4a49cc7b6421848a0eac1efbff2cb9e4913bb367
[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     global lock is held. Automatically recover from the transaction
80     recovery area if needed, then continue with the open as
81     usual. This allows for smooth crash recovery with no administrator
82     intervention.
83
84   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85     still available, but no transaction recovery area is used and no
86     fsync/msync calls are made.
87
88   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89     tdb_add_flags() transaction is enabled.
90     The default is that transaction nesting is not allowed and an attempt
91     to create a nested transaction will fail with TDB_ERR_NESTING.
92
93     Beware. when transactions are nested a transaction successfully
94     completed with tdb_transaction_commit() can be silently unrolled later.
95 */
96
97
98 /*
99   hold the context of any current transaction
100 */
101 struct tdb_transaction {
102         /* we keep a mirrored copy of the tdb hash heads here so
103            tdb_next_hash_chain() can operate efficiently */
104         uint32_t *hash_heads;
105
106         /* the original io methods - used to do IOs to the real db */
107         const struct tdb_methods *io_methods;
108
109         /* the list of transaction blocks. When a block is first
110            written to, it gets created in this list */
111         uint8_t **blocks;
112         uint32_t num_blocks;
113         uint32_t block_size;      /* bytes in each block */
114         uint32_t last_block_size; /* number of valid bytes in the last block */
115
116         /* non-zero when an internal transaction error has
117            occurred. All write operations will then fail until the
118            transaction is ended */
119         int transaction_error;
120
121         /* when inside a transaction we need to keep track of any
122            nested tdb_transaction_start() calls, as these are allowed,
123            but don't create a new transaction */
124         int nesting;
125
126         /* set when a prepare has already occurred */
127         bool prepared;
128         tdb_off_t magic_offset;
129
130         /* old file size before transaction */
131         tdb_len_t old_map_size;
132
133         /* we should re-pack on commit */
134         bool need_repack;
135 };
136
137
138 /*
139   read while in a transaction. We need to check first if the data is in our list
140   of transaction elements, then if not do a real read
141 */
142 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
143                             tdb_len_t len, int cv)
144 {
145         uint32_t blk;
146
147         /* break it down into block sized ops */
148         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
149                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
150                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
151                         return -1;
152                 }
153                 len -= len2;
154                 off += len2;
155                 buf = (void *)(len2 + (char *)buf);
156         }
157
158         if (len == 0) {
159                 return 0;
160         }
161
162         blk = off / tdb->transaction->block_size;
163
164         /* see if we have it in the block list */
165         if (tdb->transaction->num_blocks <= blk ||
166             tdb->transaction->blocks[blk] == NULL) {
167                 /* nope, do a real read */
168                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
169                         goto fail;
170                 }
171                 return 0;
172         }
173
174         /* it is in the block list. Now check for the last block */
175         if (blk == tdb->transaction->num_blocks-1) {
176                 if (len > tdb->transaction->last_block_size) {
177                         goto fail;
178                 }
179         }
180         
181         /* now copy it out of this block */
182         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
183         if (cv) {
184                 tdb_convert(buf, len);
185         }
186         return 0;
187
188 fail:
189         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
190         tdb->ecode = TDB_ERR_IO;
191         tdb->transaction->transaction_error = 1;
192         return -1;
193 }
194
195
196 /*
197   write while in a transaction
198 */
199 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
200                              const void *buf, tdb_len_t len)
201 {
202         uint32_t blk;
203
204         /* Only a commit is allowed on a prepared transaction */
205         if (tdb->transaction->prepared) {
206                 tdb->ecode = TDB_ERR_EINVAL;
207                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
208                 tdb->transaction->transaction_error = 1;
209                 return -1;
210         }
211
212         /* if the write is to a hash head, then update the transaction
213            hash heads */
214         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
215             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
216                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
217                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
218         }
219
220         /* break it up into block sized chunks */
221         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
222                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
223                 if (transaction_write(tdb, off, buf, len2) != 0) {
224                         return -1;
225                 }
226                 len -= len2;
227                 off += len2;
228                 if (buf != NULL) {
229                         buf = (const void *)(len2 + (const char *)buf);
230                 }
231         }
232
233         if (len == 0) {
234                 return 0;
235         }
236
237         blk = off / tdb->transaction->block_size;
238         off = off % tdb->transaction->block_size;
239
240         if (tdb->transaction->num_blocks <= blk) {
241                 uint8_t **new_blocks;
242                 /* expand the blocks array */
243                 if (tdb->transaction->blocks == NULL) {
244                         new_blocks = (uint8_t **)malloc(
245                                 (blk+1)*sizeof(uint8_t *));
246                 } else {
247                         new_blocks = (uint8_t **)realloc(
248                                 tdb->transaction->blocks,
249                                 (blk+1)*sizeof(uint8_t *));
250                 }
251                 if (new_blocks == NULL) {
252                         tdb->ecode = TDB_ERR_OOM;
253                         goto fail;
254                 }
255                 memset(&new_blocks[tdb->transaction->num_blocks], 0, 
256                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
257                 tdb->transaction->blocks = new_blocks;
258                 tdb->transaction->num_blocks = blk+1;
259                 tdb->transaction->last_block_size = 0;
260         }
261
262         /* allocate and fill a block? */
263         if (tdb->transaction->blocks[blk] == NULL) {
264                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
265                 if (tdb->transaction->blocks[blk] == NULL) {
266                         tdb->ecode = TDB_ERR_OOM;
267                         tdb->transaction->transaction_error = 1;
268                         return -1;                      
269                 }
270                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
271                         tdb_len_t len2 = tdb->transaction->block_size;
272                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
273                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
274                         }
275                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
276                                                                    tdb->transaction->blocks[blk], 
277                                                                    len2, 0) != 0) {
278                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
279                                 tdb->ecode = TDB_ERR_IO;
280                                 goto fail;
281                         }
282                         if (blk == tdb->transaction->num_blocks-1) {
283                                 tdb->transaction->last_block_size = len2;
284                         }                       
285                 }
286         }
287         
288         /* overwrite part of an existing block */
289         if (buf == NULL) {
290                 memset(tdb->transaction->blocks[blk] + off, 0, len);
291         } else {
292                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
293         }
294         if (blk == tdb->transaction->num_blocks-1) {
295                 if (len + off > tdb->transaction->last_block_size) {
296                         tdb->transaction->last_block_size = len + off;
297                 }
298         }
299
300         return 0;
301
302 fail:
303         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", 
304                  (blk*tdb->transaction->block_size) + off, len));
305         tdb->transaction->transaction_error = 1;
306         return -1;
307 }
308
309
310 /*
311   write while in a transaction - this varient never expands the transaction blocks, it only
312   updates existing blocks. This means it cannot change the recovery size
313 */
314 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
315                                       const void *buf, tdb_len_t len)
316 {
317         uint32_t blk;
318
319         /* break it up into block sized chunks */
320         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
321                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
322                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
323                         return -1;
324                 }
325                 len -= len2;
326                 off += len2;
327                 if (buf != NULL) {
328                         buf = (const void *)(len2 + (const char *)buf);
329                 }
330         }
331
332         if (len == 0) {
333                 return 0;
334         }
335
336         blk = off / tdb->transaction->block_size;
337         off = off % tdb->transaction->block_size;
338
339         if (tdb->transaction->num_blocks <= blk ||
340             tdb->transaction->blocks[blk] == NULL) {
341                 return 0;
342         }
343
344         if (blk == tdb->transaction->num_blocks-1 &&
345             off + len > tdb->transaction->last_block_size) {
346                 if (off >= tdb->transaction->last_block_size) {
347                         return 0;
348                 }
349                 len = tdb->transaction->last_block_size - off;
350         }
351
352         /* overwrite part of an existing block */
353         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
354
355         return 0;
356 }
357
358
359 /*
360   accelerated hash chain head search, using the cached hash heads
361 */
362 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
363 {
364         uint32_t h = *chain;
365         for (;h < tdb->header.hash_size;h++) {
366                 /* the +1 takes account of the freelist */
367                 if (0 != tdb->transaction->hash_heads[h+1]) {
368                         break;
369                 }
370         }
371         (*chain) = h;
372 }
373
374 /*
375   out of bounds check during a transaction
376 */
377 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
378 {
379         if (len <= tdb->map_size) {
380                 return 0;
381         }
382         tdb->ecode = TDB_ERR_IO;
383         return -1;
384 }
385
386 /*
387   transaction version of tdb_expand().
388 */
389 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
390                                    tdb_off_t addition)
391 {
392         /* add a write to the transaction elements, so subsequent
393            reads see the zero data */
394         if (transaction_write(tdb, size, NULL, addition) != 0) {
395                 return -1;
396         }
397
398         tdb->transaction->need_repack = true;
399
400         return 0;
401 }
402
403 /*
404   brlock during a transaction - ignore them
405 */
406 static int transaction_brlock(struct tdb_context *tdb,
407                               int rw_type, tdb_off_t offset, size_t len,
408                               enum tdb_lock_flags flags)
409 {
410         return 0;
411 }
412
413 static int transaction_brunlock(struct tdb_context *tdb,
414                                 int rw_type, tdb_off_t offset, size_t len)
415 {
416         return 0;
417 }
418
419 static const struct tdb_methods transaction_methods = {
420         transaction_read,
421         transaction_write,
422         transaction_next_hash_chain,
423         transaction_oob,
424         transaction_expand_file,
425         transaction_brlock,
426         transaction_brunlock
427 };
428
429 /*
430   sync to disk
431 */
432 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
433 {       
434         if (tdb->flags & TDB_NOSYNC) {
435                 return 0;
436         }
437
438         if (fsync(tdb->fd) != 0) {
439                 tdb->ecode = TDB_ERR_IO;
440                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
441                 return -1;
442         }
443 #ifdef MS_SYNC
444         if (tdb->map_ptr) {
445                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
446                 if (msync(moffset + (char *)tdb->map_ptr, 
447                           length + (offset - moffset), MS_SYNC) != 0) {
448                         tdb->ecode = TDB_ERR_IO;
449                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
450                                  strerror(errno)));
451                         return -1;
452                 }
453         }
454 #endif
455         return 0;
456 }
457
458 /* ltype is F_WRLCK after prepare. */
459 int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype)
460 {
461         int i, ret = 0;
462
463         if (tdb->transaction == NULL) {
464                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
465                 return -1;
466         }
467
468         if (tdb->transaction->nesting != 0) {
469                 tdb->transaction->transaction_error = 1;
470                 tdb->transaction->nesting--;
471                 return 0;
472         }               
473
474         tdb->map_size = tdb->transaction->old_map_size;
475
476         /* free all the transaction blocks */
477         for (i=0;i<tdb->transaction->num_blocks;i++) {
478                 if (tdb->transaction->blocks[i] != NULL) {
479                         free(tdb->transaction->blocks[i]);
480                 }
481         }
482         SAFE_FREE(tdb->transaction->blocks);
483
484         if (tdb->transaction->magic_offset) {
485                 const struct tdb_methods *methods = tdb->transaction->io_methods;
486                 uint32_t zero = 0;
487
488                 /* remove the recovery marker */
489                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
490                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
491                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
492                         ret = -1;
493                 }
494         }
495
496         /* remove any global lock created during the transaction */
497         if (tdb->global_lock.count != 0) {
498                 tdb_brunlock(tdb, tdb->global_lock.ltype,
499                              FREELIST_TOP, 4*tdb->header.hash_size);
500                 tdb->global_lock.count = 0;
501         }
502
503         /* remove any locks created during the transaction */
504         if (tdb->num_locks != 0) {
505                 for (i=0;i<tdb->num_lockrecs;i++) {
506                         tdb_brunlock(tdb, tdb->lockrecs[i].ltype,
507                                      FREELIST_TOP+4*tdb->lockrecs[i].list, 1);
508                 }
509                 tdb->num_locks = 0;
510                 tdb->num_lockrecs = 0;
511                 SAFE_FREE(tdb->lockrecs);
512         }
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->num_locks != 0 || tdb->global_lock.count) {
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 list_struct 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 list_struct *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 list_struct *)data;
790         memset(rec, 0, sizeof(*rec));
791
792         rec->magic    = 0;
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 list_struct, 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 #ifdef TDB_TRACE
923         /* store seqnum now, before reading becomes illegal. */
924         tdb_ofs_read(tdb, TDB_SEQNUM_OFS, &tdb->transaction_prepare_seqnum);
925 #endif
926
927         /* check for a null transaction */
928         if (tdb->transaction->blocks == NULL) {
929                 return 0;
930         }
931
932         methods = tdb->transaction->io_methods;
933         
934         /* if there are any locks pending then the caller has not
935            nested their locks properly, so fail the transaction */
936         if (tdb->num_locks || tdb->global_lock.count) {
937                 tdb->ecode = TDB_ERR_LOCK;
938                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
939                 _tdb_transaction_cancel(tdb, F_RDLCK);
940                 return -1;
941         }
942
943         /* upgrade the main transaction lock region to a write lock */
944         if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
945                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
946                 tdb->ecode = TDB_ERR_LOCK;
947                 _tdb_transaction_cancel(tdb, F_RDLCK);
948                 return -1;
949         }
950
951         /* get the global lock - this prevents new users attaching to the database
952            during the commit */
953         if (tdb_brlock(tdb, F_WRLCK, GLOBAL_LOCK, 1, TDB_LOCK_WAIT) == -1) {
954                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
955                 tdb->ecode = TDB_ERR_LOCK;
956                 _tdb_transaction_cancel(tdb, F_WRLCK);
957                 return -1;
958         }
959
960         if (!(tdb->flags & TDB_NOSYNC)) {
961                 /* write the recovery data to the end of the file */
962                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
963                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
964                         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
965                         _tdb_transaction_cancel(tdb, F_WRLCK);
966                         return -1;
967                 }
968         }
969
970         tdb->transaction->prepared = true;
971
972         /* expand the file to the new size if needed */
973         if (tdb->map_size != tdb->transaction->old_map_size) {
974                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
975                                              tdb->map_size - 
976                                              tdb->transaction->old_map_size) == -1) {
977                         tdb->ecode = TDB_ERR_IO;
978                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
979                         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
980                         _tdb_transaction_cancel(tdb, F_WRLCK);
981                         return -1;
982                 }
983                 tdb->map_size = tdb->transaction->old_map_size;
984                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
985         }
986
987         /* Keep the global lock until the actual commit */
988
989         return 0;
990 }
991
992 /*
993    prepare to commit the current transaction
994 */
995 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
996 {       
997         tdb_trace(tdb, "tdb_transaction_prepare_commit");
998         return _tdb_transaction_prepare_commit(tdb);
999 }
1000
1001 /*
1002   commit the current transaction
1003 */
1004 int tdb_transaction_commit(struct tdb_context *tdb)
1005 {       
1006         const struct tdb_methods *methods;
1007         int i;
1008         bool need_repack;
1009
1010         if (tdb->transaction == NULL) {
1011                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1012                 return -1;
1013         }
1014
1015         /* If we've prepared, can't read seqnum. */
1016         if (tdb->transaction->prepared) {
1017                 tdb_trace_seqnum(tdb, tdb->transaction_prepare_seqnum,
1018                                  "tdb_transaction_commit");
1019         } else {
1020                 tdb_trace(tdb, "tdb_transaction_commit");
1021         }
1022
1023         if (tdb->transaction->transaction_error) {
1024                 tdb->ecode = TDB_ERR_IO;
1025                 tdb_transaction_cancel(tdb);
1026                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1027                 return -1;
1028         }
1029
1030
1031         if (tdb->transaction->nesting != 0) {
1032                 tdb->transaction->nesting--;
1033                 return 0;
1034         }
1035
1036         /* check for a null transaction */
1037         if (tdb->transaction->blocks == NULL) {
1038                 _tdb_transaction_cancel(tdb, F_RDLCK);
1039                 return 0;
1040         }
1041
1042         if (!tdb->transaction->prepared) {
1043                 int ret = _tdb_transaction_prepare_commit(tdb);
1044                 if (ret)
1045                         return ret;
1046         }
1047
1048         methods = tdb->transaction->io_methods;
1049
1050         /* perform all the writes */
1051         for (i=0;i<tdb->transaction->num_blocks;i++) {
1052                 tdb_off_t offset;
1053                 tdb_len_t length;
1054
1055                 if (tdb->transaction->blocks[i] == NULL) {
1056                         continue;
1057                 }
1058
1059                 offset = i * tdb->transaction->block_size;
1060                 length = tdb->transaction->block_size;
1061                 if (i == tdb->transaction->num_blocks-1) {
1062                         length = tdb->transaction->last_block_size;
1063                 }
1064
1065                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1066                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1067                         
1068                         /* we've overwritten part of the data and
1069                            possibly expanded the file, so we need to
1070                            run the crash recovery code */
1071                         tdb->methods = methods;
1072                         tdb_transaction_recover(tdb); 
1073
1074                         _tdb_transaction_cancel(tdb, F_WRLCK);
1075                         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
1076
1077                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1078                         return -1;
1079                 }
1080                 SAFE_FREE(tdb->transaction->blocks[i]);
1081         } 
1082
1083         SAFE_FREE(tdb->transaction->blocks);
1084         tdb->transaction->num_blocks = 0;
1085
1086         /* ensure the new data is on disk */
1087         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1088                 return -1;
1089         }
1090
1091         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
1092
1093         /*
1094           TODO: maybe write to some dummy hdr field, or write to magic
1095           offset without mmap, before the last sync, instead of the
1096           utime() call
1097         */
1098
1099         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1100            don't change the mtime of the file, this means the file may
1101            not be backed up (as tdb rounding to block sizes means that
1102            file size changes are quite rare too). The following forces
1103            mtime changes when a transaction completes */
1104 #if HAVE_UTIME
1105         utime(tdb->name, NULL);
1106 #endif
1107
1108         need_repack = tdb->transaction->need_repack;
1109
1110         /* use a transaction cancel to free memory and remove the
1111            transaction locks */
1112         _tdb_transaction_cancel(tdb, F_WRLCK);
1113
1114         if (need_repack) {
1115                 return tdb_repack(tdb);
1116         }
1117
1118         return 0;
1119 }
1120
1121
1122 /*
1123   recover from an aborted transaction. Must be called with exclusive
1124   database write access already established (including the global
1125   lock to prevent new processes attaching)
1126 */
1127 int tdb_transaction_recover(struct tdb_context *tdb)
1128 {
1129         tdb_off_t recovery_head, recovery_eof;
1130         unsigned char *data, *p;
1131         uint32_t zero = 0;
1132         struct list_struct rec;
1133
1134         /* find the recovery area */
1135         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1136                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1137                 tdb->ecode = TDB_ERR_IO;
1138                 return -1;
1139         }
1140
1141         if (recovery_head == 0) {
1142                 /* we have never allocated a recovery record */
1143                 return 0;
1144         }
1145
1146         /* read the recovery record */
1147         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1148                                    sizeof(rec), DOCONV()) == -1) {
1149                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1150                 tdb->ecode = TDB_ERR_IO;
1151                 return -1;
1152         }
1153
1154         if (rec.magic != TDB_RECOVERY_MAGIC) {
1155                 /* there is no valid recovery data */
1156                 return 0;
1157         }
1158
1159         if (tdb->read_only) {
1160                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1161                 tdb->ecode = TDB_ERR_CORRUPT;
1162                 return -1;
1163         }
1164
1165         recovery_eof = rec.key_len;
1166
1167         data = (unsigned char *)malloc(rec.data_len);
1168         if (data == NULL) {
1169                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1170                 tdb->ecode = TDB_ERR_OOM;
1171                 return -1;
1172         }
1173
1174         /* read the full recovery data */
1175         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1176                                    rec.data_len, 0) == -1) {
1177                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1178                 tdb->ecode = TDB_ERR_IO;
1179                 return -1;
1180         }
1181
1182         /* recover the file data */
1183         p = data;
1184         while (p+8 < data + rec.data_len) {
1185                 uint32_t ofs, len;
1186                 if (DOCONV()) {
1187                         tdb_convert(p, 8);
1188                 }
1189                 memcpy(&ofs, p, 4);
1190                 memcpy(&len, p+4, 4);
1191
1192                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1193                         free(data);
1194                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1195                         tdb->ecode = TDB_ERR_IO;
1196                         return -1;
1197                 }
1198                 p += 8 + len;
1199         }
1200
1201         free(data);
1202
1203         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1204                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1205                 tdb->ecode = TDB_ERR_IO;
1206                 return -1;
1207         }
1208
1209         /* if the recovery area is after the recovered eof then remove it */
1210         if (recovery_eof <= recovery_head) {
1211                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1212                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1213                         tdb->ecode = TDB_ERR_IO;
1214                         return -1;                      
1215                 }
1216         }
1217
1218         /* remove the recovery magic */
1219         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic), 
1220                           &zero) == -1) {
1221                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1222                 tdb->ecode = TDB_ERR_IO;
1223                 return -1;                      
1224         }
1225         
1226         /* reduce the file size to the old size */
1227         tdb_munmap(tdb);
1228         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1229                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1230                 tdb->ecode = TDB_ERR_IO;
1231                 return -1;                      
1232         }
1233         tdb->map_size = recovery_eof;
1234         tdb_mmap(tdb);
1235
1236         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1237                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1238                 tdb->ecode = TDB_ERR_IO;
1239                 return -1;
1240         }
1241
1242         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1243                  recovery_eof));
1244
1245         /* all done */
1246         return 0;
1247 }