tdb: cleanup: tdb_release_extra_locks() helper
[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         /* set when the OPEN_LOCK has been taken */
139         bool open_lock_taken;
140
141         /* old file size before transaction */
142         tdb_len_t old_map_size;
143
144         /* we should re-pack on commit */
145         bool need_repack;
146 };
147
148
149 /*
150   read while in a transaction. We need to check first if the data is in our list
151   of transaction elements, then if not do a real read
152 */
153 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
154                             tdb_len_t len, int cv)
155 {
156         uint32_t blk;
157
158         /* break it down into block sized ops */
159         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
160                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
161                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
162                         return -1;
163                 }
164                 len -= len2;
165                 off += len2;
166                 buf = (void *)(len2 + (char *)buf);
167         }
168
169         if (len == 0) {
170                 return 0;
171         }
172
173         blk = off / tdb->transaction->block_size;
174
175         /* see if we have it in the block list */
176         if (tdb->transaction->num_blocks <= blk ||
177             tdb->transaction->blocks[blk] == NULL) {
178                 /* nope, do a real read */
179                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
180                         goto fail;
181                 }
182                 return 0;
183         }
184
185         /* it is in the block list. Now check for the last block */
186         if (blk == tdb->transaction->num_blocks-1) {
187                 if (len > tdb->transaction->last_block_size) {
188                         goto fail;
189                 }
190         }
191         
192         /* now copy it out of this block */
193         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
194         if (cv) {
195                 tdb_convert(buf, len);
196         }
197         return 0;
198
199 fail:
200         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
201         tdb->ecode = TDB_ERR_IO;
202         tdb->transaction->transaction_error = 1;
203         return -1;
204 }
205
206
207 /*
208   write while in a transaction
209 */
210 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
211                              const void *buf, tdb_len_t len)
212 {
213         uint32_t blk;
214
215         /* Only a commit is allowed on a prepared transaction */
216         if (tdb->transaction->prepared) {
217                 tdb->ecode = TDB_ERR_EINVAL;
218                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
219                 tdb->transaction->transaction_error = 1;
220                 return -1;
221         }
222
223         /* if the write is to a hash head, then update the transaction
224            hash heads */
225         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
226             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
227                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
228                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
229         }
230
231         /* break it up into block sized chunks */
232         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
233                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
234                 if (transaction_write(tdb, off, buf, len2) != 0) {
235                         return -1;
236                 }
237                 len -= len2;
238                 off += len2;
239                 if (buf != NULL) {
240                         buf = (const void *)(len2 + (const char *)buf);
241                 }
242         }
243
244         if (len == 0) {
245                 return 0;
246         }
247
248         blk = off / tdb->transaction->block_size;
249         off = off % tdb->transaction->block_size;
250
251         if (tdb->transaction->num_blocks <= blk) {
252                 uint8_t **new_blocks;
253                 /* expand the blocks array */
254                 if (tdb->transaction->blocks == NULL) {
255                         new_blocks = (uint8_t **)malloc(
256                                 (blk+1)*sizeof(uint8_t *));
257                 } else {
258                         new_blocks = (uint8_t **)realloc(
259                                 tdb->transaction->blocks,
260                                 (blk+1)*sizeof(uint8_t *));
261                 }
262                 if (new_blocks == NULL) {
263                         tdb->ecode = TDB_ERR_OOM;
264                         goto fail;
265                 }
266                 memset(&new_blocks[tdb->transaction->num_blocks], 0, 
267                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
268                 tdb->transaction->blocks = new_blocks;
269                 tdb->transaction->num_blocks = blk+1;
270                 tdb->transaction->last_block_size = 0;
271         }
272
273         /* allocate and fill a block? */
274         if (tdb->transaction->blocks[blk] == NULL) {
275                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
276                 if (tdb->transaction->blocks[blk] == NULL) {
277                         tdb->ecode = TDB_ERR_OOM;
278                         tdb->transaction->transaction_error = 1;
279                         return -1;                      
280                 }
281                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
282                         tdb_len_t len2 = tdb->transaction->block_size;
283                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
284                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
285                         }
286                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
287                                                                    tdb->transaction->blocks[blk], 
288                                                                    len2, 0) != 0) {
289                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
290                                 tdb->ecode = TDB_ERR_IO;
291                                 goto fail;
292                         }
293                         if (blk == tdb->transaction->num_blocks-1) {
294                                 tdb->transaction->last_block_size = len2;
295                         }                       
296                 }
297         }
298         
299         /* overwrite part of an existing block */
300         if (buf == NULL) {
301                 memset(tdb->transaction->blocks[blk] + off, 0, len);
302         } else {
303                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
304         }
305         if (blk == tdb->transaction->num_blocks-1) {
306                 if (len + off > tdb->transaction->last_block_size) {
307                         tdb->transaction->last_block_size = len + off;
308                 }
309         }
310
311         return 0;
312
313 fail:
314         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", 
315                  (blk*tdb->transaction->block_size) + off, len));
316         tdb->transaction->transaction_error = 1;
317         return -1;
318 }
319
320
321 /*
322   write while in a transaction - this varient never expands the transaction blocks, it only
323   updates existing blocks. This means it cannot change the recovery size
324 */
325 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
326                                       const void *buf, tdb_len_t len)
327 {
328         uint32_t blk;
329
330         /* break it up into block sized chunks */
331         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
332                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
333                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
334                         return -1;
335                 }
336                 len -= len2;
337                 off += len2;
338                 if (buf != NULL) {
339                         buf = (const void *)(len2 + (const char *)buf);
340                 }
341         }
342
343         if (len == 0) {
344                 return 0;
345         }
346
347         blk = off / tdb->transaction->block_size;
348         off = off % tdb->transaction->block_size;
349
350         if (tdb->transaction->num_blocks <= blk ||
351             tdb->transaction->blocks[blk] == NULL) {
352                 return 0;
353         }
354
355         if (blk == tdb->transaction->num_blocks-1 &&
356             off + len > tdb->transaction->last_block_size) {
357                 if (off >= tdb->transaction->last_block_size) {
358                         return 0;
359                 }
360                 len = tdb->transaction->last_block_size - off;
361         }
362
363         /* overwrite part of an existing block */
364         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
365
366         return 0;
367 }
368
369
370 /*
371   accelerated hash chain head search, using the cached hash heads
372 */
373 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
374 {
375         uint32_t h = *chain;
376         for (;h < tdb->header.hash_size;h++) {
377                 /* the +1 takes account of the freelist */
378                 if (0 != tdb->transaction->hash_heads[h+1]) {
379                         break;
380                 }
381         }
382         (*chain) = h;
383 }
384
385 /*
386   out of bounds check during a transaction
387 */
388 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
389 {
390         if (len <= tdb->map_size) {
391                 return 0;
392         }
393         tdb->ecode = TDB_ERR_IO;
394         return -1;
395 }
396
397 /*
398   transaction version of tdb_expand().
399 */
400 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
401                                    tdb_off_t addition)
402 {
403         /* add a write to the transaction elements, so subsequent
404            reads see the zero data */
405         if (transaction_write(tdb, size, NULL, addition) != 0) {
406                 return -1;
407         }
408
409         tdb->transaction->need_repack = true;
410
411         return 0;
412 }
413
414 /*
415   brlock during a transaction - ignore them
416 */
417 static int transaction_brlock(struct tdb_context *tdb,
418                               int rw_type, tdb_off_t offset, size_t len,
419                               enum tdb_lock_flags flags)
420 {
421         return 0;
422 }
423
424 static int transaction_brunlock(struct tdb_context *tdb,
425                                 int rw_type, tdb_off_t offset, size_t len)
426 {
427         return 0;
428 }
429
430 static const struct tdb_methods transaction_methods = {
431         transaction_read,
432         transaction_write,
433         transaction_next_hash_chain,
434         transaction_oob,
435         transaction_expand_file,
436         transaction_brlock,
437         transaction_brunlock
438 };
439
440 /*
441   sync to disk
442 */
443 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
444 {       
445         if (tdb->flags & TDB_NOSYNC) {
446                 return 0;
447         }
448
449         if (fsync(tdb->fd) != 0) {
450                 tdb->ecode = TDB_ERR_IO;
451                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
452                 return -1;
453         }
454 #ifdef MS_SYNC
455         if (tdb->map_ptr) {
456                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
457                 if (msync(moffset + (char *)tdb->map_ptr, 
458                           length + (offset - moffset), MS_SYNC) != 0) {
459                         tdb->ecode = TDB_ERR_IO;
460                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
461                                  strerror(errno)));
462                         return -1;
463                 }
464         }
465 #endif
466         return 0;
467 }
468
469 /* ltype is F_WRLCK after prepare. */
470 int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype)
471 {
472         int i, ret = 0;
473
474         if (tdb->transaction == NULL) {
475                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
476                 return -1;
477         }
478
479         if (tdb->transaction->nesting != 0) {
480                 tdb->transaction->transaction_error = 1;
481                 tdb->transaction->nesting--;
482                 return 0;
483         }               
484
485         tdb->map_size = tdb->transaction->old_map_size;
486
487         /* free all the transaction blocks */
488         for (i=0;i<tdb->transaction->num_blocks;i++) {
489                 if (tdb->transaction->blocks[i] != NULL) {
490                         free(tdb->transaction->blocks[i]);
491                 }
492         }
493         SAFE_FREE(tdb->transaction->blocks);
494
495         if (tdb->transaction->magic_offset) {
496                 const struct tdb_methods *methods = tdb->transaction->io_methods;
497                 uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
498
499                 /* remove the recovery marker */
500                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
501                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
502                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
503                         ret = -1;
504                 }
505         }
506
507         if (tdb->transaction->open_lock_taken) {
508                 tdb_brunlock(tdb, F_WRLCK, OPEN_LOCK, 1);
509                 tdb->transaction->open_lock_taken = false;
510         }
511
512         tdb_release_extra_locks(tdb);
513
514         /* restore the normal io methods */
515         tdb->methods = tdb->transaction->io_methods;
516
517         tdb_brunlock(tdb, ltype, FREELIST_TOP, 0);
518         tdb_transaction_unlock(tdb, F_WRLCK);
519         SAFE_FREE(tdb->transaction->hash_heads);
520         SAFE_FREE(tdb->transaction);
521         
522         return ret;
523 }
524
525 /*
526   start a tdb transaction. No token is returned, as only a single
527   transaction is allowed to be pending per tdb_context
528 */
529 int tdb_transaction_start(struct tdb_context *tdb)
530 {
531         /* some sanity checks */
532         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
533                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
534                 tdb->ecode = TDB_ERR_EINVAL;
535                 return -1;
536         }
537
538         /* cope with nested tdb_transaction_start() calls */
539         if (tdb->transaction != NULL) {
540                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
541                         tdb->ecode = TDB_ERR_NESTING;
542                         return -1;
543                 }
544                 tdb_trace(tdb, "tdb_transaction_start");
545                 tdb->transaction->nesting++;
546                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
547                          tdb->transaction->nesting));
548                 return 0;
549         }
550
551         if (tdb_have_extra_locks(tdb)) {
552                 /* the caller must not have any locks when starting a
553                    transaction as otherwise we'll be screwed by lack
554                    of nested locks in posix */
555                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
556                 tdb->ecode = TDB_ERR_LOCK;
557                 return -1;
558         }
559
560         if (tdb->travlocks.next != NULL) {
561                 /* you cannot use transactions inside a traverse (although you can use
562                    traverse inside a transaction) as otherwise you can end up with
563                    deadlock */
564                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
565                 tdb->ecode = TDB_ERR_LOCK;
566                 return -1;
567         }
568
569         tdb->transaction = (struct tdb_transaction *)
570                 calloc(sizeof(struct tdb_transaction), 1);
571         if (tdb->transaction == NULL) {
572                 tdb->ecode = TDB_ERR_OOM;
573                 return -1;
574         }
575
576         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
577         tdb->transaction->block_size = tdb->page_size;
578
579         /* get the transaction write lock. This is a blocking lock. As
580            discussed with Volker, there are a number of ways we could
581            make this async, which we will probably do in the future */
582         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
583                 SAFE_FREE(tdb->transaction->blocks);
584                 SAFE_FREE(tdb->transaction);
585                 return -1;
586         }
587         
588         /* get a read lock from the freelist to the end of file. This
589            is upgraded to a write lock during the commit */
590         if (tdb_brlock(tdb, F_RDLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT) == -1) {
591                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
592                 tdb->ecode = TDB_ERR_LOCK;
593                 goto fail;
594         }
595
596         /* setup a copy of the hash table heads so the hash scan in
597            traverse can be fast */
598         tdb->transaction->hash_heads = (uint32_t *)
599                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
600         if (tdb->transaction->hash_heads == NULL) {
601                 tdb->ecode = TDB_ERR_OOM;
602                 goto fail;
603         }
604         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
605                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
606                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
607                 tdb->ecode = TDB_ERR_IO;
608                 goto fail;
609         }
610
611         /* make sure we know about any file expansions already done by
612            anyone else */
613         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
614         tdb->transaction->old_map_size = tdb->map_size;
615
616         /* finally hook the io methods, replacing them with
617            transaction specific methods */
618         tdb->transaction->io_methods = tdb->methods;
619         tdb->methods = &transaction_methods;
620
621         /* Trace at the end, so we get sequence number correct. */
622         tdb_trace(tdb, "tdb_transaction_start");
623         return 0;
624         
625 fail:
626         tdb_brunlock(tdb, F_RDLCK, FREELIST_TOP, 0);
627         tdb_transaction_unlock(tdb, F_WRLCK);
628         SAFE_FREE(tdb->transaction->blocks);
629         SAFE_FREE(tdb->transaction->hash_heads);
630         SAFE_FREE(tdb->transaction);
631         return -1;
632 }
633
634
635 /*
636   cancel the current transaction
637 */
638 int tdb_transaction_cancel(struct tdb_context *tdb)
639 {
640         int ltype = F_RDLCK;
641         tdb_trace(tdb, "tdb_transaction_cancel");
642         if (tdb->transaction && tdb->transaction->prepared)
643                 ltype = F_WRLCK;
644         return _tdb_transaction_cancel(tdb, ltype);
645 }
646
647 /*
648   work out how much space the linearised recovery data will consume
649 */
650 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
651 {
652         tdb_len_t recovery_size = 0;
653         int i;
654
655         recovery_size = sizeof(uint32_t);
656         for (i=0;i<tdb->transaction->num_blocks;i++) {
657                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
658                         break;
659                 }
660                 if (tdb->transaction->blocks[i] == NULL) {
661                         continue;
662                 }
663                 recovery_size += 2*sizeof(tdb_off_t);
664                 if (i == tdb->transaction->num_blocks-1) {
665                         recovery_size += tdb->transaction->last_block_size;
666                 } else {
667                         recovery_size += tdb->transaction->block_size;
668                 }
669         }       
670
671         return recovery_size;
672 }
673
674 /*
675   allocate the recovery area, or use an existing recovery area if it is
676   large enough
677 */
678 static int tdb_recovery_allocate(struct tdb_context *tdb, 
679                                  tdb_len_t *recovery_size,
680                                  tdb_off_t *recovery_offset,
681                                  tdb_len_t *recovery_max_size)
682 {
683         struct tdb_record rec;
684         const struct tdb_methods *methods = tdb->transaction->io_methods;
685         tdb_off_t recovery_head;
686
687         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
688                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
689                 return -1;
690         }
691
692         rec.rec_len = 0;
693
694         if (recovery_head != 0 && 
695             methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
696                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
697                 return -1;
698         }
699
700         *recovery_size = tdb_recovery_size(tdb);
701
702         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
703                 /* it fits in the existing area */
704                 *recovery_max_size = rec.rec_len;
705                 *recovery_offset = recovery_head;
706                 return 0;
707         }
708
709         /* we need to free up the old recovery area, then allocate a
710            new one at the end of the file. Note that we cannot use
711            tdb_allocate() to allocate the new one as that might return
712            us an area that is being currently used (as of the start of
713            the transaction) */
714         if (recovery_head != 0) {
715                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
716                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
717                         return -1;
718                 }
719         }
720
721         /* the tdb_free() call might have increased the recovery size */
722         *recovery_size = tdb_recovery_size(tdb);
723
724         /* round up to a multiple of page size */
725         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
726         *recovery_offset = tdb->map_size;
727         recovery_head = *recovery_offset;
728
729         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
730                                      (tdb->map_size - tdb->transaction->old_map_size) +
731                                      sizeof(rec) + *recovery_max_size) == -1) {
732                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
733                 return -1;
734         }
735
736         /* remap the file (if using mmap) */
737         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
738
739         /* we have to reset the old map size so that we don't try to expand the file
740            again in the transaction commit, which would destroy the recovery area */
741         tdb->transaction->old_map_size = tdb->map_size;
742
743         /* write the recovery header offset and sync - we can sync without a race here
744            as the magic ptr in the recovery record has not been set */
745         CONVERT(recovery_head);
746         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
747                                &recovery_head, sizeof(tdb_off_t)) == -1) {
748                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
749                 return -1;
750         }
751         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
752                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
753                 return -1;
754         }
755
756         return 0;
757 }
758
759
760 /*
761   setup the recovery data that will be used on a crash during commit
762 */
763 static int transaction_setup_recovery(struct tdb_context *tdb, 
764                                       tdb_off_t *magic_offset)
765 {
766         tdb_len_t recovery_size;
767         unsigned char *data, *p;
768         const struct tdb_methods *methods = tdb->transaction->io_methods;
769         struct tdb_record *rec;
770         tdb_off_t recovery_offset, recovery_max_size;
771         tdb_off_t old_map_size = tdb->transaction->old_map_size;
772         uint32_t magic, tailer;
773         int i;
774
775         /*
776           check that the recovery area has enough space
777         */
778         if (tdb_recovery_allocate(tdb, &recovery_size, 
779                                   &recovery_offset, &recovery_max_size) == -1) {
780                 return -1;
781         }
782
783         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
784         if (data == NULL) {
785                 tdb->ecode = TDB_ERR_OOM;
786                 return -1;
787         }
788
789         rec = (struct tdb_record *)data;
790         memset(rec, 0, sizeof(*rec));
791
792         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
793         rec->data_len = recovery_size;
794         rec->rec_len  = recovery_max_size;
795         rec->key_len  = old_map_size;
796         CONVERT(rec);
797
798         /* build the recovery data into a single blob to allow us to do a single
799            large write, which should be more efficient */
800         p = data + sizeof(*rec);
801         for (i=0;i<tdb->transaction->num_blocks;i++) {
802                 tdb_off_t offset;
803                 tdb_len_t length;
804
805                 if (tdb->transaction->blocks[i] == NULL) {
806                         continue;
807                 }
808
809                 offset = i * tdb->transaction->block_size;
810                 length = tdb->transaction->block_size;
811                 if (i == tdb->transaction->num_blocks-1) {
812                         length = tdb->transaction->last_block_size;
813                 }
814                 
815                 if (offset >= old_map_size) {
816                         continue;
817                 }
818                 if (offset + length > tdb->transaction->old_map_size) {
819                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
820                         free(data);
821                         tdb->ecode = TDB_ERR_CORRUPT;
822                         return -1;
823                 }
824                 memcpy(p, &offset, 4);
825                 memcpy(p+4, &length, 4);
826                 if (DOCONV()) {
827                         tdb_convert(p, 8);
828                 }
829                 /* the recovery area contains the old data, not the
830                    new data, so we have to call the original tdb_read
831                    method to get it */
832                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
833                         free(data);
834                         tdb->ecode = TDB_ERR_IO;
835                         return -1;
836                 }
837                 p += 8 + length;
838         }
839
840         /* and the tailer */
841         tailer = sizeof(*rec) + recovery_max_size;
842         memcpy(p, &tailer, 4);
843         CONVERT(p);
844
845         /* write the recovery data to the recovery area */
846         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
847                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
848                 free(data);
849                 tdb->ecode = TDB_ERR_IO;
850                 return -1;
851         }
852         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
853                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
854                 free(data);
855                 tdb->ecode = TDB_ERR_IO;
856                 return -1;
857         }
858
859         /* as we don't have ordered writes, we have to sync the recovery
860            data before we update the magic to indicate that the recovery
861            data is present */
862         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
863                 free(data);
864                 return -1;
865         }
866
867         free(data);
868
869         magic = TDB_RECOVERY_MAGIC;
870         CONVERT(magic);
871
872         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
873
874         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
875                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
876                 tdb->ecode = TDB_ERR_IO;
877                 return -1;
878         }
879         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
880                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
881                 tdb->ecode = TDB_ERR_IO;
882                 return -1;
883         }
884
885         /* ensure the recovery magic marker is on disk */
886         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
887                 return -1;
888         }
889
890         return 0;
891 }
892
893 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
894 {       
895         const struct tdb_methods *methods;
896
897         if (tdb->transaction == NULL) {
898                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
899                 return -1;
900         }
901
902         if (tdb->transaction->prepared) {
903                 tdb->ecode = TDB_ERR_EINVAL;
904                 _tdb_transaction_cancel(tdb, F_WRLCK);
905                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
906                 return -1;
907         }
908
909         if (tdb->transaction->transaction_error) {
910                 tdb->ecode = TDB_ERR_IO;
911                 _tdb_transaction_cancel(tdb, F_RDLCK);
912                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
913                 return -1;
914         }
915
916
917         if (tdb->transaction->nesting != 0) {
918                 tdb->transaction->nesting--;
919                 return 0;
920         }               
921
922         /* check for a null transaction */
923         if (tdb->transaction->blocks == NULL) {
924                 return 0;
925         }
926
927         methods = tdb->transaction->io_methods;
928         
929         /* if there are any locks pending then the caller has not
930            nested their locks properly, so fail the transaction */
931         if (tdb_have_extra_locks(tdb)) {
932                 tdb->ecode = TDB_ERR_LOCK;
933                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
934                 _tdb_transaction_cancel(tdb, F_RDLCK);
935                 return -1;
936         }
937
938         /* upgrade the main transaction lock region to a write lock */
939         if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
940                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
941                 tdb->ecode = TDB_ERR_LOCK;
942                 _tdb_transaction_cancel(tdb, F_RDLCK);
943                 return -1;
944         }
945
946         /* get the open lock - this prevents new users attaching to the database
947            during the commit */
948         if (tdb_brlock(tdb, F_WRLCK, OPEN_LOCK, 1, TDB_LOCK_WAIT) == -1) {
949                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
950                 tdb->ecode = TDB_ERR_LOCK;
951                 _tdb_transaction_cancel(tdb, F_WRLCK);
952                 return -1;
953         }
954
955         tdb->transaction->open_lock_taken = true;
956
957         if (!(tdb->flags & TDB_NOSYNC)) {
958                 /* write the recovery data to the end of the file */
959                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
960                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
961                         _tdb_transaction_cancel(tdb, F_WRLCK);
962                         return -1;
963                 }
964         }
965
966         tdb->transaction->prepared = true;
967
968         /* expand the file to the new size if needed */
969         if (tdb->map_size != tdb->transaction->old_map_size) {
970                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
971                                              tdb->map_size - 
972                                              tdb->transaction->old_map_size) == -1) {
973                         tdb->ecode = TDB_ERR_IO;
974                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
975                         _tdb_transaction_cancel(tdb, F_WRLCK);
976                         return -1;
977                 }
978                 tdb->map_size = tdb->transaction->old_map_size;
979                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
980         }
981
982         /* Keep the open lock until the actual commit */
983
984         return 0;
985 }
986
987 /*
988    prepare to commit the current transaction
989 */
990 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
991 {       
992         tdb_trace(tdb, "tdb_transaction_prepare_commit");
993         return _tdb_transaction_prepare_commit(tdb);
994 }
995
996 /*
997   commit the current transaction
998 */
999 int tdb_transaction_commit(struct tdb_context *tdb)
1000 {       
1001         const struct tdb_methods *methods;
1002         int i;
1003         bool need_repack;
1004
1005         if (tdb->transaction == NULL) {
1006                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1007                 return -1;
1008         }
1009
1010         tdb_trace(tdb, "tdb_transaction_commit");
1011
1012         if (tdb->transaction->transaction_error) {
1013                 tdb->ecode = TDB_ERR_IO;
1014                 tdb_transaction_cancel(tdb);
1015                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1016                 return -1;
1017         }
1018
1019
1020         if (tdb->transaction->nesting != 0) {
1021                 tdb->transaction->nesting--;
1022                 return 0;
1023         }
1024
1025         /* check for a null transaction */
1026         if (tdb->transaction->blocks == NULL) {
1027                 _tdb_transaction_cancel(tdb, F_RDLCK);
1028                 return 0;
1029         }
1030
1031         if (!tdb->transaction->prepared) {
1032                 int ret = _tdb_transaction_prepare_commit(tdb);
1033                 if (ret)
1034                         return ret;
1035         }
1036
1037         methods = tdb->transaction->io_methods;
1038
1039         /* perform all the writes */
1040         for (i=0;i<tdb->transaction->num_blocks;i++) {
1041                 tdb_off_t offset;
1042                 tdb_len_t length;
1043
1044                 if (tdb->transaction->blocks[i] == NULL) {
1045                         continue;
1046                 }
1047
1048                 offset = i * tdb->transaction->block_size;
1049                 length = tdb->transaction->block_size;
1050                 if (i == tdb->transaction->num_blocks-1) {
1051                         length = tdb->transaction->last_block_size;
1052                 }
1053
1054                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1055                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1056                         
1057                         /* we've overwritten part of the data and
1058                            possibly expanded the file, so we need to
1059                            run the crash recovery code */
1060                         tdb->methods = methods;
1061                         tdb_transaction_recover(tdb); 
1062
1063                         _tdb_transaction_cancel(tdb, F_WRLCK);
1064
1065                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1066                         return -1;
1067                 }
1068                 SAFE_FREE(tdb->transaction->blocks[i]);
1069         } 
1070
1071         SAFE_FREE(tdb->transaction->blocks);
1072         tdb->transaction->num_blocks = 0;
1073
1074         /* ensure the new data is on disk */
1075         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1076                 return -1;
1077         }
1078
1079         /*
1080           TODO: maybe write to some dummy hdr field, or write to magic
1081           offset without mmap, before the last sync, instead of the
1082           utime() call
1083         */
1084
1085         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1086            don't change the mtime of the file, this means the file may
1087            not be backed up (as tdb rounding to block sizes means that
1088            file size changes are quite rare too). The following forces
1089            mtime changes when a transaction completes */
1090 #if HAVE_UTIME
1091         utime(tdb->name, NULL);
1092 #endif
1093
1094         need_repack = tdb->transaction->need_repack;
1095
1096         /* use a transaction cancel to free memory and remove the
1097            transaction locks */
1098         _tdb_transaction_cancel(tdb, F_WRLCK);
1099
1100         if (need_repack) {
1101                 return tdb_repack(tdb);
1102         }
1103
1104         return 0;
1105 }
1106
1107
1108 /*
1109   recover from an aborted transaction. Must be called with exclusive
1110   database write access already established (including the open
1111   lock to prevent new processes attaching)
1112 */
1113 int tdb_transaction_recover(struct tdb_context *tdb)
1114 {
1115         tdb_off_t recovery_head, recovery_eof;
1116         unsigned char *data, *p;
1117         uint32_t zero = 0;
1118         struct tdb_record rec;
1119
1120         /* find the recovery area */
1121         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1122                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1123                 tdb->ecode = TDB_ERR_IO;
1124                 return -1;
1125         }
1126
1127         if (recovery_head == 0) {
1128                 /* we have never allocated a recovery record */
1129                 return 0;
1130         }
1131
1132         /* read the recovery record */
1133         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1134                                    sizeof(rec), DOCONV()) == -1) {
1135                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1136                 tdb->ecode = TDB_ERR_IO;
1137                 return -1;
1138         }
1139
1140         if (rec.magic != TDB_RECOVERY_MAGIC) {
1141                 /* there is no valid recovery data */
1142                 return 0;
1143         }
1144
1145         if (tdb->read_only) {
1146                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1147                 tdb->ecode = TDB_ERR_CORRUPT;
1148                 return -1;
1149         }
1150
1151         recovery_eof = rec.key_len;
1152
1153         data = (unsigned char *)malloc(rec.data_len);
1154         if (data == NULL) {
1155                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1156                 tdb->ecode = TDB_ERR_OOM;
1157                 return -1;
1158         }
1159
1160         /* read the full recovery data */
1161         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1162                                    rec.data_len, 0) == -1) {
1163                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1164                 tdb->ecode = TDB_ERR_IO;
1165                 return -1;
1166         }
1167
1168         /* recover the file data */
1169         p = data;
1170         while (p+8 < data + rec.data_len) {
1171                 uint32_t ofs, len;
1172                 if (DOCONV()) {
1173                         tdb_convert(p, 8);
1174                 }
1175                 memcpy(&ofs, p, 4);
1176                 memcpy(&len, p+4, 4);
1177
1178                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1179                         free(data);
1180                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1181                         tdb->ecode = TDB_ERR_IO;
1182                         return -1;
1183                 }
1184                 p += 8 + len;
1185         }
1186
1187         free(data);
1188
1189         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1190                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1191                 tdb->ecode = TDB_ERR_IO;
1192                 return -1;
1193         }
1194
1195         /* if the recovery area is after the recovered eof then remove it */
1196         if (recovery_eof <= recovery_head) {
1197                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1198                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1199                         tdb->ecode = TDB_ERR_IO;
1200                         return -1;                      
1201                 }
1202         }
1203
1204         /* remove the recovery magic */
1205         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1206                           &zero) == -1) {
1207                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1208                 tdb->ecode = TDB_ERR_IO;
1209                 return -1;                      
1210         }
1211         
1212         /* reduce the file size to the old size */
1213         tdb_munmap(tdb);
1214         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1215                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1216                 tdb->ecode = TDB_ERR_IO;
1217                 return -1;                      
1218         }
1219         tdb->map_size = recovery_eof;
1220         tdb_mmap(tdb);
1221
1222         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1223                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1224                 tdb->ecode = TDB_ERR_IO;
1225                 return -1;
1226         }
1227
1228         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1229                  recovery_eof));
1230
1231         /* all done */
1232         return 0;
1233 }