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)
597 if (tdb->transaction == NULL) {
598 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
602 if (tdb->transaction->nesting != 0) {
603 tdb->transaction->transaction_error = 1;
604 tdb->transaction->nesting--;
608 tdb->map_size = tdb->transaction->old_map_size;
610 /* free all the transaction blocks */
611 for (i=0;i<tdb->transaction->num_blocks;i++) {
612 if (tdb->transaction->blocks[i] != NULL) {
613 free(tdb->transaction->blocks[i]);
616 SAFE_FREE(tdb->transaction->blocks);
618 if (tdb->transaction->magic_offset) {
619 const struct tdb_methods *methods = tdb->transaction->io_methods;
620 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
622 /* remove the recovery marker */
623 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
624 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
625 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
630 /* This also removes the OPEN_LOCK, if we have it. */
631 tdb_release_transaction_locks(tdb);
633 /* restore the normal io methods */
634 tdb->methods = tdb->transaction->io_methods;
636 SAFE_FREE(tdb->transaction->hash_heads);
637 SAFE_FREE(tdb->transaction);
643 cancel the current transaction
645 _PUBLIC_ int tdb_transaction_cancel(struct tdb_context *tdb)
647 tdb_trace(tdb, "tdb_transaction_cancel");
648 return _tdb_transaction_cancel(tdb);
652 work out how much space the linearised recovery data will consume
654 static bool tdb_recovery_size(struct tdb_context *tdb, tdb_len_t *result)
656 tdb_len_t recovery_size = 0;
659 recovery_size = sizeof(uint32_t);
660 for (i=0;i<tdb->transaction->num_blocks;i++) {
661 tdb_len_t block_size;
662 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
665 if (tdb->transaction->blocks[i] == NULL) {
668 if (!tdb_add_len_t(recovery_size, 2*sizeof(tdb_off_t),
672 if (i == tdb->transaction->num_blocks-1) {
673 block_size = tdb->transaction->last_block_size;
675 block_size = tdb->transaction->block_size;
677 if (!tdb_add_len_t(recovery_size, block_size,
683 *result = recovery_size;
687 int tdb_recovery_area(struct tdb_context *tdb,
688 const struct tdb_methods *methods,
689 tdb_off_t *recovery_offset,
690 struct tdb_record *rec)
692 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, recovery_offset) == -1) {
696 if (*recovery_offset == 0) {
701 if (methods->tdb_read(tdb, *recovery_offset, rec, sizeof(*rec),
706 /* ignore invalid recovery regions: can happen in crash */
707 if (rec->magic != TDB_RECOVERY_MAGIC &&
708 rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
709 *recovery_offset = 0;
716 allocate the recovery area, or use an existing recovery area if it is
719 static int tdb_recovery_allocate(struct tdb_context *tdb,
720 tdb_len_t *recovery_size,
721 tdb_off_t *recovery_offset,
722 tdb_len_t *recovery_max_size)
724 struct tdb_record rec;
725 const struct tdb_methods *methods = tdb->transaction->io_methods;
726 tdb_off_t recovery_head, new_end;
728 if (tdb_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
729 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
733 if (!tdb_recovery_size(tdb, recovery_size)) {
734 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
735 "overflow recovery size\n"));
739 /* Existing recovery area? */
740 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
741 /* it fits in the existing area */
742 *recovery_max_size = rec.rec_len;
743 *recovery_offset = recovery_head;
747 /* If recovery area in middle of file, we need a new one. */
748 if (recovery_head == 0
749 || recovery_head + sizeof(rec) + rec.rec_len != tdb->map_size) {
750 /* we need to free up the old recovery area, then allocate a
751 new one at the end of the file. Note that we cannot use
752 tdb_allocate() to allocate the new one as that might return
753 us an area that is being currently used (as of the start of
756 if (tdb_free(tdb, recovery_head, &rec) == -1) {
757 TDB_LOG((tdb, TDB_DEBUG_FATAL,
758 "tdb_recovery_allocate: failed to"
759 " free previous recovery area\n"));
763 /* the tdb_free() call might have increased
764 * the recovery size */
765 if (!tdb_recovery_size(tdb, recovery_size)) {
766 TDB_LOG((tdb, TDB_DEBUG_FATAL,
767 "tdb_recovery_allocate: "
768 "overflow recovery size\n"));
773 /* New head will be at end of file. */
774 recovery_head = tdb->map_size;
777 /* Now we know where it will be. */
778 *recovery_offset = recovery_head;
780 /* Expand by more than we need, so we don't do it often. */
781 *recovery_max_size = tdb_expand_adjust(tdb->map_size,
786 if (!tdb_add_off_t(recovery_head, sizeof(rec), &new_end) ||
787 !tdb_add_off_t(new_end, *recovery_max_size, &new_end)) {
788 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
789 "overflow recovery area\n"));
793 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
794 new_end - tdb->transaction->old_map_size)
796 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
800 /* remap the file (if using mmap) */
801 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
803 /* we have to reset the old map size so that we don't try to expand the file
804 again in the transaction commit, which would destroy the recovery area */
805 tdb->transaction->old_map_size = tdb->map_size;
807 /* write the recovery header offset and sync - we can sync without a race here
808 as the magic ptr in the recovery record has not been set */
809 CONVERT(recovery_head);
810 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
811 &recovery_head, sizeof(tdb_off_t)) == -1) {
812 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
815 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
816 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
825 setup the recovery data that will be used on a crash during commit
827 static int transaction_setup_recovery(struct tdb_context *tdb,
828 tdb_off_t *magic_offset)
830 tdb_len_t recovery_size;
831 unsigned char *data, *p;
832 const struct tdb_methods *methods = tdb->transaction->io_methods;
833 struct tdb_record *rec;
834 tdb_off_t recovery_offset, recovery_max_size;
835 tdb_off_t old_map_size = tdb->transaction->old_map_size;
836 uint32_t magic, tailer;
840 check that the recovery area has enough space
842 if (tdb_recovery_allocate(tdb, &recovery_size,
843 &recovery_offset, &recovery_max_size) == -1) {
847 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
849 tdb->ecode = TDB_ERR_OOM;
853 rec = (struct tdb_record *)data;
854 memset(rec, 0, sizeof(*rec));
856 rec->magic = TDB_RECOVERY_INVALID_MAGIC;
857 rec->data_len = recovery_size;
858 rec->rec_len = recovery_max_size;
859 rec->key_len = old_map_size;
862 /* build the recovery data into a single blob to allow us to do a single
863 large write, which should be more efficient */
864 p = data + sizeof(*rec);
865 for (i=0;i<tdb->transaction->num_blocks;i++) {
869 if (tdb->transaction->blocks[i] == NULL) {
873 offset = i * tdb->transaction->block_size;
874 length = tdb->transaction->block_size;
875 if (i == tdb->transaction->num_blocks-1) {
876 length = tdb->transaction->last_block_size;
879 if (offset >= old_map_size) {
882 if (offset + length > tdb->transaction->old_map_size) {
883 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
885 tdb->ecode = TDB_ERR_CORRUPT;
888 memcpy(p, &offset, 4);
889 memcpy(p+4, &length, 4);
893 /* the recovery area contains the old data, not the
894 new data, so we have to call the original tdb_read
896 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
898 tdb->ecode = TDB_ERR_IO;
905 tailer = sizeof(*rec) + recovery_max_size;
906 memcpy(p, &tailer, 4);
911 /* write the recovery data to the recovery area */
912 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
913 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
915 tdb->ecode = TDB_ERR_IO;
918 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
919 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
921 tdb->ecode = TDB_ERR_IO;
925 /* as we don't have ordered writes, we have to sync the recovery
926 data before we update the magic to indicate that the recovery
928 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
935 magic = TDB_RECOVERY_MAGIC;
938 *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
940 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
941 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
942 tdb->ecode = TDB_ERR_IO;
945 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
946 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
947 tdb->ecode = TDB_ERR_IO;
951 /* ensure the recovery magic marker is on disk */
952 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
959 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
961 const struct tdb_methods *methods;
963 if (tdb->transaction == NULL) {
964 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
968 if (tdb->transaction->prepared) {
969 tdb->ecode = TDB_ERR_EINVAL;
970 _tdb_transaction_cancel(tdb);
971 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
975 if (tdb->transaction->transaction_error) {
976 tdb->ecode = TDB_ERR_IO;
977 _tdb_transaction_cancel(tdb);
978 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
983 if (tdb->transaction->nesting != 0) {
987 /* check for a null transaction */
988 if (tdb->transaction->blocks == NULL) {
992 methods = tdb->transaction->io_methods;
994 /* if there are any locks pending then the caller has not
995 nested their locks properly, so fail the transaction */
996 if (tdb_have_extra_locks(tdb)) {
997 tdb->ecode = TDB_ERR_LOCK;
998 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
999 _tdb_transaction_cancel(tdb);
1003 /* upgrade the main transaction lock region to a write lock */
1004 if (tdb_allrecord_upgrade(tdb) == -1) {
1005 if (tdb->ecode == TDB_ERR_RDONLY && tdb->read_only) {
1006 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1007 "tdb_transaction_prepare_commit: "
1008 "failed to upgrade hash locks: "
1009 "database is read only\n"));
1010 } else if (tdb->ecode == TDB_ERR_RDONLY
1011 && tdb->traverse_read) {
1012 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1013 "tdb_transaction_prepare_commit: "
1014 "failed to upgrade hash locks: "
1015 "a database traverse is in progress\n"));
1017 TDB_LOG((tdb, TDB_DEBUG_ERROR,
1018 "tdb_transaction_prepare_commit: "
1019 "failed to upgrade hash locks: %s\n",
1020 tdb_errorstr(tdb)));
1022 _tdb_transaction_cancel(tdb);
1026 /* get the open lock - this prevents new users attaching to the database
1027 during the commit */
1028 if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
1029 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
1030 _tdb_transaction_cancel(tdb);
1034 /* write the recovery data to the end of the file */
1035 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
1036 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
1037 _tdb_transaction_cancel(tdb);
1041 tdb->transaction->prepared = true;
1043 /* expand the file to the new size if needed */
1044 if (tdb->map_size != tdb->transaction->old_map_size) {
1045 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
1047 tdb->transaction->old_map_size) == -1) {
1048 tdb->ecode = TDB_ERR_IO;
1049 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
1050 _tdb_transaction_cancel(tdb);
1053 tdb->map_size = tdb->transaction->old_map_size;
1054 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
1057 /* Keep the open lock until the actual commit */
1063 prepare to commit the current transaction
1065 _PUBLIC_ int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1067 tdb_trace(tdb, "tdb_transaction_prepare_commit");
1068 return _tdb_transaction_prepare_commit(tdb);
1071 /* A repack is worthwhile if the largest is less than half total free. */
1072 static bool repack_worthwhile(struct tdb_context *tdb)
1075 struct tdb_record rec;
1076 tdb_len_t total = 0, largest = 0;
1078 if (tdb_ofs_read(tdb, FREELIST_TOP, &ptr) == -1) {
1082 while (ptr != 0 && tdb_rec_free_read(tdb, ptr, &rec) == 0) {
1083 total += rec.rec_len;
1084 if (rec.rec_len > largest) {
1085 largest = rec.rec_len;
1090 return total > largest * 2;
1094 commit the current transaction
1096 _PUBLIC_ int tdb_transaction_commit(struct tdb_context *tdb)
1098 const struct tdb_methods *methods;
1100 bool need_repack = false;
1102 if (tdb->transaction == NULL) {
1103 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1107 tdb_trace(tdb, "tdb_transaction_commit");
1109 if (tdb->transaction->transaction_error) {
1110 tdb->ecode = TDB_ERR_IO;
1111 _tdb_transaction_cancel(tdb);
1112 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1117 if (tdb->transaction->nesting != 0) {
1118 tdb->transaction->nesting--;
1122 /* check for a null transaction */
1123 if (tdb->transaction->blocks == NULL) {
1124 _tdb_transaction_cancel(tdb);
1128 if (!tdb->transaction->prepared) {
1129 int ret = _tdb_transaction_prepare_commit(tdb);
1134 methods = tdb->transaction->io_methods;
1136 /* perform all the writes */
1137 for (i=0;i<tdb->transaction->num_blocks;i++) {
1141 if (tdb->transaction->blocks[i] == NULL) {
1145 offset = i * tdb->transaction->block_size;
1146 length = tdb->transaction->block_size;
1147 if (i == tdb->transaction->num_blocks-1) {
1148 length = tdb->transaction->last_block_size;
1151 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1152 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1154 /* we've overwritten part of the data and
1155 possibly expanded the file, so we need to
1156 run the crash recovery code */
1157 tdb->methods = methods;
1158 tdb_transaction_recover(tdb);
1160 _tdb_transaction_cancel(tdb);
1162 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1165 SAFE_FREE(tdb->transaction->blocks[i]);
1168 /* Do this before we drop lock or blocks. */
1169 if (tdb->transaction->expanded) {
1170 need_repack = repack_worthwhile(tdb);
1173 SAFE_FREE(tdb->transaction->blocks);
1174 tdb->transaction->num_blocks = 0;
1176 /* ensure the new data is on disk */
1177 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1182 TODO: maybe write to some dummy hdr field, or write to magic
1183 offset without mmap, before the last sync, instead of the
1187 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1188 don't change the mtime of the file, this means the file may
1189 not be backed up (as tdb rounding to block sizes means that
1190 file size changes are quite rare too). The following forces
1191 mtime changes when a transaction completes */
1193 utime(tdb->name, NULL);
1196 /* use a transaction cancel to free memory and remove the
1197 transaction locks */
1198 _tdb_transaction_cancel(tdb);
1201 return tdb_repack(tdb);
1209 recover from an aborted transaction. Must be called with exclusive
1210 database write access already established (including the open
1211 lock to prevent new processes attaching)
1213 int tdb_transaction_recover(struct tdb_context *tdb)
1215 tdb_off_t recovery_head, recovery_eof;
1216 unsigned char *data, *p;
1218 struct tdb_record rec;
1220 /* find the recovery area */
1221 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1222 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1223 tdb->ecode = TDB_ERR_IO;
1227 if (recovery_head == 0) {
1228 /* we have never allocated a recovery record */
1232 /* read the recovery record */
1233 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1234 sizeof(rec), DOCONV()) == -1) {
1235 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1236 tdb->ecode = TDB_ERR_IO;
1240 if (rec.magic != TDB_RECOVERY_MAGIC) {
1241 /* there is no valid recovery data */
1245 if (tdb->read_only) {
1246 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1247 tdb->ecode = TDB_ERR_CORRUPT;
1251 recovery_eof = rec.key_len;
1253 data = (unsigned char *)malloc(rec.data_len);
1255 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1256 tdb->ecode = TDB_ERR_OOM;
1260 /* read the full recovery data */
1261 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1262 rec.data_len, 0) == -1) {
1263 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1264 tdb->ecode = TDB_ERR_IO;
1268 /* recover the file data */
1270 while (p+8 < data + rec.data_len) {
1276 memcpy(&len, p+4, 4);
1278 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1280 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %u bytes at offset %u\n", len, ofs));
1281 tdb->ecode = TDB_ERR_IO;
1289 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1290 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1291 tdb->ecode = TDB_ERR_IO;
1295 /* if the recovery area is after the recovered eof then remove it */
1296 if (recovery_eof <= recovery_head) {
1297 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1298 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1299 tdb->ecode = TDB_ERR_IO;
1304 /* remove the recovery magic */
1305 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1307 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1308 tdb->ecode = TDB_ERR_IO;
1312 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1313 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1314 tdb->ecode = TDB_ERR_IO;
1318 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %u byte database\n",
1325 /* Any I/O failures we say "needs recovery". */
1326 bool tdb_needs_recovery(struct tdb_context *tdb)
1328 tdb_off_t recovery_head;
1329 struct tdb_record rec;
1331 /* find the recovery area */
1332 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1336 if (recovery_head == 0) {
1337 /* we have never allocated a recovery record */
1341 /* read the recovery record */
1342 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1343 sizeof(rec), DOCONV()) == -1) {
1347 return (rec.magic == TDB_RECOVERY_MAGIC);