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 all writes 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 open 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 fsync/msync calls are made. This means we
86 are still proof against a process dying during transaction commit,
87 but not against machine reboot.
89 - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
90 tdb_add_flags() transaction nesting is enabled.
91 It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
92 The default is that transaction nesting is allowed.
93 Note: this default may change in future versions of tdb.
95 Beware. when transactions are nested a transaction successfully
96 completed with tdb_transaction_commit() can be silently unrolled later.
98 - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
99 tdb_add_flags() transaction nesting is disabled.
100 It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
101 An attempt create a nested transaction will fail with TDB_ERR_NESTING.
102 The default is that transaction nesting is allowed.
103 Note: this default may change in future versions of tdb.
108 hold the context of any current transaction
110 struct tdb_transaction {
111 /* we keep a mirrored copy of the tdb hash heads here so
112 tdb_next_hash_chain() can operate efficiently */
113 uint32_t *hash_heads;
115 /* the original io methods - used to do IOs to the real db */
116 const struct tdb_methods *io_methods;
118 /* the list of transaction blocks. When a block is first
119 written to, it gets created in this list */
122 uint32_t block_size; /* bytes in each block */
123 uint32_t last_block_size; /* number of valid bytes in the last block */
125 /* non-zero when an internal transaction error has
126 occurred. All write operations will then fail until the
127 transaction is ended */
128 int transaction_error;
130 /* when inside a transaction we need to keep track of any
131 nested tdb_transaction_start() calls, as these are allowed,
132 but don't create a new transaction */
135 /* set when a prepare has already occurred */
137 tdb_off_t magic_offset;
139 /* old file size before transaction */
140 tdb_len_t old_map_size;
142 /* did we expand in this transaction */
148 read while in a transaction. We need to check first if the data is in our list
149 of transaction elements, then if not do a real read
151 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
152 tdb_len_t len, int cv)
156 /* break it down into block sized ops */
157 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
158 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
159 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
164 buf = (void *)(len2 + (char *)buf);
171 blk = off / tdb->transaction->block_size;
173 /* see if we have it in the block list */
174 if (tdb->transaction->num_blocks <= blk ||
175 tdb->transaction->blocks[blk] == NULL) {
176 /* nope, do a real read */
177 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
183 /* it is in the block list. Now check for the last block */
184 if (blk == tdb->transaction->num_blocks-1) {
185 if (len > tdb->transaction->last_block_size) {
190 /* now copy it out of this block */
191 memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
193 tdb_convert(buf, len);
198 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%u len=%u\n", off, len));
199 tdb->ecode = TDB_ERR_IO;
200 tdb->transaction->transaction_error = 1;
206 write while in a transaction
208 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
209 const void *buf, tdb_len_t len)
217 /* Only a commit is allowed on a prepared transaction */
218 if (tdb->transaction->prepared) {
219 tdb->ecode = TDB_ERR_EINVAL;
220 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
221 tdb->transaction->transaction_error = 1;
225 /* if the write is to a hash head, then update the transaction
227 if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
228 off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
229 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
230 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
233 /* break it up into block sized chunks */
234 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
235 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
236 if (transaction_write(tdb, off, buf, len2) != 0) {
241 buf = (const void *)(len2 + (const char *)buf);
248 blk = off / tdb->transaction->block_size;
249 off = off % tdb->transaction->block_size;
251 if (tdb->transaction->num_blocks <= blk) {
252 uint8_t **new_blocks;
253 /* expand the blocks array */
254 new_blocks = (uint8_t **)realloc(tdb->transaction->blocks,
255 (blk+1)*sizeof(uint8_t *));
256 if (new_blocks == NULL) {
257 tdb->ecode = TDB_ERR_OOM;
260 memset(&new_blocks[tdb->transaction->num_blocks], 0,
261 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
262 tdb->transaction->blocks = new_blocks;
263 tdb->transaction->num_blocks = blk+1;
264 tdb->transaction->last_block_size = 0;
267 /* allocate and fill a block? */
268 if (tdb->transaction->blocks[blk] == NULL) {
269 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
270 if (tdb->transaction->blocks[blk] == NULL) {
271 tdb->ecode = TDB_ERR_OOM;
272 tdb->transaction->transaction_error = 1;
275 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
276 tdb_len_t len2 = tdb->transaction->block_size;
277 if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
278 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
280 if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
281 tdb->transaction->blocks[blk],
283 SAFE_FREE(tdb->transaction->blocks[blk]);
284 tdb->ecode = TDB_ERR_IO;
287 if (blk == tdb->transaction->num_blocks-1) {
288 tdb->transaction->last_block_size = len2;
293 /* overwrite part of an existing block */
294 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
295 if (blk == tdb->transaction->num_blocks-1) {
296 if (len + off > tdb->transaction->last_block_size) {
297 tdb->transaction->last_block_size = len + off;
304 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%u len=%u\n",
305 (blk*tdb->transaction->block_size) + off, len));
306 tdb->transaction->transaction_error = 1;
312 write while in a transaction - this variant never expands the transaction blocks, it only
313 updates existing blocks. This means it cannot change the recovery size
315 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
316 const void *buf, tdb_len_t len)
320 /* break it up into block sized chunks */
321 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
322 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
323 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
329 buf = (const void *)(len2 + (const char *)buf);
337 blk = off / tdb->transaction->block_size;
338 off = off % tdb->transaction->block_size;
340 if (tdb->transaction->num_blocks <= blk ||
341 tdb->transaction->blocks[blk] == NULL) {
345 if (blk == tdb->transaction->num_blocks-1 &&
346 off + len > tdb->transaction->last_block_size) {
347 if (off >= tdb->transaction->last_block_size) {
350 len = tdb->transaction->last_block_size - off;
353 /* overwrite part of an existing block */
354 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
361 accelerated hash chain head search, using the cached hash heads
363 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
366 for (;h < tdb->hash_size;h++) {
367 /* the +1 takes account of the freelist */
368 if (0 != tdb->transaction->hash_heads[h+1]) {
376 out of bounds check during a transaction
378 static int transaction_oob(struct tdb_context *tdb, tdb_off_t off,
379 tdb_len_t len, int probe)
381 if (off + len >= off && off + len <= tdb->map_size) {
384 tdb->ecode = TDB_ERR_IO;
389 transaction version of tdb_expand().
391 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
394 const char buf_zero[8192] = {0};
395 size_t buf_len = sizeof(buf_zero);
397 while (addition > 0) {
398 size_t n = MIN(addition, buf_len);
401 ret = transaction_write(tdb, size, buf_zero, n);
410 tdb->transaction->expanded = true;
415 static const struct tdb_methods transaction_methods = {
418 transaction_next_hash_chain,
420 transaction_expand_file,
424 * Is a transaction currently active on this context?
427 _PUBLIC_ bool tdb_transaction_active(struct tdb_context *tdb)
429 return (tdb->transaction != NULL);
433 start a tdb transaction. No token is returned, as only a single
434 transaction is allowed to be pending per tdb_context
436 static int _tdb_transaction_start(struct tdb_context *tdb,
437 enum tdb_lock_flags lockflags)
439 /* some sanity checks */
440 if (tdb->read_only || (tdb->flags & TDB_INTERNAL)
441 || tdb->traverse_read) {
442 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
443 tdb->ecode = TDB_ERR_EINVAL;
447 /* cope with nested tdb_transaction_start() calls */
448 if (tdb->transaction != NULL) {
449 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
450 tdb->ecode = TDB_ERR_NESTING;
453 tdb->transaction->nesting++;
454 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
455 tdb->transaction->nesting));
459 if (tdb_have_extra_locks(tdb)) {
460 /* the caller must not have any locks when starting a
461 transaction as otherwise we'll be screwed by lack
462 of nested locks in posix */
463 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
464 tdb->ecode = TDB_ERR_LOCK;
468 if (tdb->travlocks.next != NULL) {
469 /* you cannot use transactions inside a traverse (although you can use
470 traverse inside a transaction) as otherwise you can end up with
472 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
473 tdb->ecode = TDB_ERR_LOCK;
477 tdb->transaction = (struct tdb_transaction *)
478 calloc(sizeof(struct tdb_transaction), 1);
479 if (tdb->transaction == NULL) {
480 tdb->ecode = TDB_ERR_OOM;
484 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
485 tdb->transaction->block_size = tdb->page_size;
487 /* get the transaction write lock. This is a blocking lock. As
488 discussed with Volker, there are a number of ways we could
489 make this async, which we will probably do in the future */
490 if (tdb_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
491 SAFE_FREE(tdb->transaction->blocks);
492 SAFE_FREE(tdb->transaction);
493 if ((lockflags & TDB_LOCK_WAIT) == 0) {
494 tdb->ecode = TDB_ERR_NOLOCK;
496 TDB_LOG((tdb, TDB_DEBUG_ERROR,
497 "tdb_transaction_start: "
498 "failed to get transaction lock\n"));
503 /* get a read lock from the freelist to the end of file. This
504 is upgraded to a write lock during the commit */
505 if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
506 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
507 goto fail_allrecord_lock;
510 /* setup a copy of the hash table heads so the hash scan in
511 traverse can be fast */
512 tdb->transaction->hash_heads = (uint32_t *)
513 calloc(tdb->hash_size+1, sizeof(uint32_t));
514 if (tdb->transaction->hash_heads == NULL) {
515 tdb->ecode = TDB_ERR_OOM;
518 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
519 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
520 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
521 tdb->ecode = TDB_ERR_IO;
525 /* make sure we know about any file expansions already done by
527 tdb->methods->tdb_oob(tdb, tdb->map_size, 1, 1);
528 tdb->transaction->old_map_size = tdb->map_size;
530 /* finally hook the io methods, replacing them with
531 transaction specific methods */
532 tdb->transaction->io_methods = tdb->methods;
533 tdb->methods = &transaction_methods;
535 /* Trace at the end, so we get sequence number correct. */
536 tdb_trace(tdb, "tdb_transaction_start");
540 tdb_allrecord_unlock(tdb, F_RDLCK, false);
542 tdb_transaction_unlock(tdb, F_WRLCK);
543 SAFE_FREE(tdb->transaction->blocks);
544 SAFE_FREE(tdb->transaction->hash_heads);
545 SAFE_FREE(tdb->transaction);
549 _PUBLIC_ int tdb_transaction_start(struct tdb_context *tdb)
551 return _tdb_transaction_start(tdb, TDB_LOCK_WAIT);
554 _PUBLIC_ int tdb_transaction_start_nonblock(struct tdb_context *tdb)
556 return _tdb_transaction_start(tdb, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
562 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
564 if (tdb->flags & TDB_NOSYNC) {
568 #ifdef HAVE_FDATASYNC
569 if (fdatasync(tdb->fd) != 0) {
571 if (fsync(tdb->fd) != 0) {
573 tdb->ecode = TDB_ERR_IO;
574 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
579 tdb_off_t moffset = offset & ~(tdb->page_size-1);
580 if (msync(moffset + (char *)tdb->map_ptr,
581 length + (offset - moffset), MS_SYNC) != 0) {
582 tdb->ecode = TDB_ERR_IO;
583 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
593 static int _tdb_transaction_cancel(struct tdb_context *tdb)
598 if (tdb->transaction == NULL) {
599 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
603 if (tdb->transaction->nesting != 0) {
604 tdb->transaction->transaction_error = 1;
605 tdb->transaction->nesting--;
609 tdb->map_size = tdb->transaction->old_map_size;
611 /* free all the transaction blocks */
612 for (i=0;i<tdb->transaction->num_blocks;i++) {
613 if (tdb->transaction->blocks[i] != NULL) {
614 free(tdb->transaction->blocks[i]);
617 SAFE_FREE(tdb->transaction->blocks);
619 if (tdb->transaction->magic_offset) {
620 const struct tdb_methods *methods = tdb->transaction->io_methods;
621 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
623 /* remove the recovery marker */
624 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
625 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
626 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
631 /* This also removes the OPEN_LOCK, if we have it. */
632 tdb_release_transaction_locks(tdb);
634 /* restore the normal io methods */
635 tdb->methods = tdb->transaction->io_methods;
637 SAFE_FREE(tdb->transaction->hash_heads);
638 SAFE_FREE(tdb->transaction);
644 cancel the current transaction
646 _PUBLIC_ int tdb_transaction_cancel(struct tdb_context *tdb)
648 tdb_trace(tdb, "tdb_transaction_cancel");
649 return _tdb_transaction_cancel(tdb);
653 work out how much space the linearised recovery data will consume
655 static bool tdb_recovery_size(struct tdb_context *tdb, tdb_len_t *result)
657 tdb_len_t recovery_size = 0;
660 recovery_size = sizeof(uint32_t);
661 for (i=0;i<tdb->transaction->num_blocks;i++) {
662 tdb_len_t block_size;
663 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
666 if (tdb->transaction->blocks[i] == NULL) {
669 if (!tdb_add_len_t(recovery_size, 2*sizeof(tdb_off_t),
673 if (i == tdb->transaction->num_blocks-1) {
674 block_size = tdb->transaction->last_block_size;
676 block_size = tdb->transaction->block_size;
678 if (!tdb_add_len_t(recovery_size, block_size,
684 *result = recovery_size;
688 int tdb_recovery_area(struct tdb_context *tdb,
689 const struct tdb_methods *methods,
690 tdb_off_t *recovery_offset,
691 struct tdb_record *rec)
695 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, recovery_offset) == -1) {
699 if (*recovery_offset == 0) {
704 if (methods->tdb_read(tdb, *recovery_offset, rec, sizeof(*rec),
709 /* ignore invalid recovery regions: can happen in crash */
710 if (rec->magic != TDB_RECOVERY_MAGIC &&
711 rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
712 *recovery_offset = 0;
716 ret = methods->tdb_oob(tdb, *recovery_offset, rec->rec_len, 1);
718 *recovery_offset = 0;
726 allocate the recovery area, or use an existing recovery area if it is
729 static int tdb_recovery_allocate(struct tdb_context *tdb,
730 tdb_len_t *recovery_size,
731 tdb_off_t *recovery_offset,
732 tdb_len_t *recovery_max_size)
734 struct tdb_record rec;
735 const struct tdb_methods *methods = tdb->transaction->io_methods;
736 tdb_off_t recovery_head, new_end;
738 if (tdb_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
739 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
743 if (!tdb_recovery_size(tdb, recovery_size)) {
744 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
745 "overflow recovery size\n"));
749 /* Existing recovery area? */
750 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
751 /* it fits in the existing area */
752 *recovery_max_size = rec.rec_len;
753 *recovery_offset = recovery_head;
757 /* If recovery area in middle of file, we need a new one. */
758 if (recovery_head == 0
759 || recovery_head + sizeof(rec) + rec.rec_len != tdb->map_size) {
760 /* we need to free up the old recovery area, then allocate a
761 new one at the end of the file. Note that we cannot use
762 tdb_allocate() to allocate the new one as that might return
763 us an area that is being currently used (as of the start of
766 if (tdb_free(tdb, recovery_head, &rec) == -1) {
767 TDB_LOG((tdb, TDB_DEBUG_FATAL,
768 "tdb_recovery_allocate: failed to"
769 " free previous recovery area\n"));
773 /* the tdb_free() call might have increased
774 * the recovery size */
775 if (!tdb_recovery_size(tdb, recovery_size)) {
776 TDB_LOG((tdb, TDB_DEBUG_FATAL,
777 "tdb_recovery_allocate: "
778 "overflow recovery size\n"));
783 /* New head will be at end of file. */
784 recovery_head = tdb->map_size;
787 /* Now we know where it will be. */
788 *recovery_offset = recovery_head;
790 /* Expand by more than we need, so we don't do it often. */
791 *recovery_max_size = tdb_expand_adjust(tdb->map_size,
796 if (!tdb_add_off_t(recovery_head, sizeof(rec), &new_end) ||
797 !tdb_add_off_t(new_end, *recovery_max_size, &new_end)) {
798 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
799 "overflow recovery area\n"));
803 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
804 new_end - tdb->transaction->old_map_size)
806 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
810 /* remap the file (if using mmap) */
811 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
813 /* we have to reset the old map size so that we don't try to expand the file
814 again in the transaction commit, which would destroy the recovery area */
815 tdb->transaction->old_map_size = tdb->map_size;
817 /* write the recovery header offset and sync - we can sync without a race here
818 as the magic ptr in the recovery record has not been set */
819 CONVERT(recovery_head);
820 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
821 &recovery_head, sizeof(tdb_off_t)) == -1) {
822 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
825 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
826 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
835 setup the recovery data that will be used on a crash during commit
837 static int transaction_setup_recovery(struct tdb_context *tdb,
838 tdb_off_t *magic_offset)
840 tdb_len_t recovery_size;
841 unsigned char *data, *p;
842 const struct tdb_methods *methods = tdb->transaction->io_methods;
843 struct tdb_record *rec;
844 tdb_off_t recovery_offset, recovery_max_size;
845 tdb_off_t old_map_size = tdb->transaction->old_map_size;
846 uint32_t magic, tailer;
850 check that the recovery area has enough space
852 if (tdb_recovery_allocate(tdb, &recovery_size,
853 &recovery_offset, &recovery_max_size) == -1) {
857 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
859 tdb->ecode = TDB_ERR_OOM;
863 rec = (struct tdb_record *)data;
864 memset(rec, 0, sizeof(*rec));
866 rec->magic = TDB_RECOVERY_INVALID_MAGIC;
867 rec->data_len = recovery_size;
868 rec->rec_len = recovery_max_size;
869 rec->key_len = old_map_size;
872 /* build the recovery data into a single blob to allow us to do a single
873 large write, which should be more efficient */
874 p = data + sizeof(*rec);
875 for (i=0;i<tdb->transaction->num_blocks;i++) {
879 if (tdb->transaction->blocks[i] == NULL) {
883 offset = i * tdb->transaction->block_size;
884 length = tdb->transaction->block_size;
885 if (i == tdb->transaction->num_blocks-1) {
886 length = tdb->transaction->last_block_size;
889 if (offset >= old_map_size) {
892 if (offset + length > tdb->transaction->old_map_size) {
893 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
895 tdb->ecode = TDB_ERR_CORRUPT;
898 memcpy(p, &offset, 4);
899 memcpy(p+4, &length, 4);
903 /* the recovery area contains the old data, not the
904 new data, so we have to call the original tdb_read
906 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
908 tdb->ecode = TDB_ERR_IO;
915 tailer = sizeof(*rec) + recovery_max_size;
916 memcpy(p, &tailer, 4);
921 /* write the recovery data to the recovery area */
922 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
923 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
925 tdb->ecode = TDB_ERR_IO;
928 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
929 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
931 tdb->ecode = TDB_ERR_IO;
935 /* as we don't have ordered writes, we have to sync the recovery
936 data before we update the magic to indicate that the recovery
938 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
945 magic = TDB_RECOVERY_MAGIC;
948 *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
950 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
951 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
952 tdb->ecode = TDB_ERR_IO;
955 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
956 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
957 tdb->ecode = TDB_ERR_IO;
961 /* ensure the recovery magic marker is on disk */
962 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
969 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
971 const struct tdb_methods *methods;
973 if (tdb->transaction == NULL) {
974 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
978 if (tdb->transaction->prepared) {
979 tdb->ecode = TDB_ERR_EINVAL;
980 _tdb_transaction_cancel(tdb);
981 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
985 if (tdb->transaction->transaction_error) {
986 tdb->ecode = TDB_ERR_IO;
987 _tdb_transaction_cancel(tdb);
988 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
993 if (tdb->transaction->nesting != 0) {
997 /* check for a null transaction */
998 if (tdb->transaction->blocks == NULL) {
1002 methods = tdb->transaction->io_methods;
1004 /* if there are any locks pending then the caller has not
1005 nested their locks properly, so fail the transaction */
1006 if (tdb_have_extra_locks(tdb)) {
1007 tdb->ecode = TDB_ERR_LOCK;
1008 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
1009 _tdb_transaction_cancel(tdb);
1013 /* upgrade the main transaction lock region to a write lock */
1014 if (tdb_allrecord_upgrade(tdb) == -1) {
1015 if (tdb->ecode == TDB_ERR_RDONLY && tdb->read_only) {
1016 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1017 "tdb_transaction_prepare_commit: "
1018 "failed to upgrade hash locks: "
1019 "database is read only\n"));
1020 } else if (tdb->ecode == TDB_ERR_RDONLY
1021 && tdb->traverse_read) {
1022 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1023 "tdb_transaction_prepare_commit: "
1024 "failed to upgrade hash locks: "
1025 "a database traverse is in progress\n"));
1027 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1028 "tdb_transaction_prepare_commit: "
1029 "failed to upgrade hash locks: %s\n",
1030 tdb_errorstr(tdb)));
1032 _tdb_transaction_cancel(tdb);
1036 /* get the open lock - this prevents new users attaching to the database
1037 during the commit */
1038 if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
1039 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
1040 _tdb_transaction_cancel(tdb);
1044 /* write the recovery data to the end of the file */
1045 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
1046 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
1047 _tdb_transaction_cancel(tdb);
1051 tdb->transaction->prepared = true;
1053 /* expand the file to the new size if needed */
1054 if (tdb->map_size != tdb->transaction->old_map_size) {
1055 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
1057 tdb->transaction->old_map_size) == -1) {
1058 tdb->ecode = TDB_ERR_IO;
1059 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
1060 _tdb_transaction_cancel(tdb);
1063 tdb->map_size = tdb->transaction->old_map_size;
1064 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
1067 /* Keep the open lock until the actual commit */
1073 prepare to commit the current transaction
1075 _PUBLIC_ int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1077 tdb_trace(tdb, "tdb_transaction_prepare_commit");
1078 return _tdb_transaction_prepare_commit(tdb);
1081 /* A repack is worthwhile if the largest is less than half total free. */
1082 static bool repack_worthwhile(struct tdb_context *tdb)
1085 struct tdb_record rec;
1086 tdb_len_t total = 0, largest = 0;
1088 if (tdb_ofs_read(tdb, FREELIST_TOP, &ptr) == -1) {
1092 while (ptr != 0 && tdb_rec_free_read(tdb, ptr, &rec) == 0) {
1093 total += rec.rec_len;
1094 if (rec.rec_len > largest) {
1095 largest = rec.rec_len;
1100 return total > largest * 2;
1104 commit the current transaction
1106 _PUBLIC_ int tdb_transaction_commit(struct tdb_context *tdb)
1108 const struct tdb_methods *methods;
1110 bool need_repack = false;
1112 if (tdb->transaction == NULL) {
1113 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1117 tdb_trace(tdb, "tdb_transaction_commit");
1119 if (tdb->transaction->transaction_error) {
1120 tdb->ecode = TDB_ERR_IO;
1121 _tdb_transaction_cancel(tdb);
1122 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1127 if (tdb->transaction->nesting != 0) {
1128 tdb->transaction->nesting--;
1132 /* check for a null transaction */
1133 if (tdb->transaction->blocks == NULL) {
1134 _tdb_transaction_cancel(tdb);
1138 if (!tdb->transaction->prepared) {
1139 int ret = _tdb_transaction_prepare_commit(tdb);
1144 methods = tdb->transaction->io_methods;
1146 /* perform all the writes */
1147 for (i=0;i<tdb->transaction->num_blocks;i++) {
1151 if (tdb->transaction->blocks[i] == NULL) {
1155 offset = i * tdb->transaction->block_size;
1156 length = tdb->transaction->block_size;
1157 if (i == tdb->transaction->num_blocks-1) {
1158 length = tdb->transaction->last_block_size;
1161 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1162 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1164 /* we've overwritten part of the data and
1165 possibly expanded the file, so we need to
1166 run the crash recovery code */
1167 tdb->methods = methods;
1168 tdb_transaction_recover(tdb);
1170 _tdb_transaction_cancel(tdb);
1172 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1175 SAFE_FREE(tdb->transaction->blocks[i]);
1178 /* Do this before we drop lock or blocks. */
1179 if (tdb->transaction->expanded) {
1180 need_repack = repack_worthwhile(tdb);
1183 SAFE_FREE(tdb->transaction->blocks);
1184 tdb->transaction->num_blocks = 0;
1186 /* ensure the new data is on disk */
1187 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1192 TODO: maybe write to some dummy hdr field, or write to magic
1193 offset without mmap, before the last sync, instead of the
1197 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1198 don't change the mtime of the file, this means the file may
1199 not be backed up (as tdb rounding to block sizes means that
1200 file size changes are quite rare too). The following forces
1201 mtime changes when a transaction completes */
1203 utime(tdb->name, NULL);
1206 /* use a transaction cancel to free memory and remove the
1207 transaction locks */
1208 _tdb_transaction_cancel(tdb);
1211 return tdb_repack(tdb);
1219 recover from an aborted transaction. Must be called with exclusive
1220 database write access already established (including the open
1221 lock to prevent new processes attaching)
1223 int tdb_transaction_recover(struct tdb_context *tdb)
1225 tdb_off_t recovery_head, recovery_eof;
1226 unsigned char *data, *p;
1228 struct tdb_record rec;
1230 /* find the recovery area */
1231 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1232 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1233 tdb->ecode = TDB_ERR_IO;
1237 if (recovery_head == 0) {
1238 /* we have never allocated a recovery record */
1242 /* read the recovery record */
1243 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1244 sizeof(rec), DOCONV()) == -1) {
1245 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1246 tdb->ecode = TDB_ERR_IO;
1250 if (rec.magic != TDB_RECOVERY_MAGIC) {
1251 /* there is no valid recovery data */
1255 if (tdb->read_only) {
1256 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1257 tdb->ecode = TDB_ERR_CORRUPT;
1261 recovery_eof = rec.key_len;
1263 data = (unsigned char *)malloc(rec.data_len);
1265 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1266 tdb->ecode = TDB_ERR_OOM;
1270 /* read the full recovery data */
1271 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1272 rec.data_len, 0) == -1) {
1273 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1274 tdb->ecode = TDB_ERR_IO;
1278 /* recover the file data */
1280 while (p+8 < data + rec.data_len) {
1286 memcpy(&len, p+4, 4);
1288 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1290 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %u bytes at offset %u\n", len, ofs));
1291 tdb->ecode = TDB_ERR_IO;
1299 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1300 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1301 tdb->ecode = TDB_ERR_IO;
1305 /* if the recovery area is after the recovered eof then remove it */
1306 if (recovery_eof <= recovery_head) {
1307 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1308 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1309 tdb->ecode = TDB_ERR_IO;
1314 /* remove the recovery magic */
1315 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1317 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1318 tdb->ecode = TDB_ERR_IO;
1322 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1323 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1324 tdb->ecode = TDB_ERR_IO;
1328 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %u byte database\n",
1335 /* Any I/O failures we say "needs recovery". */
1336 bool tdb_needs_recovery(struct tdb_context *tdb)
1338 tdb_off_t recovery_head;
1339 struct tdb_record rec;
1341 /* find the recovery area */
1342 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1346 if (recovery_head == 0) {
1347 /* we have never allocated a recovery record */
1351 /* read the recovery record */
1352 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1353 sizeof(rec), DOCONV()) == -1) {
1357 return (rec.magic == TDB_RECOVERY_MAGIC);