2 Unix SMB/CIFS implementation.
4 trivial database library
6 Copyright (C) Andrew Tridgell 2005
8 ** NOTE! The following LGPL license applies to the tdb
9 ** library. This does NOT imply that all of Samba is released
12 This library is free software; you can redistribute it and/or
13 modify it under the terms of the GNU Lesser General Public
14 License as published by the Free Software Foundation; either
15 version 3 of the License, or (at your option) any later version.
17 This library is distributed in the hope that it will be useful,
18 but WITHOUT ANY WARRANTY; without even the implied warranty of
19 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 Lesser General Public License for more details.
22 You should have received a copy of the GNU Lesser General Public
23 License along with this library; if not, see <http://www.gnu.org/licenses/>.
26 #include "tdb_private.h"
31 - only allow a single transaction at a time per database. This makes
32 using the transaction API simpler, as otherwise the caller would
33 have to cope with temporary failures in transactions that conflict
34 with other current transactions
36 - keep the transaction recovery information in the same file as the
37 database, using a special 'transaction recovery' record pointed at
38 by the header. This removes the need for extra journal files as
39 used by some other databases
41 - dynamically allocated the transaction recover record, re-using it
42 for subsequent transactions. If a larger record is needed then
43 tdb_free() the old record to place it on the normal tdb freelist
44 before allocating the new record
46 - during transactions, keep a linked list of writes all that have
47 been performed by intercepting all tdb_write() calls. The hooked
48 transaction versions of tdb_read() and tdb_write() check this
49 linked list and try to use the elements of the list in preference
52 - don't allow any locks to be held when a transaction starts,
53 otherwise we can end up with deadlock (plus lack of lock nesting
54 in posix locks would mean the lock is lost)
56 - if the caller gains a lock during the transaction but doesn't
57 release it then fail the commit
59 - allow for nested calls to tdb_transaction_start(), re-using the
60 existing transaction record. If the inner transaction is cancelled
61 then a subsequent commit will fail
63 - keep a mirrored copy of the tdb hash chain heads to allow for the
64 fast hash heads scan on traverse, updating the mirrored copy in
65 the transaction version of tdb_write
67 - allow callers to mix transaction and non-transaction use of tdb,
68 although once a transaction is started then an exclusive lock is
69 gained until the transaction is committed or cancelled
71 - the commit stategy involves first saving away all modified data
72 into a linearised buffer in the transaction recovery area, then
73 marking the transaction recovery area with a magic value to
74 indicate a valid recovery record. In total 4 fsync/msync calls are
75 needed per commit to prevent race conditions. It might be possible
76 to reduce this to 3 or even 2 with some more work.
78 - check for a valid recovery record on open of the tdb, while the
79 global lock is held. Automatically recover from the transaction
80 recovery area if needed, then continue with the open as
81 usual. This allows for smooth crash recovery with no administrator
84 - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85 still available, but no transaction recovery area is used and no
86 fsync/msync calls are made.
88 - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89 tdb_add_flags() transaction nesting is enabled.
90 It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
91 The default is that transaction nesting is allowed.
92 Note: this default may change in future versions of tdb.
94 Beware. when transactions are nested a transaction successfully
95 completed with tdb_transaction_commit() can be silently unrolled later.
97 - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
98 tdb_add_flags() transaction nesting is disabled.
99 It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
100 An attempt create a nested transaction will fail with TDB_ERR_NESTING.
101 The default is that transaction nesting is allowed.
102 Note: this default may change in future versions of tdb.
107 hold the context of any current transaction
109 struct tdb_transaction {
110 /* we keep a mirrored copy of the tdb hash heads here so
111 tdb_next_hash_chain() can operate efficiently */
112 uint32_t *hash_heads;
114 /* the original io methods - used to do IOs to the real db */
115 const struct tdb_methods *io_methods;
117 /* the list of transaction blocks. When a block is first
118 written to, it gets created in this list */
121 uint32_t block_size; /* bytes in each block */
122 uint32_t last_block_size; /* number of valid bytes in the last block */
124 /* non-zero when an internal transaction error has
125 occurred. All write operations will then fail until the
126 transaction is ended */
127 int transaction_error;
129 /* when inside a transaction we need to keep track of any
130 nested tdb_transaction_start() calls, as these are allowed,
131 but don't create a new transaction */
134 /* set when a prepare has already occurred */
136 tdb_off_t magic_offset;
138 /* old file size before transaction */
139 tdb_len_t old_map_size;
141 /* we should re-pack on commit */
147 read while in a transaction. We need to check first if the data is in our list
148 of transaction elements, then if not do a real read
150 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
151 tdb_len_t len, int cv)
155 /* break it down into block sized ops */
156 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
157 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
158 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
163 buf = (void *)(len2 + (char *)buf);
170 blk = off / tdb->transaction->block_size;
172 /* see if we have it in the block list */
173 if (tdb->transaction->num_blocks <= blk ||
174 tdb->transaction->blocks[blk] == NULL) {
175 /* nope, do a real read */
176 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
182 /* it is in the block list. Now check for the last block */
183 if (blk == tdb->transaction->num_blocks-1) {
184 if (len > tdb->transaction->last_block_size) {
189 /* now copy it out of this block */
190 memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
192 tdb_convert(buf, len);
197 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
198 tdb->ecode = TDB_ERR_IO;
199 tdb->transaction->transaction_error = 1;
205 write while in a transaction
207 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
208 const void *buf, tdb_len_t len)
212 /* Only a commit is allowed on a prepared transaction */
213 if (tdb->transaction->prepared) {
214 tdb->ecode = TDB_ERR_EINVAL;
215 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
216 tdb->transaction->transaction_error = 1;
220 /* if the write is to a hash head, then update the transaction
222 if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
223 off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
224 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
225 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
228 /* break it up into block sized chunks */
229 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
230 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
231 if (transaction_write(tdb, off, buf, len2) != 0) {
237 buf = (const void *)(len2 + (const char *)buf);
245 blk = off / tdb->transaction->block_size;
246 off = off % tdb->transaction->block_size;
248 if (tdb->transaction->num_blocks <= blk) {
249 uint8_t **new_blocks;
250 /* expand the blocks array */
251 if (tdb->transaction->blocks == NULL) {
252 new_blocks = (uint8_t **)malloc(
253 (blk+1)*sizeof(uint8_t *));
255 new_blocks = (uint8_t **)realloc(
256 tdb->transaction->blocks,
257 (blk+1)*sizeof(uint8_t *));
259 if (new_blocks == NULL) {
260 tdb->ecode = TDB_ERR_OOM;
263 memset(&new_blocks[tdb->transaction->num_blocks], 0,
264 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
265 tdb->transaction->blocks = new_blocks;
266 tdb->transaction->num_blocks = blk+1;
267 tdb->transaction->last_block_size = 0;
270 /* allocate and fill a block? */
271 if (tdb->transaction->blocks[blk] == NULL) {
272 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
273 if (tdb->transaction->blocks[blk] == NULL) {
274 tdb->ecode = TDB_ERR_OOM;
275 tdb->transaction->transaction_error = 1;
278 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
279 tdb_len_t len2 = tdb->transaction->block_size;
280 if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
281 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
283 if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
284 tdb->transaction->blocks[blk],
286 SAFE_FREE(tdb->transaction->blocks[blk]);
287 tdb->ecode = TDB_ERR_IO;
290 if (blk == tdb->transaction->num_blocks-1) {
291 tdb->transaction->last_block_size = len2;
296 /* overwrite part of an existing block */
298 memset(tdb->transaction->blocks[blk] + off, 0, len);
300 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
302 if (blk == tdb->transaction->num_blocks-1) {
303 if (len + off > tdb->transaction->last_block_size) {
304 tdb->transaction->last_block_size = len + off;
311 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
312 (blk*tdb->transaction->block_size) + off, len));
313 tdb->transaction->transaction_error = 1;
319 write while in a transaction - this varient never expands the transaction blocks, it only
320 updates existing blocks. This means it cannot change the recovery size
322 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
323 const void *buf, tdb_len_t len)
327 /* break it up into block sized chunks */
328 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
329 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
330 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
336 buf = (const void *)(len2 + (const char *)buf);
344 blk = off / tdb->transaction->block_size;
345 off = off % tdb->transaction->block_size;
347 if (tdb->transaction->num_blocks <= blk ||
348 tdb->transaction->blocks[blk] == NULL) {
352 if (blk == tdb->transaction->num_blocks-1 &&
353 off + len > tdb->transaction->last_block_size) {
354 if (off >= tdb->transaction->last_block_size) {
357 len = tdb->transaction->last_block_size - off;
360 /* overwrite part of an existing block */
361 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
368 accelerated hash chain head search, using the cached hash heads
370 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
373 for (;h < tdb->header.hash_size;h++) {
374 /* the +1 takes account of the freelist */
375 if (0 != tdb->transaction->hash_heads[h+1]) {
383 out of bounds check during a transaction
385 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
387 if (len <= tdb->map_size) {
390 tdb->ecode = TDB_ERR_IO;
395 transaction version of tdb_expand().
397 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
400 /* add a write to the transaction elements, so subsequent
401 reads see the zero data */
402 if (transaction_write(tdb, size, NULL, addition) != 0) {
406 tdb->transaction->need_repack = true;
412 brlock during a transaction - ignore them
414 static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset,
415 int rw_type, int lck_type, int probe, size_t len)
420 static const struct tdb_methods transaction_methods = {
423 transaction_next_hash_chain,
425 transaction_expand_file,
431 start a tdb transaction. No token is returned, as only a single
432 transaction is allowed to be pending per tdb_context
434 int tdb_transaction_start(struct tdb_context *tdb)
436 /* some sanity checks */
437 if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
438 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
439 tdb->ecode = TDB_ERR_EINVAL;
443 /* cope with nested tdb_transaction_start() calls */
444 if (tdb->transaction != NULL) {
445 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
446 tdb->ecode = TDB_ERR_NESTING;
449 tdb->transaction->nesting++;
450 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
451 tdb->transaction->nesting));
455 if (tdb->num_locks != 0 || tdb->global_lock.count) {
456 /* the caller must not have any locks when starting a
457 transaction as otherwise we'll be screwed by lack
458 of nested locks in posix */
459 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
460 tdb->ecode = TDB_ERR_LOCK;
464 if (tdb->travlocks.next != NULL) {
465 /* you cannot use transactions inside a traverse (although you can use
466 traverse inside a transaction) as otherwise you can end up with
468 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
469 tdb->ecode = TDB_ERR_LOCK;
473 tdb->transaction = (struct tdb_transaction *)
474 calloc(sizeof(struct tdb_transaction), 1);
475 if (tdb->transaction == NULL) {
476 tdb->ecode = TDB_ERR_OOM;
480 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
481 tdb->transaction->block_size = tdb->page_size;
483 /* get the transaction write lock. This is a blocking lock. As
484 discussed with Volker, there are a number of ways we could
485 make this async, which we will probably do in the future */
486 if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
487 SAFE_FREE(tdb->transaction->blocks);
488 SAFE_FREE(tdb->transaction);
492 /* get a read lock from the freelist to the end of file. This
493 is upgraded to a write lock during the commit */
494 if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
495 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
496 tdb->ecode = TDB_ERR_LOCK;
500 /* setup a copy of the hash table heads so the hash scan in
501 traverse can be fast */
502 tdb->transaction->hash_heads = (uint32_t *)
503 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
504 if (tdb->transaction->hash_heads == NULL) {
505 tdb->ecode = TDB_ERR_OOM;
508 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
509 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
510 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
511 tdb->ecode = TDB_ERR_IO;
515 /* make sure we know about any file expansions already done by
517 tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
518 tdb->transaction->old_map_size = tdb->map_size;
520 /* finally hook the io methods, replacing them with
521 transaction specific methods */
522 tdb->transaction->io_methods = tdb->methods;
523 tdb->methods = &transaction_methods;
525 /* Trace at the end, so we get sequence number correct. */
526 tdb_trace(tdb, "tdb_transaction_start");
530 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
531 tdb_transaction_unlock(tdb);
532 SAFE_FREE(tdb->transaction->blocks);
533 SAFE_FREE(tdb->transaction->hash_heads);
534 SAFE_FREE(tdb->transaction);
542 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
544 if (tdb->flags & TDB_NOSYNC) {
548 if (fsync(tdb->fd) != 0) {
549 tdb->ecode = TDB_ERR_IO;
550 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
555 tdb_off_t moffset = offset & ~(tdb->page_size-1);
556 if (msync(moffset + (char *)tdb->map_ptr,
557 length + (offset - moffset), MS_SYNC) != 0) {
558 tdb->ecode = TDB_ERR_IO;
559 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
569 int _tdb_transaction_cancel(struct tdb_context *tdb)
573 if (tdb->transaction == NULL) {
574 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
578 if (tdb->transaction->nesting != 0) {
579 tdb->transaction->transaction_error = 1;
580 tdb->transaction->nesting--;
584 tdb->map_size = tdb->transaction->old_map_size;
586 /* free all the transaction blocks */
587 for (i=0;i<tdb->transaction->num_blocks;i++) {
588 if (tdb->transaction->blocks[i] != NULL) {
589 free(tdb->transaction->blocks[i]);
592 SAFE_FREE(tdb->transaction->blocks);
594 if (tdb->transaction->magic_offset) {
595 const struct tdb_methods *methods = tdb->transaction->io_methods;
598 /* remove the recovery marker */
599 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
600 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
601 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
606 /* remove any global lock created during the transaction */
607 if (tdb->global_lock.count != 0) {
608 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
609 tdb->global_lock.count = 0;
612 /* remove any locks created during the transaction */
613 if (tdb->num_locks != 0) {
614 for (i=0;i<tdb->num_lockrecs;i++) {
615 tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
616 F_UNLCK,F_SETLKW, 0, 1);
619 tdb->num_lockrecs = 0;
620 SAFE_FREE(tdb->lockrecs);
623 /* restore the normal io methods */
624 tdb->methods = tdb->transaction->io_methods;
626 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
627 tdb_transaction_unlock(tdb);
628 SAFE_FREE(tdb->transaction->hash_heads);
629 SAFE_FREE(tdb->transaction);
635 cancel the current transaction
637 int tdb_transaction_cancel(struct tdb_context *tdb)
639 tdb_trace(tdb, "tdb_transaction_cancel");
640 return _tdb_transaction_cancel(tdb);
644 work out how much space the linearised recovery data will consume
646 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
648 tdb_len_t recovery_size = 0;
651 recovery_size = sizeof(uint32_t);
652 for (i=0;i<tdb->transaction->num_blocks;i++) {
653 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
656 if (tdb->transaction->blocks[i] == NULL) {
659 recovery_size += 2*sizeof(tdb_off_t);
660 if (i == tdb->transaction->num_blocks-1) {
661 recovery_size += tdb->transaction->last_block_size;
663 recovery_size += tdb->transaction->block_size;
667 return recovery_size;
671 allocate the recovery area, or use an existing recovery area if it is
674 static int tdb_recovery_allocate(struct tdb_context *tdb,
675 tdb_len_t *recovery_size,
676 tdb_off_t *recovery_offset,
677 tdb_len_t *recovery_max_size)
679 struct tdb_record rec;
680 const struct tdb_methods *methods = tdb->transaction->io_methods;
681 tdb_off_t recovery_head;
683 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
684 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
690 if (recovery_head != 0 &&
691 methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
692 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
696 *recovery_size = tdb_recovery_size(tdb);
698 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
699 /* it fits in the existing area */
700 *recovery_max_size = rec.rec_len;
701 *recovery_offset = recovery_head;
705 /* we need to free up the old recovery area, then allocate a
706 new one at the end of the file. Note that we cannot use
707 tdb_allocate() to allocate the new one as that might return
708 us an area that is being currently used (as of the start of
710 if (recovery_head != 0) {
711 if (tdb_free(tdb, recovery_head, &rec) == -1) {
712 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
717 /* the tdb_free() call might have increased the recovery size */
718 *recovery_size = tdb_recovery_size(tdb);
720 /* round up to a multiple of page size */
721 *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
722 *recovery_offset = tdb->map_size;
723 recovery_head = *recovery_offset;
725 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
726 (tdb->map_size - tdb->transaction->old_map_size) +
727 sizeof(rec) + *recovery_max_size) == -1) {
728 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
732 /* remap the file (if using mmap) */
733 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
735 /* we have to reset the old map size so that we don't try to expand the file
736 again in the transaction commit, which would destroy the recovery area */
737 tdb->transaction->old_map_size = tdb->map_size;
739 /* write the recovery header offset and sync - we can sync without a race here
740 as the magic ptr in the recovery record has not been set */
741 CONVERT(recovery_head);
742 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
743 &recovery_head, sizeof(tdb_off_t)) == -1) {
744 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
747 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
748 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
757 setup the recovery data that will be used on a crash during commit
759 static int transaction_setup_recovery(struct tdb_context *tdb,
760 tdb_off_t *magic_offset)
762 tdb_len_t recovery_size;
763 unsigned char *data, *p;
764 const struct tdb_methods *methods = tdb->transaction->io_methods;
765 struct tdb_record *rec;
766 tdb_off_t recovery_offset, recovery_max_size;
767 tdb_off_t old_map_size = tdb->transaction->old_map_size;
768 uint32_t magic, tailer;
772 check that the recovery area has enough space
774 if (tdb_recovery_allocate(tdb, &recovery_size,
775 &recovery_offset, &recovery_max_size) == -1) {
779 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
781 tdb->ecode = TDB_ERR_OOM;
785 rec = (struct tdb_record *)data;
786 memset(rec, 0, sizeof(*rec));
789 rec->data_len = recovery_size;
790 rec->rec_len = recovery_max_size;
791 rec->key_len = old_map_size;
794 /* build the recovery data into a single blob to allow us to do a single
795 large write, which should be more efficient */
796 p = data + sizeof(*rec);
797 for (i=0;i<tdb->transaction->num_blocks;i++) {
801 if (tdb->transaction->blocks[i] == NULL) {
805 offset = i * tdb->transaction->block_size;
806 length = tdb->transaction->block_size;
807 if (i == tdb->transaction->num_blocks-1) {
808 length = tdb->transaction->last_block_size;
811 if (offset >= old_map_size) {
814 if (offset + length > tdb->transaction->old_map_size) {
815 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
817 tdb->ecode = TDB_ERR_CORRUPT;
820 memcpy(p, &offset, 4);
821 memcpy(p+4, &length, 4);
825 /* the recovery area contains the old data, not the
826 new data, so we have to call the original tdb_read
828 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
830 tdb->ecode = TDB_ERR_IO;
837 tailer = sizeof(*rec) + recovery_max_size;
838 memcpy(p, &tailer, 4);
841 /* write the recovery data to the recovery area */
842 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
843 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
845 tdb->ecode = TDB_ERR_IO;
848 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
849 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
851 tdb->ecode = TDB_ERR_IO;
855 /* as we don't have ordered writes, we have to sync the recovery
856 data before we update the magic to indicate that the recovery
858 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
865 magic = TDB_RECOVERY_MAGIC;
868 *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
870 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
871 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
872 tdb->ecode = TDB_ERR_IO;
875 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
876 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
877 tdb->ecode = TDB_ERR_IO;
881 /* ensure the recovery magic marker is on disk */
882 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
889 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
891 const struct tdb_methods *methods;
893 if (tdb->transaction == NULL) {
894 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
898 if (tdb->transaction->prepared) {
899 tdb->ecode = TDB_ERR_EINVAL;
900 _tdb_transaction_cancel(tdb);
901 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
905 if (tdb->transaction->transaction_error) {
906 tdb->ecode = TDB_ERR_IO;
907 _tdb_transaction_cancel(tdb);
908 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
913 if (tdb->transaction->nesting != 0) {
917 /* check for a null transaction */
918 if (tdb->transaction->blocks == NULL) {
922 methods = tdb->transaction->io_methods;
924 /* if there are any locks pending then the caller has not
925 nested their locks properly, so fail the transaction */
926 if (tdb->num_locks || tdb->global_lock.count) {
927 tdb->ecode = TDB_ERR_LOCK;
928 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
929 _tdb_transaction_cancel(tdb);
933 /* upgrade the main transaction lock region to a write lock */
934 if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
935 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
936 tdb->ecode = TDB_ERR_LOCK;
937 _tdb_transaction_cancel(tdb);
941 /* get the global lock - this prevents new users attaching to the database
943 if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
944 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
945 tdb->ecode = TDB_ERR_LOCK;
946 _tdb_transaction_cancel(tdb);
950 if (!(tdb->flags & TDB_NOSYNC)) {
951 /* write the recovery data to the end of the file */
952 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
953 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
954 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
955 _tdb_transaction_cancel(tdb);
960 tdb->transaction->prepared = true;
962 /* expand the file to the new size if needed */
963 if (tdb->map_size != tdb->transaction->old_map_size) {
964 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
966 tdb->transaction->old_map_size) == -1) {
967 tdb->ecode = TDB_ERR_IO;
968 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
969 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
970 _tdb_transaction_cancel(tdb);
973 tdb->map_size = tdb->transaction->old_map_size;
974 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
977 /* Keep the global lock until the actual commit */
983 prepare to commit the current transaction
985 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
987 tdb_trace(tdb, "tdb_transaction_prepare_commit");
988 return _tdb_transaction_prepare_commit(tdb);
992 commit the current transaction
994 int tdb_transaction_commit(struct tdb_context *tdb)
996 const struct tdb_methods *methods;
1000 if (tdb->transaction == NULL) {
1001 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1005 tdb_trace(tdb, "tdb_transaction_commit");
1007 if (tdb->transaction->transaction_error) {
1008 tdb->ecode = TDB_ERR_IO;
1009 _tdb_transaction_cancel(tdb);
1010 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1015 if (tdb->transaction->nesting != 0) {
1016 tdb->transaction->nesting--;
1020 /* check for a null transaction */
1021 if (tdb->transaction->blocks == NULL) {
1022 _tdb_transaction_cancel(tdb);
1026 if (!tdb->transaction->prepared) {
1027 int ret = _tdb_transaction_prepare_commit(tdb);
1032 methods = tdb->transaction->io_methods;
1034 /* perform all the writes */
1035 for (i=0;i<tdb->transaction->num_blocks;i++) {
1039 if (tdb->transaction->blocks[i] == NULL) {
1043 offset = i * tdb->transaction->block_size;
1044 length = tdb->transaction->block_size;
1045 if (i == tdb->transaction->num_blocks-1) {
1046 length = tdb->transaction->last_block_size;
1049 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1050 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1052 /* we've overwritten part of the data and
1053 possibly expanded the file, so we need to
1054 run the crash recovery code */
1055 tdb->methods = methods;
1056 tdb_transaction_recover(tdb);
1058 _tdb_transaction_cancel(tdb);
1059 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1061 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1064 SAFE_FREE(tdb->transaction->blocks[i]);
1067 SAFE_FREE(tdb->transaction->blocks);
1068 tdb->transaction->num_blocks = 0;
1070 /* ensure the new data is on disk */
1071 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1075 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1078 TODO: maybe write to some dummy hdr field, or write to magic
1079 offset without mmap, before the last sync, instead of the
1083 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1084 don't change the mtime of the file, this means the file may
1085 not be backed up (as tdb rounding to block sizes means that
1086 file size changes are quite rare too). The following forces
1087 mtime changes when a transaction completes */
1089 utime(tdb->name, NULL);
1092 need_repack = tdb->transaction->need_repack;
1094 /* use a transaction cancel to free memory and remove the
1095 transaction locks */
1096 _tdb_transaction_cancel(tdb);
1099 return tdb_repack(tdb);
1107 recover from an aborted transaction. Must be called with exclusive
1108 database write access already established (including the global
1109 lock to prevent new processes attaching)
1111 int tdb_transaction_recover(struct tdb_context *tdb)
1113 tdb_off_t recovery_head, recovery_eof;
1114 unsigned char *data, *p;
1116 struct tdb_record rec;
1118 /* find the recovery area */
1119 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1120 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1121 tdb->ecode = TDB_ERR_IO;
1125 if (recovery_head == 0) {
1126 /* we have never allocated a recovery record */
1130 /* read the recovery record */
1131 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1132 sizeof(rec), DOCONV()) == -1) {
1133 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1134 tdb->ecode = TDB_ERR_IO;
1138 if (rec.magic != TDB_RECOVERY_MAGIC) {
1139 /* there is no valid recovery data */
1143 if (tdb->read_only) {
1144 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1145 tdb->ecode = TDB_ERR_CORRUPT;
1149 recovery_eof = rec.key_len;
1151 data = (unsigned char *)malloc(rec.data_len);
1153 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1154 tdb->ecode = TDB_ERR_OOM;
1158 /* read the full recovery data */
1159 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1160 rec.data_len, 0) == -1) {
1161 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1162 tdb->ecode = TDB_ERR_IO;
1166 /* recover the file data */
1168 while (p+8 < data + rec.data_len) {
1174 memcpy(&len, p+4, 4);
1176 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1178 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1179 tdb->ecode = TDB_ERR_IO;
1187 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1188 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1189 tdb->ecode = TDB_ERR_IO;
1193 /* if the recovery area is after the recovered eof then remove it */
1194 if (recovery_eof <= recovery_head) {
1195 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1196 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1197 tdb->ecode = TDB_ERR_IO;
1202 /* remove the recovery magic */
1203 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1205 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1206 tdb->ecode = TDB_ERR_IO;
1210 /* reduce the file size to the old size */
1212 if (ftruncate(tdb->fd, recovery_eof) != 0) {
1213 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1214 tdb->ecode = TDB_ERR_IO;
1217 tdb->map_size = recovery_eof;
1220 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1221 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1222 tdb->ecode = TDB_ERR_IO;
1226 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",