdb_allrecord_upgrade(tdb);
916 if (ecode != TDB_SUCCESS) {
917 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
918 "tdb_transaction_prepare_commit:"
919 " failed to upgrade hash locks");
920 _tdb_transaction_cancel(tdb);
924 /* get the open lock - this prevents new users attaching to the database
926 ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
927 if (ecode != TDB_SUCCESS) {
928 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
929 "tdb_transaction_prepare_commit:"
930 " failed to get open lock");
931 _tdb_transaction_cancel(tdb);
935 /* Since we have whole db locked, we don't need the expansion lock. */
936 if (!(tdb->flags & TDB_NOSYNC)) {
937 /* write the recovery data to the end of the file */
938 ecode = transaction_setup_recovery(tdb,
941 if (ecode != TDB_SUCCESS) {
942 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
943 "tdb_transaction_prepare_commit:"
944 " failed to setup recovery data");
945 _tdb_transaction_cancel(tdb);
950 tdb->transaction->prepared = true;
952 /* expand the file to the new size if needed */
953 if (tdb->map_size != tdb->transaction->old_map_size) {
954 tdb_len_t add = tdb->map_size - tdb->transaction->old_map_size;
955 /* Restore original map size for tdb_expand_file */
956 tdb->map_size = tdb->transaction->old_map_size;
957 ecode = methods->expand_file(tdb, add);
958 if (ecode != TDB_SUCCESS) {
959 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
960 "tdb_transaction_prepare_commit:"
961 " expansion failed");
962 _tdb_transaction_cancel(tdb);
967 /* Keep the open lock until the actual commit */
972 prepare to commit the current transaction
974 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
976 tdb->ecode = _tdb_transaction_prepare_commit(tdb);
977 if (tdb->ecode != TDB_SUCCESS)
983 commit the current transaction
985 int tdb_transaction_commit(struct tdb_context *tdb)
987 const struct tdb_methods *methods;
989 enum TDB_ERROR ecode;
991 if (tdb->transaction == NULL) {
992 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
993 "tdb_transaction_commit: no transaction");
997 tdb_trace(tdb, "tdb_transaction_commit");
999 if (tdb->transaction->nesting != 0) {
1000 tdb->transaction->nesting--;
1004 /* check for a null transaction */
1005 if (tdb->transaction->blocks == NULL) {
1006 _tdb_transaction_cancel(tdb);
1010 if (!tdb->transaction->prepared) {
1011 tdb->ecode = _tdb_transaction_prepare_commit(tdb);
1012 if (tdb->ecode != TDB_SUCCESS)
1016 methods = tdb->transaction->io_methods;
1018 /* perform all the writes */
1019 for (i=0;i<tdb->transaction->num_blocks;i++) {
1023 if (tdb->transaction->blocks[i] == NULL) {
1027 offset = i * getpagesize();
1028 length = getpagesize();
1029 if (i == tdb->transaction->num_blocks-1) {
1030 length = tdb->transaction->last_block_size;
1033 ecode = methods->twrite(tdb, offset,
1034 tdb->transaction->blocks[i], length);
1035 if (ecode != TDB_SUCCESS) {
1036 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1037 "tdb_transaction_commit:"
1038 " write failed during commit");
1040 /* we've overwritten part of the data and
1041 possibly expanded the file, so we need to
1042 run the crash recovery code */
1043 tdb->methods = methods;
1044 tdb_transaction_recover(tdb);
1046 _tdb_transaction_cancel(tdb);
1050 SAFE_FREE(tdb->transaction->blocks[i]);
1053 SAFE_FREE(tdb->transaction->blocks);
1054 tdb->transaction->num_blocks = 0;
1056 /* ensure the new data is on disk */
1057 ecode = transaction_sync(tdb, 0, tdb->map_size);
1058 if (ecode != TDB_SUCCESS) {
1064 TODO: maybe write to some dummy hdr field, or write to magic
1065 offset without mmap, before the last sync, instead of the
1069 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1070 don't change the mtime of the file, this means the file may
1071 not be backed up (as tdb rounding to block sizes means that
1072 file size changes are quite rare too). The following forces
1073 mtime changes when a transaction completes */
1075 utime(tdb->name, NULL);
1078 /* use a transaction cancel to free memory and remove the
1079 transaction locks */
1080 _tdb_transaction_cancel(tdb);
1087 recover from an aborted transaction. Must be called with exclusive
1088 database write access already established (including the open
1089 lock to prevent new processes attaching)
1091 enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb)
1093 tdb_off_t recovery_head, recovery_eof;
1094 unsigned char *data, *p;
1095 struct tdb_recovery_record rec;
1096 enum TDB_ERROR ecode;
1098 /* find the recovery area */
1099 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1100 if (TDB_OFF_IS_ERR(recovery_head)) {
1101 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
1102 "tdb_transaction_recover:"
1103 " failed to read recovery head");
1106 if (recovery_head == 0) {
1107 /* we have never allocated a recovery record */
1111 /* read the recovery record */
1112 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1113 if (ecode != TDB_SUCCESS) {
1114 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1115 "tdb_transaction_recover:"
1116 " failed to read recovery record");
1119 if (rec.magic != TDB_RECOVERY_MAGIC) {
1120 /* there is no valid recovery data */
1124 if (tdb->read_only) {
1125 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
1126 "tdb_transaction_recover:"
1127 " attempt to recover read only database");
1130 recovery_eof = rec.eof;
1132 data = (unsigned char *)malloc(rec.len);
1134 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
1135 "tdb_transaction_recover:"
1136 " failed to allocate recovery data");
1139 /* read the full recovery data */
1140 ecode = tdb->methods->tread(tdb, recovery_head + sizeof(rec), data,
1142 if (ecode != TDB_SUCCESS) {
1143 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1144 "tdb_transaction_recover:"
1145 " failed to read recovery data");
1148 /* recover the file data */
1150 while (p+sizeof(tdb_off_t)+sizeof(tdb_len_t) < data + rec.len) {
1153 tdb_convert(tdb, p, sizeof(ofs) + sizeof(len));
1154 memcpy(&ofs, p, sizeof(ofs));
1155 memcpy(&len, p + sizeof(ofs), sizeof(len));
1156 p += sizeof(ofs) + sizeof(len);
1158 ecode = tdb->methods->twrite(tdb, ofs, p, len);
1159 if (ecode != TDB_SUCCESS) {
1161 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1162 "tdb_transaction_recover:"
1163 " failed to recover %zu bytes"
1165 (size_t)len, (size_t)ofs);
1172 ecode = transaction_sync(tdb, 0, tdb->map_size);
1173 if (ecode != TDB_SUCCESS) {
1174 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1175 "tdb_transaction_recover:"
1176 " failed to sync recovery");
1179 /* if the recovery area is after the recovered eof then remove it */
1180 if (recovery_eof <= recovery_head) {
1181 ecode = tdb_write_off(tdb, offsetof(struct tdb_header,
1184 if (ecode != TDB_SUCCESS) {
1185 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1186 "tdb_transaction_recover:"
1187 " failed to remove recovery head");
1191 /* remove the recovery magic */
1192 ecode = tdb_write_off(tdb,
1194 + offsetof(struct tdb_recovery_record, magic),
1195 TDB_RECOVERY_INVALID_MAGIC);
1196 if (ecode != TDB_SUCCESS) {
1197 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1198 "tdb_transaction_recover:"
1199 " failed to remove recovery magic");
1202 ecode = transaction_sync(tdb, 0, recovery_eof);
1203 if (ecode != TDB_SUCCESS) {
1204 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1205 "tdb_transaction_recover:"
1206 " failed to sync2 recovery");
1209 tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
1210 "tdb_transaction_recover: recovered %zu byte database",
1211 (size_t)recovery_eof);
1217 tdb_bool_err tdb_needs_recovery(struct tdb_context *tdb)
1219 tdb_off_t recovery_head;
1220 struct tdb_recovery_record rec;
1221 enum TDB_ERROR ecode;
1223 /* find the recovery area */
1224 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1225 if (TDB_OFF_IS_ERR(recovery_head)) {
1226 return recovery_head;
1229 if (recovery_head == 0) {
1230 /* we have never allocated a recovery record */
1234 /* read the recovery record */
1235 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1236 if (ecode != TDB_SUCCESS) {
1240 return (rec.magic == TDB_RECOVERY_MAGIC);