tdb: tdb_allrecord_lock/tdb_allrecord_unlock/tdb_allrecord_upgrade
[ccan] / ccan / tdb / transaction.c
1  /* 
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              2005
7
8      ** NOTE! The following LGPL license applies to the tdb
9      ** library. This does NOT imply that all of Samba is released
10      ** under the LGPL
11    
12    This library is free software; you can redistribute it and/or
13    modify it under the terms of the GNU Lesser General Public
14    License as published by the Free Software Foundation; either
15    version 3 of the License, or (at your option) any later version.
16
17    This library is distributed in the hope that it will be useful,
18    but WITHOUT ANY WARRANTY; without even the implied warranty of
19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
20    Lesser General Public License for more details.
21
22    You should have received a copy of the GNU Lesser General Public
23    License along with this library; if not, see <http://www.gnu.org/licenses/>.
24 */
25
26 #include "tdb_private.h"
27
28 /*
29   transaction design:
30
31   - only allow a single transaction at a time per database. This makes
32     using the transaction API simpler, as otherwise the caller would
33     have to cope with temporary failures in transactions that conflict
34     with other current transactions
35
36   - keep the transaction recovery information in the same file as the
37     database, using a special 'transaction recovery' record pointed at
38     by the header. This removes the need for extra journal files as
39     used by some other databases
40
41   - dynamically allocated the transaction recover record, re-using it
42     for subsequent transactions. If a larger record is needed then
43     tdb_free() the old record to place it on the normal tdb freelist
44     before allocating the new record
45
46   - during transactions, keep a linked list of writes all that have
47     been performed by intercepting all tdb_write() calls. The hooked
48     transaction versions of tdb_read() and tdb_write() check this
49     linked list and try to use the elements of the list in preference
50     to the real database.
51
52   - don't allow any locks to be held when a transaction starts,
53     otherwise we can end up with deadlock (plus lack of lock nesting
54     in posix locks would mean the lock is lost)
55
56   - if the caller gains a lock during the transaction but doesn't
57     release it then fail the commit
58
59   - allow for nested calls to tdb_transaction_start(), re-using the
60     existing transaction record. If the inner transaction is cancelled
61     then a subsequent commit will fail
62  
63   - keep a mirrored copy of the tdb hash chain heads to allow for the
64     fast hash heads scan on traverse, updating the mirrored copy in
65     the transaction version of tdb_write
66
67   - allow callers to mix transaction and non-transaction use of tdb,
68     although once a transaction is started then an exclusive lock is
69     gained until the transaction is committed or cancelled
70
71   - the commit stategy involves first saving away all modified data
72     into a linearised buffer in the transaction recovery area, then
73     marking the transaction recovery area with a magic value to
74     indicate a valid recovery record. In total 4 fsync/msync calls are
75     needed per commit to prevent race conditions. It might be possible
76     to reduce this to 3 or even 2 with some more work.
77
78   - check for a valid recovery record on open of the tdb, while the
79     open lock is held. Automatically recover from the transaction
80     recovery area if needed, then continue with the open as
81     usual. This allows for smooth crash recovery with no administrator
82     intervention.
83
84   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85     still available, but no transaction recovery area is used and no
86     fsync/msync calls are made.
87
88   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89     tdb_add_flags() transaction nesting is enabled.
90     It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
91     The default is that transaction nesting is allowed.
92     Note: this default may change in future versions of tdb.
93
94     Beware. when transactions are nested a transaction successfully
95     completed with tdb_transaction_commit() can be silently unrolled later.
96
97   - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
98     tdb_add_flags() transaction nesting is disabled.
99     It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
100     An attempt create a nested transaction will fail with TDB_ERR_NESTING.
101     The default is that transaction nesting is allowed.
102     Note: this default may change in future versions of tdb.
103 */
104
105
106 /*
107   hold the context of any current transaction
108 */
109 struct tdb_transaction {
110         /* we keep a mirrored copy of the tdb hash heads here so
111            tdb_next_hash_chain() can operate efficiently */
112         uint32_t *hash_heads;
113
114         /* the original io methods - used to do IOs to the real db */
115         const struct tdb_methods *io_methods;
116
117         /* the list of transaction blocks. When a block is first
118            written to, it gets created in this list */
119         uint8_t **blocks;
120         uint32_t num_blocks;
121         uint32_t block_size;      /* bytes in each block */
122         uint32_t last_block_size; /* number of valid bytes in the last block */
123
124         /* non-zero when an internal transaction error has
125            occurred. All write operations will then fail until the
126            transaction is ended */
127         int transaction_error;
128
129         /* when inside a transaction we need to keep track of any
130            nested tdb_transaction_start() calls, as these are allowed,
131            but don't create a new transaction */
132         int nesting;
133
134         /* set when a prepare has already occurred */
135         bool prepared;
136         tdb_off_t magic_offset;
137
138         /* old file size before transaction */
139         tdb_len_t old_map_size;
140
141         /* we should re-pack on commit */
142         bool need_repack;
143 };
144
145
146 /*
147   read while in a transaction. We need to check first if the data is in our list
148   of transaction elements, then if not do a real read
149 */
150 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
151                             tdb_len_t len, int cv)
152 {
153         uint32_t blk;
154
155         /* break it down into block sized ops */
156         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
157                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
158                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
159                         return -1;
160                 }
161                 len -= len2;
162                 off += len2;
163                 buf = (void *)(len2 + (char *)buf);
164         }
165
166         if (len == 0) {
167                 return 0;
168         }
169
170         blk = off / tdb->transaction->block_size;
171
172         /* see if we have it in the block list */
173         if (tdb->transaction->num_blocks <= blk ||
174             tdb->transaction->blocks[blk] == NULL) {
175                 /* nope, do a real read */
176                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
177                         goto fail;
178                 }
179                 return 0;
180         }
181
182         /* it is in the block list. Now check for the last block */
183         if (blk == tdb->transaction->num_blocks-1) {
184                 if (len > tdb->transaction->last_block_size) {
185                         goto fail;
186                 }
187         }
188         
189         /* now copy it out of this block */
190         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
191         if (cv) {
192                 tdb_convert(buf, len);
193         }
194         return 0;
195
196 fail:
197         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
198         tdb->ecode = TDB_ERR_IO;
199         tdb->transaction->transaction_error = 1;
200         return -1;
201 }
202
203
204 /*
205   write while in a transaction
206 */
207 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
208                              const void *buf, tdb_len_t len)
209 {
210         uint32_t blk;
211
212         /* Only a commit is allowed on a prepared transaction */
213         if (tdb->transaction->prepared) {
214                 tdb->ecode = TDB_ERR_EINVAL;
215                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
216                 tdb->transaction->transaction_error = 1;
217                 return -1;
218         }
219
220         /* if the write is to a hash head, then update the transaction
221            hash heads */
222         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
223             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
224                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
225                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
226         }
227
228         /* break it up into block sized chunks */
229         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
230                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
231                 if (transaction_write(tdb, off, buf, len2) != 0) {
232                         return -1;
233                 }
234                 len -= len2;
235                 off += len2;
236                 if (buf != NULL) {
237                         buf = (const void *)(len2 + (const char *)buf);
238                 }
239         }
240
241         if (len == 0) {
242                 return 0;
243         }
244
245         blk = off / tdb->transaction->block_size;
246         off = off % tdb->transaction->block_size;
247
248         if (tdb->transaction->num_blocks <= blk) {
249                 uint8_t **new_blocks;
250                 /* expand the blocks array */
251                 if (tdb->transaction->blocks == NULL) {
252                         new_blocks = (uint8_t **)malloc(
253                                 (blk+1)*sizeof(uint8_t *));
254                 } else {
255                         new_blocks = (uint8_t **)realloc(
256                                 tdb->transaction->blocks,
257                                 (blk+1)*sizeof(uint8_t *));
258                 }
259                 if (new_blocks == NULL) {
260                         tdb->ecode = TDB_ERR_OOM;
261                         goto fail;
262                 }
263                 memset(&new_blocks[tdb->transaction->num_blocks], 0, 
264                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
265                 tdb->transaction->blocks = new_blocks;
266                 tdb->transaction->num_blocks = blk+1;
267                 tdb->transaction->last_block_size = 0;
268         }
269
270         /* allocate and fill a block? */
271         if (tdb->transaction->blocks[blk] == NULL) {
272                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
273                 if (tdb->transaction->blocks[blk] == NULL) {
274                         tdb->ecode = TDB_ERR_OOM;
275                         tdb->transaction->transaction_error = 1;
276                         return -1;                      
277                 }
278                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
279                         tdb_len_t len2 = tdb->transaction->block_size;
280                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
281                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
282                         }
283                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
284                                                                    tdb->transaction->blocks[blk], 
285                                                                    len2, 0) != 0) {
286                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
287                                 tdb->ecode = TDB_ERR_IO;
288                                 goto fail;
289                         }
290                         if (blk == tdb->transaction->num_blocks-1) {
291                                 tdb->transaction->last_block_size = len2;
292                         }                       
293                 }
294         }
295         
296         /* overwrite part of an existing block */
297         if (buf == NULL) {
298                 memset(tdb->transaction->blocks[blk] + off, 0, len);
299         } else {
300                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
301         }
302         if (blk == tdb->transaction->num_blocks-1) {
303                 if (len + off > tdb->transaction->last_block_size) {
304                         tdb->transaction->last_block_size = len + off;
305                 }
306         }
307
308         return 0;
309
310 fail:
311         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", 
312                  (blk*tdb->transaction->block_size) + off, len));
313         tdb->transaction->transaction_error = 1;
314         return -1;
315 }
316
317
318 /*
319   write while in a transaction - this varient never expands the transaction blocks, it only
320   updates existing blocks. This means it cannot change the recovery size
321 */
322 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
323                                       const void *buf, tdb_len_t len)
324 {
325         uint32_t blk;
326
327         /* break it up into block sized chunks */
328         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
329                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
330                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
331                         return -1;
332                 }
333                 len -= len2;
334                 off += len2;
335                 if (buf != NULL) {
336                         buf = (const void *)(len2 + (const char *)buf);
337                 }
338         }
339
340         if (len == 0) {
341                 return 0;
342         }
343
344         blk = off / tdb->transaction->block_size;
345         off = off % tdb->transaction->block_size;
346
347         if (tdb->transaction->num_blocks <= blk ||
348             tdb->transaction->blocks[blk] == NULL) {
349                 return 0;
350         }
351
352         if (blk == tdb->transaction->num_blocks-1 &&
353             off + len > tdb->transaction->last_block_size) {
354                 if (off >= tdb->transaction->last_block_size) {
355                         return 0;
356                 }
357                 len = tdb->transaction->last_block_size - off;
358         }
359
360         /* overwrite part of an existing block */
361         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
362
363         return 0;
364 }
365
366
367 /*
368   accelerated hash chain head search, using the cached hash heads
369 */
370 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
371 {
372         uint32_t h = *chain;
373         for (;h < tdb->header.hash_size;h++) {
374                 /* the +1 takes account of the freelist */
375                 if (0 != tdb->transaction->hash_heads[h+1]) {
376                         break;
377                 }
378         }
379         (*chain) = h;
380 }
381
382 /*
383   out of bounds check during a transaction
384 */
385 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
386 {
387         if (len <= tdb->map_size) {
388                 return 0;
389         }
390         tdb->ecode = TDB_ERR_IO;
391         return -1;
392 }
393
394 /*
395   transaction version of tdb_expand().
396 */
397 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
398                                    tdb_off_t addition)
399 {
400         /* add a write to the transaction elements, so subsequent
401            reads see the zero data */
402         if (transaction_write(tdb, size, NULL, addition) != 0) {
403                 return -1;
404         }
405
406         tdb->transaction->need_repack = true;
407
408         return 0;
409 }
410
411 /*
412   brlock during a transaction - ignore them
413 */
414 static int transaction_brlock(struct tdb_context *tdb,
415                               int rw_type, tdb_off_t offset, size_t len,
416                               enum tdb_lock_flags flags)
417 {
418         /* FIXME: We actually grab the open lock during a transaction. */
419         if (offset == OPEN_LOCK)
420                 return tdb_brlock(tdb, rw_type, offset, len, flags);
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         /* This also removes the OPEN_LOCK, if we have it. */
508         tdb_release_extra_locks(tdb);
509
510         /* restore the normal io methods */
511         tdb->methods = tdb->transaction->io_methods;
512
513         tdb_transaction_unlock(tdb, F_WRLCK);
514         SAFE_FREE(tdb->transaction->hash_heads);
515         SAFE_FREE(tdb->transaction);
516         
517         return ret;
518 }
519
520 /*
521   start a tdb transaction. No token is returned, as only a single
522   transaction is allowed to be pending per tdb_context
523 */
524 int tdb_transaction_start(struct tdb_context *tdb)
525 {
526         /* some sanity checks */
527         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
528                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
529                 tdb->ecode = TDB_ERR_EINVAL;
530                 return -1;
531         }
532
533         /* cope with nested tdb_transaction_start() calls */
534         if (tdb->transaction != NULL) {
535                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
536                         tdb->ecode = TDB_ERR_NESTING;
537                         return -1;
538                 }
539                 tdb_trace(tdb, "tdb_transaction_start");
540                 tdb->transaction->nesting++;
541                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
542                          tdb->transaction->nesting));
543                 return 0;
544         }
545
546         if (tdb_have_extra_locks(tdb)) {
547                 /* the caller must not have any locks when starting a
548                    transaction as otherwise we'll be screwed by lack
549                    of nested locks in posix */
550                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
551                 tdb->ecode = TDB_ERR_LOCK;
552                 return -1;
553         }
554
555         if (tdb->travlocks.next != NULL) {
556                 /* you cannot use transactions inside a traverse (although you can use
557                    traverse inside a transaction) as otherwise you can end up with
558                    deadlock */
559                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
560                 tdb->ecode = TDB_ERR_LOCK;
561                 return -1;
562         }
563
564         tdb->transaction = (struct tdb_transaction *)
565                 calloc(sizeof(struct tdb_transaction), 1);
566         if (tdb->transaction == NULL) {
567                 tdb->ecode = TDB_ERR_OOM;
568                 return -1;
569         }
570
571         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
572         tdb->transaction->block_size = tdb->page_size;
573
574         /* get the transaction write lock. This is a blocking lock. As
575            discussed with Volker, there are a number of ways we could
576            make this async, which we will probably do in the future */
577         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
578                 SAFE_FREE(tdb->transaction->blocks);
579                 SAFE_FREE(tdb->transaction);
580                 return -1;
581         }
582         
583         /* get a read lock from the freelist to the end of file. This
584            is upgraded to a write lock during the commit */
585         if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
586                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
587                 goto fail_allrecord_lock;
588         }
589
590         /* setup a copy of the hash table heads so the hash scan in
591            traverse can be fast */
592         tdb->transaction->hash_heads = (uint32_t *)
593                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
594         if (tdb->transaction->hash_heads == NULL) {
595                 tdb->ecode = TDB_ERR_OOM;
596                 goto fail;
597         }
598         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
599                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
600                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
601                 tdb->ecode = TDB_ERR_IO;
602                 goto fail;
603         }
604
605         /* make sure we know about any file expansions already done by
606            anyone else */
607         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
608         tdb->transaction->old_map_size = tdb->map_size;
609
610         /* finally hook the io methods, replacing them with
611            transaction specific methods */
612         tdb->transaction->io_methods = tdb->methods;
613         tdb->methods = &transaction_methods;
614
615         /* Trace at the end, so we get sequence number correct. */
616         tdb_trace(tdb, "tdb_transaction_start");
617         return 0;
618         
619 fail:
620         tdb_allrecord_unlock(tdb, F_RDLCK, false);
621 fail_allrecord_lock:
622         tdb_transaction_unlock(tdb, F_WRLCK);
623         SAFE_FREE(tdb->transaction->blocks);
624         SAFE_FREE(tdb->transaction->hash_heads);
625         SAFE_FREE(tdb->transaction);
626         return -1;
627 }
628
629
630 /*
631   cancel the current transaction
632 */
633 int tdb_transaction_cancel(struct tdb_context *tdb)
634 {
635         int ltype = F_RDLCK;
636         tdb_trace(tdb, "tdb_transaction_cancel");
637         if (tdb->transaction && tdb->transaction->prepared)
638                 ltype = F_WRLCK;
639         return _tdb_transaction_cancel(tdb, ltype);
640 }
641
642 /*
643   work out how much space the linearised recovery data will consume
644 */
645 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
646 {
647         tdb_len_t recovery_size = 0;
648         int i;
649
650         recovery_size = sizeof(uint32_t);
651         for (i=0;i<tdb->transaction->num_blocks;i++) {
652                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
653                         break;
654                 }
655                 if (tdb->transaction->blocks[i] == NULL) {
656                         continue;
657                 }
658                 recovery_size += 2*sizeof(tdb_off_t);
659                 if (i == tdb->transaction->num_blocks-1) {
660                         recovery_size += tdb->transaction->last_block_size;
661                 } else {
662                         recovery_size += tdb->transaction->block_size;
663                 }
664         }       
665
666         return recovery_size;
667 }
668
669 /*
670   allocate the recovery area, or use an existing recovery area if it is
671   large enough
672 */
673 static int tdb_recovery_allocate(struct tdb_context *tdb, 
674                                  tdb_len_t *recovery_size,
675                                  tdb_off_t *recovery_offset,
676                                  tdb_len_t *recovery_max_size)
677 {
678         struct tdb_record rec;
679         const struct tdb_methods *methods = tdb->transaction->io_methods;
680         tdb_off_t recovery_head;
681
682         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
683                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
684                 return -1;
685         }
686
687         rec.rec_len = 0;
688
689         if (recovery_head != 0 && 
690             methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
691                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
692                 return -1;
693         }
694
695         *recovery_size = tdb_recovery_size(tdb);
696
697         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
698                 /* it fits in the existing area */
699                 *recovery_max_size = rec.rec_len;
700                 *recovery_offset = recovery_head;
701                 return 0;
702         }
703
704         /* we need to free up the old recovery area, then allocate a
705            new one at the end of the file. Note that we cannot use
706            tdb_allocate() to allocate the new one as that might return
707            us an area that is being currently used (as of the start of
708            the transaction) */
709         if (recovery_head != 0) {
710                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
711                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
712                         return -1;
713                 }
714         }
715
716         /* the tdb_free() call might have increased the recovery size */
717         *recovery_size = tdb_recovery_size(tdb);
718
719         /* round up to a multiple of page size */
720         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
721         *recovery_offset = tdb->map_size;
722         recovery_head = *recovery_offset;
723
724         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
725                                      (tdb->map_size - tdb->transaction->old_map_size) +
726                                      sizeof(rec) + *recovery_max_size) == -1) {
727                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
728                 return -1;
729         }
730
731         /* remap the file (if using mmap) */
732         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
733
734         /* we have to reset the old map size so that we don't try to expand the file
735            again in the transaction commit, which would destroy the recovery area */
736         tdb->transaction->old_map_size = tdb->map_size;
737
738         /* write the recovery header offset and sync - we can sync without a race here
739            as the magic ptr in the recovery record has not been set */
740         CONVERT(recovery_head);
741         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
742                                &recovery_head, sizeof(tdb_off_t)) == -1) {
743                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
744                 return -1;
745         }
746         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
747                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
748                 return -1;
749         }
750
751         return 0;
752 }
753
754
755 /*
756   setup the recovery data that will be used on a crash during commit
757 */
758 static int transaction_setup_recovery(struct tdb_context *tdb, 
759                                       tdb_off_t *magic_offset)
760 {
761         tdb_len_t recovery_size;
762         unsigned char *data, *p;
763         const struct tdb_methods *methods = tdb->transaction->io_methods;
764         struct tdb_record *rec;
765         tdb_off_t recovery_offset, recovery_max_size;
766         tdb_off_t old_map_size = tdb->transaction->old_map_size;
767         uint32_t magic, tailer;
768         int i;
769
770         /*
771           check that the recovery area has enough space
772         */
773         if (tdb_recovery_allocate(tdb, &recovery_size, 
774                                   &recovery_offset, &recovery_max_size) == -1) {
775                 return -1;
776         }
777
778         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
779         if (data == NULL) {
780                 tdb->ecode = TDB_ERR_OOM;
781                 return -1;
782         }
783
784         rec = (struct tdb_record *)data;
785         memset(rec, 0, sizeof(*rec));
786
787         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
788         rec->data_len = recovery_size;
789         rec->rec_len  = recovery_max_size;
790         rec->key_len  = old_map_size;
791         CONVERT(rec);
792
793         /* build the recovery data into a single blob to allow us to do a single
794            large write, which should be more efficient */
795         p = data + sizeof(*rec);
796         for (i=0;i<tdb->transaction->num_blocks;i++) {
797                 tdb_off_t offset;
798                 tdb_len_t length;
799
800                 if (tdb->transaction->blocks[i] == NULL) {
801                         continue;
802                 }
803
804                 offset = i * tdb->transaction->block_size;
805                 length = tdb->transaction->block_size;
806                 if (i == tdb->transaction->num_blocks-1) {
807                         length = tdb->transaction->last_block_size;
808                 }
809                 
810                 if (offset >= old_map_size) {
811                         continue;
812                 }
813                 if (offset + length > tdb->transaction->old_map_size) {
814                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
815                         free(data);
816                         tdb->ecode = TDB_ERR_CORRUPT;
817                         return -1;
818                 }
819                 memcpy(p, &offset, 4);
820                 memcpy(p+4, &length, 4);
821                 if (DOCONV()) {
822                         tdb_convert(p, 8);
823                 }
824                 /* the recovery area contains the old data, not the
825                    new data, so we have to call the original tdb_read
826                    method to get it */
827                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
828                         free(data);
829                         tdb->ecode = TDB_ERR_IO;
830                         return -1;
831                 }
832                 p += 8 + length;
833         }
834
835         /* and the tailer */
836         tailer = sizeof(*rec) + recovery_max_size;
837         memcpy(p, &tailer, 4);
838         CONVERT(p);
839
840         /* write the recovery data to the recovery area */
841         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
842                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
843                 free(data);
844                 tdb->ecode = TDB_ERR_IO;
845                 return -1;
846         }
847         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
848                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
849                 free(data);
850                 tdb->ecode = TDB_ERR_IO;
851                 return -1;
852         }
853
854         /* as we don't have ordered writes, we have to sync the recovery
855            data before we update the magic to indicate that the recovery
856            data is present */
857         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
858                 free(data);
859                 return -1;
860         }
861
862         free(data);
863
864         magic = TDB_RECOVERY_MAGIC;
865         CONVERT(magic);
866
867         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
868
869         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
870                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
871                 tdb->ecode = TDB_ERR_IO;
872                 return -1;
873         }
874         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
875                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
876                 tdb->ecode = TDB_ERR_IO;
877                 return -1;
878         }
879
880         /* ensure the recovery magic marker is on disk */
881         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
882                 return -1;
883         }
884
885         return 0;
886 }
887
888 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
889 {       
890         const struct tdb_methods *methods;
891
892         if (tdb->transaction == NULL) {
893                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
894                 return -1;
895         }
896
897         if (tdb->transaction->prepared) {
898                 tdb->ecode = TDB_ERR_EINVAL;
899                 _tdb_transaction_cancel(tdb, F_WRLCK);
900                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
901                 return -1;
902         }
903
904         if (tdb->transaction->transaction_error) {
905                 tdb->ecode = TDB_ERR_IO;
906                 _tdb_transaction_cancel(tdb, F_RDLCK);
907                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
908                 return -1;
909         }
910
911
912         if (tdb->transaction->nesting != 0) {
913                 tdb->transaction->nesting--;
914                 return 0;
915         }               
916
917         /* check for a null transaction */
918         if (tdb->transaction->blocks == NULL) {
919                 return 0;
920         }
921
922         methods = tdb->transaction->io_methods;
923         
924         /* if there are any locks pending then the caller has not
925            nested their locks properly, so fail the transaction */
926         if (tdb_have_extra_locks(tdb)) {
927                 tdb->ecode = TDB_ERR_LOCK;
928                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
929                 _tdb_transaction_cancel(tdb, F_RDLCK);
930                 return -1;
931         }
932
933         /* upgrade the main transaction lock region to a write lock */
934         if (tdb_allrecord_upgrade(tdb) == -1) {
935                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
936                 _tdb_transaction_cancel(tdb, F_RDLCK);
937                 return -1;
938         }
939
940         /* get the open lock - this prevents new users attaching to the database
941            during the commit */
942         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
943                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
944                 _tdb_transaction_cancel(tdb, F_WRLCK);
945                 return -1;
946         }
947
948         if (!(tdb->flags & TDB_NOSYNC)) {
949                 /* write the recovery data to the end of the file */
950                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
951                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
952                         _tdb_transaction_cancel(tdb, F_WRLCK);
953                         return -1;
954                 }
955         }
956
957         tdb->transaction->prepared = true;
958
959         /* expand the file to the new size if needed */
960         if (tdb->map_size != tdb->transaction->old_map_size) {
961                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
962                                              tdb->map_size - 
963                                              tdb->transaction->old_map_size) == -1) {
964                         tdb->ecode = TDB_ERR_IO;
965                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
966                         _tdb_transaction_cancel(tdb, F_WRLCK);
967                         return -1;
968                 }
969                 tdb->map_size = tdb->transaction->old_map_size;
970                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
971         }
972
973         /* Keep the open lock until the actual commit */
974
975         return 0;
976 }
977
978 /*
979    prepare to commit the current transaction
980 */
981 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
982 {       
983         tdb_trace(tdb, "tdb_transaction_prepare_commit");
984         return _tdb_transaction_prepare_commit(tdb);
985 }
986
987 /*
988   commit the current transaction
989 */
990 int tdb_transaction_commit(struct tdb_context *tdb)
991 {       
992         const struct tdb_methods *methods;
993         int i;
994         bool need_repack;
995
996         if (tdb->transaction == NULL) {
997                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
998                 return -1;
999         }
1000
1001         tdb_trace(tdb, "tdb_transaction_commit");
1002
1003         if (tdb->transaction->transaction_error) {
1004                 tdb->ecode = TDB_ERR_IO;
1005                 tdb_transaction_cancel(tdb);
1006                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1007                 return -1;
1008         }
1009
1010
1011         if (tdb->transaction->nesting != 0) {
1012                 tdb->transaction->nesting--;
1013                 return 0;
1014         }
1015
1016         /* check for a null transaction */
1017         if (tdb->transaction->blocks == NULL) {
1018                 _tdb_transaction_cancel(tdb, F_RDLCK);
1019                 return 0;
1020         }
1021
1022         if (!tdb->transaction->prepared) {
1023                 int ret = _tdb_transaction_prepare_commit(tdb);
1024                 if (ret)
1025                         return ret;
1026         }
1027
1028         methods = tdb->transaction->io_methods;
1029
1030         /* perform all the writes */
1031         for (i=0;i<tdb->transaction->num_blocks;i++) {
1032                 tdb_off_t offset;
1033                 tdb_len_t length;
1034
1035                 if (tdb->transaction->blocks[i] == NULL) {
1036                         continue;
1037                 }
1038
1039                 offset = i * tdb->transaction->block_size;
1040                 length = tdb->transaction->block_size;
1041                 if (i == tdb->transaction->num_blocks-1) {
1042                         length = tdb->transaction->last_block_size;
1043                 }
1044
1045                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1046                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1047                         
1048                         /* we've overwritten part of the data and
1049                            possibly expanded the file, so we need to
1050                            run the crash recovery code */
1051                         tdb->methods = methods;
1052                         tdb_transaction_recover(tdb); 
1053
1054                         _tdb_transaction_cancel(tdb, F_WRLCK);
1055
1056                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1057                         return -1;
1058                 }
1059                 SAFE_FREE(tdb->transaction->blocks[i]);
1060         } 
1061
1062         SAFE_FREE(tdb->transaction->blocks);
1063         tdb->transaction->num_blocks = 0;
1064
1065         /* ensure the new data is on disk */
1066         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1067                 return -1;
1068         }
1069
1070         /*
1071           TODO: maybe write to some dummy hdr field, or write to magic
1072           offset without mmap, before the last sync, instead of the
1073           utime() call
1074         */
1075
1076         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1077            don't change the mtime of the file, this means the file may
1078            not be backed up (as tdb rounding to block sizes means that
1079            file size changes are quite rare too). The following forces
1080            mtime changes when a transaction completes */
1081 #if HAVE_UTIME
1082         utime(tdb->name, NULL);
1083 #endif
1084
1085         need_repack = tdb->transaction->need_repack;
1086
1087         /* use a transaction cancel to free memory and remove the
1088            transaction locks */
1089         _tdb_transaction_cancel(tdb, F_WRLCK);
1090
1091         if (need_repack) {
1092                 return tdb_repack(tdb);
1093         }
1094
1095         return 0;
1096 }
1097
1098
1099 /*
1100   recover from an aborted transaction. Must be called with exclusive
1101   database write access already established (including the open
1102   lock to prevent new processes attaching)
1103 */
1104 int tdb_transaction_recover(struct tdb_context *tdb)
1105 {
1106         tdb_off_t recovery_head, recovery_eof;
1107         unsigned char *data, *p;
1108         uint32_t zero = 0;
1109         struct tdb_record rec;
1110
1111         /* find the recovery area */
1112         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1113                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1114                 tdb->ecode = TDB_ERR_IO;
1115                 return -1;
1116         }
1117
1118         if (recovery_head == 0) {
1119                 /* we have never allocated a recovery record */
1120                 return 0;
1121         }
1122
1123         /* read the recovery record */
1124         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1125                                    sizeof(rec), DOCONV()) == -1) {
1126                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1127                 tdb->ecode = TDB_ERR_IO;
1128                 return -1;
1129         }
1130
1131         if (rec.magic != TDB_RECOVERY_MAGIC) {
1132                 /* there is no valid recovery data */
1133                 return 0;
1134         }
1135
1136         if (tdb->read_only) {
1137                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1138                 tdb->ecode = TDB_ERR_CORRUPT;
1139                 return -1;
1140         }
1141
1142         recovery_eof = rec.key_len;
1143
1144         data = (unsigned char *)malloc(rec.data_len);
1145         if (data == NULL) {
1146                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1147                 tdb->ecode = TDB_ERR_OOM;
1148                 return -1;
1149         }
1150
1151         /* read the full recovery data */
1152         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1153                                    rec.data_len, 0) == -1) {
1154                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1155                 tdb->ecode = TDB_ERR_IO;
1156                 return -1;
1157         }
1158
1159         /* recover the file data */
1160         p = data;
1161         while (p+8 < data + rec.data_len) {
1162                 uint32_t ofs, len;
1163                 if (DOCONV()) {
1164                         tdb_convert(p, 8);
1165                 }
1166                 memcpy(&ofs, p, 4);
1167                 memcpy(&len, p+4, 4);
1168
1169                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1170                         free(data);
1171                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1172                         tdb->ecode = TDB_ERR_IO;
1173                         return -1;
1174                 }
1175                 p += 8 + len;
1176         }
1177
1178         free(data);
1179
1180         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1181                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1182                 tdb->ecode = TDB_ERR_IO;
1183                 return -1;
1184         }
1185
1186         /* if the recovery area is after the recovered eof then remove it */
1187         if (recovery_eof <= recovery_head) {
1188                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1189                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1190                         tdb->ecode = TDB_ERR_IO;
1191                         return -1;                      
1192                 }
1193         }
1194
1195         /* remove the recovery magic */
1196         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1197                           &zero) == -1) {
1198                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1199                 tdb->ecode = TDB_ERR_IO;
1200                 return -1;                      
1201         }
1202         
1203         /* reduce the file size to the old size */
1204         tdb_munmap(tdb);
1205         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1206                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1207                 tdb->ecode = TDB_ERR_IO;
1208                 return -1;                      
1209         }
1210         tdb->map_size = recovery_eof;
1211         tdb_mmap(tdb);
1212
1213         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1214                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1215                 tdb->ecode = TDB_ERR_IO;
1216                 return -1;
1217         }
1218
1219         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1220                  recovery_eof));
1221
1222         /* all done */
1223         return 0;
1224 }