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 TDB_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 TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89 tdb1_add_flags() transaction nesting is enabled.
90 The default is that transaction nesting is NOT allowed.
92 Beware. when transactions are nested a transaction successfully
93 completed with tdb1_transaction_commit() can be silently unrolled later.
98 hold the context of any current transaction
100 struct tdb1_transaction {
101 /* we keep a mirrored copy of the tdb hash heads here so
102 tdb1_next_hash_chain() can operate efficiently */
103 uint32_t *hash_heads;
105 /* the original io methods - used to do IOs to the real db */
106 const struct tdb1_methods *io_methods;
108 /* the list of transaction blocks. When a block is first
109 written to, it gets created in this list */
112 uint32_t block_size; /* bytes in each block */
113 uint32_t last_block_size; /* number of valid bytes in the last block */
115 /* non-zero when an internal transaction error has
116 occurred. All write operations will then fail until the
117 transaction is ended */
118 int transaction_error;
120 /* when inside a transaction we need to keep track of any
121 nested tdb1_transaction_start() calls, as these are allowed,
122 but don't create a new transaction */
125 /* set when a prepare has already occurred */
127 tdb1_off_t magic_offset;
129 /* old file size before transaction */
130 tdb1_len_t old_map_size;
132 /* did we expand in this transaction */
138 read while in a transaction. We need to check first if the data is in our list
139 of transaction elements, then if not do a real read
141 static int transaction1_read(struct tdb_context *tdb, tdb1_off_t off, void *buf,
142 tdb1_len_t len, int cv)
146 /* break it down into block sized ops */
147 while (len + (off % tdb->tdb1.transaction->block_size) > tdb->tdb1.transaction->block_size) {
148 tdb1_len_t len2 = tdb->tdb1.transaction->block_size - (off % tdb->tdb1.transaction->block_size);
149 if (transaction1_read(tdb, off, buf, len2, cv) != 0) {
154 buf = (void *)(len2 + (char *)buf);
161 blk = off / tdb->tdb1.transaction->block_size;
163 /* see if we have it in the block list */
164 if (tdb->tdb1.transaction->num_blocks <= blk ||
165 tdb->tdb1.transaction->blocks[blk] == NULL) {
166 /* nope, do a real read */
167 if (tdb->tdb1.transaction->io_methods->tdb1_read(tdb, off, buf, len, cv) != 0) {
173 /* it is in the block list. Now check for the last block */
174 if (blk == tdb->tdb1.transaction->num_blocks-1) {
175 if (len > tdb->tdb1.transaction->last_block_size) {
180 /* now copy it out of this block */
181 memcpy(buf, tdb->tdb1.transaction->blocks[blk] + (off % tdb->tdb1.transaction->block_size), len);
183 tdb1_convert(buf, len);
188 tdb->last_error = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
189 "transaction_read: failed at off=%d len=%d",
191 tdb->tdb1.transaction->transaction_error = 1;
197 write while in a transaction
199 static int transaction1_write(struct tdb_context *tdb, tdb1_off_t off,
200 const void *buf, tdb1_len_t len)
204 /* Only a commit is allowed on a prepared transaction */
205 if (tdb->tdb1.transaction->prepared) {
206 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
207 "transaction_write: transaction already"
208 " prepared, write not allowed");
209 tdb->tdb1.transaction->transaction_error = 1;
213 /* if the write is to a hash head, then update the transaction
215 if (len == sizeof(tdb1_off_t) && off >= TDB1_FREELIST_TOP &&
216 off < TDB1_FREELIST_TOP+TDB1_HASHTABLE_SIZE(tdb)) {
217 uint32_t chain = (off-TDB1_FREELIST_TOP) / sizeof(tdb1_off_t);
218 memcpy(&tdb->tdb1.transaction->hash_heads[chain], buf, len);
221 /* break it up into block sized chunks */
222 while (len + (off % tdb->tdb1.transaction->block_size) > tdb->tdb1.transaction->block_size) {
223 tdb1_len_t len2 = tdb->tdb1.transaction->block_size - (off % tdb->tdb1.transaction->block_size);
224 if (transaction1_write(tdb, off, buf, len2) != 0) {
230 buf = (const void *)(len2 + (const char *)buf);
238 blk = off / tdb->tdb1.transaction->block_size;
239 off = off % tdb->tdb1.transaction->block_size;
241 if (tdb->tdb1.transaction->num_blocks <= blk) {
242 uint8_t **new_blocks;
243 /* expand the blocks array */
244 if (tdb->tdb1.transaction->blocks == NULL) {
245 new_blocks = (uint8_t **)malloc(
246 (blk+1)*sizeof(uint8_t *));
248 new_blocks = (uint8_t **)realloc(
249 tdb->tdb1.transaction->blocks,
250 (blk+1)*sizeof(uint8_t *));
252 if (new_blocks == NULL) {
253 tdb->last_error = TDB_ERR_OOM;
256 memset(&new_blocks[tdb->tdb1.transaction->num_blocks], 0,
257 (1+(blk - tdb->tdb1.transaction->num_blocks))*sizeof(uint8_t *));
258 tdb->tdb1.transaction->blocks = new_blocks;
259 tdb->tdb1.transaction->num_blocks = blk+1;
260 tdb->tdb1.transaction->last_block_size = 0;
263 /* allocate and fill a block? */
264 if (tdb->tdb1.transaction->blocks[blk] == NULL) {
265 tdb->tdb1.transaction->blocks[blk] = (uint8_t *)calloc(tdb->tdb1.transaction->block_size, 1);
266 if (tdb->tdb1.transaction->blocks[blk] == NULL) {
267 tdb->last_error = TDB_ERR_OOM;
268 tdb->tdb1.transaction->transaction_error = 1;
271 if (tdb->tdb1.transaction->old_map_size > blk * tdb->tdb1.transaction->block_size) {
272 tdb1_len_t len2 = tdb->tdb1.transaction->block_size;
273 if (len2 + (blk * tdb->tdb1.transaction->block_size) > tdb->tdb1.transaction->old_map_size) {
274 len2 = tdb->tdb1.transaction->old_map_size - (blk * tdb->tdb1.transaction->block_size);
276 if (tdb->tdb1.transaction->io_methods->tdb1_read(tdb, blk * tdb->tdb1.transaction->block_size,
277 tdb->tdb1.transaction->blocks[blk],
279 SAFE_FREE(tdb->tdb1.transaction->blocks[blk]);
280 tdb->last_error = TDB_ERR_IO;
283 if (blk == tdb->tdb1.transaction->num_blocks-1) {
284 tdb->tdb1.transaction->last_block_size = len2;
289 /* overwrite part of an existing block */
291 memset(tdb->tdb1.transaction->blocks[blk] + off, 0, len);
293 memcpy(tdb->tdb1.transaction->blocks[blk] + off, buf, len);
295 if (blk == tdb->tdb1.transaction->num_blocks-1) {
296 if (len + off > tdb->tdb1.transaction->last_block_size) {
297 tdb->tdb1.transaction->last_block_size = len + off;
304 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
305 "transaction_write: failed at off=%d len=%d",
306 (blk*tdb->tdb1.transaction->block_size) + off, len);
307 tdb->tdb1.transaction->transaction_error = 1;
313 write while in a transaction - this varient never expands the transaction blocks, it only
314 updates existing blocks. This means it cannot change the recovery size
316 static int transaction1_write_existing(struct tdb_context *tdb, tdb1_off_t off,
317 const void *buf, tdb1_len_t len)
321 /* break it up into block sized chunks */
322 while (len + (off % tdb->tdb1.transaction->block_size) > tdb->tdb1.transaction->block_size) {
323 tdb1_len_t len2 = tdb->tdb1.transaction->block_size - (off % tdb->tdb1.transaction->block_size);
324 if (transaction1_write_existing(tdb, off, buf, len2) != 0) {
330 buf = (const void *)(len2 + (const char *)buf);
338 blk = off / tdb->tdb1.transaction->block_size;
339 off = off % tdb->tdb1.transaction->block_size;
341 if (tdb->tdb1.transaction->num_blocks <= blk ||
342 tdb->tdb1.transaction->blocks[blk] == NULL) {
346 if (blk == tdb->tdb1.transaction->num_blocks-1 &&
347 off + len > tdb->tdb1.transaction->last_block_size) {
348 if (off >= tdb->tdb1.transaction->last_block_size) {
351 len = tdb->tdb1.transaction->last_block_size - off;
354 /* overwrite part of an existing block */
355 memcpy(tdb->tdb1.transaction->blocks[blk] + off, buf, len);
362 accelerated hash chain head search, using the cached hash heads
364 static void transaction1_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
367 for (;h < tdb->tdb1.header.hash_size;h++) {
368 /* the +1 takes account of the freelist */
369 if (0 != tdb->tdb1.transaction->hash_heads[h+1]) {
377 out of bounds check during a transaction
379 static int transaction1_oob(struct tdb_context *tdb, tdb1_off_t off, tdb1_off_t len, int probe)
381 if (off + len >= off && off + len <= tdb->file->map_size) {
384 tdb->last_error = TDB_ERR_IO;
389 transaction version of tdb1_expand().
391 static int transaction1_expand_file(struct tdb_context *tdb, tdb1_off_t size,
394 /* add a write to the transaction elements, so subsequent
395 reads see the zero data */
396 if (transaction1_write(tdb, size, NULL, addition) != 0) {
400 tdb->tdb1.transaction->expanded = true;
405 static const struct tdb1_methods transaction1_methods = {
408 transaction1_next_hash_chain,
410 transaction1_expand_file,
415 start a tdb transaction. No token is returned, as only a single
416 transaction is allowed to be pending per tdb_context
418 static int _tdb1_transaction_start(struct tdb_context *tdb)
420 /* some sanity checks */
421 if (tdb->flags & TDB_INTERNAL) {
422 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
424 "tdb1_transaction_start:"
431 if ((tdb->flags & TDB_RDONLY) || tdb->tdb1.traverse_read) {
432 tdb->last_error = tdb_logerr(tdb, TDB_ERR_RDONLY,
434 "tdb_transaction_start:"
441 /* cope with nested tdb1_transaction_start() calls */
442 if (tdb->tdb1.transaction != NULL) {
443 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
445 = tdb_logerr(tdb, TDB_ERR_EINVAL,
447 "tdb_transaction_start:"
448 " already inside transaction");
451 tdb->stats.transaction_nest++;
452 tdb->tdb1.transaction->nesting++;
456 if (tdb1_have_extra_locks(tdb)) {
457 /* the caller must not have any locks when starting a
458 transaction as otherwise we'll be screwed by lack
459 of nested locks in posix */
460 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
461 "tdb1_transaction_start: cannot start a"
462 " transaction with locks held");
466 if (tdb->tdb1.travlocks.next != NULL) {
467 /* you cannot use transactions inside a traverse (although you can use
468 traverse inside a transaction) as otherwise you can end up with
470 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
471 "tdb1_transaction_start: cannot start a"
472 " transaction within a traverse");
476 tdb->tdb1.transaction = (struct tdb1_transaction *)
477 calloc(sizeof(struct tdb1_transaction), 1);
478 if (tdb->tdb1.transaction == NULL) {
479 tdb->last_error = TDB_ERR_OOM;
483 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
484 tdb->tdb1.transaction->block_size = tdb->tdb1.page_size;
486 /* get the transaction write lock. This is a blocking lock. As
487 discussed with Volker, there are a number of ways we could
488 make this async, which we will probably do in the future */
489 if (tdb1_transaction_lock(tdb, F_WRLCK, TDB_LOCK_WAIT) == -1) {
490 SAFE_FREE(tdb->tdb1.transaction->blocks);
491 SAFE_FREE(tdb->tdb1.transaction);
495 /* get a read lock from the freelist to the end of file. This
496 is upgraded to a write lock during the commit */
497 if (tdb1_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
498 if (errno != EAGAIN && errno != EINTR) {
499 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
500 "tdb1_transaction_start:"
501 " failed to get hash locks");
503 goto fail_allrecord_lock;
506 /* setup a copy of the hash table heads so the hash scan in
507 traverse can be fast */
508 tdb->tdb1.transaction->hash_heads = (uint32_t *)
509 calloc(tdb->tdb1.header.hash_size+1, sizeof(uint32_t));
510 if (tdb->tdb1.transaction->hash_heads == NULL) {
511 tdb->last_error = TDB_ERR_OOM;
514 if (tdb->tdb1.io->tdb1_read(tdb, TDB1_FREELIST_TOP, tdb->tdb1.transaction->hash_heads,
515 TDB1_HASHTABLE_SIZE(tdb), 0) != 0) {
516 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
517 "tdb1_transaction_start: failed to read hash heads");
521 /* make sure we know about any file expansions already done by
523 tdb->tdb1.io->tdb1_oob(tdb, tdb->file->map_size, 1, 1);
524 tdb->tdb1.transaction->old_map_size = tdb->file->map_size;
526 /* finally hook the io methods, replacing them with
527 transaction specific methods */
528 tdb->tdb1.transaction->io_methods = tdb->tdb1.io;
529 tdb->tdb1.io = &transaction1_methods;
531 tdb->stats.transactions++;
535 tdb1_allrecord_unlock(tdb, F_RDLCK);
537 tdb1_transaction_unlock(tdb, F_WRLCK);
538 SAFE_FREE(tdb->tdb1.transaction->blocks);
539 SAFE_FREE(tdb->tdb1.transaction->hash_heads);
540 SAFE_FREE(tdb->tdb1.transaction);
544 int tdb1_transaction_start(struct tdb_context *tdb)
546 return _tdb1_transaction_start(tdb);
552 static int transaction1_sync(struct tdb_context *tdb, tdb1_off_t offset, tdb1_len_t length)
554 if (tdb->flags & TDB_NOSYNC) {
559 if (fdatasync(tdb->file->fd) != 0) {
561 if (fsync(tdb->file->fd) != 0) {
563 tdb->last_error = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
564 "tdb1_transaction: fsync failed");
568 if (tdb->file->map_ptr) {
569 tdb1_off_t moffset = offset & ~(tdb->tdb1.page_size-1);
570 if (msync(moffset + (char *)tdb->file->map_ptr,
571 length + (offset - moffset), MS_SYNC) != 0) {
572 tdb->last_error = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
574 " msync failed - %s",
584 static int _tdb1_transaction_cancel(struct tdb_context *tdb)
588 if (tdb->tdb1.transaction == NULL) {
589 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
590 "tdb1_transaction_cancel:"
595 if (tdb->tdb1.transaction->nesting != 0) {
596 tdb->tdb1.transaction->transaction_error = 1;
597 tdb->tdb1.transaction->nesting--;
601 tdb->file->map_size = tdb->tdb1.transaction->old_map_size;
603 /* free all the transaction blocks */
604 for (i=0;i<tdb->tdb1.transaction->num_blocks;i++) {
605 if (tdb->tdb1.transaction->blocks[i] != NULL) {
606 free(tdb->tdb1.transaction->blocks[i]);
609 SAFE_FREE(tdb->tdb1.transaction->blocks);
611 if (tdb->tdb1.transaction->magic_offset) {
612 const struct tdb1_methods *methods = tdb->tdb1.transaction->io_methods;
613 const uint32_t invalid = TDB1_RECOVERY_INVALID_MAGIC;
615 /* remove the recovery marker */
616 if (methods->tdb1_write(tdb, tdb->tdb1.transaction->magic_offset, &invalid, 4) == -1 ||
617 transaction1_sync(tdb, tdb->tdb1.transaction->magic_offset, 4) == -1) {
618 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
619 "tdb1_transaction_cancel: failed to"
620 " remove recovery magic");
625 /* This also removes the OPEN_LOCK, if we have it. */
626 tdb1_release_transaction_locks(tdb);
628 /* restore the normal io methods */
629 tdb->tdb1.io = tdb->tdb1.transaction->io_methods;
631 SAFE_FREE(tdb->tdb1.transaction->hash_heads);
632 SAFE_FREE(tdb->tdb1.transaction);
638 cancel the current transaction
640 int tdb1_transaction_cancel(struct tdb_context *tdb)
642 tdb->stats.transaction_cancel++;
643 return _tdb1_transaction_cancel(tdb);
647 work out how much space the linearised recovery data will consume
649 static tdb1_len_t tdb1_recovery_size(struct tdb_context *tdb)
651 tdb1_len_t recovery_size = 0;
654 recovery_size = sizeof(uint32_t);
655 for (i=0;i<tdb->tdb1.transaction->num_blocks;i++) {
656 if (i * tdb->tdb1.transaction->block_size >= tdb->tdb1.transaction->old_map_size) {
659 if (tdb->tdb1.transaction->blocks[i] == NULL) {
662 recovery_size += 2*sizeof(tdb1_off_t);
663 if (i == tdb->tdb1.transaction->num_blocks-1) {
664 recovery_size += tdb->tdb1.transaction->last_block_size;
666 recovery_size += tdb->tdb1.transaction->block_size;
670 return recovery_size;
673 int tdb1_recovery_area(struct tdb_context *tdb,
674 const struct tdb1_methods *methods,
675 tdb1_off_t *recovery_offset,
676 struct tdb1_record *rec)
678 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, recovery_offset) == -1) {
682 if (*recovery_offset == 0) {
687 if (methods->tdb1_read(tdb, *recovery_offset, rec, sizeof(*rec),
688 TDB1_DOCONV()) == -1) {
692 /* ignore invalid recovery regions: can happen in crash */
693 if (rec->magic != TDB1_RECOVERY_MAGIC &&
694 rec->magic != TDB1_RECOVERY_INVALID_MAGIC) {
695 *recovery_offset = 0;
702 allocate the recovery area, or use an existing recovery area if it is
705 static int tdb1_recovery_allocate(struct tdb_context *tdb,
706 tdb1_len_t *recovery_size,
707 tdb1_off_t *recovery_offset,
708 tdb1_len_t *recovery_max_size)
710 struct tdb1_record rec;
711 const struct tdb1_methods *methods = tdb->tdb1.transaction->io_methods;
712 tdb1_off_t recovery_head;
714 if (tdb1_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
715 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
716 "tdb1_recovery_allocate:"
717 " failed to read recovery head");
721 *recovery_size = tdb1_recovery_size(tdb);
723 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
724 /* it fits in the existing area */
725 *recovery_max_size = rec.rec_len;
726 *recovery_offset = recovery_head;
730 /* we need to free up the old recovery area, then allocate a
731 new one at the end of the file. Note that we cannot use
732 tdb1_allocate() to allocate the new one as that might return
733 us an area that is being currently used (as of the start of
735 if (recovery_head != 0) {
736 if (tdb1_free(tdb, recovery_head, &rec) == -1) {
737 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
738 "tdb1_recovery_allocate: failed to free"
739 " previous recovery area");
744 /* the tdb1_free() call might have increased the recovery size */
745 *recovery_size = tdb1_recovery_size(tdb);
747 /* round up to a multiple of page size */
748 *recovery_max_size = tdb1_expand_adjust(tdb->file->map_size,
753 *recovery_offset = tdb->file->map_size;
754 recovery_head = *recovery_offset;
756 if (methods->tdb1_expand_file(tdb, tdb->tdb1.transaction->old_map_size,
757 (tdb->file->map_size - tdb->tdb1.transaction->old_map_size) +
758 sizeof(rec) + *recovery_max_size) == -1) {
759 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
760 "tdb1_recovery_allocate:"
761 " failed to create recovery area");
764 tdb->stats.transaction_expand_file++;
766 /* remap the file (if using mmap) */
767 methods->tdb1_oob(tdb, tdb->file->map_size, 1, 1);
769 /* we have to reset the old map size so that we don't try to expand the file
770 again in the transaction commit, which would destroy the recovery area */
771 tdb->tdb1.transaction->old_map_size = tdb->file->map_size;
773 /* write the recovery header offset and sync - we can sync without a race here
774 as the magic ptr in the recovery record has not been set */
775 TDB1_CONV(recovery_head);
776 if (methods->tdb1_write(tdb, TDB1_RECOVERY_HEAD,
777 &recovery_head, sizeof(tdb1_off_t)) == -1) {
778 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
779 "tdb1_recovery_allocate:"
780 " failed to write recovery head");
783 if (transaction1_write_existing(tdb, TDB1_RECOVERY_HEAD, &recovery_head, sizeof(tdb1_off_t)) == -1) {
784 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
785 "tdb1_recovery_allocate:"
786 " failed to write recovery head");
795 setup the recovery data that will be used on a crash during commit
797 static int transaction1_setup_recovery(struct tdb_context *tdb,
798 tdb1_off_t *magic_offset)
800 tdb1_len_t recovery_size;
801 unsigned char *data, *p;
802 const struct tdb1_methods *methods = tdb->tdb1.transaction->io_methods;
803 struct tdb1_record *rec;
804 tdb1_off_t recovery_offset, recovery_max_size;
805 tdb1_off_t old_map_size = tdb->tdb1.transaction->old_map_size;
806 uint32_t magic, tailer;
810 check that the recovery area has enough space
812 if (tdb1_recovery_allocate(tdb, &recovery_size,
813 &recovery_offset, &recovery_max_size) == -1) {
817 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
819 tdb->last_error = TDB_ERR_OOM;
823 rec = (struct tdb1_record *)data;
824 memset(rec, 0, sizeof(*rec));
826 rec->magic = TDB1_RECOVERY_INVALID_MAGIC;
827 rec->data_len = recovery_size;
828 rec->rec_len = recovery_max_size;
829 rec->key_len = old_map_size;
832 /* build the recovery data into a single blob to allow us to do a single
833 large write, which should be more efficient */
834 p = data + sizeof(*rec);
835 for (i=0;i<tdb->tdb1.transaction->num_blocks;i++) {
839 if (tdb->tdb1.transaction->blocks[i] == NULL) {
843 offset = i * tdb->tdb1.transaction->block_size;
844 length = tdb->tdb1.transaction->block_size;
845 if (i == tdb->tdb1.transaction->num_blocks-1) {
846 length = tdb->tdb1.transaction->last_block_size;
849 if (offset >= old_map_size) {
852 if (offset + length > tdb->tdb1.transaction->old_map_size) {
853 tdb->last_error = tdb_logerr(tdb, TDB_ERR_CORRUPT,
855 "tdb1_transaction_setup_recovery: transaction data over new region boundary");
859 memcpy(p, &offset, 4);
860 memcpy(p+4, &length, 4);
864 /* the recovery area contains the old data, not the
865 new data, so we have to call the original tdb1_read
867 if (methods->tdb1_read(tdb, offset, p + 8, length, 0) != 0) {
869 tdb->last_error = TDB_ERR_IO;
876 tailer = sizeof(*rec) + recovery_max_size;
877 memcpy(p, &tailer, 4);
882 /* write the recovery data to the recovery area */
883 if (methods->tdb1_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
884 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
885 "tdb1_transaction_setup_recovery:"
886 " failed to write recovery data");
890 if (transaction1_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
891 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
892 "tdb1_transaction_setup_recovery: failed to write"
893 " secondary recovery data");
898 /* as we don't have ordered writes, we have to sync the recovery
899 data before we update the magic to indicate that the recovery
901 if (transaction1_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
908 magic = TDB1_RECOVERY_MAGIC;
911 *magic_offset = recovery_offset + offsetof(struct tdb1_record, magic);
913 if (methods->tdb1_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
914 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
915 "tdb1_transaction_setup_recovery:"
916 " failed to write recovery magic");
919 if (transaction1_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
920 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
921 "tdb1_transaction_setup_recovery:"
922 " failed to write secondary recovery magic");
926 /* ensure the recovery magic marker is on disk */
927 if (transaction1_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
934 static int _tdb1_transaction_prepare_commit(struct tdb_context *tdb)
936 const struct tdb1_methods *methods;
938 if (tdb->tdb1.transaction == NULL) {
939 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
940 "tdb1_transaction_prepare_commit:"
945 if (tdb->tdb1.transaction->prepared) {
946 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
947 "tdb1_transaction_prepare_commit:"
948 " transaction already prepared");
949 _tdb1_transaction_cancel(tdb);
953 if (tdb->tdb1.transaction->transaction_error) {
954 tdb->last_error = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
955 "tdb1_transaction_prepare_commit:"
956 " transaction error pending");
957 _tdb1_transaction_cancel(tdb);
962 if (tdb->tdb1.transaction->nesting != 0) {
966 /* check for a null transaction */
967 if (tdb->tdb1.transaction->blocks == NULL) {
971 methods = tdb->tdb1.transaction->io_methods;
973 /* if there are any locks pending then the caller has not
974 nested their locks properly, so fail the transaction */
975 if (tdb1_have_extra_locks(tdb)) {
976 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
977 "tdb1_transaction_prepare_commit:"
978 " locks pending on commit");
979 _tdb1_transaction_cancel(tdb);
983 /* upgrade the main transaction lock region to a write lock */
984 if (tdb1_allrecord_upgrade(tdb) == -1) {
985 if (errno != EAGAIN && errno != EINTR) {
986 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
987 "tdb1_transaction_prepare_commit:"
988 " failed to upgrade hash locks");
993 /* get the open lock - this prevents new users attaching to the database
995 if (tdb1_nest_lock(tdb, TDB1_OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
996 if (errno != EAGAIN && errno != EINTR) {
997 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
998 "tdb1_transaction_prepare_commit:"
999 " failed to get open lock");
1004 if (!(tdb->flags & TDB_NOSYNC)) {
1005 /* write the recovery data to the end of the file */
1006 if (transaction1_setup_recovery(tdb, &tdb->tdb1.transaction->magic_offset) == -1) {
1007 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1008 "tdb1_transaction_prepare_commit:"
1009 " failed to setup recovery data");
1014 tdb->tdb1.transaction->prepared = true;
1016 /* expand the file to the new size if needed */
1017 if (tdb->file->map_size != tdb->tdb1.transaction->old_map_size) {
1018 if (methods->tdb1_expand_file(tdb, tdb->tdb1.transaction->old_map_size,
1019 tdb->file->map_size -
1020 tdb->tdb1.transaction->old_map_size) == -1) {
1021 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1022 "tdb1_transaction_prepare_commit:"
1023 " expansion failed");
1026 tdb->stats.transaction_expand_file++;
1027 tdb->file->map_size = tdb->tdb1.transaction->old_map_size;
1028 methods->tdb1_oob(tdb, tdb->file->map_size, 1, 1);
1031 /* Keep the open lock until the actual commit */
1037 prepare to commit the current transaction
1039 int tdb1_transaction_prepare_commit(struct tdb_context *tdb)
1041 return _tdb1_transaction_prepare_commit(tdb);
1044 /* A repack is worthwhile if the largest is less than half total free. */
1045 static bool repack_worthwhile(struct tdb_context *tdb)
1048 struct tdb1_record rec;
1049 tdb1_len_t total = 0, largest = 0;
1051 if (tdb1_ofs_read(tdb, TDB1_FREELIST_TOP, &ptr) == -1) {
1055 while (ptr != 0 && tdb1_rec_free_read(tdb, ptr, &rec) == 0) {
1056 total += rec.rec_len;
1057 if (rec.rec_len > largest) {
1058 largest = rec.rec_len;
1063 return total > largest * 2;
1067 commit the current transaction
1069 int tdb1_transaction_commit(struct tdb_context *tdb)
1071 const struct tdb1_methods *methods;
1073 bool need_repack = false;
1075 if (tdb->tdb1.transaction == NULL) {
1076 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
1077 "tdb1_transaction_commit:"
1082 if (tdb->tdb1.transaction->transaction_error) {
1083 tdb->last_error = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
1084 "tdb1_transaction_commit:"
1085 " transaction error pending");
1086 _tdb1_transaction_cancel(tdb);
1091 if (tdb->tdb1.transaction->nesting != 0) {
1092 tdb->tdb1.transaction->nesting--;
1096 /* check for a null transaction */
1097 if (tdb->tdb1.transaction->blocks == NULL) {
1098 _tdb1_transaction_cancel(tdb);
1102 if (!tdb->tdb1.transaction->prepared) {
1103 int ret = _tdb1_transaction_prepare_commit(tdb);
1105 _tdb1_transaction_cancel(tdb);
1110 methods = tdb->tdb1.transaction->io_methods;
1112 /* perform all the writes */
1113 for (i=0;i<tdb->tdb1.transaction->num_blocks;i++) {
1117 if (tdb->tdb1.transaction->blocks[i] == NULL) {
1121 offset = i * tdb->tdb1.transaction->block_size;
1122 length = tdb->tdb1.transaction->block_size;
1123 if (i == tdb->tdb1.transaction->num_blocks-1) {
1124 length = tdb->tdb1.transaction->last_block_size;
1127 if (methods->tdb1_write(tdb, offset, tdb->tdb1.transaction->blocks[i], length) == -1) {
1128 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1129 "tdb1_transaction_commit:"
1130 " write failed during commit");
1132 /* we've overwritten part of the data and
1133 possibly expanded the file, so we need to
1134 run the crash recovery code */
1135 tdb->tdb1.io = methods;
1136 tdb1_transaction_recover(tdb);
1138 _tdb1_transaction_cancel(tdb);
1140 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1141 "tdb1_transaction_commit: write failed");
1144 SAFE_FREE(tdb->tdb1.transaction->blocks[i]);
1147 /* Do this before we drop lock or blocks. */
1148 if (tdb->tdb1.transaction->expanded) {
1149 need_repack = repack_worthwhile(tdb);
1152 SAFE_FREE(tdb->tdb1.transaction->blocks);
1153 tdb->tdb1.transaction->num_blocks = 0;
1155 /* ensure the new data is on disk */
1156 if (transaction1_sync(tdb, 0, tdb->file->map_size) == -1) {
1161 TODO: maybe write to some dummy hdr field, or write to magic
1162 offset without mmap, before the last sync, instead of the
1166 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1167 don't change the mtime of the file, this means the file may
1168 not be backed up (as tdb rounding to block sizes means that
1169 file size changes are quite rare too). The following forces
1170 mtime changes when a transaction completes */
1172 utime(tdb->name, NULL);
1175 /* use a transaction cancel to free memory and remove the
1176 transaction locks */
1177 _tdb1_transaction_cancel(tdb);
1180 if (tdb_repack(tdb) != 0)
1189 recover from an aborted transaction. Must be called with exclusive
1190 database write access already established (including the open
1191 lock to prevent new processes attaching)
1193 int tdb1_transaction_recover(struct tdb_context *tdb)
1195 tdb1_off_t recovery_head, recovery_eof;
1196 unsigned char *data, *p;
1198 struct tdb1_record rec;
1200 /* find the recovery area */
1201 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1202 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1203 "tdb1_transaction_recover:"
1204 " failed to read recovery head");
1208 if (recovery_head == 0) {
1209 /* we have never allocated a recovery record */
1213 /* read the recovery record */
1214 if (tdb->tdb1.io->tdb1_read(tdb, recovery_head, &rec,
1215 sizeof(rec), TDB1_DOCONV()) == -1) {
1216 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1217 "tdb1_transaction_recover:"
1218 " failed to read recovery record");
1222 if (rec.magic != TDB1_RECOVERY_MAGIC) {
1223 /* there is no valid recovery data */
1227 if (tdb->flags & TDB_RDONLY) {
1228 tdb->last_error = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
1229 "tdb1_transaction_recover:"
1230 " attempt to recover read only"
1235 recovery_eof = rec.key_len;
1237 data = (unsigned char *)malloc(rec.data_len);
1239 tdb->last_error = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
1240 "tdb1_transaction_recover:"
1241 " failed to allocate recovery data");
1245 /* read the full recovery data */
1246 if (tdb->tdb1.io->tdb1_read(tdb, recovery_head + sizeof(rec), data,
1247 rec.data_len, 0) == -1) {
1248 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1249 "tdb1_transaction_recover:"
1250 " failed to read recovery data");
1254 /* recover the file data */
1256 while (p+8 < data + rec.data_len) {
1258 if (TDB1_DOCONV()) {
1262 memcpy(&len, p+4, 4);
1264 if (tdb->tdb1.io->tdb1_write(tdb, ofs, p+8, len) == -1) {
1266 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1267 "tdb1_transaction_recover: failed to recover"
1268 " %d bytes at offset %d", len, ofs);
1276 if (transaction1_sync(tdb, 0, tdb->file->map_size) == -1) {
1277 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1278 "tdb1_transaction_recover: failed to sync recovery");
1282 /* if the recovery area is after the recovered eof then remove it */
1283 if (recovery_eof <= recovery_head) {
1284 if (tdb1_ofs_write(tdb, TDB1_RECOVERY_HEAD, &zero) == -1) {
1285 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1286 "tdb1_transaction_recover: failed to remove"
1292 /* remove the recovery magic */
1293 if (tdb1_ofs_write(tdb, recovery_head + offsetof(struct tdb1_record, magic),
1295 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1296 "tdb1_transaction_recover: failed to remove"
1301 if (transaction1_sync(tdb, 0, recovery_eof) == -1) {
1302 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
1303 "tdb1_transaction_recover:"
1304 " failed to sync2 recovery");
1308 tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
1309 "tdb1_transaction_recover: recovered %d byte database",
1316 /* Any I/O failures we say "needs recovery". */
1317 tdb_bool_err tdb1_needs_recovery(struct tdb_context *tdb)
1319 tdb1_off_t recovery_head;
1320 struct tdb1_record rec;
1322 /* find the recovery area */
1323 if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1324 return TDB_ERR_TO_OFF(tdb->last_error);
1327 if (recovery_head == 0) {
1328 /* we have never allocated a recovery record */
1332 /* read the recovery record */
1333 if (tdb->tdb1.io->tdb1_read(tdb, recovery_head, &rec,
1334 sizeof(rec), TDB1_DOCONV()) == -1) {
1335 return TDB_ERR_TO_OFF(tdb->last_error);
1338 return (rec.magic == TDB1_RECOVERY_MAGIC);