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 canceled
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 canceled
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.
92 hold the context of any current transaction
94 struct tdb_transaction {
95 /* the original io methods - used to do IOs to the real db */
96 const struct tdb_methods *io_methods;
98 /* the list of transaction blocks. When a block is first
99 written to, it gets created in this list */
102 size_t last_block_size; /* number of valid bytes in the last block */
104 /* non-zero when an internal transaction error has
105 occurred. All write operations will then fail until the
106 transaction is ended */
107 int transaction_error;
109 /* when inside a transaction we need to keep track of any
110 nested tdb_transaction_start() calls, as these are allowed,
111 but don't create a new transaction */
112 unsigned int nesting;
114 /* set when a prepare has already occurred */
116 tdb_off_t magic_offset;
118 /* old file size before transaction */
119 tdb_len_t old_map_size;
122 /* This doesn't really need to be pagesize, but we use it for similar reasons. */
123 #define PAGESIZE 65536
126 read while in a transaction. We need to check first if the data is in our list
127 of transaction elements, then if not do a real read
129 static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
130 void *buf, tdb_len_t len)
133 enum TDB_ERROR ecode;
135 /* break it down into block sized ops */
136 while (len + (off % PAGESIZE) > PAGESIZE) {
137 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
138 ecode = transaction_read(tdb, off, buf, len2);
139 if (ecode != TDB_SUCCESS) {
144 buf = (void *)(len2 + (char *)buf);
151 blk = off / PAGESIZE;
153 /* see if we have it in the block list */
154 if (tdb->transaction->num_blocks <= blk ||
155 tdb->transaction->blocks[blk] == NULL) {
156 /* nope, do a real read */
157 ecode = tdb->transaction->io_methods->tread(tdb, off, buf, len);
158 if (ecode != TDB_SUCCESS) {
164 /* it is in the block list. Now check for the last block */
165 if (blk == tdb->transaction->num_blocks-1) {
166 if (len > tdb->transaction->last_block_size) {
172 /* now copy it out of this block */
173 memcpy(buf, tdb->transaction->blocks[blk] + (off % PAGESIZE), len);
177 tdb->transaction->transaction_error = 1;
178 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
179 "transaction_read: failed at off=%zu len=%zu",
180 (size_t)off, (size_t)len);
185 write while in a transaction
187 static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
188 const void *buf, tdb_len_t len)
191 enum TDB_ERROR ecode;
193 /* Only a commit is allowed on a prepared transaction */
194 if (tdb->transaction->prepared) {
195 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
196 "transaction_write: transaction already"
197 " prepared, write not allowed");
201 /* break it up into block sized chunks */
202 while (len + (off % PAGESIZE) > PAGESIZE) {
203 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
204 ecode = transaction_write(tdb, off, buf, len2);
205 if (ecode != TDB_SUCCESS) {
211 buf = (const void *)(len2 + (const char *)buf);
219 blk = off / PAGESIZE;
220 off = off % PAGESIZE;
222 if (tdb->transaction->num_blocks <= blk) {
223 uint8_t **new_blocks;
224 /* expand the blocks array */
225 if (tdb->transaction->blocks == NULL) {
226 new_blocks = (uint8_t **)malloc(
227 (blk+1)*sizeof(uint8_t *));
229 new_blocks = (uint8_t **)realloc(
230 tdb->transaction->blocks,
231 (blk+1)*sizeof(uint8_t *));
233 if (new_blocks == NULL) {
234 ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
236 " failed to allocate");
239 memset(&new_blocks[tdb->transaction->num_blocks], 0,
240 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
241 tdb->transaction->blocks = new_blocks;
242 tdb->transaction->num_blocks = blk+1;
243 tdb->transaction->last_block_size = 0;
246 /* allocate and fill a block? */
247 if (tdb->transaction->blocks[blk] == NULL) {
248 tdb->transaction->blocks[blk] = (uint8_t *)calloc(PAGESIZE, 1);
249 if (tdb->transaction->blocks[blk] == NULL) {
250 ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
252 " failed to allocate");
255 if (tdb->transaction->old_map_size > blk * PAGESIZE) {
256 tdb_len_t len2 = PAGESIZE;
257 if (len2 + (blk * PAGESIZE) > tdb->transaction->old_map_size) {
258 len2 = tdb->transaction->old_map_size - (blk * PAGESIZE);
260 ecode = tdb->transaction->io_methods->tread(tdb,
262 tdb->transaction->blocks[blk],
264 if (ecode != TDB_SUCCESS) {
265 ecode = tdb_logerr(tdb, ecode,
269 " read old block: %s",
271 SAFE_FREE(tdb->transaction->blocks[blk]);
274 if (blk == tdb->transaction->num_blocks-1) {
275 tdb->transaction->last_block_size = len2;
280 /* overwrite part of an existing block */
282 memset(tdb->transaction->blocks[blk] + off, 0, len);
284 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
286 if (blk == tdb->transaction->num_blocks-1) {
287 if (len + off > tdb->transaction->last_block_size) {
288 tdb->transaction->last_block_size = len + off;
295 tdb->transaction->transaction_error = 1;
301 write while in a transaction - this variant never expands the transaction blocks, it only
302 updates existing blocks. This means it cannot change the recovery size
304 static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
305 const void *buf, tdb_len_t len)
309 /* break it up into block sized chunks */
310 while (len + (off % PAGESIZE) > PAGESIZE) {
311 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
312 transaction_write_existing(tdb, off, buf, len2);
316 buf = (const void *)(len2 + (const char *)buf);
324 blk = off / PAGESIZE;
325 off = off % PAGESIZE;
327 if (tdb->transaction->num_blocks <= blk ||
328 tdb->transaction->blocks[blk] == NULL) {
332 if (blk == tdb->transaction->num_blocks-1 &&
333 off + len > tdb->transaction->last_block_size) {
334 if (off >= tdb->transaction->last_block_size) {
337 len = tdb->transaction->last_block_size - off;
340 /* overwrite part of an existing block */
341 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
346 out of bounds check during a transaction
348 static enum TDB_ERROR transaction_oob(struct tdb_context *tdb, tdb_off_t len,
351 if (len <= tdb->file->map_size) {
355 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
356 "tdb_oob len %lld beyond transaction size %lld",
358 (long long)tdb->file->map_size);
364 transaction version of tdb_expand().
366 static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
369 enum TDB_ERROR ecode;
371 /* add a write to the transaction elements, so subsequent
372 reads see the zero data */
373 ecode = transaction_write(tdb, tdb->file->map_size, NULL, addition);
374 if (ecode == TDB_SUCCESS) {
375 tdb->file->map_size += addition;
380 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
381 size_t len, bool write_mode)
383 size_t blk = off / PAGESIZE, end_blk;
385 /* This is wrong for zero-length blocks, but will fail gracefully */
386 end_blk = (off + len - 1) / PAGESIZE;
388 /* Can only do direct if in single block and we've already copied. */
392 if (blk >= tdb->transaction->num_blocks)
394 if (tdb->transaction->blocks[blk] == NULL)
396 return tdb->transaction->blocks[blk] + off % PAGESIZE;
399 /* Single which we have copied? */
401 && blk < tdb->transaction->num_blocks
402 && tdb->transaction->blocks[blk])
403 return tdb->transaction->blocks[blk] + off % PAGESIZE;
405 /* Otherwise must be all not copied. */
406 while (blk <= end_blk) {
407 if (blk >= tdb->transaction->num_blocks)
409 if (tdb->transaction->blocks[blk])
413 return tdb->transaction->io_methods->direct(tdb, off, len, false);
416 static const struct tdb_methods transaction_methods = {
420 transaction_expand_file,
427 static enum TDB_ERROR transaction_sync(struct tdb_context *tdb,
428 tdb_off_t offset, tdb_len_t length)
430 if (tdb->flags & TDB_NOSYNC) {
434 if (fsync(tdb->file->fd) != 0) {
435 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
436 "tdb_transaction: fsync failed: %s",
440 if (tdb->file->map_ptr) {
441 tdb_off_t moffset = offset & ~(PAGESIZE-1);
442 if (msync(moffset + (char *)tdb->file->map_ptr,
443 length + (offset - moffset), MS_SYNC) != 0) {
444 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
445 "tdb_transaction: msync failed: %s",
454 static void _tdb_transaction_cancel(struct tdb_context *tdb)
457 enum TDB_ERROR ecode;
459 if (tdb->transaction == NULL) {
460 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
461 "tdb_transaction_cancel: no transaction");
465 if (tdb->transaction->nesting != 0) {
466 tdb->transaction->transaction_error = 1;
467 tdb->transaction->nesting--;
471 tdb->file->map_size = tdb->transaction->old_map_size;
473 /* free all the transaction blocks */
474 for (i=0;i<tdb->transaction->num_blocks;i++) {
475 if (tdb->transaction->blocks[i] != NULL) {
476 free(tdb->transaction->blocks[i]);
479 SAFE_FREE(tdb->transaction->blocks);
481 if (tdb->transaction->magic_offset) {
482 const struct tdb_methods *methods = tdb->transaction->io_methods;
483 uint64_t invalid = TDB_RECOVERY_INVALID_MAGIC;
485 /* remove the recovery marker */
486 ecode = methods->twrite(tdb, tdb->transaction->magic_offset,
487 &invalid, sizeof(invalid));
488 if (ecode == TDB_SUCCESS)
489 ecode = transaction_sync(tdb,
490 tdb->transaction->magic_offset,
492 if (ecode != TDB_SUCCESS) {
493 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
494 "tdb_transaction_cancel: failed to remove"
499 if (tdb->file->allrecord_lock.count)
500 tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
502 /* restore the normal io methods */
503 tdb->methods = tdb->transaction->io_methods;
505 tdb_transaction_unlock(tdb, F_WRLCK);
507 if (tdb_has_open_lock(tdb))
508 tdb_unlock_open(tdb);
510 SAFE_FREE(tdb->transaction);
514 start a tdb transaction. No token is returned, as only a single
515 transaction is allowed to be pending per tdb_context
517 enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
519 enum TDB_ERROR ecode;
521 /* some sanity checks */
522 if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
523 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
525 "tdb_transaction_start:"
528 "read-only or internal db");
531 /* cope with nested tdb_transaction_start() calls */
532 if (tdb->transaction != NULL) {
533 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
534 return tdb->last_error
535 = tdb_logerr(tdb, TDB_ERR_IO,
537 "tdb_transaction_start:"
538 " already inside transaction");
540 tdb->transaction->nesting++;
544 if (tdb_has_hash_locks(tdb)) {
545 /* the caller must not have any locks when starting a
546 transaction as otherwise we'll be screwed by lack
547 of nested locks in POSIX */
548 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
550 "tdb_transaction_start:"
552 " transaction with locks"
556 tdb->transaction = (struct tdb_transaction *)
557 calloc(sizeof(struct tdb_transaction), 1);
558 if (tdb->transaction == NULL) {
559 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_OOM,
561 "tdb_transaction_start:"
565 /* get the transaction write lock. This is a blocking lock. As
566 discussed with Volker, there are a number of ways we could
567 make this async, which we will probably do in the future */
568 ecode = tdb_transaction_lock(tdb, F_WRLCK);
569 if (ecode != TDB_SUCCESS) {
570 SAFE_FREE(tdb->transaction->blocks);
571 SAFE_FREE(tdb->transaction);
572 return tdb->last_error = ecode;
575 /* get a read lock over entire file. This is upgraded to a write
576 lock during the commit */
577 ecode = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true);
578 if (ecode != TDB_SUCCESS) {
579 goto fail_allrecord_lock;
582 /* make sure we know about any file expansions already done by
584 tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
585 tdb->transaction->old_map_size = tdb->file->map_size;
587 /* finally hook the io methods, replacing them with
588 transaction specific methods */
589 tdb->transaction->io_methods = tdb->methods;
590 tdb->methods = &transaction_methods;
591 return tdb->last_error = TDB_SUCCESS;
594 tdb_transaction_unlock(tdb, F_WRLCK);
595 SAFE_FREE(tdb->transaction->blocks);
596 SAFE_FREE(tdb->transaction);
597 return tdb->last_error = ecode;
602 cancel the current transaction
604 void tdb_transaction_cancel(struct tdb_context *tdb)
606 _tdb_transaction_cancel(tdb);
610 work out how much space the linearised recovery data will consume (worst case)
612 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
614 tdb_len_t recovery_size = 0;
618 for (i=0;i<tdb->transaction->num_blocks;i++) {
619 if (i * PAGESIZE >= tdb->transaction->old_map_size) {
622 if (tdb->transaction->blocks[i] == NULL) {
625 recovery_size += 2*sizeof(tdb_off_t);
626 if (i == tdb->transaction->num_blocks-1) {
627 recovery_size += tdb->transaction->last_block_size;
629 recovery_size += PAGESIZE;
633 return recovery_size;
636 static enum TDB_ERROR tdb_recovery_area(struct tdb_context *tdb,
637 const struct tdb_methods *methods,
638 tdb_off_t *recovery_offset,
639 struct tdb_recovery_record *rec)
641 enum TDB_ERROR ecode;
643 *recovery_offset = tdb_read_off(tdb,
644 offsetof(struct tdb_header, recovery));
645 if (TDB_OFF_IS_ERR(*recovery_offset)) {
646 return *recovery_offset;
649 if (*recovery_offset == 0) {
654 ecode = methods->tread(tdb, *recovery_offset, rec, sizeof(*rec));
655 if (ecode != TDB_SUCCESS)
658 tdb_convert(tdb, rec, sizeof(*rec));
659 /* ignore invalid recovery regions: can happen in crash */
660 if (rec->magic != TDB_RECOVERY_MAGIC &&
661 rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
662 *recovery_offset = 0;
668 static unsigned int same(const unsigned char *new,
669 const unsigned char *old,
674 for (i = 0; i < length; i++) {
675 if (new[i] != old[i])
681 static unsigned int different(const unsigned char *new,
682 const unsigned char *old,
684 unsigned int min_same,
685 unsigned int *samelen)
690 for (i = 0; i < length; i++) {
691 if (new[i] == old[i]) {
694 if (*samelen >= min_same) {
701 if (*samelen < min_same)
703 return length - *samelen;
706 /* Allocates recovery blob, without tdb_recovery_record at head set up. */
707 static struct tdb_recovery_record *alloc_recovery(struct tdb_context *tdb,
710 struct tdb_recovery_record *rec;
712 enum TDB_ERROR ecode;
714 const struct tdb_methods *old_methods = tdb->methods;
716 rec = malloc(sizeof(*rec) + tdb_recovery_size(tdb));
718 tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
719 "transaction_setup_recovery:"
721 return TDB_ERR_PTR(TDB_ERR_OOM);
724 /* We temporarily revert to the old I/O methods, so we can use
726 tdb->methods = tdb->transaction->io_methods;
728 /* build the recovery data into a single blob to allow us to do a single
729 large write, which should be more efficient */
730 p = (unsigned char *)(rec + 1);
731 for (i=0;i<tdb->transaction->num_blocks;i++) {
735 const unsigned char *buffer;
737 if (tdb->transaction->blocks[i] == NULL) {
741 offset = i * PAGESIZE;
743 if (i == tdb->transaction->num_blocks-1) {
744 length = tdb->transaction->last_block_size;
747 if (offset >= tdb->transaction->old_map_size) {
751 if (offset + length > tdb->file->map_size) {
752 ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
753 "tdb_transaction_setup_recovery:"
754 " transaction data over new region"
758 if (offset + length > tdb->transaction->old_map_size) {
759 /* Short read at EOF. */
760 length = tdb->transaction->old_map_size - offset;
762 buffer = tdb_access_read(tdb, offset, length, false);
763 if (TDB_PTR_IS_ERR(buffer)) {
764 ecode = TDB_PTR_ERR(buffer);
768 /* Skip over anything the same at the start. */
769 off = same(tdb->transaction->blocks[i], buffer, length);
772 while (off < length) {
774 unsigned int samelen;
776 len = different(tdb->transaction->blocks[i] + off,
777 buffer + off, length - off,
778 sizeof(offset) + sizeof(len) + 1,
781 memcpy(p, &offset, sizeof(offset));
782 memcpy(p + sizeof(offset), &len, sizeof(len));
783 tdb_convert(tdb, p, sizeof(offset) + sizeof(len));
784 p += sizeof(offset) + sizeof(len);
785 memcpy(p, buffer + off, len);
787 off += len + samelen;
788 offset += len + samelen;
790 tdb_access_release(tdb, buffer);
793 *len = p - (unsigned char *)(rec + 1);
794 tdb->methods = old_methods;
799 tdb->methods = old_methods;
800 return TDB_ERR_PTR(ecode);
803 static tdb_off_t create_recovery_area(struct tdb_context *tdb,
804 tdb_len_t rec_length,
805 struct tdb_recovery_record *rec)
807 tdb_off_t off, recovery_off;
809 enum TDB_ERROR ecode;
810 const struct tdb_methods *methods = tdb->transaction->io_methods;
812 /* round up to a multiple of page size. Overallocate, since each
813 * such allocation forces us to expand the file. */
815 = (((sizeof(*rec) + rec_length + rec_length / 2)
816 + PAGESIZE-1) & ~(PAGESIZE-1))
818 off = tdb->file->map_size;
820 /* Restore ->map_size before calling underlying expand_file.
821 Also so that we don't try to expand the file again in the
822 transaction commit, which would destroy the recovery
824 addition = (tdb->file->map_size - tdb->transaction->old_map_size) +
825 sizeof(*rec) + rec->max_len;
826 tdb->file->map_size = tdb->transaction->old_map_size;
827 ecode = methods->expand_file(tdb, addition);
828 if (ecode != TDB_SUCCESS) {
829 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
830 "tdb_recovery_allocate:"
831 " failed to create recovery area");
834 /* we have to reset the old map size so that we don't try to
835 expand the file again in the transaction commit, which
836 would destroy the recovery area */
837 tdb->transaction->old_map_size = tdb->file->map_size;
839 /* write the recovery header offset and sync - we can sync without a race here
840 as the magic ptr in the recovery record has not been set */
842 tdb_convert(tdb, &recovery_off, sizeof(recovery_off));
843 ecode = methods->twrite(tdb, offsetof(struct tdb_header, recovery),
844 &recovery_off, sizeof(tdb_off_t));
845 if (ecode != TDB_SUCCESS) {
846 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
847 "tdb_recovery_allocate:"
848 " failed to write recovery head");
850 transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
857 setup the recovery data that will be used on a crash during commit
859 static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb)
861 tdb_len_t recovery_size = 0;
862 tdb_off_t recovery_off = 0;
863 tdb_off_t old_map_size = tdb->transaction->old_map_size;
864 struct tdb_recovery_record *recovery;
865 const struct tdb_methods *methods = tdb->transaction->io_methods;
867 enum TDB_ERROR ecode;
869 recovery = alloc_recovery(tdb, &recovery_size);
870 if (TDB_PTR_IS_ERR(recovery))
871 return TDB_PTR_ERR(recovery);
873 ecode = tdb_recovery_area(tdb, methods, &recovery_off, recovery);
879 if (recovery->max_len < recovery_size) {
880 /* Not large enough. Free up old recovery area. */
883 ecode = add_free_record(tdb, recovery_off,
886 TDB_LOCK_WAIT, true);
888 if (ecode != TDB_SUCCESS) {
889 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
890 "tdb_recovery_allocate:"
891 " failed to free previous"
895 /* Refresh recovery after add_free_record above. */
896 recovery = alloc_recovery(tdb, &recovery_size);
897 if (TDB_PTR_IS_ERR(recovery))
898 return TDB_PTR_ERR(recovery);
901 recovery_off = create_recovery_area(tdb, recovery_size,
903 if (TDB_OFF_IS_ERR(recovery_off)) {
909 /* Now we know size, convert rec header. */
910 recovery->magic = TDB_RECOVERY_INVALID_MAGIC;
911 recovery->len = recovery_size;
912 recovery->eof = old_map_size;
913 tdb_convert(tdb, recovery, sizeof(*recovery));
915 /* write the recovery data to the recovery area */
916 ecode = methods->twrite(tdb, recovery_off, recovery, recovery_size);
917 if (ecode != TDB_SUCCESS) {
919 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
920 "tdb_transaction_setup_recovery:"
921 " failed to write recovery data");
923 transaction_write_existing(tdb, recovery_off, recovery, recovery_size);
927 /* as we don't have ordered writes, we have to sync the recovery
928 data before we update the magic to indicate that the recovery
930 ecode = transaction_sync(tdb, recovery_off, recovery_size);
931 if (ecode != TDB_SUCCESS)
934 magic = TDB_RECOVERY_MAGIC;
935 tdb_convert(tdb, &magic, sizeof(magic));
937 tdb->transaction->magic_offset
938 = recovery_off + offsetof(struct tdb_recovery_record, magic);
940 ecode = methods->twrite(tdb, tdb->transaction->magic_offset,
941 &magic, sizeof(magic));
942 if (ecode != TDB_SUCCESS) {
943 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
944 "tdb_transaction_setup_recovery:"
945 " failed to write recovery magic");
947 transaction_write_existing(tdb, tdb->transaction->magic_offset,
948 &magic, sizeof(magic));
950 /* ensure the recovery magic marker is on disk */
951 return transaction_sync(tdb, tdb->transaction->magic_offset,
955 static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
957 const struct tdb_methods *methods;
958 enum TDB_ERROR ecode;
960 if (tdb->transaction == NULL) {
961 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
962 "tdb_transaction_prepare_commit:"
966 if (tdb->transaction->prepared) {
967 _tdb_transaction_cancel(tdb);
968 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
969 "tdb_transaction_prepare_commit:"
970 " transaction already prepared");
973 if (tdb->transaction->transaction_error) {
974 _tdb_transaction_cancel(tdb);
975 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
976 "tdb_transaction_prepare_commit:"
977 " transaction error pending");
981 if (tdb->transaction->nesting != 0) {
985 /* check for a null transaction */
986 if (tdb->transaction->blocks == NULL) {
990 methods = tdb->transaction->io_methods;
992 /* upgrade the main transaction lock region to a write lock */
993 ecode = tdb_allrecord_upgrade(tdb);
994 if (ecode != TDB_SUCCESS) {
998 /* get the open lock - this prevents new users attaching to the database
1000 ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
1001 if (ecode != TDB_SUCCESS) {
1005 /* Since we have whole db locked, we don't need the expansion lock. */
1006 if (!(tdb->flags & TDB_NOSYNC)) {
1007 /* Sets up tdb->transaction->recovery and
1008 * tdb->transaction->magic_offset. */
1009 ecode = transaction_setup_recovery(tdb);
1010 if (ecode != TDB_SUCCESS) {
1015 tdb->transaction->prepared = true;
1017 /* expand the file to the new size if needed */
1018 if (tdb->file->map_size != tdb->transaction->old_map_size) {
1021 add = tdb->file->map_size - tdb->transaction->old_map_size;
1022 /* Restore original map size for tdb_expand_file */
1023 tdb->file->map_size = tdb->transaction->old_map_size;
1024 ecode = methods->expand_file(tdb, add);
1025 if (ecode != TDB_SUCCESS) {
1030 /* Keep the open lock until the actual commit */
1035 prepare to commit the current transaction
1037 enum TDB_ERROR tdb_transaction_prepare_commit(struct tdb_context *tdb)
1039 return _tdb_transaction_prepare_commit(tdb);
1043 commit the current transaction
1045 enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
1047 const struct tdb_methods *methods;
1049 enum TDB_ERROR ecode;
1051 if (tdb->transaction == NULL) {
1052 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
1054 "tdb_transaction_commit:"
1058 tdb_trace(tdb, "tdb_transaction_commit");
1060 if (tdb->transaction->nesting != 0) {
1061 tdb->transaction->nesting--;
1062 return tdb->last_error = TDB_SUCCESS;
1065 /* check for a null transaction */
1066 if (tdb->transaction->blocks == NULL) {
1067 _tdb_transaction_cancel(tdb);
1068 return tdb->last_error = TDB_SUCCESS;
1071 if (!tdb->transaction->prepared) {
1072 ecode = _tdb_transaction_prepare_commit(tdb);
1073 if (ecode != TDB_SUCCESS) {
1074 _tdb_transaction_cancel(tdb);
1075 return tdb->last_error = ecode;
1079 methods = tdb->transaction->io_methods;
1081 /* perform all the writes */
1082 for (i=0;i<tdb->transaction->num_blocks;i++) {
1086 if (tdb->transaction->blocks[i] == NULL) {
1090 offset = i * PAGESIZE;
1092 if (i == tdb->transaction->num_blocks-1) {
1093 length = tdb->transaction->last_block_size;
1096 ecode = methods->twrite(tdb, offset,
1097 tdb->transaction->blocks[i], length);
1098 if (ecode != TDB_SUCCESS) {
1099 /* we've overwritten part of the data and
1100 possibly expanded the file, so we need to
1101 run the crash recovery code */
1102 tdb->methods = methods;
1103 tdb_transaction_recover(tdb);
1105 _tdb_transaction_cancel(tdb);
1107 return tdb->last_error = ecode;
1109 SAFE_FREE(tdb->transaction->blocks[i]);
1112 SAFE_FREE(tdb->transaction->blocks);
1113 tdb->transaction->num_blocks = 0;
1115 /* ensure the new data is on disk */
1116 ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1117 if (ecode != TDB_SUCCESS) {
1118 return tdb->last_error = ecode;
1122 TODO: maybe write to some dummy hdr field, or write to magic
1123 offset without mmap, before the last sync, instead of the
1127 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1128 don't change the mtime of the file, this means the file may
1129 not be backed up (as tdb rounding to block sizes means that
1130 file size changes are quite rare too). The following forces
1131 mtime changes when a transaction completes */
1133 utime(tdb->name, NULL);
1136 /* use a transaction cancel to free memory and remove the
1137 transaction locks: it "restores" map_size, too. */
1138 tdb->transaction->old_map_size = tdb->file->map_size;
1139 _tdb_transaction_cancel(tdb);
1141 return tdb->last_error = TDB_SUCCESS;
1146 recover from an aborted transaction. Must be called with exclusive
1147 database write access already established (including the open
1148 lock to prevent new processes attaching)
1150 enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb)
1152 tdb_off_t recovery_head, recovery_eof;
1153 unsigned char *data, *p;
1154 struct tdb_recovery_record rec;
1155 enum TDB_ERROR ecode;
1157 /* find the recovery area */
1158 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1159 if (TDB_OFF_IS_ERR(recovery_head)) {
1160 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
1161 "tdb_transaction_recover:"
1162 " failed to read recovery head");
1165 if (recovery_head == 0) {
1166 /* we have never allocated a recovery record */
1170 /* read the recovery record */
1171 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1172 if (ecode != TDB_SUCCESS) {
1173 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1174 "tdb_transaction_recover:"
1175 " failed to read recovery record");
1178 if (rec.magic != TDB_RECOVERY_MAGIC) {
1179 /* there is no valid recovery data */
1183 if (tdb->read_only) {
1184 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
1185 "tdb_transaction_recover:"
1186 " attempt to recover read only database");
1189 recovery_eof = rec.eof;
1191 data = (unsigned char *)malloc(rec.len);
1193 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
1194 "tdb_transaction_recover:"
1195 " failed to allocate recovery data");
1198 /* read the full recovery data */
1199 ecode = tdb->methods->tread(tdb, recovery_head + sizeof(rec), data,
1201 if (ecode != TDB_SUCCESS) {
1202 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1203 "tdb_transaction_recover:"
1204 " failed to read recovery data");
1207 /* recover the file data */
1209 while (p+sizeof(tdb_off_t)+sizeof(tdb_len_t) < data + rec.len) {
1212 tdb_convert(tdb, p, sizeof(ofs) + sizeof(len));
1213 memcpy(&ofs, p, sizeof(ofs));
1214 memcpy(&len, p + sizeof(ofs), sizeof(len));
1215 p += sizeof(ofs) + sizeof(len);
1217 ecode = tdb->methods->twrite(tdb, ofs, p, len);
1218 if (ecode != TDB_SUCCESS) {
1220 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1221 "tdb_transaction_recover:"
1222 " failed to recover %zu bytes"
1224 (size_t)len, (size_t)ofs);
1231 ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1232 if (ecode != TDB_SUCCESS) {
1233 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1234 "tdb_transaction_recover:"
1235 " failed to sync recovery");
1238 /* if the recovery area is after the recovered eof then remove it */
1239 if (recovery_eof <= recovery_head) {
1240 ecode = tdb_write_off(tdb, offsetof(struct tdb_header,
1243 if (ecode != TDB_SUCCESS) {
1244 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1245 "tdb_transaction_recover:"
1246 " failed to remove recovery head");
1250 /* remove the recovery magic */
1251 ecode = tdb_write_off(tdb,
1253 + offsetof(struct tdb_recovery_record, magic),
1254 TDB_RECOVERY_INVALID_MAGIC);
1255 if (ecode != TDB_SUCCESS) {
1256 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1257 "tdb_transaction_recover:"
1258 " failed to remove recovery magic");
1261 ecode = transaction_sync(tdb, 0, recovery_eof);
1262 if (ecode != TDB_SUCCESS) {
1263 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1264 "tdb_transaction_recover:"
1265 " failed to sync2 recovery");
1268 tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
1269 "tdb_transaction_recover: recovered %zu byte database",
1270 (size_t)recovery_eof);
1276 tdb_bool_err tdb_needs_recovery(struct tdb_context *tdb)
1278 tdb_off_t recovery_head;
1279 struct tdb_recovery_record rec;
1280 enum TDB_ERROR ecode;
1282 /* find the recovery area */
1283 recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1284 if (TDB_OFF_IS_ERR(recovery_head)) {
1285 return recovery_head;
1288 if (recovery_head == 0) {
1289 /* we have never allocated a recovery record */
1293 /* read the recovery record */
1294 ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1295 if (ecode != TDB_SUCCESS) {
1299 return (rec.magic == TDB_RECOVERY_MAGIC);