tdb: use same expansion factor logic when expanding for new recovery area.
[ccan] / ccan / tdb / 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     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
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   - 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.
93
94     Beware. when transactions are nested a transaction successfully
95     completed with tdb_transaction_commit() can be silently unrolled later.
96
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.
103 */
104
105
106 /*
107   hold the context of any current transaction
108 */
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;
113
114         /* the original io methods - used to do IOs to the real db */
115         const struct tdb_methods *io_methods;
116
117         /* the list of transaction blocks. When a block is first
118            written to, it gets created in this list */
119         uint8_t **blocks;
120         uint32_t num_blocks;
121         uint32_t block_size;      /* bytes in each block */
122         uint32_t last_block_size; /* number of valid bytes in the last block */
123
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;
128
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 */
132         int nesting;
133
134         /* set when a prepare has already occurred */
135         bool prepared;
136         tdb_off_t magic_offset;
137
138         /* old file size before transaction */
139         tdb_len_t old_map_size;
140
141         /* we should re-pack on commit */
142         bool need_repack;
143 };
144
145
146 /*
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
149 */
150 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
151                             tdb_len_t len, int cv)
152 {
153         uint32_t blk;
154
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) {
159                         return -1;
160                 }
161                 len -= len2;
162                 off += len2;
163                 buf = (void *)(len2 + (char *)buf);
164         }
165
166         if (len == 0) {
167                 return 0;
168         }
169
170         blk = off / tdb->transaction->block_size;
171
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) {
177                         goto fail;
178                 }
179                 return 0;
180         }
181
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) {
185                         goto fail;
186                 }
187         }
188         
189         /* now copy it out of this block */
190         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
191         if (cv) {
192                 tdb_convert(buf, len);
193         }
194         return 0;
195
196 fail:
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;
200         return -1;
201 }
202
203
204 /*
205   write while in a transaction
206 */
207 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
208                              const void *buf, tdb_len_t len)
209 {
210         uint32_t blk;
211
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;
217                 return -1;
218         }
219
220         /* if the write is to a hash head, then update the transaction
221            hash heads */
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);
226         }
227
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) {
232                         return -1;
233                 }
234                 len -= len2;
235                 off += len2;
236                 if (buf != NULL) {
237                         buf = (const void *)(len2 + (const char *)buf);
238                 }
239         }
240
241         if (len == 0) {
242                 return 0;
243         }
244
245         blk = off / tdb->transaction->block_size;
246         off = off % tdb->transaction->block_size;
247
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 *));
254                 } else {
255                         new_blocks = (uint8_t **)realloc(
256                                 tdb->transaction->blocks,
257                                 (blk+1)*sizeof(uint8_t *));
258                 }
259                 if (new_blocks == NULL) {
260                         tdb->ecode = TDB_ERR_OOM;
261                         goto fail;
262                 }
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;
268         }
269
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;
276                         return -1;                      
277                 }
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);
282                         }
283                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
284                                                                    tdb->transaction->blocks[blk], 
285                                                                    len2, 0) != 0) {
286                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
287                                 tdb->ecode = TDB_ERR_IO;
288                                 goto fail;
289                         }
290                         if (blk == tdb->transaction->num_blocks-1) {
291                                 tdb->transaction->last_block_size = len2;
292                         }                       
293                 }
294         }
295         
296         /* overwrite part of an existing block */
297         if (buf == NULL) {
298                 memset(tdb->transaction->blocks[blk] + off, 0, len);
299         } else {
300                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
301         }
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;
305                 }
306         }
307
308         return 0;
309
310 fail:
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;
314         return -1;
315 }
316
317
318 /*
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
321 */
322 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
323                                       const void *buf, tdb_len_t len)
324 {
325         uint32_t blk;
326
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) {
331                         return -1;
332                 }
333                 len -= len2;
334                 off += len2;
335                 if (buf != NULL) {
336                         buf = (const void *)(len2 + (const char *)buf);
337                 }
338         }
339
340         if (len == 0) {
341                 return 0;
342         }
343
344         blk = off / tdb->transaction->block_size;
345         off = off % tdb->transaction->block_size;
346
347         if (tdb->transaction->num_blocks <= blk ||
348             tdb->transaction->blocks[blk] == NULL) {
349                 return 0;
350         }
351
352         if (blk == tdb->transaction->num_blocks-1 &&
353             off + len > tdb->transaction->last_block_size) {
354                 if (off >= tdb->transaction->last_block_size) {
355                         return 0;
356                 }
357                 len = tdb->transaction->last_block_size - off;
358         }
359
360         /* overwrite part of an existing block */
361         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
362
363         return 0;
364 }
365
366
367 /*
368   accelerated hash chain head search, using the cached hash heads
369 */
370 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
371 {
372         uint32_t h = *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]) {
376                         break;
377                 }
378         }
379         (*chain) = h;
380 }
381
382 /*
383   out of bounds check during a transaction
384 */
385 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
386 {
387         if (len <= tdb->map_size) {
388                 return 0;
389         }
390         tdb->ecode = TDB_ERR_IO;
391         return -1;
392 }
393
394 /*
395   transaction version of tdb_expand().
396 */
397 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
398                                    tdb_off_t addition)
399 {
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) {
403                 return -1;
404         }
405
406         tdb->transaction->need_repack = true;
407
408         return 0;
409 }
410
411 static const struct tdb_methods transaction_methods = {
412         transaction_read,
413         transaction_write,
414         transaction_next_hash_chain,
415         transaction_oob,
416         transaction_expand_file,
417 };
418
419 /*
420   sync to disk
421 */
422 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
423 {       
424         if (tdb->flags & TDB_NOSYNC) {
425                 return 0;
426         }
427
428         if (fsync(tdb->fd) != 0) {
429                 tdb->ecode = TDB_ERR_IO;
430                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
431                 return -1;
432         }
433 #ifdef MS_SYNC
434         if (tdb->map_ptr) {
435                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
436                 if (msync(moffset + (char *)tdb->map_ptr, 
437                           length + (offset - moffset), MS_SYNC) != 0) {
438                         tdb->ecode = TDB_ERR_IO;
439                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
440                                  strerror(errno)));
441                         return -1;
442                 }
443         }
444 #endif
445         return 0;
446 }
447
448
449 static int _tdb_transaction_cancel(struct tdb_context *tdb)
450 {
451         int i, ret = 0;
452
453         if (tdb->transaction == NULL) {
454                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
455                 return -1;
456         }
457
458         if (tdb->transaction->nesting != 0) {
459                 tdb->transaction->transaction_error = 1;
460                 tdb->transaction->nesting--;
461                 return 0;
462         }               
463
464         tdb->map_size = tdb->transaction->old_map_size;
465
466         /* free all the transaction blocks */
467         for (i=0;i<tdb->transaction->num_blocks;i++) {
468                 if (tdb->transaction->blocks[i] != NULL) {
469                         free(tdb->transaction->blocks[i]);
470                 }
471         }
472         SAFE_FREE(tdb->transaction->blocks);
473
474         if (tdb->transaction->magic_offset) {
475                 const struct tdb_methods *methods = tdb->transaction->io_methods;
476                 uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
477
478                 /* remove the recovery marker */
479                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
480                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
481                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
482                         ret = -1;
483                 }
484         }
485
486         /* This also removes the OPEN_LOCK, if we have it. */
487         tdb_release_extra_locks(tdb);
488
489         /* restore the normal io methods */
490         tdb->methods = tdb->transaction->io_methods;
491
492         tdb_transaction_unlock(tdb, F_WRLCK);
493         SAFE_FREE(tdb->transaction->hash_heads);
494         SAFE_FREE(tdb->transaction);
495         
496         return ret;
497 }
498
499 /*
500   start a tdb transaction. No token is returned, as only a single
501   transaction is allowed to be pending per tdb_context
502 */
503 int tdb_transaction_start(struct tdb_context *tdb)
504 {
505         /* some sanity checks */
506         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
507                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
508                 tdb->ecode = TDB_ERR_EINVAL;
509                 return -1;
510         }
511
512         /* cope with nested tdb_transaction_start() calls */
513         if (tdb->transaction != NULL) {
514                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
515                         tdb->ecode = TDB_ERR_NESTING;
516                         return -1;
517                 }
518                 tdb_trace(tdb, "tdb_transaction_start");
519                 tdb->transaction->nesting++;
520                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
521                          tdb->transaction->nesting));
522                 return 0;
523         }
524
525         if (tdb_have_extra_locks(tdb)) {
526                 /* the caller must not have any locks when starting a
527                    transaction as otherwise we'll be screwed by lack
528                    of nested locks in posix */
529                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
530                 tdb->ecode = TDB_ERR_LOCK;
531                 return -1;
532         }
533
534         if (tdb->travlocks.next != NULL) {
535                 /* you cannot use transactions inside a traverse (although you can use
536                    traverse inside a transaction) as otherwise you can end up with
537                    deadlock */
538                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
539                 tdb->ecode = TDB_ERR_LOCK;
540                 return -1;
541         }
542
543         tdb->transaction = (struct tdb_transaction *)
544                 calloc(sizeof(struct tdb_transaction), 1);
545         if (tdb->transaction == NULL) {
546                 tdb->ecode = TDB_ERR_OOM;
547                 return -1;
548         }
549
550         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
551         tdb->transaction->block_size = tdb->page_size;
552
553         /* get the transaction write lock. This is a blocking lock. As
554            discussed with Volker, there are a number of ways we could
555            make this async, which we will probably do in the future */
556         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
557                 SAFE_FREE(tdb->transaction->blocks);
558                 SAFE_FREE(tdb->transaction);
559                 return -1;
560         }
561         
562         /* get a read lock from the freelist to the end of file. This
563            is upgraded to a write lock during the commit */
564         if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
565                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
566                 goto fail_allrecord_lock;
567         }
568
569         /* setup a copy of the hash table heads so the hash scan in
570            traverse can be fast */
571         tdb->transaction->hash_heads = (uint32_t *)
572                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
573         if (tdb->transaction->hash_heads == NULL) {
574                 tdb->ecode = TDB_ERR_OOM;
575                 goto fail;
576         }
577         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
578                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
579                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
580                 tdb->ecode = TDB_ERR_IO;
581                 goto fail;
582         }
583
584         /* make sure we know about any file expansions already done by
585            anyone else */
586         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
587         tdb->transaction->old_map_size = tdb->map_size;
588
589         /* finally hook the io methods, replacing them with
590            transaction specific methods */
591         tdb->transaction->io_methods = tdb->methods;
592         tdb->methods = &transaction_methods;
593
594         /* Trace at the end, so we get sequence number correct. */
595         tdb_trace(tdb, "tdb_transaction_start");
596         return 0;
597         
598 fail:
599         tdb_allrecord_unlock(tdb, F_RDLCK, false);
600 fail_allrecord_lock:
601         tdb_transaction_unlock(tdb, F_WRLCK);
602         SAFE_FREE(tdb->transaction->blocks);
603         SAFE_FREE(tdb->transaction->hash_heads);
604         SAFE_FREE(tdb->transaction);
605         return -1;
606 }
607
608
609 /*
610   cancel the current transaction
611 */
612 int tdb_transaction_cancel(struct tdb_context *tdb)
613 {
614         tdb_trace(tdb, "tdb_transaction_cancel");
615         return _tdb_transaction_cancel(tdb);
616 }
617
618 /*
619   work out how much space the linearised recovery data will consume
620 */
621 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
622 {
623         tdb_len_t recovery_size = 0;
624         int i;
625
626         recovery_size = sizeof(uint32_t);
627         for (i=0;i<tdb->transaction->num_blocks;i++) {
628                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
629                         break;
630                 }
631                 if (tdb->transaction->blocks[i] == NULL) {
632                         continue;
633                 }
634                 recovery_size += 2*sizeof(tdb_off_t);
635                 if (i == tdb->transaction->num_blocks-1) {
636                         recovery_size += tdb->transaction->last_block_size;
637                 } else {
638                         recovery_size += tdb->transaction->block_size;
639                 }
640         }       
641
642         return recovery_size;
643 }
644
645 /*
646   allocate the recovery area, or use an existing recovery area if it is
647   large enough
648 */
649 static int tdb_recovery_allocate(struct tdb_context *tdb, 
650                                  tdb_len_t *recovery_size,
651                                  tdb_off_t *recovery_offset,
652                                  tdb_len_t *recovery_max_size)
653 {
654         struct tdb_record rec;
655         const struct tdb_methods *methods = tdb->transaction->io_methods;
656         tdb_off_t recovery_head;
657
658         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
659                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
660                 return -1;
661         }
662
663         rec.rec_len = 0;
664
665         if (recovery_head != 0) {
666                 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
667                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
668                         return -1;
669                 }
670                 /* ignore invalid recovery regions: can happen in crash */
671                 if (rec.magic != TDB_RECOVERY_MAGIC &&
672                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
673                         recovery_head = 0;
674                 }
675         }
676
677         *recovery_size = tdb_recovery_size(tdb);
678
679         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
680                 /* it fits in the existing area */
681                 *recovery_max_size = rec.rec_len;
682                 *recovery_offset = recovery_head;
683                 return 0;
684         }
685
686         /* we need to free up the old recovery area, then allocate a
687            new one at the end of the file. Note that we cannot use
688            tdb_allocate() to allocate the new one as that might return
689            us an area that is being currently used (as of the start of
690            the transaction) */
691         if (recovery_head != 0) {
692                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
693                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
694                         return -1;
695                 }
696         }
697
698         /* the tdb_free() call might have increased the recovery size */
699         *recovery_size = tdb_recovery_size(tdb);
700
701         /* round up to a multiple of page size */
702         *recovery_max_size = tdb_expand_adjust(tdb->map_size,
703                                                *recovery_size,
704                                                tdb->page_size)
705                 - sizeof(rec);
706
707         *recovery_offset = tdb->map_size;
708         recovery_head = *recovery_offset;
709
710         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
711                                      (tdb->map_size - tdb->transaction->old_map_size) +
712                                      sizeof(rec) + *recovery_max_size) == -1) {
713                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
714                 return -1;
715         }
716
717         /* remap the file (if using mmap) */
718         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
719
720         /* we have to reset the old map size so that we don't try to expand the file
721            again in the transaction commit, which would destroy the recovery area */
722         tdb->transaction->old_map_size = tdb->map_size;
723
724         /* write the recovery header offset and sync - we can sync without a race here
725            as the magic ptr in the recovery record has not been set */
726         CONVERT(recovery_head);
727         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
728                                &recovery_head, sizeof(tdb_off_t)) == -1) {
729                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
730                 return -1;
731         }
732         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
733                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
734                 return -1;
735         }
736
737         return 0;
738 }
739
740
741 /*
742   setup the recovery data that will be used on a crash during commit
743 */
744 static int transaction_setup_recovery(struct tdb_context *tdb, 
745                                       tdb_off_t *magic_offset)
746 {
747         tdb_len_t recovery_size;
748         unsigned char *data, *p;
749         const struct tdb_methods *methods = tdb->transaction->io_methods;
750         struct tdb_record *rec;
751         tdb_off_t recovery_offset, recovery_max_size;
752         tdb_off_t old_map_size = tdb->transaction->old_map_size;
753         uint32_t magic, tailer;
754         int i;
755
756         /*
757           check that the recovery area has enough space
758         */
759         if (tdb_recovery_allocate(tdb, &recovery_size, 
760                                   &recovery_offset, &recovery_max_size) == -1) {
761                 return -1;
762         }
763
764         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
765         if (data == NULL) {
766                 tdb->ecode = TDB_ERR_OOM;
767                 return -1;
768         }
769
770         rec = (struct tdb_record *)data;
771         memset(rec, 0, sizeof(*rec));
772
773         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
774         rec->data_len = recovery_size;
775         rec->rec_len  = recovery_max_size;
776         rec->key_len  = old_map_size;
777         CONVERT(rec);
778
779         /* build the recovery data into a single blob to allow us to do a single
780            large write, which should be more efficient */
781         p = data + sizeof(*rec);
782         for (i=0;i<tdb->transaction->num_blocks;i++) {
783                 tdb_off_t offset;
784                 tdb_len_t length;
785
786                 if (tdb->transaction->blocks[i] == NULL) {
787                         continue;
788                 }
789
790                 offset = i * tdb->transaction->block_size;
791                 length = tdb->transaction->block_size;
792                 if (i == tdb->transaction->num_blocks-1) {
793                         length = tdb->transaction->last_block_size;
794                 }
795                 
796                 if (offset >= old_map_size) {
797                         continue;
798                 }
799                 if (offset + length > tdb->transaction->old_map_size) {
800                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
801                         free(data);
802                         tdb->ecode = TDB_ERR_CORRUPT;
803                         return -1;
804                 }
805                 memcpy(p, &offset, 4);
806                 memcpy(p+4, &length, 4);
807                 if (DOCONV()) {
808                         tdb_convert(p, 8);
809                 }
810                 /* the recovery area contains the old data, not the
811                    new data, so we have to call the original tdb_read
812                    method to get it */
813                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
814                         free(data);
815                         tdb->ecode = TDB_ERR_IO;
816                         return -1;
817                 }
818                 p += 8 + length;
819         }
820
821         /* and the tailer */
822         tailer = sizeof(*rec) + recovery_max_size;
823         memcpy(p, &tailer, 4);
824         CONVERT(p);
825
826         /* write the recovery data to the recovery area */
827         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
828                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
829                 free(data);
830                 tdb->ecode = TDB_ERR_IO;
831                 return -1;
832         }
833         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
834                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
835                 free(data);
836                 tdb->ecode = TDB_ERR_IO;
837                 return -1;
838         }
839
840         /* as we don't have ordered writes, we have to sync the recovery
841            data before we update the magic to indicate that the recovery
842            data is present */
843         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
844                 free(data);
845                 return -1;
846         }
847
848         free(data);
849
850         magic = TDB_RECOVERY_MAGIC;
851         CONVERT(magic);
852
853         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
854
855         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
856                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
857                 tdb->ecode = TDB_ERR_IO;
858                 return -1;
859         }
860         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
861                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
862                 tdb->ecode = TDB_ERR_IO;
863                 return -1;
864         }
865
866         /* ensure the recovery magic marker is on disk */
867         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
868                 return -1;
869         }
870
871         return 0;
872 }
873
874 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
875 {       
876         const struct tdb_methods *methods;
877
878         if (tdb->transaction == NULL) {
879                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
880                 return -1;
881         }
882
883         if (tdb->transaction->prepared) {
884                 tdb->ecode = TDB_ERR_EINVAL;
885                 _tdb_transaction_cancel(tdb);
886                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
887                 return -1;
888         }
889
890         if (tdb->transaction->transaction_error) {
891                 tdb->ecode = TDB_ERR_IO;
892                 _tdb_transaction_cancel(tdb);
893                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
894                 return -1;
895         }
896
897
898         if (tdb->transaction->nesting != 0) {
899                 tdb->transaction->nesting--;
900                 return 0;
901         }               
902
903         /* check for a null transaction */
904         if (tdb->transaction->blocks == NULL) {
905                 return 0;
906         }
907
908         methods = tdb->transaction->io_methods;
909         
910         /* if there are any locks pending then the caller has not
911            nested their locks properly, so fail the transaction */
912         if (tdb_have_extra_locks(tdb)) {
913                 tdb->ecode = TDB_ERR_LOCK;
914                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
915                 _tdb_transaction_cancel(tdb);
916                 return -1;
917         }
918
919         /* upgrade the main transaction lock region to a write lock */
920         if (tdb_allrecord_upgrade(tdb) == -1) {
921                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
922                 _tdb_transaction_cancel(tdb);
923                 return -1;
924         }
925
926         /* get the open lock - this prevents new users attaching to the database
927            during the commit */
928         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
929                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
930                 _tdb_transaction_cancel(tdb);
931                 return -1;
932         }
933
934         if (!(tdb->flags & TDB_NOSYNC)) {
935                 /* write the recovery data to the end of the file */
936                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
937                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
938                         _tdb_transaction_cancel(tdb);
939                         return -1;
940                 }
941         }
942
943         tdb->transaction->prepared = true;
944
945         /* expand the file to the new size if needed */
946         if (tdb->map_size != tdb->transaction->old_map_size) {
947                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
948                                              tdb->map_size - 
949                                              tdb->transaction->old_map_size) == -1) {
950                         tdb->ecode = TDB_ERR_IO;
951                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
952                         _tdb_transaction_cancel(tdb);
953                         return -1;
954                 }
955                 tdb->map_size = tdb->transaction->old_map_size;
956                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
957         }
958
959         /* Keep the open lock until the actual commit */
960
961         return 0;
962 }
963
964 /*
965    prepare to commit the current transaction
966 */
967 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
968 {       
969         tdb_trace(tdb, "tdb_transaction_prepare_commit");
970         return _tdb_transaction_prepare_commit(tdb);
971 }
972
973 /*
974   commit the current transaction
975 */
976 int tdb_transaction_commit(struct tdb_context *tdb)
977 {       
978         const struct tdb_methods *methods;
979         int i;
980         bool need_repack;
981
982         if (tdb->transaction == NULL) {
983                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
984                 return -1;
985         }
986
987         tdb_trace(tdb, "tdb_transaction_commit");
988
989         if (tdb->transaction->transaction_error) {
990                 tdb->ecode = TDB_ERR_IO;
991                 tdb_transaction_cancel(tdb);
992                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
993                 return -1;
994         }
995
996
997         if (tdb->transaction->nesting != 0) {
998                 tdb->transaction->nesting--;
999                 return 0;
1000         }
1001
1002         /* check for a null transaction */
1003         if (tdb->transaction->blocks == NULL) {
1004                 _tdb_transaction_cancel(tdb);
1005                 return 0;
1006         }
1007
1008         if (!tdb->transaction->prepared) {
1009                 int ret = _tdb_transaction_prepare_commit(tdb);
1010                 if (ret)
1011                         return ret;
1012         }
1013
1014         methods = tdb->transaction->io_methods;
1015
1016         /* perform all the writes */
1017         for (i=0;i<tdb->transaction->num_blocks;i++) {
1018                 tdb_off_t offset;
1019                 tdb_len_t length;
1020
1021                 if (tdb->transaction->blocks[i] == NULL) {
1022                         continue;
1023                 }
1024
1025                 offset = i * tdb->transaction->block_size;
1026                 length = tdb->transaction->block_size;
1027                 if (i == tdb->transaction->num_blocks-1) {
1028                         length = tdb->transaction->last_block_size;
1029                 }
1030
1031                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1032                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1033                         
1034                         /* we've overwritten part of the data and
1035                            possibly expanded the file, so we need to
1036                            run the crash recovery code */
1037                         tdb->methods = methods;
1038                         tdb_transaction_recover(tdb); 
1039
1040                         _tdb_transaction_cancel(tdb);
1041
1042                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1043                         return -1;
1044                 }
1045                 SAFE_FREE(tdb->transaction->blocks[i]);
1046         } 
1047
1048         SAFE_FREE(tdb->transaction->blocks);
1049         tdb->transaction->num_blocks = 0;
1050
1051         /* ensure the new data is on disk */
1052         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1053                 return -1;
1054         }
1055
1056         /*
1057           TODO: maybe write to some dummy hdr field, or write to magic
1058           offset without mmap, before the last sync, instead of the
1059           utime() call
1060         */
1061
1062         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1063            don't change the mtime of the file, this means the file may
1064            not be backed up (as tdb rounding to block sizes means that
1065            file size changes are quite rare too). The following forces
1066            mtime changes when a transaction completes */
1067 #if HAVE_UTIME
1068         utime(tdb->name, NULL);
1069 #endif
1070
1071         need_repack = tdb->transaction->need_repack;
1072
1073         /* use a transaction cancel to free memory and remove the
1074            transaction locks */
1075         _tdb_transaction_cancel(tdb);
1076
1077         if (need_repack) {
1078                 return tdb_repack(tdb);
1079         }
1080
1081         return 0;
1082 }
1083
1084
1085 /*
1086   recover from an aborted transaction. Must be called with exclusive
1087   database write access already established (including the open
1088   lock to prevent new processes attaching)
1089 */
1090 int tdb_transaction_recover(struct tdb_context *tdb)
1091 {
1092         tdb_off_t recovery_head, recovery_eof;
1093         unsigned char *data, *p;
1094         uint32_t zero = 0;
1095         struct tdb_record rec;
1096
1097         /* find the recovery area */
1098         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1099                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1100                 tdb->ecode = TDB_ERR_IO;
1101                 return -1;
1102         }
1103
1104         if (recovery_head == 0) {
1105                 /* we have never allocated a recovery record */
1106                 return 0;
1107         }
1108
1109         /* read the recovery record */
1110         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1111                                    sizeof(rec), DOCONV()) == -1) {
1112                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1113                 tdb->ecode = TDB_ERR_IO;
1114                 return -1;
1115         }
1116
1117         if (rec.magic != TDB_RECOVERY_MAGIC) {
1118                 /* there is no valid recovery data */
1119                 return 0;
1120         }
1121
1122         if (tdb->read_only) {
1123                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1124                 tdb->ecode = TDB_ERR_CORRUPT;
1125                 return -1;
1126         }
1127
1128         recovery_eof = rec.key_len;
1129
1130         data = (unsigned char *)malloc(rec.data_len);
1131         if (data == NULL) {
1132                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1133                 tdb->ecode = TDB_ERR_OOM;
1134                 return -1;
1135         }
1136
1137         /* read the full recovery data */
1138         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1139                                    rec.data_len, 0) == -1) {
1140                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1141                 tdb->ecode = TDB_ERR_IO;
1142                 return -1;
1143         }
1144
1145         /* recover the file data */
1146         p = data;
1147         while (p+8 < data + rec.data_len) {
1148                 uint32_t ofs, len;
1149                 if (DOCONV()) {
1150                         tdb_convert(p, 8);
1151                 }
1152                 memcpy(&ofs, p, 4);
1153                 memcpy(&len, p+4, 4);
1154
1155                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1156                         free(data);
1157                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1158                         tdb->ecode = TDB_ERR_IO;
1159                         return -1;
1160                 }
1161                 p += 8 + len;
1162         }
1163
1164         free(data);
1165
1166         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1167                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1168                 tdb->ecode = TDB_ERR_IO;
1169                 return -1;
1170         }
1171
1172         /* if the recovery area is after the recovered eof then remove it */
1173         if (recovery_eof <= recovery_head) {
1174                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1175                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1176                         tdb->ecode = TDB_ERR_IO;
1177                         return -1;                      
1178                 }
1179         }
1180
1181         /* remove the recovery magic */
1182         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1183                           &zero) == -1) {
1184                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1185                 tdb->ecode = TDB_ERR_IO;
1186                 return -1;                      
1187         }
1188         
1189         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1190                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1191                 tdb->ecode = TDB_ERR_IO;
1192                 return -1;
1193         }
1194
1195         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1196                  recovery_eof));
1197
1198         /* all done */
1199         return 0;
1200 }
1201
1202 /* Any I/O failures we say "needs recovery". */
1203 bool tdb_needs_recovery(struct tdb_context *tdb)
1204 {
1205         tdb_off_t recovery_head;
1206         struct tdb_record rec;
1207
1208         /* find the recovery area */
1209         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1210                 return true;
1211         }
1212
1213         if (recovery_head == 0) {
1214                 /* we have never allocated a recovery record */
1215                 return false;
1216         }
1217
1218         /* read the recovery record */
1219         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1220                                    sizeof(rec), DOCONV()) == -1) {
1221                 return true;
1222         }
1223
1224         return (rec.magic == TDB_RECOVERY_MAGIC);
1225 }