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 "tdb1_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 tdb1_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 tdb1_write() calls. The hooked
48 transaction versions of tdb1_read() and tdb1_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 tdb1_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 tdb1_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 TDB1_NOSYNC is passed to flags in tdb1_open then transactions are
85 still available, but no transaction recovery area is used and no
86 fsync/msync calls are made.
88 - if TDB1_ALLOW_NESTING is passed to flags in tdb open, or added using
89 tdb1_add_flags() transaction nesting is enabled.
90 It resets the TDB1_DISALLOW_NESTING flag, as both cannot be used together.
91 The default is that transaction nesting is allowed.
92 Note: this default may change in future versions of tdb.
94 Beware. when transactions are nested a transaction successfully
95 completed with tdb1_transaction_commit() can be silently unrolled later.
97 - if TDB1_DISALLOW_NESTING is passed to flags in tdb open, or added using
98 tdb1_add_flags() transaction nesting is disabled.
99 It resets the TDB1_ALLOW_NESTING flag, as both cannot be used together.
100 An attempt create a nested transaction will fail with TDB1_ERR_NESTING.
101 The default is that transaction nesting is allowed.
102 Note: this default may change in future versions of tdb.
107 hold the context of any current transaction
109 struct tdb1_transaction {
110 /* we keep a mirrored copy of the tdb hash heads here so
111 tdb1_next_hash_chain() can operate efficiently */
112 uint32_t *hash_heads;
114 /* the original io methods - used to do IOs to the real db */
115 const struct tdb1_methods *io_methods;
117 /* the list of transaction blocks. When a block is first
118 written to, it gets created in this list */
121 uint32_t block_size; /* bytes in each block */
122 uint32_t last_block_size; /* number of valid bytes in the last block */
124 /* non-zero when an internal transaction error has
125 occurred. All write operations will then fail until the
126 transaction is ended */
127 int transaction_error;
129 /* when inside a transaction we need to keep track of any
130 nested tdb1_transaction_start() calls, as these are allowed,
131 but don't create a new transaction */
134 /* set when a prepare has already occurred */
136 tdb1_off_t magic_offset;
138 /* old file size before transaction */
139 tdb1_len_t old_map_size;
141 /* did we expand in this transaction */
147 read while in a transaction. We need to check first if the data is in our list
148 of transaction elements, then if not do a real read
150 static int transaction1_read(struct tdb1_context *tdb, tdb1_off_t off, void *buf,
151 tdb1_len_t len, int cv)
155 /* break it down into block sized ops */
156 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
157 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
158 if (transaction1_read(tdb, off, buf, len2, cv) != 0) {
163 buf = (void *)(len2 + (char *)buf);
170 blk = off / tdb->transaction->block_size;
172 /* see if we have it in the block list */
173 if (tdb->transaction->num_blocks <= blk ||
174 tdb->transaction->blocks[blk] == NULL) {
175 /* nope, do a real read */
176 if (tdb->transaction->io_methods->tdb1_read(tdb, off, buf, len, cv) != 0) {
182 /* it is in the block list. Now check for the last block */
183 if (blk == tdb->transaction->num_blocks-1) {
184 if (len > tdb->transaction->last_block_size) {
189 /* now copy it out of this block */
190 memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
192 tdb1_convert(buf, len);
197 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
198 tdb->ecode = TDB1_ERR_IO;
199 tdb->transaction->transaction_error = 1;
205 write while in a transaction
207 static int transaction1_write(struct tdb1_context *tdb, tdb1_off_t off,
208 const void *buf, tdb1_len_t len)
212 /* Only a commit is allowed on a prepared transaction */
213 if (tdb->transaction->prepared) {
214 tdb->ecode = TDB1_ERR_EINVAL;
215 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
216 tdb->transaction->transaction_error = 1;
220 /* if the write is to a hash head, then update the transaction
222 if (len == sizeof(tdb1_off_t) && off >= TDB1_FREELIST_TOP &&
223 off < TDB1_FREELIST_TOP+TDB1_HASHTABLE_SIZE(tdb)) {
224 uint32_t chain = (off-TDB1_FREELIST_TOP) / sizeof(tdb1_off_t);
225 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
228 /* break it up into block sized chunks */
229 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
230 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
231 if (transaction1_write(tdb, off, buf, len2) != 0) {
237 buf = (const void *)(len2 + (const char *)buf);
245 blk = off / tdb->transaction->block_size;
246 off = off % tdb->transaction->block_size;
248 if (tdb->transaction->num_blocks <= blk) {
249 uint8_t **new_blocks;
250 /* expand the blocks array */
251 if (tdb->transaction->blocks == NULL) {
252 new_blocks = (uint8_t **)malloc(
253 (blk+1)*sizeof(uint8_t *));
255 new_blocks = (uint8_t **)realloc(
256 tdb->transaction->blocks,
257 (blk+1)*sizeof(uint8_t *));
259 if (new_blocks == NULL) {
260 tdb->ecode = TDB1_ERR_OOM;
263 memset(&new_blocks[tdb->transaction->num_blocks], 0,
264 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
265 tdb->transaction->blocks = new_blocks;
266 tdb->transaction->num_blocks = blk+1;
267 tdb->transaction->last_block_size = 0;
270 /* allocate and fill a block? */
271 if (tdb->transaction->blocks[blk] == NULL) {
272 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
273 if (tdb->transaction->blocks[blk] == NULL) {
274 tdb->ecode = TDB1_ERR_OOM;
275 tdb->transaction->transaction_error = 1;
278 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
279 tdb1_len_t len2 = tdb->transaction->block_size;
280 if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
281 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
283 if (tdb->transaction->io_methods->tdb1_read(tdb, blk * tdb->transaction->block_size,
284 tdb->transaction->blocks[blk],
286 SAFE_FREE(tdb->transaction->blocks[blk]);
287 tdb->ecode = TDB1_ERR_IO;
290 if (blk == tdb->transaction->num_blocks-1) {
291 tdb->transaction->last_block_size = len2;
296 /* overwrite part of an existing block */
298 memset(tdb->transaction->blocks[blk] + off, 0, len);
300 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
302 if (blk == tdb->transaction->num_blocks-1) {
303 if (len + off > tdb->transaction->last_block_size) {
304 tdb->transaction->last_block_size = len + off;
311 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
312 (blk*tdb->transaction->block_size) + off, len));
313 tdb->transaction->transaction_error = 1;
319 write while in a transaction - this varient never expands the transaction blocks, it only
320 updates existing blocks. This means it cannot change the recovery size
322 static int transaction1_write_existing(struct tdb1_context *tdb, tdb1_off_t off,
323 const void *buf, tdb1_len_t len)
327 /* break it up into block sized chunks */
328 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
329 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
330 if (transaction1_write_existing(tdb, off, buf, len2) != 0) {
336 buf = (const void *)(len2 + (const char *)buf);
344 blk = off / tdb->transaction->block_size;
345 off = off % tdb->transaction->block_size;
347 if (tdb->transaction->num_blocks <= blk ||
348 tdb->transaction->blocks[blk] == NULL) {
352 if (blk == tdb->transaction->num_blocks-1 &&
353 off + len > tdb->transaction->last_block_size) {
354 if (off >= tdb->transaction->last_block_size) {
357 len = tdb->transaction->last_block_size - off;
360 /* overwrite part of an existing block */
361 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
368 accelerated hash chain head search, using the cached hash heads
370 static void transaction1_next_hash_chain(struct tdb1_context *tdb, uint32_t *chain)
373 for (;h < tdb->header.hash_size;h++) {
374 /* the +1 takes account of the freelist */
375 if (0 != tdb->transaction->hash_heads[h+1]) {
383 out of bounds check during a transaction
385 static int transaction1_oob(struct tdb1_context *tdb, tdb1_off_t len, int probe)
387 if (len <= tdb->map_size) {
390 tdb->ecode = TDB1_ERR_IO;
395 transaction version of tdb1_expand().
397 static int transaction1_expand_file(struct tdb1_context *tdb, tdb1_off_t size,
400 /* add a write to the transaction elements, so subsequent
401 reads see the zero data */
402 if (transaction1_write(tdb, size, NULL, addition) != 0) {
406 tdb->transaction->expanded = true;
411 static const struct tdb1_methods transaction1_methods = {
414 transaction1_next_hash_chain,
416 transaction1_expand_file,
421 start a tdb transaction. No token is returned, as only a single
422 transaction is allowed to be pending per tdb1_context
424 static int _tdb1_transaction_start(struct tdb1_context *tdb,
425 enum tdb1_lock_flags lockflags)
427 /* some sanity checks */
428 if (tdb->read_only || (tdb->flags & TDB1_INTERNAL) || tdb->traverse_read) {
429 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction on a read-only or internal db\n"));
430 tdb->ecode = TDB1_ERR_EINVAL;
434 /* cope with nested tdb1_transaction_start() calls */
435 if (tdb->transaction != NULL) {
436 if (!(tdb->flags & TDB1_ALLOW_NESTING)) {
437 tdb->ecode = TDB1_ERR_NESTING;
440 tdb->transaction->nesting++;
441 TDB1_LOG((tdb, TDB1_DEBUG_TRACE, "tdb1_transaction_start: nesting %d\n",
442 tdb->transaction->nesting));
446 if (tdb1_have_extra_locks(tdb)) {
447 /* the caller must not have any locks when starting a
448 transaction as otherwise we'll be screwed by lack
449 of nested locks in posix */
450 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction with locks held\n"));
451 tdb->ecode = TDB1_ERR_LOCK;
455 if (tdb->travlocks.next != NULL) {
456 /* you cannot use transactions inside a traverse (although you can use
457 traverse inside a transaction) as otherwise you can end up with
459 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction within a traverse\n"));
460 tdb->ecode = TDB1_ERR_LOCK;
464 tdb->transaction = (struct tdb1_transaction *)
465 calloc(sizeof(struct tdb1_transaction), 1);
466 if (tdb->transaction == NULL) {
467 tdb->ecode = TDB1_ERR_OOM;
471 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
472 tdb->transaction->block_size = tdb->page_size;
474 /* get the transaction write lock. This is a blocking lock. As
475 discussed with Volker, there are a number of ways we could
476 make this async, which we will probably do in the future */
477 if (tdb1_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
478 SAFE_FREE(tdb->transaction->blocks);
479 SAFE_FREE(tdb->transaction);
480 if ((lockflags & TDB1_LOCK_WAIT) == 0) {
481 tdb->ecode = TDB1_ERR_NOLOCK;
486 /* get a read lock from the freelist to the end of file. This
487 is upgraded to a write lock during the commit */
488 if (tdb1_allrecord_lock(tdb, F_RDLCK, TDB1_LOCK_WAIT, true) == -1) {
489 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: failed to get hash locks\n"));
490 goto fail_allrecord_lock;
493 /* setup a copy of the hash table heads so the hash scan in
494 traverse can be fast */
495 tdb->transaction->hash_heads = (uint32_t *)
496 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
497 if (tdb->transaction->hash_heads == NULL) {
498 tdb->ecode = TDB1_ERR_OOM;
501 if (tdb->methods->tdb1_read(tdb, TDB1_FREELIST_TOP, tdb->transaction->hash_heads,
502 TDB1_HASHTABLE_SIZE(tdb), 0) != 0) {
503 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_start: failed to read hash heads\n"));
504 tdb->ecode = TDB1_ERR_IO;
508 /* make sure we know about any file expansions already done by
510 tdb->methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
511 tdb->transaction->old_map_size = tdb->map_size;
513 /* finally hook the io methods, replacing them with
514 transaction specific methods */
515 tdb->transaction->io_methods = tdb->methods;
516 tdb->methods = &transaction1_methods;
521 tdb1_allrecord_unlock(tdb, F_RDLCK, false);
523 tdb1_transaction_unlock(tdb, F_WRLCK);
524 SAFE_FREE(tdb->transaction->blocks);
525 SAFE_FREE(tdb->transaction->hash_heads);
526 SAFE_FREE(tdb->transaction);
530 _PUBLIC_ int tdb1_transaction_start(struct tdb1_context *tdb)
532 return _tdb1_transaction_start(tdb, TDB1_LOCK_WAIT);
535 _PUBLIC_ int tdb1_transaction_start_nonblock(struct tdb1_context *tdb)
537 return _tdb1_transaction_start(tdb, TDB1_LOCK_NOWAIT|TDB1_LOCK_PROBE);
543 static int transaction1_sync(struct tdb1_context *tdb, tdb1_off_t offset, tdb1_len_t length)
545 if (tdb->flags & TDB1_NOSYNC) {
550 if (fdatasync(tdb->fd) != 0) {
552 if (fsync(tdb->fd) != 0) {
554 tdb->ecode = TDB1_ERR_IO;
555 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction: fsync failed\n"));
560 tdb1_off_t moffset = offset & ~(tdb->page_size-1);
561 if (msync(moffset + (char *)tdb->map_ptr,
562 length + (offset - moffset), MS_SYNC) != 0) {
563 tdb->ecode = TDB1_ERR_IO;
564 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction: msync failed - %s\n",
574 static int _tdb1_transaction_cancel(struct tdb1_context *tdb)
578 if (tdb->transaction == NULL) {
579 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_cancel: no transaction\n"));
583 if (tdb->transaction->nesting != 0) {
584 tdb->transaction->transaction_error = 1;
585 tdb->transaction->nesting--;
589 tdb->map_size = tdb->transaction->old_map_size;
591 /* free all the transaction blocks */
592 for (i=0;i<tdb->transaction->num_blocks;i++) {
593 if (tdb->transaction->blocks[i] != NULL) {
594 free(tdb->transaction->blocks[i]);
597 SAFE_FREE(tdb->transaction->blocks);
599 if (tdb->transaction->magic_offset) {
600 const struct tdb1_methods *methods = tdb->transaction->io_methods;
601 const uint32_t invalid = TDB1_RECOVERY_INVALID_MAGIC;
603 /* remove the recovery marker */
604 if (methods->tdb1_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
605 transaction1_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
606 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_cancel: failed to remove recovery magic\n"));
611 /* This also removes the OPEN_LOCK, if we have it. */
612 tdb1_release_transaction_locks(tdb);
614 /* restore the normal io methods */
615 tdb->methods = tdb->transaction->io_methods;
617 SAFE_FREE(tdb->transaction->hash_heads);
618 SAFE_FREE(tdb->transaction);
624 cancel the current transaction
626 _PUBLIC_ int tdb1_transaction_cancel(struct tdb1_context *tdb)
628 return _tdb1_transaction_cancel(tdb);
632 work out how much space the linearised recovery data will consume
634 static tdb1_len_t tdb1_recovery_size(struct tdb1_context *tdb)
636 tdb1_len_t recovery_size = 0;
639 recovery_size = sizeof(uint32_t);
640 for (i=0;i<tdb->transaction->num_blocks;i++) {
641 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
644 if (tdb->transaction->blocks[i] == NULL) {
647 recovery_size += 2*sizeof(tdb1_off_t);
648 if (i == tdb->transaction->num_blocks-1) {
649 recovery_size += tdb->transaction->last_block_size;
651 recovery_size += tdb->transaction->block_size;
655 return recovery_size;
658 int tdb1_recovery_area(struct tdb1_context *tdb,
659 const struct tdb1_methods *methods,
660 tdb1_off_t *recovery_offset,
661 struct tdb1_record *rec)
663 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, recovery_offset) == -1) {
667 if (*recovery_offset == 0) {
672 if (methods->tdb1_read(tdb, *recovery_offset, rec, sizeof(*rec),
673 TDB1_DOCONV()) == -1) {
677 /* ignore invalid recovery regions: can happen in crash */
678 if (rec->magic != TDB1_RECOVERY_MAGIC &&
679 rec->magic != TDB1_RECOVERY_INVALID_MAGIC) {
680 *recovery_offset = 0;
687 allocate the recovery area, or use an existing recovery area if it is
690 static int tdb1_recovery_allocate(struct tdb1_context *tdb,
691 tdb1_len_t *recovery_size,
692 tdb1_off_t *recovery_offset,
693 tdb1_len_t *recovery_max_size)
695 struct tdb1_record rec;
696 const struct tdb1_methods *methods = tdb->transaction->io_methods;
697 tdb1_off_t recovery_head;
699 if (tdb1_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
700 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to read recovery head\n"));
704 *recovery_size = tdb1_recovery_size(tdb);
706 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
707 /* it fits in the existing area */
708 *recovery_max_size = rec.rec_len;
709 *recovery_offset = recovery_head;
713 /* we need to free up the old recovery area, then allocate a
714 new one at the end of the file. Note that we cannot use
715 tdb1_allocate() to allocate the new one as that might return
716 us an area that is being currently used (as of the start of
718 if (recovery_head != 0) {
719 if (tdb1_free(tdb, recovery_head, &rec) == -1) {
720 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to free previous recovery area\n"));
725 /* the tdb1_free() call might have increased the recovery size */
726 *recovery_size = tdb1_recovery_size(tdb);
728 /* round up to a multiple of page size */
729 *recovery_max_size = TDB1_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
730 *recovery_offset = tdb->map_size;
731 recovery_head = *recovery_offset;
733 if (methods->tdb1_expand_file(tdb, tdb->transaction->old_map_size,
734 (tdb->map_size - tdb->transaction->old_map_size) +
735 sizeof(rec) + *recovery_max_size) == -1) {
736 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to create recovery area\n"));
740 /* remap the file (if using mmap) */
741 methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
743 /* we have to reset the old map size so that we don't try to expand the file
744 again in the transaction commit, which would destroy the recovery area */
745 tdb->transaction->old_map_size = tdb->map_size;
747 /* write the recovery header offset and sync - we can sync without a race here
748 as the magic ptr in the recovery record has not been set */
749 TDB1_CONV(recovery_head);
750 if (methods->tdb1_write(tdb, TDB1_RECOVERY_HEAD,
751 &recovery_head, sizeof(tdb1_off_t)) == -1) {
752 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to write recovery head\n"));
755 if (transaction1_write_existing(tdb, TDB1_RECOVERY_HEAD, &recovery_head, sizeof(tdb1_off_t)) == -1) {
756 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to write recovery head\n"));
765 setup the recovery data that will be used on a crash during commit
767 static int transaction1_setup_recovery(struct tdb1_context *tdb,
768 tdb1_off_t *magic_offset)
770 tdb1_len_t recovery_size;
771 unsigned char *data, *p;
772 const struct tdb1_methods *methods = tdb->transaction->io_methods;
773 struct tdb1_record *rec;
774 tdb1_off_t recovery_offset, recovery_max_size;
775 tdb1_off_t old_map_size = tdb->transaction->old_map_size;
776 uint32_t magic, tailer;
780 check that the recovery area has enough space
782 if (tdb1_recovery_allocate(tdb, &recovery_size,
783 &recovery_offset, &recovery_max_size) == -1) {
787 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
789 tdb->ecode = TDB1_ERR_OOM;
793 rec = (struct tdb1_record *)data;
794 memset(rec, 0, sizeof(*rec));
796 rec->magic = TDB1_RECOVERY_INVALID_MAGIC;
797 rec->data_len = recovery_size;
798 rec->rec_len = recovery_max_size;
799 rec->key_len = old_map_size;
802 /* build the recovery data into a single blob to allow us to do a single
803 large write, which should be more efficient */
804 p = data + sizeof(*rec);
805 for (i=0;i<tdb->transaction->num_blocks;i++) {
809 if (tdb->transaction->blocks[i] == NULL) {
813 offset = i * tdb->transaction->block_size;
814 length = tdb->transaction->block_size;
815 if (i == tdb->transaction->num_blocks-1) {
816 length = tdb->transaction->last_block_size;
819 if (offset >= old_map_size) {
822 if (offset + length > tdb->transaction->old_map_size) {
823 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: transaction data over new region boundary\n"));
825 tdb->ecode = TDB1_ERR_CORRUPT;
828 memcpy(p, &offset, 4);
829 memcpy(p+4, &length, 4);
833 /* the recovery area contains the old data, not the
834 new data, so we have to call the original tdb1_read
836 if (methods->tdb1_read(tdb, offset, p + 8, length, 0) != 0) {
838 tdb->ecode = TDB1_ERR_IO;
845 tailer = sizeof(*rec) + recovery_max_size;
846 memcpy(p, &tailer, 4);
851 /* write the recovery data to the recovery area */
852 if (methods->tdb1_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
853 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write recovery data\n"));
855 tdb->ecode = TDB1_ERR_IO;
858 if (transaction1_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
859 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write secondary recovery data\n"));
861 tdb->ecode = TDB1_ERR_IO;
865 /* as we don't have ordered writes, we have to sync the recovery
866 data before we update the magic to indicate that the recovery
868 if (transaction1_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
875 magic = TDB1_RECOVERY_MAGIC;
878 *magic_offset = recovery_offset + offsetof(struct tdb1_record, magic);
880 if (methods->tdb1_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
881 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write recovery magic\n"));
882 tdb->ecode = TDB1_ERR_IO;
885 if (transaction1_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
886 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write secondary recovery magic\n"));
887 tdb->ecode = TDB1_ERR_IO;
891 /* ensure the recovery magic marker is on disk */
892 if (transaction1_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
899 static int _tdb1_transaction_prepare_commit(struct tdb1_context *tdb)
901 const struct tdb1_methods *methods;
903 if (tdb->transaction == NULL) {
904 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: no transaction\n"));
908 if (tdb->transaction->prepared) {
909 tdb->ecode = TDB1_ERR_EINVAL;
910 _tdb1_transaction_cancel(tdb);
911 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: transaction already prepared\n"));
915 if (tdb->transaction->transaction_error) {
916 tdb->ecode = TDB1_ERR_IO;
917 _tdb1_transaction_cancel(tdb);
918 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: transaction error pending\n"));
923 if (tdb->transaction->nesting != 0) {
927 /* check for a null transaction */
928 if (tdb->transaction->blocks == NULL) {
932 methods = tdb->transaction->io_methods;
934 /* if there are any locks pending then the caller has not
935 nested their locks properly, so fail the transaction */
936 if (tdb1_have_extra_locks(tdb)) {
937 tdb->ecode = TDB1_ERR_LOCK;
938 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: locks pending on commit\n"));
939 _tdb1_transaction_cancel(tdb);
943 /* upgrade the main transaction lock region to a write lock */
944 if (tdb1_allrecord_upgrade(tdb) == -1) {
945 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: failed to upgrade hash locks\n"));
946 _tdb1_transaction_cancel(tdb);
950 /* get the open lock - this prevents new users attaching to the database
952 if (tdb1_nest_lock(tdb, TDB1_OPEN_LOCK, F_WRLCK, TDB1_LOCK_WAIT) == -1) {
953 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: failed to get open lock\n"));
954 _tdb1_transaction_cancel(tdb);
958 if (!(tdb->flags & TDB1_NOSYNC)) {
959 /* write the recovery data to the end of the file */
960 if (transaction1_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
961 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_prepare_commit: failed to setup recovery data\n"));
962 _tdb1_transaction_cancel(tdb);
967 tdb->transaction->prepared = true;
969 /* expand the file to the new size if needed */
970 if (tdb->map_size != tdb->transaction->old_map_size) {
971 if (methods->tdb1_expand_file(tdb, tdb->transaction->old_map_size,
973 tdb->transaction->old_map_size) == -1) {
974 tdb->ecode = TDB1_ERR_IO;
975 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_prepare_commit: expansion failed\n"));
976 _tdb1_transaction_cancel(tdb);
979 tdb->map_size = tdb->transaction->old_map_size;
980 methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
983 /* Keep the open lock until the actual commit */
989 prepare to commit the current transaction
991 _PUBLIC_ int tdb1_transaction_prepare_commit(struct tdb1_context *tdb)
993 return _tdb1_transaction_prepare_commit(tdb);
996 /* A repack is worthwhile if the largest is less than half total free. */
997 static bool repack_worthwhile(struct tdb1_context *tdb)
1000 struct tdb1_record rec;
1001 tdb1_len_t total = 0, largest = 0;
1003 if (tdb1_ofs_read(tdb, TDB1_FREELIST_TOP, &ptr) == -1) {
1007 while (ptr != 0 && tdb1_rec_free_read(tdb, ptr, &rec) == 0) {
1008 total += rec.rec_len;
1009 if (rec.rec_len > largest) {
1010 largest = rec.rec_len;
1015 return total > largest * 2;
1019 commit the current transaction
1021 _PUBLIC_ int tdb1_transaction_commit(struct tdb1_context *tdb)
1023 const struct tdb1_methods *methods;
1025 bool need_repack = false;
1027 if (tdb->transaction == NULL) {
1028 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_commit: no transaction\n"));
1032 if (tdb->transaction->transaction_error) {
1033 tdb->ecode = TDB1_ERR_IO;
1034 _tdb1_transaction_cancel(tdb);
1035 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_commit: transaction error pending\n"));
1040 if (tdb->transaction->nesting != 0) {
1041 tdb->transaction->nesting--;
1045 /* check for a null transaction */
1046 if (tdb->transaction->blocks == NULL) {
1047 _tdb1_transaction_cancel(tdb);
1051 if (!tdb->transaction->prepared) {
1052 int ret = _tdb1_transaction_prepare_commit(tdb);
1057 methods = tdb->transaction->io_methods;
1059 /* perform all the writes */
1060 for (i=0;i<tdb->transaction->num_blocks;i++) {
1064 if (tdb->transaction->blocks[i] == NULL) {
1068 offset = i * tdb->transaction->block_size;
1069 length = tdb->transaction->block_size;
1070 if (i == tdb->transaction->num_blocks-1) {
1071 length = tdb->transaction->last_block_size;
1074 if (methods->tdb1_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1075 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_commit: write failed during commit\n"));
1077 /* we've overwritten part of the data and
1078 possibly expanded the file, so we need to
1079 run the crash recovery code */
1080 tdb->methods = methods;
1081 tdb1_transaction_recover(tdb);
1083 _tdb1_transaction_cancel(tdb);
1085 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_commit: write failed\n"));
1088 SAFE_FREE(tdb->transaction->blocks[i]);
1091 /* Do this before we drop lock or blocks. */
1092 if (tdb->transaction->expanded) {
1093 need_repack = repack_worthwhile(tdb);
1096 SAFE_FREE(tdb->transaction->blocks);
1097 tdb->transaction->num_blocks = 0;
1099 /* ensure the new data is on disk */
1100 if (transaction1_sync(tdb, 0, tdb->map_size) == -1) {
1105 TODO: maybe write to some dummy hdr field, or write to magic
1106 offset without mmap, before the last sync, instead of the
1110 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1111 don't change the mtime of the file, this means the file may
1112 not be backed up (as tdb rounding to block sizes means that
1113 file size changes are quite rare too). The following forces
1114 mtime changes when a transaction completes */
1116 utime(tdb->name, NULL);
1119 /* use a transaction cancel to free memory and remove the
1120 transaction locks */
1121 _tdb1_transaction_cancel(tdb);
1124 return tdb1_repack(tdb);
1132 recover from an aborted transaction. Must be called with exclusive
1133 database write access already established (including the open
1134 lock to prevent new processes attaching)
1136 int tdb1_transaction_recover(struct tdb1_context *tdb)
1138 tdb1_off_t recovery_head, recovery_eof;
1139 unsigned char *data, *p;
1141 struct tdb1_record rec;
1143 /* find the recovery area */
1144 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1145 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery head\n"));
1146 tdb->ecode = TDB1_ERR_IO;
1150 if (recovery_head == 0) {
1151 /* we have never allocated a recovery record */
1155 /* read the recovery record */
1156 if (tdb->methods->tdb1_read(tdb, recovery_head, &rec,
1157 sizeof(rec), TDB1_DOCONV()) == -1) {
1158 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery record\n"));
1159 tdb->ecode = TDB1_ERR_IO;
1163 if (rec.magic != TDB1_RECOVERY_MAGIC) {
1164 /* there is no valid recovery data */
1168 if (tdb->read_only) {
1169 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: attempt to recover read only database\n"));
1170 tdb->ecode = TDB1_ERR_CORRUPT;
1174 recovery_eof = rec.key_len;
1176 data = (unsigned char *)malloc(rec.data_len);
1178 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to allocate recovery data\n"));
1179 tdb->ecode = TDB1_ERR_OOM;
1183 /* read the full recovery data */
1184 if (tdb->methods->tdb1_read(tdb, recovery_head + sizeof(rec), data,
1185 rec.data_len, 0) == -1) {
1186 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery data\n"));
1187 tdb->ecode = TDB1_ERR_IO;
1191 /* recover the file data */
1193 while (p+8 < data + rec.data_len) {
1195 if (TDB1_DOCONV()) {
1199 memcpy(&len, p+4, 4);
1201 if (tdb->methods->tdb1_write(tdb, ofs, p+8, len) == -1) {
1203 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1204 tdb->ecode = TDB1_ERR_IO;
1212 if (transaction1_sync(tdb, 0, tdb->map_size) == -1) {
1213 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to sync recovery\n"));
1214 tdb->ecode = TDB1_ERR_IO;
1218 /* if the recovery area is after the recovered eof then remove it */
1219 if (recovery_eof <= recovery_head) {
1220 if (tdb1_ofs_write(tdb, TDB1_RECOVERY_HEAD, &zero) == -1) {
1221 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to remove recovery head\n"));
1222 tdb->ecode = TDB1_ERR_IO;
1227 /* remove the recovery magic */
1228 if (tdb1_ofs_write(tdb, recovery_head + offsetof(struct tdb1_record, magic),
1230 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to remove recovery magic\n"));
1231 tdb->ecode = TDB1_ERR_IO;
1235 if (transaction1_sync(tdb, 0, recovery_eof) == -1) {
1236 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to sync2 recovery\n"));
1237 tdb->ecode = TDB1_ERR_IO;
1241 TDB1_LOG((tdb, TDB1_DEBUG_TRACE, "tdb1_transaction_recover: recovered %d byte database\n",
1248 /* Any I/O failures we say "needs recovery". */
1249 bool tdb1_needs_recovery(struct tdb1_context *tdb)
1251 tdb1_off_t recovery_head;
1252 struct tdb1_record rec;
1254 /* find the recovery area */
1255 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1259 if (recovery_head == 0) {
1260 /* we have never allocated a recovery record */
1264 /* read the recovery record */
1265 if (tdb->methods->tdb1_read(tdb, recovery_head, &rec,
1266 sizeof(rec), TDB1_DOCONV()) == -1) {
1270 return (rec.magic == TDB1_RECOVERY_MAGIC);