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 tdb->ecode = TDB_ERR_IO;
387 transaction version of tdb_expand().
389 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
392 /* add a write to the transaction elements, so subsequent
393 reads see the zero data */
394 if (transaction_write(tdb, size, NULL, addition) != 0) {
398 tdb->transaction->need_repack = true;
404 brlock during a transaction - ignore them
406 static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset,
407 int rw_type, int lck_type, int probe, size_t len)
412 static const struct tdb_methods transaction_methods = {
415 transaction_next_hash_chain,
417 transaction_expand_file,
423 start a tdb transaction. No token is returned, as only a single
424 transaction is allowed to be pending per tdb_context
426 int tdb_transaction_start(struct tdb_context *tdb)
428 /* some sanity checks */
429 if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
430 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
431 tdb->ecode = TDB_ERR_EINVAL;
435 /* cope with nested tdb_transaction_start() calls */
436 if (tdb->transaction != NULL) {
437 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
438 tdb->ecode = TDB_ERR_NESTING;
441 tdb->transaction->nesting++;
442 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
443 tdb->transaction->nesting));
447 if (tdb->num_locks != 0 || tdb->global_lock.count) {
448 /* the caller must not have any locks when starting a
449 transaction as otherwise we'll be screwed by lack
450 of nested locks in posix */
451 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
452 tdb->ecode = TDB_ERR_LOCK;
456 if (tdb->travlocks.next != NULL) {
457 /* you cannot use transactions inside a traverse (although you can use
458 traverse inside a transaction) as otherwise you can end up with
460 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
461 tdb->ecode = TDB_ERR_LOCK;
465 tdb->transaction = (struct tdb_transaction *)
466 calloc(sizeof(struct tdb_transaction), 1);
467 if (tdb->transaction == NULL) {
468 tdb->ecode = TDB_ERR_OOM;
472 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
473 tdb->transaction->block_size = tdb->page_size;
475 /* get the transaction write lock. This is a blocking lock. As
476 discussed with Volker, there are a number of ways we could
477 make this async, which we will probably do in the future */
478 if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
479 SAFE_FREE(tdb->transaction->blocks);
480 SAFE_FREE(tdb->transaction);
484 /* get a read lock from the freelist to the end of file. This
485 is upgraded to a write lock during the commit */
486 if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
487 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
488 tdb->ecode = TDB_ERR_LOCK;
492 /* setup a copy of the hash table heads so the hash scan in
493 traverse can be fast */
494 tdb->transaction->hash_heads = (uint32_t *)
495 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
496 if (tdb->transaction->hash_heads == NULL) {
497 tdb->ecode = TDB_ERR_OOM;
500 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
501 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
502 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
503 tdb->ecode = TDB_ERR_IO;
507 /* make sure we know about any file expansions already done by
509 tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
510 tdb->transaction->old_map_size = tdb->map_size;
512 /* finally hook the io methods, replacing them with
513 transaction specific methods */
514 tdb->transaction->io_methods = tdb->methods;
515 tdb->methods = &transaction_methods;
517 /* Trace at the end, so we get sequence number correct. */
518 tdb_trace(tdb, "tdb_transaction_start");
522 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
523 tdb_transaction_unlock(tdb);
524 SAFE_FREE(tdb->transaction->blocks);
525 SAFE_FREE(tdb->transaction->hash_heads);
526 SAFE_FREE(tdb->transaction);
534 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
536 if (tdb->flags & TDB_NOSYNC) {
540 if (fsync(tdb->fd) != 0) {
541 tdb->ecode = TDB_ERR_IO;
542 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
547 tdb_off_t moffset = offset & ~(tdb->page_size-1);
548 if (msync(moffset + (char *)tdb->map_ptr,
549 length + (offset - moffset), MS_SYNC) != 0) {
550 tdb->ecode = TDB_ERR_IO;
551 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
561 int _tdb_transaction_cancel(struct tdb_context *tdb)
565 if (tdb->transaction == NULL) {
566 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
570 if (tdb->transaction->nesting != 0) {
571 tdb->transaction->transaction_error = 1;
572 tdb->transaction->nesting--;
576 tdb->map_size = tdb->transaction->old_map_size;
578 /* free all the transaction blocks */
579 for (i=0;i<tdb->transaction->num_blocks;i++) {
580 if (tdb->transaction->blocks[i] != NULL) {
581 free(tdb->transaction->blocks[i]);
584 SAFE_FREE(tdb->transaction->blocks);
586 if (tdb->transaction->magic_offset) {
587 const struct tdb_methods *methods = tdb->transaction->io_methods;
590 /* remove the recovery marker */
591 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
592 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
593 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
598 /* remove any global lock created during the transaction */
599 if (tdb->global_lock.count != 0) {
600 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
601 tdb->global_lock.count = 0;
604 /* remove any locks created during the transaction */
605 if (tdb->num_locks != 0) {
606 for (i=0;i<tdb->num_lockrecs;i++) {
607 tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
608 F_UNLCK,F_SETLKW, 0, 1);
611 tdb->num_lockrecs = 0;
612 SAFE_FREE(tdb->lockrecs);
615 /* restore the normal io methods */
616 tdb->methods = tdb->transaction->io_methods;
618 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
619 tdb_transaction_unlock(tdb);
620 SAFE_FREE(tdb->transaction->hash_heads);
621 SAFE_FREE(tdb->transaction);
627 cancel the current transaction
629 int tdb_transaction_cancel(struct tdb_context *tdb)
631 tdb_trace(tdb, "tdb_transaction_cancel");
632 return _tdb_transaction_cancel(tdb);
636 work out how much space the linearised recovery data will consume
638 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
640 tdb_len_t recovery_size = 0;
643 recovery_size = sizeof(uint32_t);
644 for (i=0;i<tdb->transaction->num_blocks;i++) {
645 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
648 if (tdb->transaction->blocks[i] == NULL) {
651 recovery_size += 2*sizeof(tdb_off_t);
652 if (i == tdb->transaction->num_blocks-1) {
653 recovery_size += tdb->transaction->last_block_size;
655 recovery_size += tdb->transaction->block_size;
659 return recovery_size;
663 allocate the recovery area, or use an existing recovery area if it is
666 static int tdb_recovery_allocate(struct tdb_context *tdb,
667 tdb_len_t *recovery_size,
668 tdb_off_t *recovery_offset,
669 tdb_len_t *recovery_max_size)
671 struct tdb_record rec;
672 const struct tdb_methods *methods = tdb->transaction->io_methods;
673 tdb_off_t recovery_head;
675 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
676 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
682 if (recovery_head != 0 &&
683 methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
684 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
688 *recovery_size = tdb_recovery_size(tdb);
690 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
691 /* it fits in the existing area */
692 *recovery_max_size = rec.rec_len;
693 *recovery_offset = recovery_head;
697 /* we need to free up the old recovery area, then allocate a
698 new one at the end of the file. Note that we cannot use
699 tdb_allocate() to allocate the new one as that might return
700 us an area that is being currently used (as of the start of
702 if (recovery_head != 0) {
703 if (tdb_free(tdb, recovery_head, &rec) == -1) {
704 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
709 /* the tdb_free() call might have increased the recovery size */
710 *recovery_size = tdb_recovery_size(tdb);
712 /* round up to a multiple of page size */
713 *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
714 *recovery_offset = tdb->map_size;
715 recovery_head = *recovery_offset;
717 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
718 (tdb->map_size - tdb->transaction->old_map_size) +
719 sizeof(rec) + *recovery_max_size) == -1) {
720 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
724 /* remap the file (if using mmap) */
725 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
727 /* we have to reset the old map size so that we don't try to expand the file
728 again in the transaction commit, which would destroy the recovery area */
729 tdb->transaction->old_map_size = tdb->map_size;
731 /* write the recovery header offset and sync - we can sync without a race here
732 as the magic ptr in the recovery record has not been set */
733 CONVERT(recovery_head);
734 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
735 &recovery_head, sizeof(tdb_off_t)) == -1) {
736 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
739 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
740 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
749 setup the recovery data that will be used on a crash during commit
751 static int transaction_setup_recovery(struct tdb_context *tdb,
752 tdb_off_t *magic_offset)
754 tdb_len_t recovery_size;
755 unsigned char *data, *p;
756 const struct tdb_methods *methods = tdb->transaction->io_methods;
757 struct tdb_record *rec;
758 tdb_off_t recovery_offset, recovery_max_size;
759 tdb_off_t old_map_size = tdb->transaction->old_map_size;
760 uint32_t magic, tailer;
764 check that the recovery area has enough space
766 if (tdb_recovery_allocate(tdb, &recovery_size,
767 &recovery_offset, &recovery_max_size) == -1) {
771 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
773 tdb->ecode = TDB_ERR_OOM;
777 rec = (struct tdb_record *)data;
778 memset(rec, 0, sizeof(*rec));
781 rec->data_len = recovery_size;
782 rec->rec_len = recovery_max_size;
783 rec->key_len = old_map_size;
786 /* build the recovery data into a single blob to allow us to do a single
787 large write, which should be more efficient */
788 p = data + sizeof(*rec);
789 for (i=0;i<tdb->transaction->num_blocks;i++) {
793 if (tdb->transaction->blocks[i] == NULL) {
797 offset = i * tdb->transaction->block_size;
798 length = tdb->transaction->block_size;
799 if (i == tdb->transaction->num_blocks-1) {
800 length = tdb->transaction->last_block_size;
803 if (offset >= old_map_size) {
806 if (offset + length > tdb->transaction->old_map_size) {
807 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
809 tdb->ecode = TDB_ERR_CORRUPT;
812 memcpy(p, &offset, 4);
813 memcpy(p+4, &length, 4);
817 /* the recovery area contains the old data, not the
818 new data, so we have to call the original tdb_read
820 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
822 tdb->ecode = TDB_ERR_IO;
829 tailer = sizeof(*rec) + recovery_max_size;
830 memcpy(p, &tailer, 4);
833 /* write the recovery data to the recovery area */
834 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
835 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
837 tdb->ecode = TDB_ERR_IO;
840 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
841 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
843 tdb->ecode = TDB_ERR_IO;
847 /* as we don't have ordered writes, we have to sync the recovery
848 data before we update the magic to indicate that the recovery
850 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
857 magic = TDB_RECOVERY_MAGIC;
860 *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
862 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
863 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
864 tdb->ecode = TDB_ERR_IO;
867 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
868 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
869 tdb->ecode = TDB_ERR_IO;
873 /* ensure the recovery magic marker is on disk */
874 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
881 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
883 const struct tdb_methods *methods;
885 if (tdb->transaction == NULL) {
886 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
890 if (tdb->transaction->prepared) {
891 tdb->ecode = TDB_ERR_EINVAL;
892 _tdb_transaction_cancel(tdb);
893 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
897 if (tdb->transaction->transaction_error) {
898 tdb->ecode = TDB_ERR_IO;
899 _tdb_transaction_cancel(tdb);
900 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
905 if (tdb->transaction->nesting != 0) {
909 /* check for a null transaction */
910 if (tdb->transaction->blocks == NULL) {
914 methods = tdb->transaction->io_methods;
916 /* if there are any locks pending then the caller has not
917 nested their locks properly, so fail the transaction */
918 if (tdb->num_locks || tdb->global_lock.count) {
919 tdb->ecode = TDB_ERR_LOCK;
920 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
921 _tdb_transaction_cancel(tdb);
925 /* upgrade the main transaction lock region to a write lock */
926 if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
927 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
928 tdb->ecode = TDB_ERR_LOCK;
929 _tdb_transaction_cancel(tdb);
933 /* get the global lock - this prevents new users attaching to the database
935 if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
936 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
937 tdb->ecode = TDB_ERR_LOCK;
938 _tdb_transaction_cancel(tdb);
942 if (!(tdb->flags & TDB_NOSYNC)) {
943 /* write the recovery data to the end of the file */
944 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
945 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
946 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
947 _tdb_transaction_cancel(tdb);
952 tdb->transaction->prepared = true;
954 /* expand the file to the new size if needed */
955 if (tdb->map_size != tdb->transaction->old_map_size) {
956 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
958 tdb->transaction->old_map_size) == -1) {
959 tdb->ecode = TDB_ERR_IO;
960 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
961 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
962 _tdb_transaction_cancel(tdb);
965 tdb->map_size = tdb->transaction->old_map_size;
966 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
969 /* Keep the global lock until the actual commit */
975 prepare to commit the current transaction
977 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
979 tdb_trace(tdb, "tdb_transaction_prepare_commit");
980 return _tdb_transaction_prepare_commit(tdb);
984 commit the current transaction
986 int tdb_transaction_commit(struct tdb_context *tdb)
988 const struct tdb_methods *methods;
992 if (tdb->transaction == NULL) {
993 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
997 tdb_trace(tdb, "tdb_transaction_commit");
999 if (tdb->transaction->transaction_error) {
1000 tdb->ecode = TDB_ERR_IO;
1001 _tdb_transaction_cancel(tdb);
1002 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1007 if (tdb->transaction->nesting != 0) {
1008 tdb->transaction->nesting--;
1012 /* check for a null transaction */
1013 if (tdb->transaction->blocks == NULL) {
1014 _tdb_transaction_cancel(tdb);
1018 if (!tdb->transaction->prepared) {
1019 int ret = _tdb_transaction_prepare_commit(tdb);
1024 methods = tdb->transaction->io_methods;
1026 /* perform all the writes */
1027 for (i=0;i<tdb->transaction->num_blocks;i++) {
1031 if (tdb->transaction->blocks[i] == NULL) {
1035 offset = i * tdb->transaction->block_size;
1036 length = tdb->transaction->block_size;
1037 if (i == tdb->transaction->num_blocks-1) {
1038 length = tdb->transaction->last_block_size;
1041 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1042 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1044 /* we've overwritten part of the data and
1045 possibly expanded the file, so we need to
1046 run the crash recovery code */
1047 tdb->methods = methods;
1048 tdb_transaction_recover(tdb);
1050 _tdb_transaction_cancel(tdb);
1051 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1053 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1056 SAFE_FREE(tdb->transaction->blocks[i]);
1059 SAFE_FREE(tdb->transaction->blocks);
1060 tdb->transaction->num_blocks = 0;
1062 /* ensure the new data is on disk */
1063 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1067 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1070 TODO: maybe write to some dummy hdr field, or write to magic
1071 offset without mmap, before the last sync, instead of the
1075 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1076 don't change the mtime of the file, this means the file may
1077 not be backed up (as tdb rounding to block sizes means that
1078 file size changes are quite rare too). The following forces
1079 mtime changes when a transaction completes */
1081 utime(tdb->name, NULL);
1084 need_repack = tdb->transaction->need_repack;
1086 /* use a transaction cancel to free memory and remove the
1087 transaction locks */
1088 _tdb_transaction_cancel(tdb);
1091 return tdb_repack(tdb);
1099 recover from an aborted transaction. Must be called with exclusive
1100 database write access already established (including the global
1101 lock to prevent new processes attaching)
1103 int tdb_transaction_recover(struct tdb_context *tdb)
1105 tdb_off_t recovery_head, recovery_eof;
1106 unsigned char *data, *p;
1108 struct tdb_record rec;
1110 /* find the recovery area */
1111 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1112 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1113 tdb->ecode = TDB_ERR_IO;
1117 if (recovery_head == 0) {
1118 /* we have never allocated a recovery record */
1122 /* read the recovery record */
1123 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1124 sizeof(rec), DOCONV()) == -1) {
1125 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1126 tdb->ecode = TDB_ERR_IO;
1130 if (rec.magic != TDB_RECOVERY_MAGIC) {
1131 /* there is no valid recovery data */
1135 if (tdb->read_only) {
1136 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1137 tdb->ecode = TDB_ERR_CORRUPT;
1141 recovery_eof = rec.key_len;
1143 data = (unsigned char *)malloc(rec.data_len);
1145 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1146 tdb->ecode = TDB_ERR_OOM;
1150 /* read the full recovery data */
1151 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1152 rec.data_len, 0) == -1) {
1153 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1154 tdb->ecode = TDB_ERR_IO;
1158 /* recover the file data */
1160 while (p+8 < data + rec.data_len) {
1166 memcpy(&len, p+4, 4);
1168 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1170 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1171 tdb->ecode = TDB_ERR_IO;
1179 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1180 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1181 tdb->ecode = TDB_ERR_IO;
1185 /* if the recovery area is after the recovered eof then remove it */
1186 if (recovery_eof <= recovery_head) {
1187 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1188 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1189 tdb->ecode = TDB_ERR_IO;
1194 /* remove the recovery magic */
1195 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1197 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1198 tdb->ecode = TDB_ERR_IO;
1202 /* reduce the file size to the old size */
1204 if (ftruncate(tdb->fd, recovery_eof) != 0) {
1205 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1206 tdb->ecode = TDB_ERR_IO;
1209 tdb->map_size = recovery_eof;
1212 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1213 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1214 tdb->ecode = TDB_ERR_IO;
1218 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",