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 is enabled.
90 The default is that transaction nesting is not allowed and an attempt
91 to create a nested transaction will fail with TDB_ERR_NESTING.
93 Beware. when transactions are nested a transaction successfully
94 completed with tdb_transaction_commit() can be silently unrolled later.
99 hold the context of any current transaction
101 struct tdb_transaction {
102 /* we keep a mirrored copy of the tdb hash heads here so
103 tdb_next_hash_chain() can operate efficiently */
104 uint32_t *hash_heads;
106 /* the original io methods - used to do IOs to the real db */
107 const struct tdb_methods *io_methods;
109 /* the list of transaction blocks. When a block is first
110 written to, it gets created in this list */
113 uint32_t block_size; /* bytes in each block */
114 uint32_t last_block_size; /* number of valid bytes in the last block */
116 /* non-zero when an internal transaction error has
117 occurred. All write operations will then fail until the
118 transaction is ended */
119 int transaction_error;
121 /* when inside a transaction we need to keep track of any
122 nested tdb_transaction_start() calls, as these are allowed,
123 but don't create a new transaction */
126 /* set when a prepare has already occurred */
128 tdb_off_t magic_offset;
130 /* old file size before transaction */
131 tdb_len_t old_map_size;
133 /* we should re-pack on commit */
139 read while in a transaction. We need to check first if the data is in our list
140 of transaction elements, then if not do a real read
142 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
143 tdb_len_t len, int cv)
147 /* break it down into block sized ops */
148 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
149 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
150 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
155 buf = (void *)(len2 + (char *)buf);
162 blk = off / tdb->transaction->block_size;
164 /* see if we have it in the block list */
165 if (tdb->transaction->num_blocks <= blk ||
166 tdb->transaction->blocks[blk] == NULL) {
167 /* nope, do a real read */
168 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
174 /* it is in the block list. Now check for the last block */
175 if (blk == tdb->transaction->num_blocks-1) {
176 if (len > tdb->transaction->last_block_size) {
181 /* now copy it out of this block */
182 memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
184 tdb_convert(buf, len);
189 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
190 tdb->ecode = TDB_ERR_IO;
191 tdb->transaction->transaction_error = 1;
197 write while in a transaction
199 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
200 const void *buf, tdb_len_t len)
204 /* Only a commit is allowed on a prepared transaction */
205 if (tdb->transaction->prepared) {
206 tdb->ecode = TDB_ERR_EINVAL;
207 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
208 tdb->transaction->transaction_error = 1;
212 /* if the write is to a hash head, then update the transaction
214 if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
215 off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
216 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
217 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
220 /* break it up into block sized chunks */
221 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
222 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
223 if (transaction_write(tdb, off, buf, len2) != 0) {
229 buf = (const void *)(len2 + (const char *)buf);
237 blk = off / tdb->transaction->block_size;
238 off = off % tdb->transaction->block_size;
240 if (tdb->transaction->num_blocks <= blk) {
241 uint8_t **new_blocks;
242 /* expand the blocks array */
243 if (tdb->transaction->blocks == NULL) {
244 new_blocks = (uint8_t **)malloc(
245 (blk+1)*sizeof(uint8_t *));
247 new_blocks = (uint8_t **)realloc(
248 tdb->transaction->blocks,
249 (blk+1)*sizeof(uint8_t *));
251 if (new_blocks == NULL) {
252 tdb->ecode = TDB_ERR_OOM;
255 memset(&new_blocks[tdb->transaction->num_blocks], 0,
256 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
257 tdb->transaction->blocks = new_blocks;
258 tdb->transaction->num_blocks = blk+1;
259 tdb->transaction->last_block_size = 0;
262 /* allocate and fill a block? */
263 if (tdb->transaction->blocks[blk] == NULL) {
264 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
265 if (tdb->transaction->blocks[blk] == NULL) {
266 tdb->ecode = TDB_ERR_OOM;
267 tdb->transaction->transaction_error = 1;
270 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
271 tdb_len_t len2 = tdb->transaction->block_size;
272 if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
273 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
275 if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
276 tdb->transaction->blocks[blk],
278 SAFE_FREE(tdb->transaction->blocks[blk]);
279 tdb->ecode = TDB_ERR_IO;
282 if (blk == tdb->transaction->num_blocks-1) {
283 tdb->transaction->last_block_size = len2;
288 /* overwrite part of an existing block */
290 memset(tdb->transaction->blocks[blk] + off, 0, len);
292 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
294 if (blk == tdb->transaction->num_blocks-1) {
295 if (len + off > tdb->transaction->last_block_size) {
296 tdb->transaction->last_block_size = len + off;
303 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
304 (blk*tdb->transaction->block_size) + off, len));
305 tdb->transaction->transaction_error = 1;
311 write while in a transaction - this varient never expands the transaction blocks, it only
312 updates existing blocks. This means it cannot change the recovery size
314 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
315 const void *buf, tdb_len_t len)
319 /* break it up into block sized chunks */
320 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
321 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
322 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
328 buf = (const void *)(len2 + (const char *)buf);
336 blk = off / tdb->transaction->block_size;
337 off = off % tdb->transaction->block_size;
339 if (tdb->transaction->num_blocks <= blk ||
340 tdb->transaction->blocks[blk] == NULL) {
344 if (blk == tdb->transaction->num_blocks-1 &&
345 off + len > tdb->transaction->last_block_size) {
346 if (off >= tdb->transaction->last_block_size) {
349 len = tdb->transaction->last_block_size - off;
352 /* overwrite part of an existing block */
353 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
360 accelerated hash chain head search, using the cached hash heads
362 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
365 for (;h < tdb->header.hash_size;h++) {
366 /* the +1 takes account of the freelist */
367 if (0 != tdb->transaction->hash_heads[h+1]) {
375 out of bounds check during a transaction
377 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
379 if (len <= tdb->map_size) {
382 return TDB_ERRCODE(TDB_ERR_IO, -1);
386 transaction version of tdb_expand().
388 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
391 /* add a write to the transaction elements, so subsequent
392 reads see the zero data */
393 if (transaction_write(tdb, size, NULL, addition) != 0) {
397 tdb->transaction->need_repack = true;
403 brlock during a transaction - ignore them
405 static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset,
406 int rw_type, int lck_type, int probe, size_t len)
411 static const struct tdb_methods transaction_methods = {
414 transaction_next_hash_chain,
416 transaction_expand_file,
422 start a tdb transaction. No token is returned, as only a single
423 transaction is allowed to be pending per tdb_context
425 int tdb_transaction_start(struct tdb_context *tdb)
427 /* some sanity checks */
428 if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
429 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
430 tdb->ecode = TDB_ERR_EINVAL;
434 /* cope with nested tdb_transaction_start() calls */
435 if (tdb->transaction != NULL) {
436 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
437 tdb->ecode = TDB_ERR_NESTING;
440 tdb->transaction->nesting++;
441 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
442 tdb->transaction->nesting));
446 if (tdb->num_locks != 0 || tdb->global_lock.count) {
447 /* the caller must not have any locks when starting a
448 transaction as otherwise we'll be screwed by lack
449 of nested locks in posix */
450 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
451 tdb->ecode = TDB_ERR_LOCK;
455 if (tdb->travlocks.next != NULL) {
456 /* you cannot use transactions inside a traverse (although you can use
457 traverse inside a transaction) as otherwise you can end up with
459 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
460 tdb->ecode = TDB_ERR_LOCK;
464 tdb->transaction = (struct tdb_transaction *)
465 calloc(sizeof(struct tdb_transaction), 1);
466 if (tdb->transaction == NULL) {
467 tdb->ecode = TDB_ERR_OOM;
471 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
472 tdb->transaction->block_size = tdb->page_size;
474 /* get the transaction write lock. This is a blocking lock. As
475 discussed with Volker, there are a number of ways we could
476 make this async, which we will probably do in the future */
477 if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
478 SAFE_FREE(tdb->transaction->blocks);
479 SAFE_FREE(tdb->transaction);
483 /* get a read lock from the freelist to the end of file. This
484 is upgraded to a write lock during the commit */
485 if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
486 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
487 tdb->ecode = TDB_ERR_LOCK;
491 /* setup a copy of the hash table heads so the hash scan in
492 traverse can be fast */
493 tdb->transaction->hash_heads = (uint32_t *)
494 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
495 if (tdb->transaction->hash_heads == NULL) {
496 tdb->ecode = TDB_ERR_OOM;
499 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
500 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
501 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
502 tdb->ecode = TDB_ERR_IO;
506 /* make sure we know about any file expansions already done by
508 tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
509 tdb->transaction->old_map_size = tdb->map_size;
511 /* finally hook the io methods, replacing them with
512 transaction specific methods */
513 tdb->transaction->io_methods = tdb->methods;
514 tdb->methods = &transaction_methods;
516 /* Trace at the end, so we get sequence number correct. */
517 tdb_trace(tdb, "tdb_transaction_start");
521 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
522 tdb_transaction_unlock(tdb);
523 SAFE_FREE(tdb->transaction->blocks);
524 SAFE_FREE(tdb->transaction->hash_heads);
525 SAFE_FREE(tdb->transaction);
533 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
535 if (tdb->flags & TDB_NOSYNC) {
539 if (fsync(tdb->fd) != 0) {
540 tdb->ecode = TDB_ERR_IO;
541 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
546 tdb_off_t moffset = offset & ~(tdb->page_size-1);
547 if (msync(moffset + (char *)tdb->map_ptr,
548 length + (offset - moffset), MS_SYNC) != 0) {
549 tdb->ecode = TDB_ERR_IO;
550 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
560 int _tdb_transaction_cancel(struct tdb_context *tdb)
564 if (tdb->transaction == NULL) {
565 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
569 if (tdb->transaction->nesting != 0) {
570 tdb->transaction->transaction_error = 1;
571 tdb->transaction->nesting--;
575 tdb->map_size = tdb->transaction->old_map_size;
577 /* free all the transaction blocks */
578 for (i=0;i<tdb->transaction->num_blocks;i++) {
579 if (tdb->transaction->blocks[i] != NULL) {
580 free(tdb->transaction->blocks[i]);
583 SAFE_FREE(tdb->transaction->blocks);
585 if (tdb->transaction->magic_offset) {
586 const struct tdb_methods *methods = tdb->transaction->io_methods;
589 /* remove the recovery marker */
590 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
591 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
592 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
597 /* remove any global lock created during the transaction */
598 if (tdb->global_lock.count != 0) {
599 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
600 tdb->global_lock.count = 0;
603 /* remove any locks created during the transaction */
604 if (tdb->num_locks != 0) {
605 for (i=0;i<tdb->num_lockrecs;i++) {
606 tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
607 F_UNLCK,F_SETLKW, 0, 1);
610 tdb->num_lockrecs = 0;
611 SAFE_FREE(tdb->lockrecs);
614 /* restore the normal io methods */
615 tdb->methods = tdb->transaction->io_methods;
617 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
618 tdb_transaction_unlock(tdb);
619 SAFE_FREE(tdb->transaction->hash_heads);
620 SAFE_FREE(tdb->transaction);
626 cancel the current transaction
628 int tdb_transaction_cancel(struct tdb_context *tdb)
630 tdb_trace(tdb, "tdb_transaction_cancel");
631 return _tdb_transaction_cancel(tdb);
635 work out how much space the linearised recovery data will consume
637 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
639 tdb_len_t recovery_size = 0;
642 recovery_size = sizeof(uint32_t);
643 for (i=0;i<tdb->transaction->num_blocks;i++) {
644 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
647 if (tdb->transaction->blocks[i] == NULL) {
650 recovery_size += 2*sizeof(tdb_off_t);
651 if (i == tdb->transaction->num_blocks-1) {
652 recovery_size += tdb->transaction->last_block_size;
654 recovery_size += tdb->transaction->block_size;
658 return recovery_size;
662 allocate the recovery area, or use an existing recovery area if it is
665 static int tdb_recovery_allocate(struct tdb_context *tdb,
666 tdb_len_t *recovery_size,
667 tdb_off_t *recovery_offset,
668 tdb_len_t *recovery_max_size)
670 struct list_struct rec;
671 const struct tdb_methods *methods = tdb->transaction->io_methods;
672 tdb_off_t recovery_head;
674 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
675 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
681 if (recovery_head != 0 &&
682 methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
683 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
687 *recovery_size = tdb_recovery_size(tdb);
689 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
690 /* it fits in the existing area */
691 *recovery_max_size = rec.rec_len;
692 *recovery_offset = recovery_head;
696 /* we need to free up the old recovery area, then allocate a
697 new one at the end of the file. Note that we cannot use
698 tdb_allocate() to allocate the new one as that might return
699 us an area that is being currently used (as of the start of
701 if (recovery_head != 0) {
702 if (tdb_free(tdb, recovery_head, &rec) == -1) {
703 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
708 /* the tdb_free() call might have increased the recovery size */
709 *recovery_size = tdb_recovery_size(tdb);
711 /* round up to a multiple of page size */
712 *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
713 *recovery_offset = tdb->map_size;
714 recovery_head = *recovery_offset;
716 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
717 (tdb->map_size - tdb->transaction->old_map_size) +
718 sizeof(rec) + *recovery_max_size) == -1) {
719 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
723 /* remap the file (if using mmap) */
724 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
726 /* we have to reset the old map size so that we don't try to expand the file
727 again in the transaction commit, which would destroy the recovery area */
728 tdb->transaction->old_map_size = tdb->map_size;
730 /* write the recovery header offset and sync - we can sync without a race here
731 as the magic ptr in the recovery record has not been set */
732 CONVERT(recovery_head);
733 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
734 &recovery_head, sizeof(tdb_off_t)) == -1) {
735 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
738 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
739 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
748 setup the recovery data that will be used on a crash during commit
750 static int transaction_setup_recovery(struct tdb_context *tdb,
751 tdb_off_t *magic_offset)
753 tdb_len_t recovery_size;
754 unsigned char *data, *p;
755 const struct tdb_methods *methods = tdb->transaction->io_methods;
756 struct list_struct *rec;
757 tdb_off_t recovery_offset, recovery_max_size;
758 tdb_off_t old_map_size = tdb->transaction->old_map_size;
759 uint32_t magic, tailer;
763 check that the recovery area has enough space
765 if (tdb_recovery_allocate(tdb, &recovery_size,
766 &recovery_offset, &recovery_max_size) == -1) {
770 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
772 tdb->ecode = TDB_ERR_OOM;
776 rec = (struct list_struct *)data;
777 memset(rec, 0, sizeof(*rec));
780 rec->data_len = recovery_size;
781 rec->rec_len = recovery_max_size;
782 rec->key_len = old_map_size;
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 * tdb->transaction->block_size;
797 length = tdb->transaction->block_size;
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->transaction->old_map_size) {
806 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
808 tdb->ecode = TDB_ERR_CORRUPT;
811 memcpy(p, &offset, 4);
812 memcpy(p+4, &length, 4);
816 /* the recovery area contains the old data, not the
817 new data, so we have to call the original tdb_read
819 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
821 tdb->ecode = TDB_ERR_IO;
828 tailer = sizeof(*rec) + recovery_max_size;
829 memcpy(p, &tailer, 4);
832 /* write the recovery data to the recovery area */
833 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
834 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
836 tdb->ecode = TDB_ERR_IO;
839 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
840 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
842 tdb->ecode = TDB_ERR_IO;
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 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
856 magic = TDB_RECOVERY_MAGIC;
859 *magic_offset = recovery_offset + offsetof(struct list_struct, magic);
861 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
862 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
863 tdb->ecode = TDB_ERR_IO;
866 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
867 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
868 tdb->ecode = TDB_ERR_IO;
872 /* ensure the recovery magic marker is on disk */
873 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
880 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
882 const struct tdb_methods *methods;
884 if (tdb->transaction == NULL) {
885 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
889 if (tdb->transaction->prepared) {
890 tdb->ecode = TDB_ERR_EINVAL;
891 _tdb_transaction_cancel(tdb);
892 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
896 if (tdb->transaction->transaction_error) {
897 tdb->ecode = TDB_ERR_IO;
898 _tdb_transaction_cancel(tdb);
899 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
904 if (tdb->transaction->nesting != 0) {
908 /* check for a null transaction */
909 if (tdb->transaction->blocks == NULL) {
913 methods = tdb->transaction->io_methods;
915 /* if there are any locks pending then the caller has not
916 nested their locks properly, so fail the transaction */
917 if (tdb->num_locks || tdb->global_lock.count) {
918 tdb->ecode = TDB_ERR_LOCK;
919 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
920 _tdb_transaction_cancel(tdb);
924 /* upgrade the main transaction lock region to a write lock */
925 if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
926 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
927 tdb->ecode = TDB_ERR_LOCK;
928 _tdb_transaction_cancel(tdb);
932 /* get the global lock - this prevents new users attaching to the database
934 if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
935 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
936 tdb->ecode = TDB_ERR_LOCK;
937 _tdb_transaction_cancel(tdb);
941 if (!(tdb->flags & TDB_NOSYNC)) {
942 /* write the recovery data to the end of the file */
943 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
944 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
945 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
946 _tdb_transaction_cancel(tdb);
951 tdb->transaction->prepared = true;
953 /* expand the file to the new size if needed */
954 if (tdb->map_size != tdb->transaction->old_map_size) {
955 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
957 tdb->transaction->old_map_size) == -1) {
958 tdb->ecode = TDB_ERR_IO;
959 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
960 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
961 _tdb_transaction_cancel(tdb);
964 tdb->map_size = tdb->transaction->old_map_size;
965 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
968 /* Keep the global lock until the actual commit */
974 prepare to commit the current transaction
976 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
978 tdb_trace(tdb, "tdb_transaction_prepare_commit");
979 return _tdb_transaction_prepare_commit(tdb);
983 commit the current transaction
985 int tdb_transaction_commit(struct tdb_context *tdb)
987 const struct tdb_methods *methods;
991 if (tdb->transaction == NULL) {
992 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
996 tdb_trace(tdb, "tdb_transaction_commit");
998 if (tdb->transaction->transaction_error) {
999 tdb->ecode = TDB_ERR_IO;
1000 _tdb_transaction_cancel(tdb);
1001 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1006 if (tdb->transaction->nesting != 0) {
1007 tdb->transaction->nesting--;
1011 /* check for a null transaction */
1012 if (tdb->transaction->blocks == NULL) {
1013 _tdb_transaction_cancel(tdb);
1017 if (!tdb->transaction->prepared) {
1018 int ret = _tdb_transaction_prepare_commit(tdb);
1023 methods = tdb->transaction->io_methods;
1025 /* perform all the writes */
1026 for (i=0;i<tdb->transaction->num_blocks;i++) {
1030 if (tdb->transaction->blocks[i] == NULL) {
1034 offset = i * tdb->transaction->block_size;
1035 length = tdb->transaction->block_size;
1036 if (i == tdb->transaction->num_blocks-1) {
1037 length = tdb->transaction->last_block_size;
1040 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1041 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1043 /* we've overwritten part of the data and
1044 possibly expanded the file, so we need to
1045 run the crash recovery code */
1046 tdb->methods = methods;
1047 tdb_transaction_recover(tdb);
1049 _tdb_transaction_cancel(tdb);
1050 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1052 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1055 SAFE_FREE(tdb->transaction->blocks[i]);
1058 SAFE_FREE(tdb->transaction->blocks);
1059 tdb->transaction->num_blocks = 0;
1061 /* ensure the new data is on disk */
1062 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1066 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1069 TODO: maybe write to some dummy hdr field, or write to magic
1070 offset without mmap, before the last sync, instead of the
1074 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1075 don't change the mtime of the file, this means the file may
1076 not be backed up (as tdb rounding to block sizes means that
1077 file size changes are quite rare too). The following forces
1078 mtime changes when a transaction completes */
1080 utime(tdb->name, NULL);
1083 need_repack = tdb->transaction->need_repack;
1085 /* use a transaction cancel to free memory and remove the
1086 transaction locks */
1087 _tdb_transaction_cancel(tdb);
1090 return tdb_repack(tdb);
1098 recover from an aborted transaction. Must be called with exclusive
1099 database write access already established (including the global
1100 lock to prevent new processes attaching)
1102 int tdb_transaction_recover(struct tdb_context *tdb)
1104 tdb_off_t recovery_head, recovery_eof;
1105 unsigned char *data, *p;
1107 struct list_struct rec;
1109 /* find the recovery area */
1110 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1111 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1112 tdb->ecode = TDB_ERR_IO;
1116 if (recovery_head == 0) {
1117 /* we have never allocated a recovery record */
1121 /* read the recovery record */
1122 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1123 sizeof(rec), DOCONV()) == -1) {
1124 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1125 tdb->ecode = TDB_ERR_IO;
1129 if (rec.magic != TDB_RECOVERY_MAGIC) {
1130 /* there is no valid recovery data */
1134 if (tdb->read_only) {
1135 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1136 tdb->ecode = TDB_ERR_CORRUPT;
1140 recovery_eof = rec.key_len;
1142 data = (unsigned char *)malloc(rec.data_len);
1144 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1145 tdb->ecode = TDB_ERR_OOM;
1149 /* read the full recovery data */
1150 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1151 rec.data_len, 0) == -1) {
1152 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1153 tdb->ecode = TDB_ERR_IO;
1157 /* recover the file data */
1159 while (p+8 < data + rec.data_len) {
1165 memcpy(&len, p+4, 4);
1167 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1169 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1170 tdb->ecode = TDB_ERR_IO;
1178 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1179 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1180 tdb->ecode = TDB_ERR_IO;
1184 /* if the recovery area is after the recovered eof then remove it */
1185 if (recovery_eof <= recovery_head) {
1186 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1187 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1188 tdb->ecode = TDB_ERR_IO;
1193 /* remove the recovery magic */
1194 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic),
1196 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1197 tdb->ecode = TDB_ERR_IO;
1201 /* reduce the file size to the old size */
1203 if (ftruncate(tdb->fd, recovery_eof) != 0) {
1204 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1205 tdb->ecode = TDB_ERR_IO;
1208 tdb->map_size = recovery_eof;
1211 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1212 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1213 tdb->ecode = TDB_ERR_IO;
1217 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",