Merge branch 'master' of ssh://git.samba.org/data/git/samba
[samba.git] / lib / tdb / common / transaction.c
1  /* 
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              2005
7
8      ** NOTE! The following LGPL license applies to the tdb
9      ** library. This does NOT imply that all of Samba is released
10      ** under the LGPL
11    
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.
16
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.
21
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/>.
24 */
25
26 #include "tdb_private.h"
27
28 /*
29   transaction design:
30
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
35
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
40
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
45
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
50     to the real database.
51
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)
55
56   - if the caller gains a lock during the transaction but doesn't
57     release it then fail the commit
58
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
62  
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
66
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
70
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.
77
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
82     intervention.
83
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.
87
88 */
89
90
91 /*
92   hold the context of any current transaction
93 */
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 */
97         uint32_t *hash_heads;
98
99         /* the original io methods - used to do IOs to the real db */
100         const struct tdb_methods *io_methods;
101
102         /* the list of transaction blocks. When a block is first
103            written to, it gets created in this list */
104         uint8_t **blocks;
105         uint32_t num_blocks;
106         uint32_t block_size;      /* bytes in each block */
107         uint32_t last_block_size; /* number of valid bytes in the last block */
108
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;
113
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 */
117         int nesting;
118
119         /* set when a prepare has already occurred */
120         bool prepared;
121         tdb_off_t magic_offset;
122
123         /* old file size before transaction */
124         tdb_len_t old_map_size;
125 };
126
127
128 /*
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
131 */
132 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
133                             tdb_len_t len, int cv)
134 {
135         uint32_t blk;
136
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;
142                 return -1;
143         }
144
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) {
149                         return -1;
150                 }
151                 len -= len2;
152                 off += len2;
153                 buf = (void *)(len2 + (char *)buf);
154         }
155
156         if (len == 0) {
157                 return 0;
158         }
159
160         blk = off / tdb->transaction->block_size;
161
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) {
167                         goto fail;
168                 }
169                 return 0;
170         }
171
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) {
175                         goto fail;
176                 }
177         }
178         
179         /* now copy it out of this block */
180         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
181         if (cv) {
182                 tdb_convert(buf, len);
183         }
184         return 0;
185
186 fail:
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;
190         return -1;
191 }
192
193
194 /*
195   write while in a transaction
196 */
197 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
198                              const void *buf, tdb_len_t len)
199 {
200         uint32_t blk;
201
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;
207                 return -1;
208         }
209
210         /* if the write is to a hash head, then update the transaction
211            hash heads */
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);
216         }
217
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) {
222                         return -1;
223                 }
224                 len -= len2;
225                 off += len2;
226                 if (buf != NULL) {
227                         buf = (const void *)(len2 + (const char *)buf);
228                 }
229         }
230
231         if (len == 0) {
232                 return 0;
233         }
234
235         blk = off / tdb->transaction->block_size;
236         off = off % tdb->transaction->block_size;
237
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 *));
244                 } else {
245                         new_blocks = (uint8_t **)realloc(
246                                 tdb->transaction->blocks,
247                                 (blk+1)*sizeof(uint8_t *));
248                 }
249                 if (new_blocks == NULL) {
250                         tdb->ecode = TDB_ERR_OOM;
251                         goto fail;
252                 }
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;
258         }
259
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;
266                         return -1;                      
267                 }
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);
272                         }
273                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
274                                                                    tdb->transaction->blocks[blk], 
275                                                                    len2, 0) != 0) {
276                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
277                                 tdb->ecode = TDB_ERR_IO;
278                                 goto fail;
279                         }
280                         if (blk == tdb->transaction->num_blocks-1) {
281                                 tdb->transaction->last_block_size = len2;
282                         }                       
283                 }
284         }
285         
286         /* overwrite part of an existing block */
287         if (buf == NULL) {
288                 memset(tdb->transaction->blocks[blk] + off, 0, len);
289         } else {
290                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
291         }
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;
295                 }
296         }
297
298         return 0;
299
300 fail:
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;
304         return -1;
305 }
306
307
308 /*
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
311 */
312 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
313                                       const void *buf, tdb_len_t len)
314 {
315         uint32_t blk;
316
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) {
321                         return -1;
322                 }
323                 len -= len2;
324                 off += len2;
325                 if (buf != NULL) {
326                         buf = (const void *)(len2 + (const char *)buf);
327                 }
328         }
329
330         if (len == 0) {
331                 return 0;
332         }
333
334         blk = off / tdb->transaction->block_size;
335         off = off % tdb->transaction->block_size;
336
337         if (tdb->transaction->num_blocks <= blk ||
338             tdb->transaction->blocks[blk] == NULL) {
339                 return 0;
340         }
341
342         if (blk == tdb->transaction->num_blocks-1 &&
343             off + len > tdb->transaction->last_block_size) {
344                 if (off >= tdb->transaction->last_block_size) {
345                         return 0;
346                 }
347                 len = tdb->transaction->last_block_size - off;
348         }
349
350         /* overwrite part of an existing block */
351         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
352
353         return 0;
354 }
355
356
357 /*
358   accelerated hash chain head search, using the cached hash heads
359 */
360 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
361 {
362         uint32_t h = *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]) {
366                         break;
367                 }
368         }
369         (*chain) = h;
370 }
371
372 /*
373   out of bounds check during a transaction
374 */
375 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
376 {
377         if (len <= tdb->map_size) {
378                 return 0;
379         }
380         return TDB_ERRCODE(TDB_ERR_IO, -1);
381 }
382
383 /*
384   transaction version of tdb_expand().
385 */
386 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
387                                    tdb_off_t addition)
388 {
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) {
392                 return -1;
393         }
394
395         return 0;
396 }
397
398 /*
399   brlock during a transaction - ignore them
400 */
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)
403 {
404         return 0;
405 }
406
407 static const struct tdb_methods transaction_methods = {
408         transaction_read,
409         transaction_write,
410         transaction_next_hash_chain,
411         transaction_oob,
412         transaction_expand_file,
413         transaction_brlock
414 };
415
416
417 /*
418   start a tdb transaction. No token is returned, as only a single
419   transaction is allowed to be pending per tdb_context
420 */
421 int tdb_transaction_start(struct tdb_context *tdb)
422 {
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;
427                 return -1;
428         }
429
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));
435                 return 0;
436         }
437
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;
444                 return -1;
445         }
446
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
450                    deadlock */
451                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
452                 tdb->ecode = TDB_ERR_LOCK;
453                 return -1;
454         }
455
456         tdb->transaction = (struct tdb_transaction *)
457                 calloc(sizeof(struct tdb_transaction), 1);
458         if (tdb->transaction == NULL) {
459                 tdb->ecode = TDB_ERR_OOM;
460                 return -1;
461         }
462
463         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
464         tdb->transaction->block_size = tdb->page_size;
465
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);
472                 return -1;
473         }
474         
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;
480                 goto fail;
481         }
482
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;
489                 goto fail;
490         }
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;
495                 goto fail;
496         }
497
498         /* make sure we know about any file expansions already done by
499            anyone else */
500         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
501         tdb->transaction->old_map_size = tdb->map_size;
502
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;
507
508         return 0;
509         
510 fail:
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);
516         return -1;
517 }
518
519
520 /*
521   sync to disk
522 */
523 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
524 {       
525         if (tdb->flags & TDB_NOSYNC) {
526                 return 0;
527         }
528
529         if (fsync(tdb->fd) != 0) {
530                 tdb->ecode = TDB_ERR_IO;
531                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
532                 return -1;
533         }
534 #ifdef HAVE_MMAP
535         if (tdb->map_ptr) {
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",
541                                  strerror(errno)));
542                         return -1;
543                 }
544         }
545 #endif
546         return 0;
547 }
548
549
550 /*
551   cancel the current transaction
552 */
553 int tdb_transaction_cancel(struct tdb_context *tdb)
554 {       
555         int i, ret = 0;
556
557         if (tdb->transaction == NULL) {
558                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
559                 return -1;
560         }
561
562         if (tdb->transaction->nesting != 0) {
563                 tdb->transaction->transaction_error = 1;
564                 tdb->transaction->nesting--;
565                 return 0;
566         }               
567
568         tdb->map_size = tdb->transaction->old_map_size;
569
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]);
574                 }
575         }
576         SAFE_FREE(tdb->transaction->blocks);
577
578         if (tdb->transaction->magic_offset) {
579                 const struct tdb_methods *methods = tdb->transaction->io_methods;
580                 uint32_t zero = 0;
581
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"));
586                         ret = -1;
587                 }
588         }
589
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;
594         }
595
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);
601                 }
602                 tdb->num_locks = 0;
603                 tdb->num_lockrecs = 0;
604                 SAFE_FREE(tdb->lockrecs);
605         }
606
607         /* restore the normal io methods */
608         tdb->methods = tdb->transaction->io_methods;
609
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);
614         
615         return ret;
616 }
617
618
619 /*
620   work out how much space the linearised recovery data will consume
621 */
622 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
623 {
624         tdb_len_t recovery_size = 0;
625         int i;
626
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) {
630                         break;
631                 }
632                 if (tdb->transaction->blocks[i] == NULL) {
633                         continue;
634                 }
635                 recovery_size += 2*sizeof(tdb_off_t);
636                 if (i == tdb->transaction->num_blocks-1) {
637                         recovery_size += tdb->transaction->last_block_size;
638                 } else {
639                         recovery_size += tdb->transaction->block_size;
640                 }
641         }       
642
643         return recovery_size;
644 }
645
646 /*
647   allocate the recovery area, or use an existing recovery area if it is
648   large enough
649 */
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)
654 {
655         struct list_struct rec;
656         const struct tdb_methods *methods = tdb->transaction->io_methods;
657         tdb_off_t recovery_head;
658
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"));
661                 return -1;
662         }
663
664         rec.rec_len = 0;
665
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"));
669                 return -1;
670         }
671
672         *recovery_size = tdb_recovery_size(tdb);
673
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;
678                 return 0;
679         }
680
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
685            the transaction) */
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"));
689                         return -1;
690                 }
691         }
692
693         /* the tdb_free() call might have increased the recovery size */
694         *recovery_size = tdb_recovery_size(tdb);
695
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;
700
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"));
705                 return -1;
706         }
707
708         /* remap the file (if using mmap) */
709         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
710
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;
714
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"));
721                 return -1;
722         }
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"));
725                 return -1;
726         }
727
728         return 0;
729 }
730
731
732 /*
733   setup the recovery data that will be used on a crash during commit
734 */
735 static int transaction_setup_recovery(struct tdb_context *tdb, 
736                                       tdb_off_t *magic_offset)
737 {
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;
745         int i;
746
747         /*
748           check that the recovery area has enough space
749         */
750         if (tdb_recovery_allocate(tdb, &recovery_size, 
751                                   &recovery_offset, &recovery_max_size) == -1) {
752                 return -1;
753         }
754
755         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
756         if (data == NULL) {
757                 tdb->ecode = TDB_ERR_OOM;
758                 return -1;
759         }
760
761         rec = (struct list_struct *)data;
762         memset(rec, 0, sizeof(*rec));
763
764         rec->magic    = 0;
765         rec->data_len = recovery_size;
766         rec->rec_len  = recovery_max_size;
767         rec->key_len  = old_map_size;
768         CONVERT(rec);
769
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++) {
774                 tdb_off_t offset;
775                 tdb_len_t length;
776
777                 if (tdb->transaction->blocks[i] == NULL) {
778                         continue;
779                 }
780
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;
785                 }
786                 
787                 if (offset >= old_map_size) {
788                         continue;
789                 }
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"));
792                         free(data);
793                         tdb->ecode = TDB_ERR_CORRUPT;
794                         return -1;
795                 }
796                 memcpy(p, &offset, 4);
797                 memcpy(p+4, &length, 4);
798                 if (DOCONV()) {
799                         tdb_convert(p, 8);
800                 }
801                 /* the recovery area contains the old data, not the
802                    new data, so we have to call the original tdb_read
803                    method to get it */
804                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
805                         free(data);
806                         tdb->ecode = TDB_ERR_IO;
807                         return -1;
808                 }
809                 p += 8 + length;
810         }
811
812         /* and the tailer */
813         tailer = sizeof(*rec) + recovery_max_size;
814         memcpy(p, &tailer, 4);
815         CONVERT(p);
816
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"));
820                 free(data);
821                 tdb->ecode = TDB_ERR_IO;
822                 return -1;
823         }
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"));
826                 free(data);
827                 tdb->ecode = TDB_ERR_IO;
828                 return -1;
829         }
830
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
833            data is present */
834         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
835                 free(data);
836                 return -1;
837         }
838
839         free(data);
840
841         magic = TDB_RECOVERY_MAGIC;
842         CONVERT(magic);
843
844         *magic_offset = recovery_offset + offsetof(struct list_struct, magic);
845
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;
849                 return -1;
850         }
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;
854                 return -1;
855         }
856
857         /* ensure the recovery magic marker is on disk */
858         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
859                 return -1;
860         }
861
862         return 0;
863 }
864
865 /*
866   prepare to commit the current transaction
867 */
868 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
869 {       
870         const struct tdb_methods *methods;
871
872         if (tdb->transaction == NULL) {
873                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
874                 return -1;
875         }
876
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"));
881                 return -1;
882         }
883
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"));
888                 return -1;
889         }
890
891
892         if (tdb->transaction->nesting != 0) {
893                 return 0;
894         }               
895
896         /* check for a null transaction */
897         if (tdb->transaction->blocks == NULL) {
898                 return 0;
899         }
900
901         methods = tdb->transaction->io_methods;
902         
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);
909                 return -1;
910         }
911
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);
917                 return -1;
918         }
919
920         /* get the global lock - this prevents new users attaching to the database
921            during the commit */
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);
926                 return -1;
927         }
928
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);
935                         return -1;
936                 }
937         }
938
939         tdb->transaction->prepared = true;
940
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, 
944                                              tdb->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);
950                         return -1;
951                 }
952                 tdb->map_size = tdb->transaction->old_map_size;
953                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
954         }
955
956         /* Keep the global lock until the actual commit */
957
958         return 0;
959 }
960
961 /*
962   commit the current transaction
963 */
964 int tdb_transaction_commit(struct tdb_context *tdb)
965 {       
966         const struct tdb_methods *methods;
967         int i;
968
969         if (tdb->transaction == NULL) {
970                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
971                 return -1;
972         }
973
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"));
978                 return -1;
979         }
980
981
982         if (tdb->transaction->nesting != 0) {
983                 tdb->transaction->nesting--;
984                 return 0;
985         }
986
987         /* check for a null transaction */
988         if (tdb->transaction->blocks == NULL) {
989                 tdb_transaction_cancel(tdb);
990                 return 0;
991         }
992
993         if (!tdb->transaction->prepared) {
994                 int ret = tdb_transaction_prepare_commit(tdb);
995                 if (ret)
996                         return ret;
997         }
998
999         methods = tdb->transaction->io_methods;
1000
1001         /* perform all the writes */
1002         for (i=0;i<tdb->transaction->num_blocks;i++) {
1003                 tdb_off_t offset;
1004                 tdb_len_t length;
1005
1006                 if (tdb->transaction->blocks[i] == NULL) {
1007                         continue;
1008                 }
1009
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;
1014                 }
1015
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"));
1018                         
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); 
1024
1025                         tdb_transaction_cancel(tdb);
1026                         tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1027
1028                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1029                         return -1;
1030                 }
1031                 SAFE_FREE(tdb->transaction->blocks[i]);
1032         } 
1033
1034         SAFE_FREE(tdb->transaction->blocks);
1035         tdb->transaction->num_blocks = 0;
1036
1037         /* ensure the new data is on disk */
1038         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1039                 return -1;
1040         }
1041
1042         tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1043
1044         /*
1045           TODO: maybe write to some dummy hdr field, or write to magic
1046           offset without mmap, before the last sync, instead of the
1047           utime() call
1048         */
1049
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 */
1055 #ifdef HAVE_UTIME
1056         utime(tdb->name, NULL);
1057 #endif
1058
1059         /* use a transaction cancel to free memory and remove the
1060            transaction locks */
1061         tdb_transaction_cancel(tdb);
1062
1063         return 0;
1064 }
1065
1066
1067 /*
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)
1071 */
1072 int tdb_transaction_recover(struct tdb_context *tdb)
1073 {
1074         tdb_off_t recovery_head, recovery_eof;
1075         unsigned char *data, *p;
1076         uint32_t zero = 0;
1077         struct list_struct rec;
1078
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;
1083                 return -1;
1084         }
1085
1086         if (recovery_head == 0) {
1087                 /* we have never allocated a recovery record */
1088                 return 0;
1089         }
1090
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;
1096                 return -1;
1097         }
1098
1099         if (rec.magic != TDB_RECOVERY_MAGIC) {
1100                 /* there is no valid recovery data */
1101                 return 0;
1102         }
1103
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;
1107                 return -1;
1108         }
1109
1110         recovery_eof = rec.key_len;
1111
1112         data = (unsigned char *)malloc(rec.data_len);
1113         if (data == NULL) {
1114                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1115                 tdb->ecode = TDB_ERR_OOM;
1116                 return -1;
1117         }
1118
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;
1124                 return -1;
1125         }
1126
1127         /* recover the file data */
1128         p = data;
1129         while (p+8 < data + rec.data_len) {
1130                 uint32_t ofs, len;
1131                 if (DOCONV()) {
1132                         tdb_convert(p, 8);
1133                 }
1134                 memcpy(&ofs, p, 4);
1135                 memcpy(&len, p+4, 4);
1136
1137                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1138                         free(data);
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;
1141                         return -1;
1142                 }
1143                 p += 8 + len;
1144         }
1145
1146         free(data);
1147
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;
1151                 return -1;
1152         }
1153
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;
1159                         return -1;                      
1160                 }
1161         }
1162
1163         /* remove the recovery magic */
1164         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic), 
1165                           &zero) == -1) {
1166                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1167                 tdb->ecode = TDB_ERR_IO;
1168                 return -1;                      
1169         }
1170         
1171         /* reduce the file size to the old size */
1172         tdb_munmap(tdb);
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;
1176                 return -1;                      
1177         }
1178         tdb->map_size = recovery_eof;
1179         tdb_mmap(tdb);
1180
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;
1184                 return -1;
1185         }
1186
1187         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1188                  recovery_eof));
1189
1190         /* all done */
1191         return 0;
1192 }