2 Unix SMB/CIFS implementation.
4 trivial database library
6 Copyright (C) Andrew Tridgell 2005
8 ** NOTE! The following LGPL license applies to the tdb
9 ** library. This does NOT imply that all of Samba is released
12 This library is free software; you can redistribute it and/or
13 modify it under the terms of the GNU Lesser General Public
14 License as published by the Free Software Foundation; either
15 version 3 of the License, or (at your option) any later version.
17 This library is distributed in the hope that it will be useful,
18 but WITHOUT ANY WARRANTY; without even the implied warranty of
19 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 Lesser General Public License for more details.
22 You should have received a copy of the GNU Lesser General Public
23 License along with this library; if not, see <http://www.gnu.org/licenses/>.
26 #include "tdb_private.h"
31 - only allow a single transaction at a time per database. This makes
32 using the transaction API simpler, as otherwise the caller would
33 have to cope with temporary failures in transactions that conflict
34 with other current transactions
36 - keep the transaction recovery information in the same file as the
37 database, using a special 'transaction recovery' record pointed at
38 by the header. This removes the need for extra journal files as
39 used by some other databases
41 - dynamically allocated the transaction recover record, re-using it
42 for subsequent transactions. If a larger record is needed then
43 tdb_free() the old record to place it on the normal tdb freelist
44 before allocating the new record
46 - during transactions, keep a linked list of writes all that have
47 been performed by intercepting all tdb_write() calls. The hooked
48 transaction versions of tdb_read() and tdb_write() check this
49 linked list and try to use the elements of the list in preference
52 - don't allow any locks to be held when a transaction starts,
53 otherwise we can end up with deadlock (plus lack of lock nesting
54 in posix locks would mean the lock is lost)
56 - if the caller gains a lock during the transaction but doesn't
57 release it then fail the commit
59 - allow for nested calls to tdb_transaction_start(), re-using the
60 existing transaction record. If the inner transaction is cancelled
61 then a subsequent commit will fail
63 - keep a mirrored copy of the tdb hash chain heads to allow for the
64 fast hash heads scan on traverse, updating the mirrored copy in
65 the transaction version of tdb_write
67 - allow callers to mix transaction and non-transaction use of tdb,
68 although once a transaction is started then an exclusive lock is
69 gained until the transaction is committed or cancelled
71 - the commit stategy involves first saving away all modified data
72 into a linearised buffer in the transaction recovery area, then
73 marking the transaction recovery area with a magic value to
74 indicate a valid recovery record. In total 4 fsync/msync calls are
75 needed per commit to prevent race conditions. It might be possible
76 to reduce this to 3 or even 2 with some more work.
78 - check for a valid recovery record on open of the tdb, while the
79 global lock is held. Automatically recover from the transaction
80 recovery area if needed, then continue with the open as
81 usual. This allows for smooth crash recovery with no administrator
84 - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85 still available, but no transaction recovery area is used and no
86 fsync/msync calls are made.
92 hold the context of any current transaction
94 struct tdb_transaction {
95 /* we keep a mirrored copy of the tdb hash heads here so
96 tdb_next_hash_chain() can operate efficiently */
99 /* the original io methods - used to do IOs to the real db */
100 const struct tdb_methods *io_methods;
102 /* the list of transaction blocks. When a block is first
103 written to, it gets created in this list */
106 uint32_t block_size; /* bytes in each block */
107 uint32_t last_block_size; /* number of valid bytes in the last block */
109 /* non-zero when an internal transaction error has
110 occurred. All write operations will then fail until the
111 transaction is ended */
112 int transaction_error;
114 /* when inside a transaction we need to keep track of any
115 nested tdb_transaction_start() calls, as these are allowed,
116 but don't create a new transaction */
119 /* set when a prepare has already occurred */
121 tdb_off_t magic_offset;
123 /* old file size before transaction */
124 tdb_len_t old_map_size;
129 read while in a transaction. We need to check first if the data is in our list
130 of transaction elements, then if not do a real read
132 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
133 tdb_len_t len, int cv)
137 /* Only a commit is allowed on a prepared transaction */
138 if (tdb->transaction->prepared) {
139 tdb->ecode = TDB_ERR_EINVAL;
140 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: transaction already prepared, read not allowed\n"));
141 tdb->transaction->transaction_error = 1;
145 /* break it down into block sized ops */
146 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
147 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
148 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
153 buf = (void *)(len2 + (char *)buf);
160 blk = off / tdb->transaction->block_size;
162 /* see if we have it in the block list */
163 if (tdb->transaction->num_blocks <= blk ||
164 tdb->transaction->blocks[blk] == NULL) {
165 /* nope, do a real read */
166 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
172 /* it is in the block list. Now check for the last block */
173 if (blk == tdb->transaction->num_blocks-1) {
174 if (len > tdb->transaction->last_block_size) {
179 /* now copy it out of this block */
180 memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
182 tdb_convert(buf, len);
187 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
188 tdb->ecode = TDB_ERR_IO;
189 tdb->transaction->transaction_error = 1;
195 write while in a transaction
197 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
198 const void *buf, tdb_len_t len)
202 /* Only a commit is allowed on a prepared transaction */
203 if (tdb->transaction->prepared) {
204 tdb->ecode = TDB_ERR_EINVAL;
205 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
206 tdb->transaction->transaction_error = 1;
210 /* if the write is to a hash head, then update the transaction
212 if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
213 off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
214 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
215 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
218 /* break it up into block sized chunks */
219 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
220 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
221 if (transaction_write(tdb, off, buf, len2) != 0) {
227 buf = (const void *)(len2 + (const char *)buf);
235 blk = off / tdb->transaction->block_size;
236 off = off % tdb->transaction->block_size;
238 if (tdb->transaction->num_blocks <= blk) {
239 uint8_t **new_blocks;
240 /* expand the blocks array */
241 if (tdb->transaction->blocks == NULL) {
242 new_blocks = (uint8_t **)malloc(
243 (blk+1)*sizeof(uint8_t *));
245 new_blocks = (uint8_t **)realloc(
246 tdb->transaction->blocks,
247 (blk+1)*sizeof(uint8_t *));
249 if (new_blocks == NULL) {
250 tdb->ecode = TDB_ERR_OOM;
253 memset(&new_blocks[tdb->transaction->num_blocks], 0,
254 (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
255 tdb->transaction->blocks = new_blocks;
256 tdb->transaction->num_blocks = blk+1;
257 tdb->transaction->last_block_size = 0;
260 /* allocate and fill a block? */
261 if (tdb->transaction->blocks[blk] == NULL) {
262 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
263 if (tdb->transaction->blocks[blk] == NULL) {
264 tdb->ecode = TDB_ERR_OOM;
265 tdb->transaction->transaction_error = 1;
268 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
269 tdb_len_t len2 = tdb->transaction->block_size;
270 if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
271 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
273 if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
274 tdb->transaction->blocks[blk],
276 SAFE_FREE(tdb->transaction->blocks[blk]);
277 tdb->ecode = TDB_ERR_IO;
280 if (blk == tdb->transaction->num_blocks-1) {
281 tdb->transaction->last_block_size = len2;
286 /* overwrite part of an existing block */
288 memset(tdb->transaction->blocks[blk] + off, 0, len);
290 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
292 if (blk == tdb->transaction->num_blocks-1) {
293 if (len + off > tdb->transaction->last_block_size) {
294 tdb->transaction->last_block_size = len + off;
301 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
302 (blk*tdb->transaction->block_size) + off, len));
303 tdb->transaction->transaction_error = 1;
309 write while in a transaction - this varient never expands the transaction blocks, it only
310 updates existing blocks. This means it cannot change the recovery size
312 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
313 const void *buf, tdb_len_t len)
317 /* break it up into block sized chunks */
318 while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
319 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
320 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
326 buf = (const void *)(len2 + (const char *)buf);
334 blk = off / tdb->transaction->block_size;
335 off = off % tdb->transaction->block_size;
337 if (tdb->transaction->num_blocks <= blk ||
338 tdb->transaction->blocks[blk] == NULL) {
342 if (blk == tdb->transaction->num_blocks-1 &&
343 off + len > tdb->transaction->last_block_size) {
344 if (off >= tdb->transaction->last_block_size) {
347 len = tdb->transaction->last_block_size - off;
350 /* overwrite part of an existing block */
351 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
358 accelerated hash chain head search, using the cached hash heads
360 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
363 for (;h < tdb->header.hash_size;h++) {
364 /* the +1 takes account of the freelist */
365 if (0 != tdb->transaction->hash_heads[h+1]) {
373 out of bounds check during a transaction
375 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
377 if (len <= tdb->map_size) {
380 return TDB_ERRCODE(TDB_ERR_IO, -1);
384 transaction version of tdb_expand().
386 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
389 /* add a write to the transaction elements, so subsequent
390 reads see the zero data */
391 if (transaction_write(tdb, size, NULL, addition) != 0) {
399 brlock during a transaction - ignore them
401 static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset,
402 int rw_type, int lck_type, int probe, size_t len)
407 static const struct tdb_methods transaction_methods = {
410 transaction_next_hash_chain,
412 transaction_expand_file,
418 start a tdb transaction. No token is returned, as only a single
419 transaction is allowed to be pending per tdb_context
421 int tdb_transaction_start(struct tdb_context *tdb)
423 /* some sanity checks */
424 if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
425 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
426 tdb->ecode = TDB_ERR_EINVAL;
430 /* cope with nested tdb_transaction_start() calls */
431 if (tdb->transaction != NULL) {
432 tdb->transaction->nesting++;
433 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
434 tdb->transaction->nesting));
438 if (tdb->num_locks != 0 || tdb->global_lock.count) {
439 /* the caller must not have any locks when starting a
440 transaction as otherwise we'll be screwed by lack
441 of nested locks in posix */
442 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
443 tdb->ecode = TDB_ERR_LOCK;
447 if (tdb->travlocks.next != NULL) {
448 /* you cannot use transactions inside a traverse (although you can use
449 traverse inside a transaction) as otherwise you can end up with
451 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
452 tdb->ecode = TDB_ERR_LOCK;
456 tdb->transaction = (struct tdb_transaction *)
457 calloc(sizeof(struct tdb_transaction), 1);
458 if (tdb->transaction == NULL) {
459 tdb->ecode = TDB_ERR_OOM;
463 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
464 tdb->transaction->block_size = tdb->page_size;
466 /* get the transaction write lock. This is a blocking lock. As
467 discussed with Volker, there are a number of ways we could
468 make this async, which we will probably do in the future */
469 if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
470 SAFE_FREE(tdb->transaction->blocks);
471 SAFE_FREE(tdb->transaction);
475 /* get a read lock from the freelist to the end of file. This
476 is upgraded to a write lock during the commit */
477 if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
478 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
479 tdb->ecode = TDB_ERR_LOCK;
483 /* setup a copy of the hash table heads so the hash scan in
484 traverse can be fast */
485 tdb->transaction->hash_heads = (uint32_t *)
486 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
487 if (tdb->transaction->hash_heads == NULL) {
488 tdb->ecode = TDB_ERR_OOM;
491 if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
492 TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
493 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
494 tdb->ecode = TDB_ERR_IO;
498 /* make sure we know about any file expansions already done by
500 tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
501 tdb->transaction->old_map_size = tdb->map_size;
503 /* finally hook the io methods, replacing them with
504 transaction specific methods */
505 tdb->transaction->io_methods = tdb->methods;
506 tdb->methods = &transaction_methods;
511 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
512 tdb_transaction_unlock(tdb);
513 SAFE_FREE(tdb->transaction->blocks);
514 SAFE_FREE(tdb->transaction->hash_heads);
515 SAFE_FREE(tdb->transaction);
523 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
525 if (tdb->flags & TDB_NOSYNC) {
529 if (fsync(tdb->fd) != 0) {
530 tdb->ecode = TDB_ERR_IO;
531 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
536 tdb_off_t moffset = offset & ~(tdb->page_size-1);
537 if (msync(moffset + (char *)tdb->map_ptr,
538 length + (offset - moffset), MS_SYNC) != 0) {
539 tdb->ecode = TDB_ERR_IO;
540 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
551 cancel the current transaction
553 int tdb_transaction_cancel(struct tdb_context *tdb)
557 if (tdb->transaction == NULL) {
558 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
562 if (tdb->transaction->nesting != 0) {
563 tdb->transaction->transaction_error = 1;
564 tdb->transaction->nesting--;
568 tdb->map_size = tdb->transaction->old_map_size;
570 /* free all the transaction blocks */
571 for (i=0;i<tdb->transaction->num_blocks;i++) {
572 if (tdb->transaction->blocks[i] != NULL) {
573 free(tdb->transaction->blocks[i]);
576 SAFE_FREE(tdb->transaction->blocks);
578 if (tdb->transaction->magic_offset) {
579 const struct tdb_methods *methods = tdb->transaction->io_methods;
582 /* remove the recovery marker */
583 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
584 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
585 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
590 /* remove any global lock created during the transaction */
591 if (tdb->global_lock.count != 0) {
592 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
593 tdb->global_lock.count = 0;
596 /* remove any locks created during the transaction */
597 if (tdb->num_locks != 0) {
598 for (i=0;i<tdb->num_lockrecs;i++) {
599 tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
600 F_UNLCK,F_SETLKW, 0, 1);
603 tdb->num_lockrecs = 0;
604 SAFE_FREE(tdb->lockrecs);
607 /* restore the normal io methods */
608 tdb->methods = tdb->transaction->io_methods;
610 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
611 tdb_transaction_unlock(tdb);
612 SAFE_FREE(tdb->transaction->hash_heads);
613 SAFE_FREE(tdb->transaction);
620 work out how much space the linearised recovery data will consume
622 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
624 tdb_len_t recovery_size = 0;
627 recovery_size = sizeof(uint32_t);
628 for (i=0;i<tdb->transaction->num_blocks;i++) {
629 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
632 if (tdb->transaction->blocks[i] == NULL) {
635 recovery_size += 2*sizeof(tdb_off_t);
636 if (i == tdb->transaction->num_blocks-1) {
637 recovery_size += tdb->transaction->last_block_size;
639 recovery_size += tdb->transaction->block_size;
643 return recovery_size;
647 allocate the recovery area, or use an existing recovery area if it is
650 static int tdb_recovery_allocate(struct tdb_context *tdb,
651 tdb_len_t *recovery_size,
652 tdb_off_t *recovery_offset,
653 tdb_len_t *recovery_max_size)
655 struct list_struct rec;
656 const struct tdb_methods *methods = tdb->transaction->io_methods;
657 tdb_off_t recovery_head;
659 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
660 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
666 if (recovery_head != 0 &&
667 methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
668 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
672 *recovery_size = tdb_recovery_size(tdb);
674 if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
675 /* it fits in the existing area */
676 *recovery_max_size = rec.rec_len;
677 *recovery_offset = recovery_head;
681 /* we need to free up the old recovery area, then allocate a
682 new one at the end of the file. Note that we cannot use
683 tdb_allocate() to allocate the new one as that might return
684 us an area that is being currently used (as of the start of
686 if (recovery_head != 0) {
687 if (tdb_free(tdb, recovery_head, &rec) == -1) {
688 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
693 /* the tdb_free() call might have increased the recovery size */
694 *recovery_size = tdb_recovery_size(tdb);
696 /* round up to a multiple of page size */
697 *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
698 *recovery_offset = tdb->map_size;
699 recovery_head = *recovery_offset;
701 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
702 (tdb->map_size - tdb->transaction->old_map_size) +
703 sizeof(rec) + *recovery_max_size) == -1) {
704 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
708 /* remap the file (if using mmap) */
709 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
711 /* we have to reset the old map size so that we don't try to expand the file
712 again in the transaction commit, which would destroy the recovery area */
713 tdb->transaction->old_map_size = tdb->map_size;
715 /* write the recovery header offset and sync - we can sync without a race here
716 as the magic ptr in the recovery record has not been set */
717 CONVERT(recovery_head);
718 if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
719 &recovery_head, sizeof(tdb_off_t)) == -1) {
720 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
723 if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
724 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
733 setup the recovery data that will be used on a crash during commit
735 static int transaction_setup_recovery(struct tdb_context *tdb,
736 tdb_off_t *magic_offset)
738 tdb_len_t recovery_size;
739 unsigned char *data, *p;
740 const struct tdb_methods *methods = tdb->transaction->io_methods;
741 struct list_struct *rec;
742 tdb_off_t recovery_offset, recovery_max_size;
743 tdb_off_t old_map_size = tdb->transaction->old_map_size;
744 uint32_t magic, tailer;
748 check that the recovery area has enough space
750 if (tdb_recovery_allocate(tdb, &recovery_size,
751 &recovery_offset, &recovery_max_size) == -1) {
755 data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
757 tdb->ecode = TDB_ERR_OOM;
761 rec = (struct list_struct *)data;
762 memset(rec, 0, sizeof(*rec));
765 rec->data_len = recovery_size;
766 rec->rec_len = recovery_max_size;
767 rec->key_len = old_map_size;
770 /* build the recovery data into a single blob to allow us to do a single
771 large write, which should be more efficient */
772 p = data + sizeof(*rec);
773 for (i=0;i<tdb->transaction->num_blocks;i++) {
777 if (tdb->transaction->blocks[i] == NULL) {
781 offset = i * tdb->transaction->block_size;
782 length = tdb->transaction->block_size;
783 if (i == tdb->transaction->num_blocks-1) {
784 length = tdb->transaction->last_block_size;
787 if (offset >= old_map_size) {
790 if (offset + length > tdb->transaction->old_map_size) {
791 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
793 tdb->ecode = TDB_ERR_CORRUPT;
796 memcpy(p, &offset, 4);
797 memcpy(p+4, &length, 4);
801 /* the recovery area contains the old data, not the
802 new data, so we have to call the original tdb_read
804 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
806 tdb->ecode = TDB_ERR_IO;
813 tailer = sizeof(*rec) + recovery_max_size;
814 memcpy(p, &tailer, 4);
817 /* write the recovery data to the recovery area */
818 if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
819 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
821 tdb->ecode = TDB_ERR_IO;
824 if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
825 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
827 tdb->ecode = TDB_ERR_IO;
831 /* as we don't have ordered writes, we have to sync the recovery
832 data before we update the magic to indicate that the recovery
834 if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
841 magic = TDB_RECOVERY_MAGIC;
844 *magic_offset = recovery_offset + offsetof(struct list_struct, magic);
846 if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
847 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
848 tdb->ecode = TDB_ERR_IO;
851 if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
852 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
853 tdb->ecode = TDB_ERR_IO;
857 /* ensure the recovery magic marker is on disk */
858 if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
866 prepare to commit the current transaction
868 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
870 const struct tdb_methods *methods;
872 if (tdb->transaction == NULL) {
873 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
877 if (tdb->transaction->prepared) {
878 tdb->ecode = TDB_ERR_EINVAL;
879 tdb_transaction_cancel(tdb);
880 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
884 if (tdb->transaction->transaction_error) {
885 tdb->ecode = TDB_ERR_IO;
886 tdb_transaction_cancel(tdb);
887 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
892 if (tdb->transaction->nesting != 0) {
896 /* check for a null transaction */
897 if (tdb->transaction->blocks == NULL) {
901 methods = tdb->transaction->io_methods;
903 /* if there are any locks pending then the caller has not
904 nested their locks properly, so fail the transaction */
905 if (tdb->num_locks || tdb->global_lock.count) {
906 tdb->ecode = TDB_ERR_LOCK;
907 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
908 tdb_transaction_cancel(tdb);
912 /* upgrade the main transaction lock region to a write lock */
913 if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
914 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
915 tdb->ecode = TDB_ERR_LOCK;
916 tdb_transaction_cancel(tdb);
920 /* get the global lock - this prevents new users attaching to the database
922 if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
923 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
924 tdb->ecode = TDB_ERR_LOCK;
925 tdb_transaction_cancel(tdb);
929 if (!(tdb->flags & TDB_NOSYNC)) {
930 /* write the recovery data to the end of the file */
931 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
932 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
933 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
934 tdb_transaction_cancel(tdb);
939 tdb->transaction->prepared = true;
941 /* expand the file to the new size if needed */
942 if (tdb->map_size != tdb->transaction->old_map_size) {
943 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
945 tdb->transaction->old_map_size) == -1) {
946 tdb->ecode = TDB_ERR_IO;
947 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
948 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
949 tdb_transaction_cancel(tdb);
952 tdb->map_size = tdb->transaction->old_map_size;
953 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
956 /* Keep the global lock until the actual commit */
962 commit the current transaction
964 int tdb_transaction_commit(struct tdb_context *tdb)
966 const struct tdb_methods *methods;
969 if (tdb->transaction == NULL) {
970 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
974 if (tdb->transaction->transaction_error) {
975 tdb->ecode = TDB_ERR_IO;
976 tdb_transaction_cancel(tdb);
977 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
982 if (tdb->transaction->nesting != 0) {
983 tdb->transaction->nesting--;
987 /* check for a null transaction */
988 if (tdb->transaction->blocks == NULL) {
989 tdb_transaction_cancel(tdb);
993 if (!tdb->transaction->prepared) {
994 int ret = tdb_transaction_prepare_commit(tdb);
999 methods = tdb->transaction->io_methods;
1001 /* perform all the writes */
1002 for (i=0;i<tdb->transaction->num_blocks;i++) {
1006 if (tdb->transaction->blocks[i] == NULL) {
1010 offset = i * tdb->transaction->block_size;
1011 length = tdb->transaction->block_size;
1012 if (i == tdb->transaction->num_blocks-1) {
1013 length = tdb->transaction->last_block_size;
1016 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1017 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1019 /* we've overwritten part of the data and
1020 possibly expanded the file, so we need to
1021 run the crash recovery code */
1022 tdb->methods = methods;
1023 tdb_transaction_recover(tdb);
1025 tdb_transaction_cancel(tdb);
1026 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1028 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1031 SAFE_FREE(tdb->transaction->blocks[i]);
1034 SAFE_FREE(tdb->transaction->blocks);
1035 tdb->transaction->num_blocks = 0;
1037 /* ensure the new data is on disk */
1038 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1042 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1045 TODO: maybe write to some dummy hdr field, or write to magic
1046 offset without mmap, before the last sync, instead of the
1050 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1051 don't change the mtime of the file, this means the file may
1052 not be backed up (as tdb rounding to block sizes means that
1053 file size changes are quite rare too). The following forces
1054 mtime changes when a transaction completes */
1056 utime(tdb->name, NULL);
1059 /* use a transaction cancel to free memory and remove the
1060 transaction locks */
1061 tdb_transaction_cancel(tdb);
1068 recover from an aborted transaction. Must be called with exclusive
1069 database write access already established (including the global
1070 lock to prevent new processes attaching)
1072 int tdb_transaction_recover(struct tdb_context *tdb)
1074 tdb_off_t recovery_head, recovery_eof;
1075 unsigned char *data, *p;
1077 struct list_struct rec;
1079 /* find the recovery area */
1080 if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1081 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1082 tdb->ecode = TDB_ERR_IO;
1086 if (recovery_head == 0) {
1087 /* we have never allocated a recovery record */
1091 /* read the recovery record */
1092 if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1093 sizeof(rec), DOCONV()) == -1) {
1094 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1095 tdb->ecode = TDB_ERR_IO;
1099 if (rec.magic != TDB_RECOVERY_MAGIC) {
1100 /* there is no valid recovery data */
1104 if (tdb->read_only) {
1105 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1106 tdb->ecode = TDB_ERR_CORRUPT;
1110 recovery_eof = rec.key_len;
1112 data = (unsigned char *)malloc(rec.data_len);
1114 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1115 tdb->ecode = TDB_ERR_OOM;
1119 /* read the full recovery data */
1120 if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1121 rec.data_len, 0) == -1) {
1122 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1123 tdb->ecode = TDB_ERR_IO;
1127 /* recover the file data */
1129 while (p+8 < data + rec.data_len) {
1135 memcpy(&len, p+4, 4);
1137 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1139 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1140 tdb->ecode = TDB_ERR_IO;
1148 if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1149 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1150 tdb->ecode = TDB_ERR_IO;
1154 /* if the recovery area is after the recovered eof then remove it */
1155 if (recovery_eof <= recovery_head) {
1156 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1157 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1158 tdb->ecode = TDB_ERR_IO;
1163 /* remove the recovery magic */
1164 if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic),
1166 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1167 tdb->ecode = TDB_ERR_IO;
1171 /* reduce the file size to the old size */
1173 if (ftruncate(tdb->fd, recovery_eof) != 0) {
1174 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1175 tdb->ecode = TDB_ERR_IO;
1178 tdb->map_size = recovery_eof;
1181 if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1182 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1183 tdb->ecode = TDB_ERR_IO;
1187 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",