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 open lock is held. Automatically recover from the transaction
80 recovery area if needed, then continue with the open as
81 usual. This allows for smooth crash recovery with no administrator
84 - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85 still available, but no fsync/msync calls are made. This means we
86 are still proof against a process dying during transaction commit,
87 but not against machine reboot.
89 - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
90 tdb_add_flags() transaction nesting is enabled.
91 It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
92 The default is that transaction nesting is allowed.
93 Note: this default may change in future versions of tdb.
95 Beware. when transactions are nested a transaction successfully
96 completed with tdb_transaction_commit() can be silently unrolled later.
98 - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
99 tdb_add_flags() transaction nesting is disabled.
100 It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
101 An attempt create a nested transaction will fail with TDB_ERR_NESTING.
102 The default is that transaction nesting is allowed.
103 Note: this default may change in future versions of tdb.
108 hold the context of any current transaction
110 struct tdb_transaction
{
111 /* we keep a mirrored copy of the tdb hash heads here so
112 tdb_next_hash_chain() can operate efficiently */
113 uint32_t *hash_heads
;
115 /* the original io methods - used to do IOs to the real db */
116 const struct tdb_methods
*io_methods
;
118 /* the list of transaction blocks. When a block is first
119 written to, it gets created in this list */
122 uint32_t block_size
; /* bytes in each block */
123 uint32_t last_block_size
; /* number of valid bytes in the last block */
125 /* non-zero when an internal transaction error has
126 occurred. All write operations will then fail until the
127 transaction is ended */
128 int transaction_error
;
130 /* when inside a transaction we need to keep track of any
131 nested tdb_transaction_start() calls, as these are allowed,
132 but don't create a new transaction */
135 /* set when a prepare has already occurred */
137 tdb_off_t magic_offset
;
139 /* old file size before transaction */
140 tdb_len_t old_map_size
;
142 /* did we expand in this transaction */
148 read while in a transaction. We need to check first if the data is in our list
149 of transaction elements, then if not do a real read
151 static int transaction_read(struct tdb_context
*tdb
, tdb_off_t off
, void *buf
,
152 tdb_len_t len
, int cv
)
156 /* break it down into block sized ops */
157 while (len
+ (off
% tdb
->transaction
->block_size
) > tdb
->transaction
->block_size
) {
158 tdb_len_t len2
= tdb
->transaction
->block_size
- (off
% tdb
->transaction
->block_size
);
159 if (transaction_read(tdb
, off
, buf
, len2
, cv
) != 0) {
164 buf
= (void *)(len2
+ (char *)buf
);
171 blk
= off
/ tdb
->transaction
->block_size
;
173 /* see if we have it in the block list */
174 if (tdb
->transaction
->num_blocks
<= blk
||
175 tdb
->transaction
->blocks
[blk
] == NULL
) {
176 /* nope, do a real read */
177 if (tdb
->transaction
->io_methods
->tdb_read(tdb
, off
, buf
, len
, cv
) != 0) {
183 /* it is in the block list. Now check for the last block */
184 if (blk
== tdb
->transaction
->num_blocks
-1) {
185 if (len
> tdb
->transaction
->last_block_size
) {
190 /* now copy it out of this block */
191 memcpy(buf
, tdb
->transaction
->blocks
[blk
] + (off
% tdb
->transaction
->block_size
), len
);
193 tdb_convert(buf
, len
);
198 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "transaction_read: failed at off=%d len=%d\n", off
, len
));
199 tdb
->ecode
= TDB_ERR_IO
;
200 tdb
->transaction
->transaction_error
= 1;
206 write while in a transaction
208 static int transaction_write(struct tdb_context
*tdb
, tdb_off_t off
,
209 const void *buf
, tdb_len_t len
)
213 /* Only a commit is allowed on a prepared transaction */
214 if (tdb
->transaction
->prepared
) {
215 tdb
->ecode
= TDB_ERR_EINVAL
;
216 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "transaction_write: transaction already prepared, write not allowed\n"));
217 tdb
->transaction
->transaction_error
= 1;
221 /* if the write is to a hash head, then update the transaction
223 if (len
== sizeof(tdb_off_t
) && off
>= FREELIST_TOP
&&
224 off
< FREELIST_TOP
+TDB_HASHTABLE_SIZE(tdb
)) {
225 uint32_t chain
= (off
-FREELIST_TOP
) / sizeof(tdb_off_t
);
226 memcpy(&tdb
->transaction
->hash_heads
[chain
], buf
, len
);
229 /* break it up into block sized chunks */
230 while (len
+ (off
% tdb
->transaction
->block_size
) > tdb
->transaction
->block_size
) {
231 tdb_len_t len2
= tdb
->transaction
->block_size
- (off
% tdb
->transaction
->block_size
);
232 if (transaction_write(tdb
, off
, buf
, len2
) != 0) {
238 buf
= (const void *)(len2
+ (const char *)buf
);
246 blk
= off
/ tdb
->transaction
->block_size
;
247 off
= off
% tdb
->transaction
->block_size
;
249 if (tdb
->transaction
->num_blocks
<= blk
) {
250 uint8_t **new_blocks
;
251 /* expand the blocks array */
252 new_blocks
= (uint8_t **)realloc(tdb
->transaction
->blocks
,
253 (blk
+1)*sizeof(uint8_t *));
254 if (new_blocks
== NULL
) {
255 tdb
->ecode
= TDB_ERR_OOM
;
258 memset(&new_blocks
[tdb
->transaction
->num_blocks
], 0,
259 (1+(blk
- tdb
->transaction
->num_blocks
))*sizeof(uint8_t *));
260 tdb
->transaction
->blocks
= new_blocks
;
261 tdb
->transaction
->num_blocks
= blk
+1;
262 tdb
->transaction
->last_block_size
= 0;
265 /* allocate and fill a block? */
266 if (tdb
->transaction
->blocks
[blk
] == NULL
) {
267 tdb
->transaction
->blocks
[blk
] = (uint8_t *)calloc(tdb
->transaction
->block_size
, 1);
268 if (tdb
->transaction
->blocks
[blk
] == NULL
) {
269 tdb
->ecode
= TDB_ERR_OOM
;
270 tdb
->transaction
->transaction_error
= 1;
273 if (tdb
->transaction
->old_map_size
> blk
* tdb
->transaction
->block_size
) {
274 tdb_len_t len2
= tdb
->transaction
->block_size
;
275 if (len2
+ (blk
* tdb
->transaction
->block_size
) > tdb
->transaction
->old_map_size
) {
276 len2
= tdb
->transaction
->old_map_size
- (blk
* tdb
->transaction
->block_size
);
278 if (tdb
->transaction
->io_methods
->tdb_read(tdb
, blk
* tdb
->transaction
->block_size
,
279 tdb
->transaction
->blocks
[blk
],
281 SAFE_FREE(tdb
->transaction
->blocks
[blk
]);
282 tdb
->ecode
= TDB_ERR_IO
;
285 if (blk
== tdb
->transaction
->num_blocks
-1) {
286 tdb
->transaction
->last_block_size
= len2
;
291 /* overwrite part of an existing block */
293 memset(tdb
->transaction
->blocks
[blk
] + off
, 0, len
);
295 memcpy(tdb
->transaction
->blocks
[blk
] + off
, buf
, len
);
297 if (blk
== tdb
->transaction
->num_blocks
-1) {
298 if (len
+ off
> tdb
->transaction
->last_block_size
) {
299 tdb
->transaction
->last_block_size
= len
+ off
;
306 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "transaction_write: failed at off=%d len=%d\n",
307 (blk
*tdb
->transaction
->block_size
) + off
, len
));
308 tdb
->transaction
->transaction_error
= 1;
314 write while in a transaction - this variant never expands the transaction blocks, it only
315 updates existing blocks. This means it cannot change the recovery size
317 static int transaction_write_existing(struct tdb_context
*tdb
, tdb_off_t off
,
318 const void *buf
, tdb_len_t len
)
322 /* break it up into block sized chunks */
323 while (len
+ (off
% tdb
->transaction
->block_size
) > tdb
->transaction
->block_size
) {
324 tdb_len_t len2
= tdb
->transaction
->block_size
- (off
% tdb
->transaction
->block_size
);
325 if (transaction_write_existing(tdb
, off
, buf
, len2
) != 0) {
331 buf
= (const void *)(len2
+ (const char *)buf
);
339 blk
= off
/ tdb
->transaction
->block_size
;
340 off
= off
% tdb
->transaction
->block_size
;
342 if (tdb
->transaction
->num_blocks
<= blk
||
343 tdb
->transaction
->blocks
[blk
] == NULL
) {
347 if (blk
== tdb
->transaction
->num_blocks
-1 &&
348 off
+ len
> tdb
->transaction
->last_block_size
) {
349 if (off
>= tdb
->transaction
->last_block_size
) {
352 len
= tdb
->transaction
->last_block_size
- off
;
355 /* overwrite part of an existing block */
356 memcpy(tdb
->transaction
->blocks
[blk
] + off
, buf
, len
);
363 accelerated hash chain head search, using the cached hash heads
365 static void transaction_next_hash_chain(struct tdb_context
*tdb
, uint32_t *chain
)
368 for (;h
< tdb
->hash_size
;h
++) {
369 /* the +1 takes account of the freelist */
370 if (0 != tdb
->transaction
->hash_heads
[h
+1]) {
378 out of bounds check during a transaction
380 static int transaction_oob(struct tdb_context
*tdb
, tdb_off_t off
,
381 tdb_len_t len
, int probe
)
383 if (off
+ len
>= off
&& off
+ len
<= tdb
->map_size
) {
386 tdb
->ecode
= TDB_ERR_IO
;
391 transaction version of tdb_expand().
393 static int transaction_expand_file(struct tdb_context
*tdb
, tdb_off_t size
,
396 /* add a write to the transaction elements, so subsequent
397 reads see the zero data */
398 if (transaction_write(tdb
, size
, NULL
, addition
) != 0) {
402 tdb
->transaction
->expanded
= true;
407 static const struct tdb_methods transaction_methods
= {
410 transaction_next_hash_chain
,
412 transaction_expand_file
,
417 start a tdb transaction. No token is returned, as only a single
418 transaction is allowed to be pending per tdb_context
420 static int _tdb_transaction_start(struct tdb_context
*tdb
,
421 enum tdb_lock_flags lockflags
)
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 if (!(tdb
->flags
& TDB_ALLOW_NESTING
)) {
433 tdb
->ecode
= TDB_ERR_NESTING
;
436 tdb
->transaction
->nesting
++;
437 TDB_LOG((tdb
, TDB_DEBUG_TRACE
, "tdb_transaction_start: nesting %d\n",
438 tdb
->transaction
->nesting
));
442 if (tdb_have_extra_locks(tdb
)) {
443 /* the caller must not have any locks when starting a
444 transaction as otherwise we'll be screwed by lack
445 of nested locks in posix */
446 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_start: cannot start a transaction with locks held\n"));
447 tdb
->ecode
= TDB_ERR_LOCK
;
451 if (tdb
->travlocks
.next
!= NULL
) {
452 /* you cannot use transactions inside a traverse (although you can use
453 traverse inside a transaction) as otherwise you can end up with
455 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
456 tdb
->ecode
= TDB_ERR_LOCK
;
460 tdb
->transaction
= (struct tdb_transaction
*)
461 calloc(sizeof(struct tdb_transaction
), 1);
462 if (tdb
->transaction
== NULL
) {
463 tdb
->ecode
= TDB_ERR_OOM
;
467 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
468 tdb
->transaction
->block_size
= tdb
->page_size
;
470 /* get the transaction write lock. This is a blocking lock. As
471 discussed with Volker, there are a number of ways we could
472 make this async, which we will probably do in the future */
473 if (tdb_transaction_lock(tdb
, F_WRLCK
, lockflags
) == -1) {
474 SAFE_FREE(tdb
->transaction
->blocks
);
475 SAFE_FREE(tdb
->transaction
);
476 if ((lockflags
& TDB_LOCK_WAIT
) == 0) {
477 tdb
->ecode
= TDB_ERR_NOLOCK
;
482 /* get a read lock from the freelist to the end of file. This
483 is upgraded to a write lock during the commit */
484 if (tdb_allrecord_lock(tdb
, F_RDLCK
, TDB_LOCK_WAIT
, true) == -1) {
485 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_start: failed to get hash locks\n"));
486 goto fail_allrecord_lock
;
489 /* setup a copy of the hash table heads so the hash scan in
490 traverse can be fast */
491 tdb
->transaction
->hash_heads
= (uint32_t *)
492 calloc(tdb
->hash_size
+1, sizeof(uint32_t));
493 if (tdb
->transaction
->hash_heads
== NULL
) {
494 tdb
->ecode
= TDB_ERR_OOM
;
497 if (tdb
->methods
->tdb_read(tdb
, FREELIST_TOP
, tdb
->transaction
->hash_heads
,
498 TDB_HASHTABLE_SIZE(tdb
), 0) != 0) {
499 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_start: failed to read hash heads\n"));
500 tdb
->ecode
= TDB_ERR_IO
;
504 /* make sure we know about any file expansions already done by
506 tdb
->methods
->tdb_oob(tdb
, tdb
->map_size
, 1, 1);
507 tdb
->transaction
->old_map_size
= tdb
->map_size
;
509 /* finally hook the io methods, replacing them with
510 transaction specific methods */
511 tdb
->transaction
->io_methods
= tdb
->methods
;
512 tdb
->methods
= &transaction_methods
;
514 /* Trace at the end, so we get sequence number correct. */
515 tdb_trace(tdb
, "tdb_transaction_start");
519 tdb_allrecord_unlock(tdb
, F_RDLCK
, false);
521 tdb_transaction_unlock(tdb
, F_WRLCK
);
522 SAFE_FREE(tdb
->transaction
->blocks
);
523 SAFE_FREE(tdb
->transaction
->hash_heads
);
524 SAFE_FREE(tdb
->transaction
);
528 _PUBLIC_
int tdb_transaction_start(struct tdb_context
*tdb
)
530 return _tdb_transaction_start(tdb
, TDB_LOCK_WAIT
);
533 _PUBLIC_
int tdb_transaction_start_nonblock(struct tdb_context
*tdb
)
535 return _tdb_transaction_start(tdb
, TDB_LOCK_NOWAIT
|TDB_LOCK_PROBE
);
541 static int transaction_sync(struct tdb_context
*tdb
, tdb_off_t offset
, tdb_len_t length
)
543 if (tdb
->flags
& TDB_NOSYNC
) {
547 #ifdef HAVE_FDATASYNC
548 if (fdatasync(tdb
->fd
) != 0) {
550 if (fsync(tdb
->fd
) != 0) {
552 tdb
->ecode
= TDB_ERR_IO
;
553 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction: fsync failed\n"));
558 tdb_off_t moffset
= offset
& ~(tdb
->page_size
-1);
559 if (msync(moffset
+ (char *)tdb
->map_ptr
,
560 length
+ (offset
- moffset
), MS_SYNC
) != 0) {
561 tdb
->ecode
= TDB_ERR_IO
;
562 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction: msync failed - %s\n",
572 static int _tdb_transaction_cancel(struct tdb_context
*tdb
)
576 if (tdb
->transaction
== NULL
) {
577 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_cancel: no transaction\n"));
581 if (tdb
->transaction
->nesting
!= 0) {
582 tdb
->transaction
->transaction_error
= 1;
583 tdb
->transaction
->nesting
--;
587 tdb
->map_size
= tdb
->transaction
->old_map_size
;
589 /* free all the transaction blocks */
590 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
591 if (tdb
->transaction
->blocks
[i
] != NULL
) {
592 free(tdb
->transaction
->blocks
[i
]);
595 SAFE_FREE(tdb
->transaction
->blocks
);
597 if (tdb
->transaction
->magic_offset
) {
598 const struct tdb_methods
*methods
= tdb
->transaction
->io_methods
;
599 const uint32_t invalid
= TDB_RECOVERY_INVALID_MAGIC
;
601 /* remove the recovery marker */
602 if (methods
->tdb_write(tdb
, tdb
->transaction
->magic_offset
, &invalid
, 4) == -1 ||
603 transaction_sync(tdb
, tdb
->transaction
->magic_offset
, 4) == -1) {
604 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_cancel: failed to remove recovery magic\n"));
609 /* This also removes the OPEN_LOCK, if we have it. */
610 tdb_release_transaction_locks(tdb
);
612 /* restore the normal io methods */
613 tdb
->methods
= tdb
->transaction
->io_methods
;
615 SAFE_FREE(tdb
->transaction
->hash_heads
);
616 SAFE_FREE(tdb
->transaction
);
622 cancel the current transaction
624 _PUBLIC_
int tdb_transaction_cancel(struct tdb_context
*tdb
)
626 tdb_trace(tdb
, "tdb_transaction_cancel");
627 return _tdb_transaction_cancel(tdb
);
631 work out how much space the linearised recovery data will consume
633 static tdb_len_t
tdb_recovery_size(struct tdb_context
*tdb
)
635 tdb_len_t recovery_size
= 0;
638 recovery_size
= sizeof(uint32_t);
639 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
640 if (i
* tdb
->transaction
->block_size
>= tdb
->transaction
->old_map_size
) {
643 if (tdb
->transaction
->blocks
[i
] == NULL
) {
646 recovery_size
+= 2*sizeof(tdb_off_t
);
647 if (i
== tdb
->transaction
->num_blocks
-1) {
648 recovery_size
+= tdb
->transaction
->last_block_size
;
650 recovery_size
+= tdb
->transaction
->block_size
;
654 return recovery_size
;
657 int tdb_recovery_area(struct tdb_context
*tdb
,
658 const struct tdb_methods
*methods
,
659 tdb_off_t
*recovery_offset
,
660 struct tdb_record
*rec
)
662 if (tdb_ofs_read(tdb
, TDB_RECOVERY_HEAD
, recovery_offset
) == -1) {
666 if (*recovery_offset
== 0) {
671 if (methods
->tdb_read(tdb
, *recovery_offset
, rec
, sizeof(*rec
),
676 /* ignore invalid recovery regions: can happen in crash */
677 if (rec
->magic
!= TDB_RECOVERY_MAGIC
&&
678 rec
->magic
!= TDB_RECOVERY_INVALID_MAGIC
) {
679 *recovery_offset
= 0;
686 allocate the recovery area, or use an existing recovery area if it is
689 static int tdb_recovery_allocate(struct tdb_context
*tdb
,
690 tdb_len_t
*recovery_size
,
691 tdb_off_t
*recovery_offset
,
692 tdb_len_t
*recovery_max_size
)
694 struct tdb_record rec
;
695 const struct tdb_methods
*methods
= tdb
->transaction
->io_methods
;
696 tdb_off_t recovery_head
, new_end
;
698 if (tdb_recovery_area(tdb
, methods
, &recovery_head
, &rec
) == -1) {
699 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to read recovery head\n"));
703 *recovery_size
= tdb_recovery_size(tdb
);
705 /* Existing recovery area? */
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 /* If recovery area in middle of file, we need a new one. */
714 if (recovery_head
== 0
715 || recovery_head
+ sizeof(rec
) + rec
.rec_len
!= tdb
->map_size
) {
716 /* we need to free up the old recovery area, then allocate a
717 new one at the end of the file. Note that we cannot use
718 tdb_allocate() to allocate the new one as that might return
719 us an area that is being currently used (as of the start of
722 if (tdb_free(tdb
, recovery_head
, &rec
) == -1) {
723 TDB_LOG((tdb
, TDB_DEBUG_FATAL
,
724 "tdb_recovery_allocate: failed to"
725 " free previous recovery area\n"));
729 /* the tdb_free() call might have increased
730 * the recovery size */
731 *recovery_size
= tdb_recovery_size(tdb
);
734 /* New head will be at end of file. */
735 recovery_head
= tdb
->map_size
;
738 /* Now we know where it will be. */
739 *recovery_offset
= recovery_head
;
741 /* Expand by more than we need, so we don't do it often. */
742 *recovery_max_size
= tdb_expand_adjust(tdb
->map_size
,
747 new_end
= recovery_head
+ sizeof(rec
) + *recovery_max_size
;
749 if (methods
->tdb_expand_file(tdb
, tdb
->transaction
->old_map_size
,
750 new_end
- tdb
->transaction
->old_map_size
)
752 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to create recovery area\n"));
756 /* remap the file (if using mmap) */
757 methods
->tdb_oob(tdb
, tdb
->map_size
, 1, 1);
759 /* we have to reset the old map size so that we don't try to expand the file
760 again in the transaction commit, which would destroy the recovery area */
761 tdb
->transaction
->old_map_size
= tdb
->map_size
;
763 /* write the recovery header offset and sync - we can sync without a race here
764 as the magic ptr in the recovery record has not been set */
765 CONVERT(recovery_head
);
766 if (methods
->tdb_write(tdb
, TDB_RECOVERY_HEAD
,
767 &recovery_head
, sizeof(tdb_off_t
)) == -1) {
768 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to write recovery head\n"));
771 if (transaction_write_existing(tdb
, TDB_RECOVERY_HEAD
, &recovery_head
, sizeof(tdb_off_t
)) == -1) {
772 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to write recovery head\n"));
781 setup the recovery data that will be used on a crash during commit
783 static int transaction_setup_recovery(struct tdb_context
*tdb
,
784 tdb_off_t
*magic_offset
)
786 tdb_len_t recovery_size
;
787 unsigned char *data
, *p
;
788 const struct tdb_methods
*methods
= tdb
->transaction
->io_methods
;
789 struct tdb_record
*rec
;
790 tdb_off_t recovery_offset
, recovery_max_size
;
791 tdb_off_t old_map_size
= tdb
->transaction
->old_map_size
;
792 uint32_t magic
, tailer
;
796 check that the recovery area has enough space
798 if (tdb_recovery_allocate(tdb
, &recovery_size
,
799 &recovery_offset
, &recovery_max_size
) == -1) {
803 data
= (unsigned char *)malloc(recovery_size
+ sizeof(*rec
));
805 tdb
->ecode
= TDB_ERR_OOM
;
809 rec
= (struct tdb_record
*)data
;
810 memset(rec
, 0, sizeof(*rec
));
812 rec
->magic
= TDB_RECOVERY_INVALID_MAGIC
;
813 rec
->data_len
= recovery_size
;
814 rec
->rec_len
= recovery_max_size
;
815 rec
->key_len
= old_map_size
;
818 /* build the recovery data into a single blob to allow us to do a single
819 large write, which should be more efficient */
820 p
= data
+ sizeof(*rec
);
821 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
825 if (tdb
->transaction
->blocks
[i
] == NULL
) {
829 offset
= i
* tdb
->transaction
->block_size
;
830 length
= tdb
->transaction
->block_size
;
831 if (i
== tdb
->transaction
->num_blocks
-1) {
832 length
= tdb
->transaction
->last_block_size
;
835 if (offset
>= old_map_size
) {
838 if (offset
+ length
> tdb
->transaction
->old_map_size
) {
839 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
841 tdb
->ecode
= TDB_ERR_CORRUPT
;
844 memcpy(p
, &offset
, 4);
845 memcpy(p
+4, &length
, 4);
849 /* the recovery area contains the old data, not the
850 new data, so we have to call the original tdb_read
852 if (methods
->tdb_read(tdb
, offset
, p
+ 8, length
, 0) != 0) {
854 tdb
->ecode
= TDB_ERR_IO
;
861 tailer
= sizeof(*rec
) + recovery_max_size
;
862 memcpy(p
, &tailer
, 4);
867 /* write the recovery data to the recovery area */
868 if (methods
->tdb_write(tdb
, recovery_offset
, data
, sizeof(*rec
) + recovery_size
) == -1) {
869 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
871 tdb
->ecode
= TDB_ERR_IO
;
874 if (transaction_write_existing(tdb
, recovery_offset
, data
, sizeof(*rec
) + recovery_size
) == -1) {
875 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
877 tdb
->ecode
= TDB_ERR_IO
;
881 /* as we don't have ordered writes, we have to sync the recovery
882 data before we update the magic to indicate that the recovery
884 if (transaction_sync(tdb
, recovery_offset
, sizeof(*rec
) + recovery_size
) == -1) {
891 magic
= TDB_RECOVERY_MAGIC
;
894 *magic_offset
= recovery_offset
+ offsetof(struct tdb_record
, magic
);
896 if (methods
->tdb_write(tdb
, *magic_offset
, &magic
, sizeof(magic
)) == -1) {
897 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
898 tdb
->ecode
= TDB_ERR_IO
;
901 if (transaction_write_existing(tdb
, *magic_offset
, &magic
, sizeof(magic
)) == -1) {
902 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
903 tdb
->ecode
= TDB_ERR_IO
;
907 /* ensure the recovery magic marker is on disk */
908 if (transaction_sync(tdb
, *magic_offset
, sizeof(magic
)) == -1) {
915 static int _tdb_transaction_prepare_commit(struct tdb_context
*tdb
)
917 const struct tdb_methods
*methods
;
919 if (tdb
->transaction
== NULL
) {
920 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: no transaction\n"));
924 if (tdb
->transaction
->prepared
) {
925 tdb
->ecode
= TDB_ERR_EINVAL
;
926 _tdb_transaction_cancel(tdb
);
927 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: transaction already prepared\n"));
931 if (tdb
->transaction
->transaction_error
) {
932 tdb
->ecode
= TDB_ERR_IO
;
933 _tdb_transaction_cancel(tdb
);
934 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: transaction error pending\n"));
939 if (tdb
->transaction
->nesting
!= 0) {
943 /* check for a null transaction */
944 if (tdb
->transaction
->blocks
== NULL
) {
948 methods
= tdb
->transaction
->io_methods
;
950 /* if there are any locks pending then the caller has not
951 nested their locks properly, so fail the transaction */
952 if (tdb_have_extra_locks(tdb
)) {
953 tdb
->ecode
= TDB_ERR_LOCK
;
954 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: locks pending on commit\n"));
955 _tdb_transaction_cancel(tdb
);
959 /* upgrade the main transaction lock region to a write lock */
960 if (tdb_allrecord_upgrade(tdb
) == -1) {
961 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
962 _tdb_transaction_cancel(tdb
);
966 /* get the open lock - this prevents new users attaching to the database
968 if (tdb_nest_lock(tdb
, OPEN_LOCK
, F_WRLCK
, TDB_LOCK_WAIT
) == -1) {
969 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: failed to get open lock\n"));
970 _tdb_transaction_cancel(tdb
);
974 /* write the recovery data to the end of the file */
975 if (transaction_setup_recovery(tdb
, &tdb
->transaction
->magic_offset
) == -1) {
976 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
977 _tdb_transaction_cancel(tdb
);
981 tdb
->transaction
->prepared
= true;
983 /* expand the file to the new size if needed */
984 if (tdb
->map_size
!= tdb
->transaction
->old_map_size
) {
985 if (methods
->tdb_expand_file(tdb
, tdb
->transaction
->old_map_size
,
987 tdb
->transaction
->old_map_size
) == -1) {
988 tdb
->ecode
= TDB_ERR_IO
;
989 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_prepare_commit: expansion failed\n"));
990 _tdb_transaction_cancel(tdb
);
993 tdb
->map_size
= tdb
->transaction
->old_map_size
;
994 methods
->tdb_oob(tdb
, tdb
->map_size
, 1, 1);
997 /* Keep the open lock until the actual commit */
1003 prepare to commit the current transaction
1005 _PUBLIC_
int tdb_transaction_prepare_commit(struct tdb_context
*tdb
)
1007 tdb_trace(tdb
, "tdb_transaction_prepare_commit");
1008 return _tdb_transaction_prepare_commit(tdb
);
1011 /* A repack is worthwhile if the largest is less than half total free. */
1012 static bool repack_worthwhile(struct tdb_context
*tdb
)
1015 struct tdb_record rec
;
1016 tdb_len_t total
= 0, largest
= 0;
1018 if (tdb_ofs_read(tdb
, FREELIST_TOP
, &ptr
) == -1) {
1022 while (ptr
!= 0 && tdb_rec_free_read(tdb
, ptr
, &rec
) == 0) {
1023 total
+= rec
.rec_len
;
1024 if (rec
.rec_len
> largest
) {
1025 largest
= rec
.rec_len
;
1030 return total
> largest
* 2;
1034 commit the current transaction
1036 _PUBLIC_
int tdb_transaction_commit(struct tdb_context
*tdb
)
1038 const struct tdb_methods
*methods
;
1040 bool need_repack
= false;
1042 if (tdb
->transaction
== NULL
) {
1043 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_commit: no transaction\n"));
1047 tdb_trace(tdb
, "tdb_transaction_commit");
1049 if (tdb
->transaction
->transaction_error
) {
1050 tdb
->ecode
= TDB_ERR_IO
;
1051 _tdb_transaction_cancel(tdb
);
1052 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_commit: transaction error pending\n"));
1057 if (tdb
->transaction
->nesting
!= 0) {
1058 tdb
->transaction
->nesting
--;
1062 /* check for a null transaction */
1063 if (tdb
->transaction
->blocks
== NULL
) {
1064 _tdb_transaction_cancel(tdb
);
1068 if (!tdb
->transaction
->prepared
) {
1069 int ret
= _tdb_transaction_prepare_commit(tdb
);
1074 methods
= tdb
->transaction
->io_methods
;
1076 /* perform all the writes */
1077 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
1081 if (tdb
->transaction
->blocks
[i
] == NULL
) {
1085 offset
= i
* tdb
->transaction
->block_size
;
1086 length
= tdb
->transaction
->block_size
;
1087 if (i
== tdb
->transaction
->num_blocks
-1) {
1088 length
= tdb
->transaction
->last_block_size
;
1091 if (methods
->tdb_write(tdb
, offset
, tdb
->transaction
->blocks
[i
], length
) == -1) {
1092 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_commit: write failed during commit\n"));
1094 /* we've overwritten part of the data and
1095 possibly expanded the file, so we need to
1096 run the crash recovery code */
1097 tdb
->methods
= methods
;
1098 tdb_transaction_recover(tdb
);
1100 _tdb_transaction_cancel(tdb
);
1102 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_commit: write failed\n"));
1105 SAFE_FREE(tdb
->transaction
->blocks
[i
]);
1108 /* Do this before we drop lock or blocks. */
1109 if (tdb
->transaction
->expanded
) {
1110 need_repack
= repack_worthwhile(tdb
);
1113 SAFE_FREE(tdb
->transaction
->blocks
);
1114 tdb
->transaction
->num_blocks
= 0;
1116 /* ensure the new data is on disk */
1117 if (transaction_sync(tdb
, 0, tdb
->map_size
) == -1) {
1122 TODO: maybe write to some dummy hdr field, or write to magic
1123 offset without mmap, before the last sync, instead of the
1127 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1128 don't change the mtime of the file, this means the file may
1129 not be backed up (as tdb rounding to block sizes means that
1130 file size changes are quite rare too). The following forces
1131 mtime changes when a transaction completes */
1133 utime(tdb
->name
, NULL
);
1136 /* use a transaction cancel to free memory and remove the
1137 transaction locks */
1138 _tdb_transaction_cancel(tdb
);
1141 return tdb_repack(tdb
);
1149 recover from an aborted transaction. Must be called with exclusive
1150 database write access already established (including the open
1151 lock to prevent new processes attaching)
1153 int tdb_transaction_recover(struct tdb_context
*tdb
)
1155 tdb_off_t recovery_head
, recovery_eof
;
1156 unsigned char *data
, *p
;
1158 struct tdb_record rec
;
1160 /* find the recovery area */
1161 if (tdb_ofs_read(tdb
, TDB_RECOVERY_HEAD
, &recovery_head
) == -1) {
1162 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to read recovery head\n"));
1163 tdb
->ecode
= TDB_ERR_IO
;
1167 if (recovery_head
== 0) {
1168 /* we have never allocated a recovery record */
1172 /* read the recovery record */
1173 if (tdb
->methods
->tdb_read(tdb
, recovery_head
, &rec
,
1174 sizeof(rec
), DOCONV()) == -1) {
1175 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to read recovery record\n"));
1176 tdb
->ecode
= TDB_ERR_IO
;
1180 if (rec
.magic
!= TDB_RECOVERY_MAGIC
) {
1181 /* there is no valid recovery data */
1185 if (tdb
->read_only
) {
1186 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: attempt to recover read only database\n"));
1187 tdb
->ecode
= TDB_ERR_CORRUPT
;
1191 recovery_eof
= rec
.key_len
;
1193 data
= (unsigned char *)malloc(rec
.data_len
);
1195 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to allocate recovery data\n"));
1196 tdb
->ecode
= TDB_ERR_OOM
;
1200 /* read the full recovery data */
1201 if (tdb
->methods
->tdb_read(tdb
, recovery_head
+ sizeof(rec
), data
,
1202 rec
.data_len
, 0) == -1) {
1203 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to read recovery data\n"));
1204 tdb
->ecode
= TDB_ERR_IO
;
1208 /* recover the file data */
1210 while (p
+8 < data
+ rec
.data_len
) {
1216 memcpy(&len
, p
+4, 4);
1218 if (tdb
->methods
->tdb_write(tdb
, ofs
, p
+8, len
) == -1) {
1220 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len
, ofs
));
1221 tdb
->ecode
= TDB_ERR_IO
;
1229 if (transaction_sync(tdb
, 0, tdb
->map_size
) == -1) {
1230 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to sync recovery\n"));
1231 tdb
->ecode
= TDB_ERR_IO
;
1235 /* if the recovery area is after the recovered eof then remove it */
1236 if (recovery_eof
<= recovery_head
) {
1237 if (tdb_ofs_write(tdb
, TDB_RECOVERY_HEAD
, &zero
) == -1) {
1238 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to remove recovery head\n"));
1239 tdb
->ecode
= TDB_ERR_IO
;
1244 /* remove the recovery magic */
1245 if (tdb_ofs_write(tdb
, recovery_head
+ offsetof(struct tdb_record
, magic
),
1247 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to remove recovery magic\n"));
1248 tdb
->ecode
= TDB_ERR_IO
;
1252 if (transaction_sync(tdb
, 0, recovery_eof
) == -1) {
1253 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to sync2 recovery\n"));
1254 tdb
->ecode
= TDB_ERR_IO
;
1258 TDB_LOG((tdb
, TDB_DEBUG_TRACE
, "tdb_transaction_recover: recovered %d byte database\n",
1265 /* Any I/O failures we say "needs recovery". */
1266 bool tdb_needs_recovery(struct tdb_context
*tdb
)
1268 tdb_off_t recovery_head
;
1269 struct tdb_record rec
;
1271 /* find the recovery area */
1272 if (tdb_ofs_read(tdb
, TDB_RECOVERY_HEAD
, &recovery_head
) == -1) {
1276 if (recovery_head
== 0) {
1277 /* we have never allocated a recovery record */
1281 /* read the recovery record */
1282 if (tdb
->methods
->tdb_read(tdb
, recovery_head
, &rec
,
1283 sizeof(rec
), DOCONV()) == -1) {
1287 return (rec
.magic
== TDB_RECOVERY_MAGIC
);