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 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 tdb_add_flags() transaction nesting is enabled.
90 It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
91 The default is that transaction nesting is allowed.
92 Note: this default may change in future versions of tdb.
94 Beware. when transactions are nested a transaction successfully
95 completed with tdb_transaction_commit() can be silently unrolled later.
97 - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
98 tdb_add_flags() transaction nesting is disabled.
99 It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
100 An attempt create a nested transaction will fail with TDB_ERR_NESTING.
101 The default is that transaction nesting is allowed.
102 Note: this default may change in future versions of tdb.
107 hold the context of any current transaction
109 struct tdb_transaction
{
110 /* we keep a mirrored copy of the tdb hash heads here so
111 tdb_next_hash_chain() can operate efficiently */
112 uint32_t *hash_heads
;
114 /* the original io methods - used to do IOs to the real db */
115 const struct tdb_methods
*io_methods
;
117 /* the list of transaction blocks. When a block is first
118 written to, it gets created in this list */
121 uint32_t block_size
; /* bytes in each block */
122 uint32_t last_block_size
; /* number of valid bytes in the last block */
124 /* non-zero when an internal transaction error has
125 occurred. All write operations will then fail until the
126 transaction is ended */
127 int transaction_error
;
129 /* when inside a transaction we need to keep track of any
130 nested tdb_transaction_start() calls, as these are allowed,
131 but don't create a new transaction */
134 /* set when a prepare has already occurred */
136 tdb_off_t magic_offset
;
138 /* old file size before transaction */
139 tdb_len_t old_map_size
;
141 /* did we expand in this transaction */
147 read while in a transaction. We need to check first if the data is in our list
148 of transaction elements, then if not do a real read
150 static int transaction_read(struct tdb_context
*tdb
, tdb_off_t off
, void *buf
,
151 tdb_len_t len
, int cv
)
155 /* break it down into block sized ops */
156 while (len
+ (off
% tdb
->transaction
->block_size
) > tdb
->transaction
->block_size
) {
157 tdb_len_t len2
= tdb
->transaction
->block_size
- (off
% tdb
->transaction
->block_size
);
158 if (transaction_read(tdb
, off
, buf
, len2
, cv
) != 0) {
163 buf
= (void *)(len2
+ (char *)buf
);
170 blk
= off
/ tdb
->transaction
->block_size
;
172 /* see if we have it in the block list */
173 if (tdb
->transaction
->num_blocks
<= blk
||
174 tdb
->transaction
->blocks
[blk
] == NULL
) {
175 /* nope, do a real read */
176 if (tdb
->transaction
->io_methods
->tdb_read(tdb
, off
, buf
, len
, cv
) != 0) {
182 /* it is in the block list. Now check for the last block */
183 if (blk
== tdb
->transaction
->num_blocks
-1) {
184 if (len
> tdb
->transaction
->last_block_size
) {
189 /* now copy it out of this block */
190 memcpy(buf
, tdb
->transaction
->blocks
[blk
] + (off
% tdb
->transaction
->block_size
), len
);
192 tdb_convert(buf
, len
);
197 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "transaction_read: failed at off=%d len=%d\n", off
, len
));
198 tdb
->ecode
= TDB_ERR_IO
;
199 tdb
->transaction
->transaction_error
= 1;
205 write while in a transaction
207 static int transaction_write(struct tdb_context
*tdb
, tdb_off_t off
,
208 const void *buf
, tdb_len_t len
)
212 /* Only a commit is allowed on a prepared transaction */
213 if (tdb
->transaction
->prepared
) {
214 tdb
->ecode
= TDB_ERR_EINVAL
;
215 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "transaction_write: transaction already prepared, write not allowed\n"));
216 tdb
->transaction
->transaction_error
= 1;
220 /* if the write is to a hash head, then update the transaction
222 if (len
== sizeof(tdb_off_t
) && off
>= FREELIST_TOP
&&
223 off
< FREELIST_TOP
+TDB_HASHTABLE_SIZE(tdb
)) {
224 uint32_t chain
= (off
-FREELIST_TOP
) / sizeof(tdb_off_t
);
225 memcpy(&tdb
->transaction
->hash_heads
[chain
], buf
, len
);
228 /* break it up into block sized chunks */
229 while (len
+ (off
% tdb
->transaction
->block_size
) > tdb
->transaction
->block_size
) {
230 tdb_len_t len2
= tdb
->transaction
->block_size
- (off
% tdb
->transaction
->block_size
);
231 if (transaction_write(tdb
, off
, buf
, len2
) != 0) {
237 buf
= (const void *)(len2
+ (const char *)buf
);
245 blk
= off
/ tdb
->transaction
->block_size
;
246 off
= off
% tdb
->transaction
->block_size
;
248 if (tdb
->transaction
->num_blocks
<= blk
) {
249 uint8_t **new_blocks
;
250 /* expand the blocks array */
251 if (tdb
->transaction
->blocks
== NULL
) {
252 new_blocks
= (uint8_t **)malloc(
253 (blk
+1)*sizeof(uint8_t *));
255 new_blocks
= (uint8_t **)realloc(
256 tdb
->transaction
->blocks
,
257 (blk
+1)*sizeof(uint8_t *));
259 if (new_blocks
== NULL
) {
260 tdb
->ecode
= TDB_ERR_OOM
;
263 memset(&new_blocks
[tdb
->transaction
->num_blocks
], 0,
264 (1+(blk
- tdb
->transaction
->num_blocks
))*sizeof(uint8_t *));
265 tdb
->transaction
->blocks
= new_blocks
;
266 tdb
->transaction
->num_blocks
= blk
+1;
267 tdb
->transaction
->last_block_size
= 0;
270 /* allocate and fill a block? */
271 if (tdb
->transaction
->blocks
[blk
] == NULL
) {
272 tdb
->transaction
->blocks
[blk
] = (uint8_t *)calloc(tdb
->transaction
->block_size
, 1);
273 if (tdb
->transaction
->blocks
[blk
] == NULL
) {
274 tdb
->ecode
= TDB_ERR_OOM
;
275 tdb
->transaction
->transaction_error
= 1;
278 if (tdb
->transaction
->old_map_size
> blk
* tdb
->transaction
->block_size
) {
279 tdb_len_t len2
= tdb
->transaction
->block_size
;
280 if (len2
+ (blk
* tdb
->transaction
->block_size
) > tdb
->transaction
->old_map_size
) {
281 len2
= tdb
->transaction
->old_map_size
- (blk
* tdb
->transaction
->block_size
);
283 if (tdb
->transaction
->io_methods
->tdb_read(tdb
, blk
* tdb
->transaction
->block_size
,
284 tdb
->transaction
->blocks
[blk
],
286 SAFE_FREE(tdb
->transaction
->blocks
[blk
]);
287 tdb
->ecode
= TDB_ERR_IO
;
290 if (blk
== tdb
->transaction
->num_blocks
-1) {
291 tdb
->transaction
->last_block_size
= len2
;
296 /* overwrite part of an existing block */
298 memset(tdb
->transaction
->blocks
[blk
] + off
, 0, len
);
300 memcpy(tdb
->transaction
->blocks
[blk
] + off
, buf
, len
);
302 if (blk
== tdb
->transaction
->num_blocks
-1) {
303 if (len
+ off
> tdb
->transaction
->last_block_size
) {
304 tdb
->transaction
->last_block_size
= len
+ off
;
311 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "transaction_write: failed at off=%d len=%d\n",
312 (blk
*tdb
->transaction
->block_size
) + off
, len
));
313 tdb
->transaction
->transaction_error
= 1;
319 write while in a transaction - this varient never expands the transaction blocks, it only
320 updates existing blocks. This means it cannot change the recovery size
322 static int transaction_write_existing(struct tdb_context
*tdb
, tdb_off_t off
,
323 const void *buf
, tdb_len_t len
)
327 /* break it up into block sized chunks */
328 while (len
+ (off
% tdb
->transaction
->block_size
) > tdb
->transaction
->block_size
) {
329 tdb_len_t len2
= tdb
->transaction
->block_size
- (off
% tdb
->transaction
->block_size
);
330 if (transaction_write_existing(tdb
, off
, buf
, len2
) != 0) {
336 buf
= (const void *)(len2
+ (const char *)buf
);
344 blk
= off
/ tdb
->transaction
->block_size
;
345 off
= off
% tdb
->transaction
->block_size
;
347 if (tdb
->transaction
->num_blocks
<= blk
||
348 tdb
->transaction
->blocks
[blk
] == NULL
) {
352 if (blk
== tdb
->transaction
->num_blocks
-1 &&
353 off
+ len
> tdb
->transaction
->last_block_size
) {
354 if (off
>= tdb
->transaction
->last_block_size
) {
357 len
= tdb
->transaction
->last_block_size
- off
;
360 /* overwrite part of an existing block */
361 memcpy(tdb
->transaction
->blocks
[blk
] + off
, buf
, len
);
368 accelerated hash chain head search, using the cached hash heads
370 static void transaction_next_hash_chain(struct tdb_context
*tdb
, uint32_t *chain
)
373 for (;h
< tdb
->header
.hash_size
;h
++) {
374 /* the +1 takes account of the freelist */
375 if (0 != tdb
->transaction
->hash_heads
[h
+1]) {
383 out of bounds check during a transaction
385 static int transaction_oob(struct tdb_context
*tdb
, tdb_off_t len
, int probe
)
387 if (len
<= tdb
->map_size
) {
390 tdb
->ecode
= TDB_ERR_IO
;
395 transaction version of tdb_expand().
397 static int transaction_expand_file(struct tdb_context
*tdb
, tdb_off_t size
,
400 /* add a write to the transaction elements, so subsequent
401 reads see the zero data */
402 if (transaction_write(tdb
, size
, NULL
, addition
) != 0) {
406 tdb
->transaction
->expanded
= true;
411 static const struct tdb_methods transaction_methods
= {
414 transaction_next_hash_chain
,
416 transaction_expand_file
,
421 start a tdb transaction. No token is returned, as only a single
422 transaction is allowed to be pending per tdb_context
424 static int _tdb_transaction_start(struct tdb_context
*tdb
,
425 enum tdb_lock_flags lockflags
)
427 /* some sanity checks */
428 if (tdb
->read_only
|| (tdb
->flags
& TDB_INTERNAL
) || tdb
->traverse_read
) {
429 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
430 tdb
->ecode
= TDB_ERR_EINVAL
;
434 /* cope with nested tdb_transaction_start() calls */
435 if (tdb
->transaction
!= NULL
) {
436 if (!(tdb
->flags
& TDB_ALLOW_NESTING
)) {
437 tdb
->ecode
= TDB_ERR_NESTING
;
440 tdb
->transaction
->nesting
++;
441 TDB_LOG((tdb
, TDB_DEBUG_TRACE
, "tdb_transaction_start: nesting %d\n",
442 tdb
->transaction
->nesting
));
446 if (tdb_have_extra_locks(tdb
)) {
447 /* the caller must not have any locks when starting a
448 transaction as otherwise we'll be screwed by lack
449 of nested locks in posix */
450 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_start: cannot start a transaction with locks held\n"));
451 tdb
->ecode
= TDB_ERR_LOCK
;
455 if (tdb
->travlocks
.next
!= NULL
) {
456 /* you cannot use transactions inside a traverse (although you can use
457 traverse inside a transaction) as otherwise you can end up with
459 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
460 tdb
->ecode
= TDB_ERR_LOCK
;
464 tdb
->transaction
= (struct tdb_transaction
*)
465 calloc(sizeof(struct tdb_transaction
), 1);
466 if (tdb
->transaction
== NULL
) {
467 tdb
->ecode
= TDB_ERR_OOM
;
471 /* a page at a time seems like a reasonable compromise between compactness and efficiency */
472 tdb
->transaction
->block_size
= tdb
->page_size
;
474 /* get the transaction write lock. This is a blocking lock. As
475 discussed with Volker, there are a number of ways we could
476 make this async, which we will probably do in the future */
477 if (tdb_transaction_lock(tdb
, F_WRLCK
, lockflags
) == -1) {
478 SAFE_FREE(tdb
->transaction
->blocks
);
479 SAFE_FREE(tdb
->transaction
);
480 if ((lockflags
& TDB_LOCK_WAIT
) == 0) {
481 tdb
->ecode
= TDB_ERR_NOLOCK
;
486 /* get a read lock from the freelist to the end of file. This
487 is upgraded to a write lock during the commit */
488 if (tdb_allrecord_lock(tdb
, F_RDLCK
, TDB_LOCK_WAIT
, true) == -1) {
489 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_start: failed to get hash locks\n"));
490 goto fail_allrecord_lock
;
493 /* setup a copy of the hash table heads so the hash scan in
494 traverse can be fast */
495 tdb
->transaction
->hash_heads
= (uint32_t *)
496 calloc(tdb
->header
.hash_size
+1, sizeof(uint32_t));
497 if (tdb
->transaction
->hash_heads
== NULL
) {
498 tdb
->ecode
= TDB_ERR_OOM
;
501 if (tdb
->methods
->tdb_read(tdb
, FREELIST_TOP
, tdb
->transaction
->hash_heads
,
502 TDB_HASHTABLE_SIZE(tdb
), 0) != 0) {
503 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_start: failed to read hash heads\n"));
504 tdb
->ecode
= TDB_ERR_IO
;
508 /* make sure we know about any file expansions already done by
510 tdb
->methods
->tdb_oob(tdb
, tdb
->map_size
+ 1, 1);
511 tdb
->transaction
->old_map_size
= tdb
->map_size
;
513 /* finally hook the io methods, replacing them with
514 transaction specific methods */
515 tdb
->transaction
->io_methods
= tdb
->methods
;
516 tdb
->methods
= &transaction_methods
;
518 /* Trace at the end, so we get sequence number correct. */
519 tdb_trace(tdb
, "tdb_transaction_start");
523 tdb_allrecord_unlock(tdb
, F_RDLCK
, false);
525 tdb_transaction_unlock(tdb
, F_WRLCK
);
526 SAFE_FREE(tdb
->transaction
->blocks
);
527 SAFE_FREE(tdb
->transaction
->hash_heads
);
528 SAFE_FREE(tdb
->transaction
);
532 _PUBLIC_
int tdb_transaction_start(struct tdb_context
*tdb
)
534 return _tdb_transaction_start(tdb
, TDB_LOCK_WAIT
);
537 _PUBLIC_
int tdb_transaction_start_nonblock(struct tdb_context
*tdb
)
539 return _tdb_transaction_start(tdb
, TDB_LOCK_NOWAIT
|TDB_LOCK_PROBE
);
545 static int transaction_sync(struct tdb_context
*tdb
, tdb_off_t offset
, tdb_len_t length
)
547 if (tdb
->flags
& TDB_NOSYNC
) {
551 #ifdef HAVE_FDATASYNC
552 if (fdatasync(tdb
->fd
) != 0) {
554 if (fsync(tdb
->fd
) != 0) {
556 tdb
->ecode
= TDB_ERR_IO
;
557 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction: fsync failed\n"));
562 tdb_off_t moffset
= offset
& ~(tdb
->page_size
-1);
563 if (msync(moffset
+ (char *)tdb
->map_ptr
,
564 length
+ (offset
- moffset
), MS_SYNC
) != 0) {
565 tdb
->ecode
= TDB_ERR_IO
;
566 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction: msync failed - %s\n",
576 static int _tdb_transaction_cancel(struct tdb_context
*tdb
)
580 if (tdb
->transaction
== NULL
) {
581 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_cancel: no transaction\n"));
585 if (tdb
->transaction
->nesting
!= 0) {
586 tdb
->transaction
->transaction_error
= 1;
587 tdb
->transaction
->nesting
--;
591 tdb
->map_size
= tdb
->transaction
->old_map_size
;
593 /* free all the transaction blocks */
594 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
595 if (tdb
->transaction
->blocks
[i
] != NULL
) {
596 free(tdb
->transaction
->blocks
[i
]);
599 SAFE_FREE(tdb
->transaction
->blocks
);
601 if (tdb
->transaction
->magic_offset
) {
602 const struct tdb_methods
*methods
= tdb
->transaction
->io_methods
;
603 const uint32_t invalid
= TDB_RECOVERY_INVALID_MAGIC
;
605 /* remove the recovery marker */
606 if (methods
->tdb_write(tdb
, tdb
->transaction
->magic_offset
, &invalid
, 4) == -1 ||
607 transaction_sync(tdb
, tdb
->transaction
->magic_offset
, 4) == -1) {
608 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_cancel: failed to remove recovery magic\n"));
613 /* This also removes the OPEN_LOCK, if we have it. */
614 tdb_release_transaction_locks(tdb
);
616 /* restore the normal io methods */
617 tdb
->methods
= tdb
->transaction
->io_methods
;
619 SAFE_FREE(tdb
->transaction
->hash_heads
);
620 SAFE_FREE(tdb
->transaction
);
626 cancel the current transaction
628 _PUBLIC_
int tdb_transaction_cancel(struct tdb_context
*tdb
)
630 tdb_trace(tdb
, "tdb_transaction_cancel");
631 return _tdb_transaction_cancel(tdb
);
635 work out how much space the linearised recovery data will consume
637 static tdb_len_t
tdb_recovery_size(struct tdb_context
*tdb
)
639 tdb_len_t recovery_size
= 0;
642 recovery_size
= sizeof(uint32_t);
643 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
644 if (i
* tdb
->transaction
->block_size
>= tdb
->transaction
->old_map_size
) {
647 if (tdb
->transaction
->blocks
[i
] == NULL
) {
650 recovery_size
+= 2*sizeof(tdb_off_t
);
651 if (i
== tdb
->transaction
->num_blocks
-1) {
652 recovery_size
+= tdb
->transaction
->last_block_size
;
654 recovery_size
+= tdb
->transaction
->block_size
;
658 return recovery_size
;
661 int tdb_recovery_area(struct tdb_context
*tdb
,
662 const struct tdb_methods
*methods
,
663 tdb_off_t
*recovery_offset
,
664 struct tdb_record
*rec
)
666 if (tdb_ofs_read(tdb
, TDB_RECOVERY_HEAD
, recovery_offset
) == -1) {
670 if (*recovery_offset
== 0) {
675 if (methods
->tdb_read(tdb
, *recovery_offset
, rec
, sizeof(*rec
),
680 /* ignore invalid recovery regions: can happen in crash */
681 if (rec
->magic
!= TDB_RECOVERY_MAGIC
&&
682 rec
->magic
!= TDB_RECOVERY_INVALID_MAGIC
) {
683 *recovery_offset
= 0;
690 allocate the recovery area, or use an existing recovery area if it is
693 static int tdb_recovery_allocate(struct tdb_context
*tdb
,
694 tdb_len_t
*recovery_size
,
695 tdb_off_t
*recovery_offset
,
696 tdb_len_t
*recovery_max_size
)
698 struct tdb_record rec
;
699 const struct tdb_methods
*methods
= tdb
->transaction
->io_methods
;
700 tdb_off_t recovery_head
;
702 if (tdb_recovery_area(tdb
, methods
, &recovery_head
, &rec
) == -1) {
703 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to read recovery head\n"));
707 *recovery_size
= tdb_recovery_size(tdb
);
709 if (recovery_head
!= 0 && *recovery_size
<= rec
.rec_len
) {
710 /* it fits in the existing area */
711 *recovery_max_size
= rec
.rec_len
;
712 *recovery_offset
= recovery_head
;
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
721 if (recovery_head
!= 0) {
722 if (tdb_free(tdb
, recovery_head
, &rec
) == -1) {
723 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to free previous recovery area\n"));
728 /* the tdb_free() call might have increased the recovery size */
729 *recovery_size
= tdb_recovery_size(tdb
);
731 /* round up to a multiple of page size */
732 *recovery_max_size
= TDB_ALIGN(sizeof(rec
) + *recovery_size
, tdb
->page_size
) - sizeof(rec
);
733 *recovery_offset
= tdb
->map_size
;
734 recovery_head
= *recovery_offset
;
736 if (methods
->tdb_expand_file(tdb
, tdb
->transaction
->old_map_size
,
737 (tdb
->map_size
- tdb
->transaction
->old_map_size
) +
738 sizeof(rec
) + *recovery_max_size
) == -1) {
739 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to create recovery area\n"));
743 /* remap the file (if using mmap) */
744 methods
->tdb_oob(tdb
, tdb
->map_size
+ 1, 1);
746 /* we have to reset the old map size so that we don't try to expand the file
747 again in the transaction commit, which would destroy the recovery area */
748 tdb
->transaction
->old_map_size
= tdb
->map_size
;
750 /* write the recovery header offset and sync - we can sync without a race here
751 as the magic ptr in the recovery record has not been set */
752 CONVERT(recovery_head
);
753 if (methods
->tdb_write(tdb
, TDB_RECOVERY_HEAD
,
754 &recovery_head
, sizeof(tdb_off_t
)) == -1) {
755 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to write recovery head\n"));
758 if (transaction_write_existing(tdb
, TDB_RECOVERY_HEAD
, &recovery_head
, sizeof(tdb_off_t
)) == -1) {
759 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_recovery_allocate: failed to write recovery head\n"));
768 setup the recovery data that will be used on a crash during commit
770 static int transaction_setup_recovery(struct tdb_context
*tdb
,
771 tdb_off_t
*magic_offset
)
773 tdb_len_t recovery_size
;
774 unsigned char *data
, *p
;
775 const struct tdb_methods
*methods
= tdb
->transaction
->io_methods
;
776 struct tdb_record
*rec
;
777 tdb_off_t recovery_offset
, recovery_max_size
;
778 tdb_off_t old_map_size
= tdb
->transaction
->old_map_size
;
779 uint32_t magic
, tailer
;
783 check that the recovery area has enough space
785 if (tdb_recovery_allocate(tdb
, &recovery_size
,
786 &recovery_offset
, &recovery_max_size
) == -1) {
790 data
= (unsigned char *)malloc(recovery_size
+ sizeof(*rec
));
792 tdb
->ecode
= TDB_ERR_OOM
;
796 rec
= (struct tdb_record
*)data
;
797 memset(rec
, 0, sizeof(*rec
));
799 rec
->magic
= TDB_RECOVERY_INVALID_MAGIC
;
800 rec
->data_len
= recovery_size
;
801 rec
->rec_len
= recovery_max_size
;
802 rec
->key_len
= old_map_size
;
805 /* build the recovery data into a single blob to allow us to do a single
806 large write, which should be more efficient */
807 p
= data
+ sizeof(*rec
);
808 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
812 if (tdb
->transaction
->blocks
[i
] == NULL
) {
816 offset
= i
* tdb
->transaction
->block_size
;
817 length
= tdb
->transaction
->block_size
;
818 if (i
== tdb
->transaction
->num_blocks
-1) {
819 length
= tdb
->transaction
->last_block_size
;
822 if (offset
>= old_map_size
) {
825 if (offset
+ length
> tdb
->transaction
->old_map_size
) {
826 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
828 tdb
->ecode
= TDB_ERR_CORRUPT
;
831 memcpy(p
, &offset
, 4);
832 memcpy(p
+4, &length
, 4);
836 /* the recovery area contains the old data, not the
837 new data, so we have to call the original tdb_read
839 if (methods
->tdb_read(tdb
, offset
, p
+ 8, length
, 0) != 0) {
841 tdb
->ecode
= TDB_ERR_IO
;
848 tailer
= sizeof(*rec
) + recovery_max_size
;
849 memcpy(p
, &tailer
, 4);
854 /* write the recovery data to the recovery area */
855 if (methods
->tdb_write(tdb
, recovery_offset
, data
, sizeof(*rec
) + recovery_size
) == -1) {
856 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
858 tdb
->ecode
= TDB_ERR_IO
;
861 if (transaction_write_existing(tdb
, recovery_offset
, data
, sizeof(*rec
) + recovery_size
) == -1) {
862 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
864 tdb
->ecode
= TDB_ERR_IO
;
868 /* as we don't have ordered writes, we have to sync the recovery
869 data before we update the magic to indicate that the recovery
871 if (transaction_sync(tdb
, recovery_offset
, sizeof(*rec
) + recovery_size
) == -1) {
878 magic
= TDB_RECOVERY_MAGIC
;
881 *magic_offset
= recovery_offset
+ offsetof(struct tdb_record
, magic
);
883 if (methods
->tdb_write(tdb
, *magic_offset
, &magic
, sizeof(magic
)) == -1) {
884 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
885 tdb
->ecode
= TDB_ERR_IO
;
888 if (transaction_write_existing(tdb
, *magic_offset
, &magic
, sizeof(magic
)) == -1) {
889 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
890 tdb
->ecode
= TDB_ERR_IO
;
894 /* ensure the recovery magic marker is on disk */
895 if (transaction_sync(tdb
, *magic_offset
, sizeof(magic
)) == -1) {
902 static int _tdb_transaction_prepare_commit(struct tdb_context
*tdb
)
904 const struct tdb_methods
*methods
;
906 if (tdb
->transaction
== NULL
) {
907 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: no transaction\n"));
911 if (tdb
->transaction
->prepared
) {
912 tdb
->ecode
= TDB_ERR_EINVAL
;
913 _tdb_transaction_cancel(tdb
);
914 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: transaction already prepared\n"));
918 if (tdb
->transaction
->transaction_error
) {
919 tdb
->ecode
= TDB_ERR_IO
;
920 _tdb_transaction_cancel(tdb
);
921 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: transaction error pending\n"));
926 if (tdb
->transaction
->nesting
!= 0) {
930 /* check for a null transaction */
931 if (tdb
->transaction
->blocks
== NULL
) {
935 methods
= tdb
->transaction
->io_methods
;
937 /* if there are any locks pending then the caller has not
938 nested their locks properly, so fail the transaction */
939 if (tdb_have_extra_locks(tdb
)) {
940 tdb
->ecode
= TDB_ERR_LOCK
;
941 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: locks pending on commit\n"));
942 _tdb_transaction_cancel(tdb
);
946 /* upgrade the main transaction lock region to a write lock */
947 if (tdb_allrecord_upgrade(tdb
) == -1) {
948 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
949 _tdb_transaction_cancel(tdb
);
953 /* get the open lock - this prevents new users attaching to the database
955 if (tdb_nest_lock(tdb
, OPEN_LOCK
, F_WRLCK
, TDB_LOCK_WAIT
) == -1) {
956 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_prepare_commit: failed to get open lock\n"));
957 _tdb_transaction_cancel(tdb
);
961 if (!(tdb
->flags
& TDB_NOSYNC
)) {
962 /* write the recovery data to the end of the file */
963 if (transaction_setup_recovery(tdb
, &tdb
->transaction
->magic_offset
) == -1) {
964 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
965 _tdb_transaction_cancel(tdb
);
970 tdb
->transaction
->prepared
= true;
972 /* expand the file to the new size if needed */
973 if (tdb
->map_size
!= tdb
->transaction
->old_map_size
) {
974 if (methods
->tdb_expand_file(tdb
, tdb
->transaction
->old_map_size
,
976 tdb
->transaction
->old_map_size
) == -1) {
977 tdb
->ecode
= TDB_ERR_IO
;
978 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_prepare_commit: expansion failed\n"));
979 _tdb_transaction_cancel(tdb
);
982 tdb
->map_size
= tdb
->transaction
->old_map_size
;
983 methods
->tdb_oob(tdb
, tdb
->map_size
+ 1, 1);
986 /* Keep the open lock until the actual commit */
992 prepare to commit the current transaction
994 _PUBLIC_
int tdb_transaction_prepare_commit(struct tdb_context
*tdb
)
996 tdb_trace(tdb
, "tdb_transaction_prepare_commit");
997 return _tdb_transaction_prepare_commit(tdb
);
1000 /* A repack is worthwhile if the largest is less than half total free. */
1001 static bool repack_worthwhile(struct tdb_context
*tdb
)
1004 struct tdb_record rec
;
1005 tdb_len_t total
= 0, largest
= 0;
1007 if (tdb_ofs_read(tdb
, FREELIST_TOP
, &ptr
) == -1) {
1011 while (ptr
!= 0 && tdb_rec_free_read(tdb
, ptr
, &rec
) == 0) {
1012 total
+= rec
.rec_len
;
1013 if (rec
.rec_len
> largest
) {
1014 largest
= rec
.rec_len
;
1019 return total
> largest
* 2;
1023 commit the current transaction
1025 _PUBLIC_
int tdb_transaction_commit(struct tdb_context
*tdb
)
1027 const struct tdb_methods
*methods
;
1029 bool need_repack
= false;
1031 if (tdb
->transaction
== NULL
) {
1032 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_commit: no transaction\n"));
1036 tdb_trace(tdb
, "tdb_transaction_commit");
1038 if (tdb
->transaction
->transaction_error
) {
1039 tdb
->ecode
= TDB_ERR_IO
;
1040 _tdb_transaction_cancel(tdb
);
1041 TDB_LOG((tdb
, TDB_DEBUG_ERROR
, "tdb_transaction_commit: transaction error pending\n"));
1046 if (tdb
->transaction
->nesting
!= 0) {
1047 tdb
->transaction
->nesting
--;
1051 /* check for a null transaction */
1052 if (tdb
->transaction
->blocks
== NULL
) {
1053 _tdb_transaction_cancel(tdb
);
1057 if (!tdb
->transaction
->prepared
) {
1058 int ret
= _tdb_transaction_prepare_commit(tdb
);
1063 methods
= tdb
->transaction
->io_methods
;
1065 /* perform all the writes */
1066 for (i
=0;i
<tdb
->transaction
->num_blocks
;i
++) {
1070 if (tdb
->transaction
->blocks
[i
] == NULL
) {
1074 offset
= i
* tdb
->transaction
->block_size
;
1075 length
= tdb
->transaction
->block_size
;
1076 if (i
== tdb
->transaction
->num_blocks
-1) {
1077 length
= tdb
->transaction
->last_block_size
;
1080 if (methods
->tdb_write(tdb
, offset
, tdb
->transaction
->blocks
[i
], length
) == -1) {
1081 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_commit: write failed during commit\n"));
1083 /* we've overwritten part of the data and
1084 possibly expanded the file, so we need to
1085 run the crash recovery code */
1086 tdb
->methods
= methods
;
1087 tdb_transaction_recover(tdb
);
1089 _tdb_transaction_cancel(tdb
);
1091 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_commit: write failed\n"));
1094 SAFE_FREE(tdb
->transaction
->blocks
[i
]);
1097 /* Do this before we drop lock or blocks. */
1098 if (tdb
->transaction
->expanded
) {
1099 need_repack
= repack_worthwhile(tdb
);
1102 SAFE_FREE(tdb
->transaction
->blocks
);
1103 tdb
->transaction
->num_blocks
= 0;
1105 /* ensure the new data is on disk */
1106 if (transaction_sync(tdb
, 0, tdb
->map_size
) == -1) {
1111 TODO: maybe write to some dummy hdr field, or write to magic
1112 offset without mmap, before the last sync, instead of the
1116 /* on some systems (like Linux 2.6.x) changes via mmap/msync
1117 don't change the mtime of the file, this means the file may
1118 not be backed up (as tdb rounding to block sizes means that
1119 file size changes are quite rare too). The following forces
1120 mtime changes when a transaction completes */
1122 utime(tdb
->name
, NULL
);
1125 /* use a transaction cancel to free memory and remove the
1126 transaction locks */
1127 _tdb_transaction_cancel(tdb
);
1130 return tdb_repack(tdb
);
1138 recover from an aborted transaction. Must be called with exclusive
1139 database write access already established (including the open
1140 lock to prevent new processes attaching)
1142 int tdb_transaction_recover(struct tdb_context
*tdb
)
1144 tdb_off_t recovery_head
, recovery_eof
;
1145 unsigned char *data
, *p
;
1147 struct tdb_record rec
;
1149 /* find the recovery area */
1150 if (tdb_ofs_read(tdb
, TDB_RECOVERY_HEAD
, &recovery_head
) == -1) {
1151 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to read recovery head\n"));
1152 tdb
->ecode
= TDB_ERR_IO
;
1156 if (recovery_head
== 0) {
1157 /* we have never allocated a recovery record */
1161 /* read the recovery record */
1162 if (tdb
->methods
->tdb_read(tdb
, recovery_head
, &rec
,
1163 sizeof(rec
), DOCONV()) == -1) {
1164 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to read recovery record\n"));
1165 tdb
->ecode
= TDB_ERR_IO
;
1169 if (rec
.magic
!= TDB_RECOVERY_MAGIC
) {
1170 /* there is no valid recovery data */
1174 if (tdb
->read_only
) {
1175 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: attempt to recover read only database\n"));
1176 tdb
->ecode
= TDB_ERR_CORRUPT
;
1180 recovery_eof
= rec
.key_len
;
1182 data
= (unsigned char *)malloc(rec
.data_len
);
1184 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to allocate recovery data\n"));
1185 tdb
->ecode
= TDB_ERR_OOM
;
1189 /* read the full recovery data */
1190 if (tdb
->methods
->tdb_read(tdb
, recovery_head
+ sizeof(rec
), data
,
1191 rec
.data_len
, 0) == -1) {
1192 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to read recovery data\n"));
1193 tdb
->ecode
= TDB_ERR_IO
;
1197 /* recover the file data */
1199 while (p
+8 < data
+ rec
.data_len
) {
1205 memcpy(&len
, p
+4, 4);
1207 if (tdb
->methods
->tdb_write(tdb
, ofs
, p
+8, len
) == -1) {
1209 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len
, ofs
));
1210 tdb
->ecode
= TDB_ERR_IO
;
1218 if (transaction_sync(tdb
, 0, tdb
->map_size
) == -1) {
1219 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to sync recovery\n"));
1220 tdb
->ecode
= TDB_ERR_IO
;
1224 /* if the recovery area is after the recovered eof then remove it */
1225 if (recovery_eof
<= recovery_head
) {
1226 if (tdb_ofs_write(tdb
, TDB_RECOVERY_HEAD
, &zero
) == -1) {
1227 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to remove recovery head\n"));
1228 tdb
->ecode
= TDB_ERR_IO
;
1233 /* remove the recovery magic */
1234 if (tdb_ofs_write(tdb
, recovery_head
+ offsetof(struct tdb_record
, magic
),
1236 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to remove recovery magic\n"));
1237 tdb
->ecode
= TDB_ERR_IO
;
1241 if (transaction_sync(tdb
, 0, recovery_eof
) == -1) {
1242 TDB_LOG((tdb
, TDB_DEBUG_FATAL
, "tdb_transaction_recover: failed to sync2 recovery\n"));
1243 tdb
->ecode
= TDB_ERR_IO
;
1247 TDB_LOG((tdb
, TDB_DEBUG_TRACE
, "tdb_transaction_recover: recovered %d byte database\n",
1254 /* Any I/O failures we say "needs recovery". */
1255 bool tdb_needs_recovery(struct tdb_context
*tdb
)
1257 tdb_off_t recovery_head
;
1258 struct tdb_record rec
;
1260 /* find the recovery area */
1261 if (tdb_ofs_read(tdb
, TDB_RECOVERY_HEAD
, &recovery_head
) == -1) {
1265 if (recovery_head
== 0) {
1266 /* we have never allocated a recovery record */
1270 /* read the recovery record */
1271 if (tdb
->methods
->tdb_read(tdb
, recovery_head
, &rec
,
1272 sizeof(rec
), DOCONV()) == -1) {
1276 return (rec
.magic
== TDB_RECOVERY_MAGIC
);