67104eff7fba47074123d6b4bed7d52821aa965b
[ira/wip.git] / lib / tdb / common / transaction.c
1  /* 
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              2005
7
8      ** NOTE! The following LGPL license applies to the tdb
9      ** library. This does NOT imply that all of Samba is released
10      ** under the LGPL
11    
12    This library is free software; you can redistribute it and/or
13    modify it under the terms of the GNU Lesser General Public
14    License as published by the Free Software Foundation; either
15    version 3 of the License, or (at your option) any later version.
16
17    This library is distributed in the hope that it will be useful,
18    but WITHOUT ANY WARRANTY; without even the implied warranty of
19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
20    Lesser General Public License for more details.
21
22    You should have received a copy of the GNU Lesser General Public
23    License along with this library; if not, see <http://www.gnu.org/licenses/>.
24 */
25
26 #include "tdb_private.h"
27
28 /*
29   transaction design:
30
31   - only allow a single transaction at a time per database. This makes
32     using the transaction API simpler, as otherwise the caller would
33     have to cope with temporary failures in transactions that conflict
34     with other current transactions
35
36   - keep the transaction recovery information in the same file as the
37     database, using a special 'transaction recovery' record pointed at
38     by the header. This removes the need for extra journal files as
39     used by some other databases
40
41   - dynamically allocated the transaction recover record, re-using it
42     for subsequent transactions. If a larger record is needed then
43     tdb_free() the old record to place it on the normal tdb freelist
44     before allocating the new record
45
46   - during transactions, keep a linked list of writes all that have
47     been performed by intercepting all tdb_write() calls. The hooked
48     transaction versions of tdb_read() and tdb_write() check this
49     linked list and try to use the elements of the list in preference
50     to the real database.
51
52   - don't allow any locks to be held when a transaction starts,
53     otherwise we can end up with deadlock (plus lack of lock nesting
54     in posix locks would mean the lock is lost)
55
56   - if the caller gains a lock during the transaction but doesn't
57     release it then fail the commit
58
59   - allow for nested calls to tdb_transaction_start(), re-using the
60     existing transaction record. If the inner transaction is cancelled
61     then a subsequent commit will fail
62  
63   - keep a mirrored copy of the tdb hash chain heads to allow for the
64     fast hash heads scan on traverse, updating the mirrored copy in
65     the transaction version of tdb_write
66
67   - allow callers to mix transaction and non-transaction use of tdb,
68     although once a transaction is started then an exclusive lock is
69     gained until the transaction is committed or cancelled
70
71   - the commit stategy involves first saving away all modified data
72     into a linearised buffer in the transaction recovery area, then
73     marking the transaction recovery area with a magic value to
74     indicate a valid recovery record. In total 4 fsync/msync calls are
75     needed per commit to prevent race conditions. It might be possible
76     to reduce this to 3 or even 2 with some more work.
77
78   - check for a valid recovery record on open of the tdb, while the
79     global lock is held. Automatically recover from the transaction
80     recovery area if needed, then continue with the open as
81     usual. This allows for smooth crash recovery with no administrator
82     intervention.
83
84   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85     still available, but no transaction recovery area is used and no
86     fsync/msync calls are made.
87
88   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89     tdb_add_flags() transaction nesting is enabled.
90     It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
91     The default is that transaction nesting is allowed.
92     Note: this default may change in future versions of tdb.
93
94     Beware. when transactions are nested a transaction successfully
95     completed with tdb_transaction_commit() can be silently unrolled later.
96
97   - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
98     tdb_add_flags() transaction nesting is disabled.
99     It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
100     An attempt create a nested transaction will fail with TDB_ERR_NESTING.
101     The default is that transaction nesting is allowed.
102     Note: this default may change in future versions of tdb.
103 */
104
105
106 /*
107   hold the context of any current transaction
108 */
109 struct tdb_transaction {
110         /* we keep a mirrored copy of the tdb hash heads here so
111            tdb_next_hash_chain() can operate efficiently */
112         uint32_t *hash_heads;
113
114         /* the original io methods - used to do IOs to the real db */
115         const struct tdb_methods *io_methods;
116
117         /* the list of transaction blocks. When a block is first
118            written to, it gets created in this list */
119         uint8_t **blocks;
120         uint32_t num_blocks;
121         uint32_t block_size;      /* bytes in each block */
122         uint32_t last_block_size; /* number of valid bytes in the last block */
123
124         /* non-zero when an internal transaction error has
125            occurred. All write operations will then fail until the
126            transaction is ended */
127         int transaction_error;
128
129         /* when inside a transaction we need to keep track of any
130            nested tdb_transaction_start() calls, as these are allowed,
131            but don't create a new transaction */
132         int nesting;
133
134         /* set when a prepare has already occurred */
135         bool prepared;
136         tdb_off_t magic_offset;
137
138         /* set when the GLOBAL_LOCK has been taken */
139         bool global_lock_taken;
140
141         /* old file size before transaction */
142         tdb_len_t old_map_size;
143
144         /* we should re-pack on commit */
145         bool need_repack;
146 };
147
148
149 /*
150   read while in a transaction. We need to check first if the data is in our list
151   of transaction elements, then if not do a real read
152 */
153 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
154                             tdb_len_t len, int cv)
155 {
156         uint32_t blk;
157
158         /* break it down into block sized ops */
159         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
160                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
161                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
162                         return -1;
163                 }
164                 len -= len2;
165                 off += len2;
166                 buf = (void *)(len2 + (char *)buf);
167         }
168
169         if (len == 0) {
170                 return 0;
171         }
172
173         blk = off / tdb->transaction->block_size;
174
175         /* see if we have it in the block list */
176         if (tdb->transaction->num_blocks <= blk ||
177             tdb->transaction->blocks[blk] == NULL) {
178                 /* nope, do a real read */
179                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
180                         goto fail;
181                 }
182                 return 0;
183         }
184
185         /* it is in the block list. Now check for the last block */
186         if (blk == tdb->transaction->num_blocks-1) {
187                 if (len > tdb->transaction->last_block_size) {
188                         goto fail;
189                 }
190         }
191         
192         /* now copy it out of this block */
193         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
194         if (cv) {
195                 tdb_convert(buf, len);
196         }
197         return 0;
198
199 fail:
200         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
201         tdb->ecode = TDB_ERR_IO;
202         tdb->transaction->transaction_error = 1;
203         return -1;
204 }
205
206
207 /*
208   write while in a transaction
209 */
210 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
211                              const void *buf, tdb_len_t len)
212 {
213         uint32_t blk;
214
215         /* Only a commit is allowed on a prepared transaction */
216         if (tdb->transaction->prepared) {
217                 tdb->ecode = TDB_ERR_EINVAL;
218                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
219                 tdb->transaction->transaction_error = 1;
220                 return -1;
221         }
222
223         /* if the write is to a hash head, then update the transaction
224            hash heads */
225         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
226             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
227                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
228                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
229         }
230
231         /* break it up into block sized chunks */
232         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
233                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
234                 if (transaction_write(tdb, off, buf, len2) != 0) {
235                         return -1;
236                 }
237                 len -= len2;
238                 off += len2;
239                 if (buf != NULL) {
240                         buf = (const void *)(len2 + (const char *)buf);
241                 }
242         }
243
244         if (len == 0) {
245                 return 0;
246         }
247
248         blk = off / tdb->transaction->block_size;
249         off = off % tdb->transaction->block_size;
250
251         if (tdb->transaction->num_blocks <= blk) {
252                 uint8_t **new_blocks;
253                 /* expand the blocks array */
254                 if (tdb->transaction->blocks == NULL) {
255                         new_blocks = (uint8_t **)malloc(
256                                 (blk+1)*sizeof(uint8_t *));
257                 } else {
258                         new_blocks = (uint8_t **)realloc(
259                                 tdb->transaction->blocks,
260                                 (blk+1)*sizeof(uint8_t *));
261                 }
262                 if (new_blocks == NULL) {
263                         tdb->ecode = TDB_ERR_OOM;
264                         goto fail;
265                 }
266                 memset(&new_blocks[tdb->transaction->num_blocks], 0, 
267                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
268                 tdb->transaction->blocks = new_blocks;
269                 tdb->transaction->num_blocks = blk+1;
270                 tdb->transaction->last_block_size = 0;
271         }
272
273         /* allocate and fill a block? */
274         if (tdb->transaction->blocks[blk] == NULL) {
275                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
276                 if (tdb->transaction->blocks[blk] == NULL) {
277                         tdb->ecode = TDB_ERR_OOM;
278                         tdb->transaction->transaction_error = 1;
279                         return -1;                      
280                 }
281                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
282                         tdb_len_t len2 = tdb->transaction->block_size;
283                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
284                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
285                         }
286                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
287                                                                    tdb->transaction->blocks[blk], 
288                                                                    len2, 0) != 0) {
289                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
290                                 tdb->ecode = TDB_ERR_IO;
291                                 goto fail;
292                         }
293                         if (blk == tdb->transaction->num_blocks-1) {
294                                 tdb->transaction->last_block_size = len2;
295                         }                       
296                 }
297         }
298         
299         /* overwrite part of an existing block */
300         if (buf == NULL) {
301                 memset(tdb->transaction->blocks[blk] + off, 0, len);
302         } else {
303                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
304         }
305         if (blk == tdb->transaction->num_blocks-1) {
306                 if (len + off > tdb->transaction->last_block_size) {
307                         tdb->transaction->last_block_size = len + off;
308                 }
309         }
310
311         return 0;
312
313 fail:
314         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", 
315                  (blk*tdb->transaction->block_size) + off, len));
316         tdb->transaction->transaction_error = 1;
317         return -1;
318 }
319
320
321 /*
322   write while in a transaction - this varient never expands the transaction blocks, it only
323   updates existing blocks. This means it cannot change the recovery size
324 */
325 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
326                                       const void *buf, tdb_len_t len)
327 {
328         uint32_t blk;
329
330         /* break it up into block sized chunks */
331         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
332                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
333                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
334                         return -1;
335                 }
336                 len -= len2;
337                 off += len2;
338                 if (buf != NULL) {
339                         buf = (const void *)(len2 + (const char *)buf);
340                 }
341         }
342
343         if (len == 0) {
344                 return 0;
345         }
346
347         blk = off / tdb->transaction->block_size;
348         off = off % tdb->transaction->block_size;
349
350         if (tdb->transaction->num_blocks <= blk ||
351             tdb->transaction->blocks[blk] == NULL) {
352                 return 0;
353         }
354
355         if (blk == tdb->transaction->num_blocks-1 &&
356             off + len > tdb->transaction->last_block_size) {
357                 if (off >= tdb->transaction->last_block_size) {
358                         return 0;
359                 }
360                 len = tdb->transaction->last_block_size - off;
361         }
362
363         /* overwrite part of an existing block */
364         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
365
366         return 0;
367 }
368
369
370 /*
371   accelerated hash chain head search, using the cached hash heads
372 */
373 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
374 {
375         uint32_t h = *chain;
376         for (;h < tdb->header.hash_size;h++) {
377                 /* the +1 takes account of the freelist */
378                 if (0 != tdb->transaction->hash_heads[h+1]) {
379                         break;
380                 }
381         }
382         (*chain) = h;
383 }
384
385 /*
386   out of bounds check during a transaction
387 */
388 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
389 {
390         if (len <= tdb->map_size) {
391                 return 0;
392         }
393         tdb->ecode = TDB_ERR_IO;
394         return -1;
395 }
396
397 /*
398   transaction version of tdb_expand().
399 */
400 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
401                                    tdb_off_t addition)
402 {
403         /* add a write to the transaction elements, so subsequent
404            reads see the zero data */
405         if (transaction_write(tdb, size, NULL, addition) != 0) {
406                 return -1;
407         }
408
409         tdb->transaction->need_repack = true;
410
411         return 0;
412 }
413
414 /*
415   brlock during a transaction - ignore them
416 */
417 static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset, 
418                               int rw_type, int lck_type, int probe, size_t len)
419 {
420         return 0;
421 }
422
423 static const struct tdb_methods transaction_methods = {
424         transaction_read,
425         transaction_write,
426         transaction_next_hash_chain,
427         transaction_oob,
428         transaction_expand_file,
429         transaction_brlock
430 };
431
432
433 /*
434   start a tdb transaction. No token is returned, as only a single
435   transaction is allowed to be pending per tdb_context
436 */
437 int tdb_transaction_start(struct tdb_context *tdb)
438 {
439         /* some sanity checks */
440         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
441                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
442                 tdb->ecode = TDB_ERR_EINVAL;
443                 return -1;
444         }
445
446         /* cope with nested tdb_transaction_start() calls */
447         if (tdb->transaction != NULL) {
448                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
449                         tdb->ecode = TDB_ERR_NESTING;
450                         return -1;
451                 }
452                 tdb->transaction->nesting++;
453                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
454                          tdb->transaction->nesting));
455                 return 0;
456         }
457
458         if (tdb->num_locks != 0 || tdb->global_lock.count) {
459                 /* the caller must not have any locks when starting a
460                    transaction as otherwise we'll be screwed by lack
461                    of nested locks in posix */
462                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
463                 tdb->ecode = TDB_ERR_LOCK;
464                 return -1;
465         }
466
467         if (tdb->travlocks.next != NULL) {
468                 /* you cannot use transactions inside a traverse (although you can use
469                    traverse inside a transaction) as otherwise you can end up with
470                    deadlock */
471                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
472                 tdb->ecode = TDB_ERR_LOCK;
473                 return -1;
474         }
475
476         tdb->transaction = (struct tdb_transaction *)
477                 calloc(sizeof(struct tdb_transaction), 1);
478         if (tdb->transaction == NULL) {
479                 tdb->ecode = TDB_ERR_OOM;
480                 return -1;
481         }
482
483         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
484         tdb->transaction->block_size = tdb->page_size;
485
486         /* get the transaction write lock. This is a blocking lock. As
487            discussed with Volker, there are a number of ways we could
488            make this async, which we will probably do in the future */
489         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
490                 SAFE_FREE(tdb->transaction->blocks);
491                 SAFE_FREE(tdb->transaction);
492                 return -1;
493         }
494         
495         /* get a read lock from the freelist to the end of file. This
496            is upgraded to a write lock during the commit */
497         if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
498                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
499                 tdb->ecode = TDB_ERR_LOCK;
500                 goto fail;
501         }
502
503         /* setup a copy of the hash table heads so the hash scan in
504            traverse can be fast */
505         tdb->transaction->hash_heads = (uint32_t *)
506                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
507         if (tdb->transaction->hash_heads == NULL) {
508                 tdb->ecode = TDB_ERR_OOM;
509                 goto fail;
510         }
511         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
512                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
513                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
514                 tdb->ecode = TDB_ERR_IO;
515                 goto fail;
516         }
517
518         /* make sure we know about any file expansions already done by
519            anyone else */
520         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
521         tdb->transaction->old_map_size = tdb->map_size;
522
523         /* finally hook the io methods, replacing them with
524            transaction specific methods */
525         tdb->transaction->io_methods = tdb->methods;
526         tdb->methods = &transaction_methods;
527
528         /* Trace at the end, so we get sequence number correct. */
529         tdb_trace(tdb, "tdb_transaction_start");
530         return 0;
531         
532 fail:
533         tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
534         tdb_transaction_unlock(tdb);
535         SAFE_FREE(tdb->transaction->blocks);
536         SAFE_FREE(tdb->transaction->hash_heads);
537         SAFE_FREE(tdb->transaction);
538         return -1;
539 }
540
541
542 /*
543   sync to disk
544 */
545 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
546 {       
547         if (tdb->flags & TDB_NOSYNC) {
548                 return 0;
549         }
550
551         if (fsync(tdb->fd) != 0) {
552                 tdb->ecode = TDB_ERR_IO;
553                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
554                 return -1;
555         }
556 #ifdef HAVE_MMAP
557         if (tdb->map_ptr) {
558                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
559                 if (msync(moffset + (char *)tdb->map_ptr, 
560                           length + (offset - moffset), MS_SYNC) != 0) {
561                         tdb->ecode = TDB_ERR_IO;
562                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
563                                  strerror(errno)));
564                         return -1;
565                 }
566         }
567 #endif
568         return 0;
569 }
570
571
572 int _tdb_transaction_cancel(struct tdb_context *tdb)
573 {       
574         int i, ret = 0;
575
576         if (tdb->transaction == NULL) {
577                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
578                 return -1;
579         }
580
581         if (tdb->transaction->nesting != 0) {
582                 tdb->transaction->transaction_error = 1;
583                 tdb->transaction->nesting--;
584                 return 0;
585         }               
586
587         tdb->map_size = tdb->transaction->old_map_size;
588
589         /* free all the transaction blocks */
590         for (i=0;i<tdb->transaction->num_blocks;i++) {
591                 if (tdb->transaction->blocks[i] != NULL) {
592                         free(tdb->transaction->blocks[i]);
593                 }
594         }
595         SAFE_FREE(tdb->transaction->blocks);
596
597         if (tdb->transaction->magic_offset) {
598                 const struct tdb_methods *methods = tdb->transaction->io_methods;
599                 uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
600
601                 /* remove the recovery marker */
602                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
603                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
604                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
605                         ret = -1;
606                 }
607         }
608
609         if (tdb->transaction->global_lock_taken) {
610                 tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
611                 tdb->transaction->global_lock_taken = false;
612         }
613
614         /* remove any global lock created during the transaction */
615         if (tdb->global_lock.count != 0) {
616                 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
617                 tdb->global_lock.count = 0;
618         }
619
620         /* remove any locks created during the transaction */
621         if (tdb->num_locks != 0) {
622                 for (i=0;i<tdb->num_lockrecs;i++) {
623                         tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
624                                    F_UNLCK,F_SETLKW, 0, 1);
625                 }
626                 tdb->num_locks = 0;
627                 tdb->num_lockrecs = 0;
628                 SAFE_FREE(tdb->lockrecs);
629         }
630
631         /* restore the normal io methods */
632         tdb->methods = tdb->transaction->io_methods;
633
634         tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
635         tdb_transaction_unlock(tdb);
636         SAFE_FREE(tdb->transaction->hash_heads);
637         SAFE_FREE(tdb->transaction);
638         
639         return ret;
640 }
641
642 /*
643   cancel the current transaction
644 */
645 int tdb_transaction_cancel(struct tdb_context *tdb)
646 {
647         tdb_trace(tdb, "tdb_transaction_cancel");
648         return _tdb_transaction_cancel(tdb);
649 }
650
651 /*
652   work out how much space the linearised recovery data will consume
653 */
654 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
655 {
656         tdb_len_t recovery_size = 0;
657         int i;
658
659         recovery_size = sizeof(uint32_t);
660         for (i=0;i<tdb->transaction->num_blocks;i++) {
661                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
662                         break;
663                 }
664                 if (tdb->transaction->blocks[i] == NULL) {
665                         continue;
666                 }
667                 recovery_size += 2*sizeof(tdb_off_t);
668                 if (i == tdb->transaction->num_blocks-1) {
669                         recovery_size += tdb->transaction->last_block_size;
670                 } else {
671                         recovery_size += tdb->transaction->block_size;
672                 }
673         }       
674
675         return recovery_size;
676 }
677
678 /*
679   allocate the recovery area, or use an existing recovery area if it is
680   large enough
681 */
682 static int tdb_recovery_allocate(struct tdb_context *tdb, 
683                                  tdb_len_t *recovery_size,
684                                  tdb_off_t *recovery_offset,
685                                  tdb_len_t *recovery_max_size)
686 {
687         struct tdb_record rec;
688         const struct tdb_methods *methods = tdb->transaction->io_methods;
689         tdb_off_t recovery_head;
690
691         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
692                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
693                 return -1;
694         }
695
696         rec.rec_len = 0;
697
698         if (recovery_head != 0) {
699                 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
700                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
701                         return -1;
702                 }
703                 /* ignore invalid recovery regions: can happen in crash */
704                 if (rec.magic != TDB_RECOVERY_MAGIC &&
705                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
706                         recovery_head = 0;
707                 }
708         }
709
710         *recovery_size = tdb_recovery_size(tdb);
711
712         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
713                 /* it fits in the existing area */
714                 *recovery_max_size = rec.rec_len;
715                 *recovery_offset = recovery_head;
716                 return 0;
717         }
718
719         /* we need to free up the old recovery area, then allocate a
720            new one at the end of the file. Note that we cannot use
721            tdb_allocate() to allocate the new one as that might return
722            us an area that is being currently used (as of the start of
723            the transaction) */
724         if (recovery_head != 0) {
725                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
726                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
727                         return -1;
728                 }
729         }
730
731         /* the tdb_free() call might have increased the recovery size */
732         *recovery_size = tdb_recovery_size(tdb);
733
734         /* round up to a multiple of page size */
735         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
736         *recovery_offset = tdb->map_size;
737         recovery_head = *recovery_offset;
738
739         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
740                                      (tdb->map_size - tdb->transaction->old_map_size) +
741                                      sizeof(rec) + *recovery_max_size) == -1) {
742                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
743                 return -1;
744         }
745
746         /* remap the file (if using mmap) */
747         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
748
749         /* we have to reset the old map size so that we don't try to expand the file
750            again in the transaction commit, which would destroy the recovery area */
751         tdb->transaction->old_map_size = tdb->map_size;
752
753         /* write the recovery header offset and sync - we can sync without a race here
754            as the magic ptr in the recovery record has not been set */
755         CONVERT(recovery_head);
756         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
757                                &recovery_head, sizeof(tdb_off_t)) == -1) {
758                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
759                 return -1;
760         }
761         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
762                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
763                 return -1;
764         }
765
766         return 0;
767 }
768
769
770 /*
771   setup the recovery data that will be used on a crash during commit
772 */
773 static int transaction_setup_recovery(struct tdb_context *tdb, 
774                                       tdb_off_t *magic_offset)
775 {
776         tdb_len_t recovery_size;
777         unsigned char *data, *p;
778         const struct tdb_methods *methods = tdb->transaction->io_methods;
779         struct tdb_record *rec;
780         tdb_off_t recovery_offset, recovery_max_size;
781         tdb_off_t old_map_size = tdb->transaction->old_map_size;
782         uint32_t magic, tailer;
783         int i;
784
785         /*
786           check that the recovery area has enough space
787         */
788         if (tdb_recovery_allocate(tdb, &recovery_size, 
789                                   &recovery_offset, &recovery_max_size) == -1) {
790                 return -1;
791         }
792
793         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
794         if (data == NULL) {
795                 tdb->ecode = TDB_ERR_OOM;
796                 return -1;
797         }
798
799         rec = (struct tdb_record *)data;
800         memset(rec, 0, sizeof(*rec));
801
802         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
803         rec->data_len = recovery_size;
804         rec->rec_len  = recovery_max_size;
805         rec->key_len  = old_map_size;
806         CONVERT(rec);
807
808         /* build the recovery data into a single blob to allow us to do a single
809            large write, which should be more efficient */
810         p = data + sizeof(*rec);
811         for (i=0;i<tdb->transaction->num_blocks;i++) {
812                 tdb_off_t offset;
813                 tdb_len_t length;
814
815                 if (tdb->transaction->blocks[i] == NULL) {
816                         continue;
817                 }
818
819                 offset = i * tdb->transaction->block_size;
820                 length = tdb->transaction->block_size;
821                 if (i == tdb->transaction->num_blocks-1) {
822                         length = tdb->transaction->last_block_size;
823                 }
824                 
825                 if (offset >= old_map_size) {
826                         continue;
827                 }
828                 if (offset + length > tdb->transaction->old_map_size) {
829                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
830                         free(data);
831                         tdb->ecode = TDB_ERR_CORRUPT;
832                         return -1;
833                 }
834                 memcpy(p, &offset, 4);
835                 memcpy(p+4, &length, 4);
836                 if (DOCONV()) {
837                         tdb_convert(p, 8);
838                 }
839                 /* the recovery area contains the old data, not the
840                    new data, so we have to call the original tdb_read
841                    method to get it */
842                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
843                         free(data);
844                         tdb->ecode = TDB_ERR_IO;
845                         return -1;
846                 }
847                 p += 8 + length;
848         }
849
850         /* and the tailer */
851         tailer = sizeof(*rec) + recovery_max_size;
852         memcpy(p, &tailer, 4);
853         CONVERT(p);
854
855         /* write the recovery data to the recovery area */
856         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
857                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
858                 free(data);
859                 tdb->ecode = TDB_ERR_IO;
860                 return -1;
861         }
862         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
863                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
864                 free(data);
865                 tdb->ecode = TDB_ERR_IO;
866                 return -1;
867         }
868
869         /* as we don't have ordered writes, we have to sync the recovery
870            data before we update the magic to indicate that the recovery
871            data is present */
872         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
873                 free(data);
874                 return -1;
875         }
876
877         free(data);
878
879         magic = TDB_RECOVERY_MAGIC;
880         CONVERT(magic);
881
882         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
883
884         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
885                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
886                 tdb->ecode = TDB_ERR_IO;
887                 return -1;
888         }
889         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
890                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
891                 tdb->ecode = TDB_ERR_IO;
892                 return -1;
893         }
894
895         /* ensure the recovery magic marker is on disk */
896         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
897                 return -1;
898         }
899
900         return 0;
901 }
902
903 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
904 {       
905         const struct tdb_methods *methods;
906
907         if (tdb->transaction == NULL) {
908                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
909                 return -1;
910         }
911
912         if (tdb->transaction->prepared) {
913                 tdb->ecode = TDB_ERR_EINVAL;
914                 _tdb_transaction_cancel(tdb);
915                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
916                 return -1;
917         }
918
919         if (tdb->transaction->transaction_error) {
920                 tdb->ecode = TDB_ERR_IO;
921                 _tdb_transaction_cancel(tdb);
922                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
923                 return -1;
924         }
925
926
927         if (tdb->transaction->nesting != 0) {
928                 return 0;
929         }               
930
931         /* check for a null transaction */
932         if (tdb->transaction->blocks == NULL) {
933                 return 0;
934         }
935
936         methods = tdb->transaction->io_methods;
937         
938         /* if there are any locks pending then the caller has not
939            nested their locks properly, so fail the transaction */
940         if (tdb->num_locks || tdb->global_lock.count) {
941                 tdb->ecode = TDB_ERR_LOCK;
942                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
943                 _tdb_transaction_cancel(tdb);
944                 return -1;
945         }
946
947         /* upgrade the main transaction lock region to a write lock */
948         if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
949                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
950                 tdb->ecode = TDB_ERR_LOCK;
951                 _tdb_transaction_cancel(tdb);
952                 return -1;
953         }
954
955         /* get the global lock - this prevents new users attaching to the database
956            during the commit */
957         if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
958                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
959                 tdb->ecode = TDB_ERR_LOCK;
960                 _tdb_transaction_cancel(tdb);
961                 return -1;
962         }
963
964         tdb->transaction->global_lock_taken = true;
965
966         if (!(tdb->flags & TDB_NOSYNC)) {
967                 /* write the recovery data to the end of the file */
968                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
969                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
970                         _tdb_transaction_cancel(tdb);
971                         return -1;
972                 }
973         }
974
975         tdb->transaction->prepared = true;
976
977         /* expand the file to the new size if needed */
978         if (tdb->map_size != tdb->transaction->old_map_size) {
979                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
980                                              tdb->map_size - 
981                                              tdb->transaction->old_map_size) == -1) {
982                         tdb->ecode = TDB_ERR_IO;
983                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
984                         _tdb_transaction_cancel(tdb);
985                         return -1;
986                 }
987                 tdb->map_size = tdb->transaction->old_map_size;
988                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
989         }
990
991         /* Keep the global lock until the actual commit */
992
993         return 0;
994 }
995
996 /*
997    prepare to commit the current transaction
998 */
999 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1000 {       
1001         tdb_trace(tdb, "tdb_transaction_prepare_commit");
1002         return _tdb_transaction_prepare_commit(tdb);
1003 }
1004
1005 /*
1006   commit the current transaction
1007 */
1008 int tdb_transaction_commit(struct tdb_context *tdb)
1009 {       
1010         const struct tdb_methods *methods;
1011         int i;
1012         bool need_repack;
1013
1014         if (tdb->transaction == NULL) {
1015                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1016                 return -1;
1017         }
1018
1019         tdb_trace(tdb, "tdb_transaction_commit");
1020
1021         if (tdb->transaction->transaction_error) {
1022                 tdb->ecode = TDB_ERR_IO;
1023                 _tdb_transaction_cancel(tdb);
1024                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1025                 return -1;
1026         }
1027
1028
1029         if (tdb->transaction->nesting != 0) {
1030                 tdb->transaction->nesting--;
1031                 return 0;
1032         }
1033
1034         /* check for a null transaction */
1035         if (tdb->transaction->blocks == NULL) {
1036                 _tdb_transaction_cancel(tdb);
1037                 return 0;
1038         }
1039
1040         if (!tdb->transaction->prepared) {
1041                 int ret = _tdb_transaction_prepare_commit(tdb);
1042                 if (ret)
1043                         return ret;
1044         }
1045
1046         methods = tdb->transaction->io_methods;
1047
1048         /* perform all the writes */
1049         for (i=0;i<tdb->transaction->num_blocks;i++) {
1050                 tdb_off_t offset;
1051                 tdb_len_t length;
1052
1053                 if (tdb->transaction->blocks[i] == NULL) {
1054                         continue;
1055                 }
1056
1057                 offset = i * tdb->transaction->block_size;
1058                 length = tdb->transaction->block_size;
1059                 if (i == tdb->transaction->num_blocks-1) {
1060                         length = tdb->transaction->last_block_size;
1061                 }
1062
1063                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1064                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1065                         
1066                         /* we've overwritten part of the data and
1067                            possibly expanded the file, so we need to
1068                            run the crash recovery code */
1069                         tdb->methods = methods;
1070                         tdb_transaction_recover(tdb); 
1071
1072                         _tdb_transaction_cancel(tdb);
1073
1074                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1075                         return -1;
1076                 }
1077                 SAFE_FREE(tdb->transaction->blocks[i]);
1078         } 
1079
1080         SAFE_FREE(tdb->transaction->blocks);
1081         tdb->transaction->num_blocks = 0;
1082
1083         /* ensure the new data is on disk */
1084         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1085                 return -1;
1086         }
1087
1088         /*
1089           TODO: maybe write to some dummy hdr field, or write to magic
1090           offset without mmap, before the last sync, instead of the
1091           utime() call
1092         */
1093
1094         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1095            don't change the mtime of the file, this means the file may
1096            not be backed up (as tdb rounding to block sizes means that
1097            file size changes are quite rare too). The following forces
1098            mtime changes when a transaction completes */
1099 #ifdef HAVE_UTIME
1100         utime(tdb->name, NULL);
1101 #endif
1102
1103         need_repack = tdb->transaction->need_repack;
1104
1105         /* use a transaction cancel to free memory and remove the
1106            transaction locks */
1107         _tdb_transaction_cancel(tdb);
1108
1109         if (need_repack) {
1110                 return tdb_repack(tdb);
1111         }
1112
1113         return 0;
1114 }
1115
1116
1117 /*
1118   recover from an aborted transaction. Must be called with exclusive
1119   database write access already established (including the global
1120   lock to prevent new processes attaching)
1121 */
1122 int tdb_transaction_recover(struct tdb_context *tdb)
1123 {
1124         tdb_off_t recovery_head, recovery_eof;
1125         unsigned char *data, *p;
1126         uint32_t zero = 0;
1127         struct tdb_record rec;
1128
1129         /* find the recovery area */
1130         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1131                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1132                 tdb->ecode = TDB_ERR_IO;
1133                 return -1;
1134         }
1135
1136         if (recovery_head == 0) {
1137                 /* we have never allocated a recovery record */
1138                 return 0;
1139         }
1140
1141         /* read the recovery record */
1142         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1143                                    sizeof(rec), DOCONV()) == -1) {
1144                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1145                 tdb->ecode = TDB_ERR_IO;
1146                 return -1;
1147         }
1148
1149         if (rec.magic != TDB_RECOVERY_MAGIC) {
1150                 /* there is no valid recovery data */
1151                 return 0;
1152         }
1153
1154         if (tdb->read_only) {
1155                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1156                 tdb->ecode = TDB_ERR_CORRUPT;
1157                 return -1;
1158         }
1159
1160         recovery_eof = rec.key_len;
1161
1162         data = (unsigned char *)malloc(rec.data_len);
1163         if (data == NULL) {
1164                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1165                 tdb->ecode = TDB_ERR_OOM;
1166                 return -1;
1167         }
1168
1169         /* read the full recovery data */
1170         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1171                                    rec.data_len, 0) == -1) {
1172                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1173                 tdb->ecode = TDB_ERR_IO;
1174                 return -1;
1175         }
1176
1177         /* recover the file data */
1178         p = data;
1179         while (p+8 < data + rec.data_len) {
1180                 uint32_t ofs, len;
1181                 if (DOCONV()) {
1182                         tdb_convert(p, 8);
1183                 }
1184                 memcpy(&ofs, p, 4);
1185                 memcpy(&len, p+4, 4);
1186
1187                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1188                         free(data);
1189                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1190                         tdb->ecode = TDB_ERR_IO;
1191                         return -1;
1192                 }
1193                 p += 8 + len;
1194         }
1195
1196         free(data);
1197
1198         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1199                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1200                 tdb->ecode = TDB_ERR_IO;
1201                 return -1;
1202         }
1203
1204         /* if the recovery area is after the recovered eof then remove it */
1205         if (recovery_eof <= recovery_head) {
1206                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1207                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1208                         tdb->ecode = TDB_ERR_IO;
1209                         return -1;                      
1210                 }
1211         }
1212
1213         /* remove the recovery magic */
1214         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1215                           &zero) == -1) {
1216                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1217                 tdb->ecode = TDB_ERR_IO;
1218                 return -1;                      
1219         }
1220         
1221         /* reduce the file size to the old size */
1222         tdb_munmap(tdb);
1223         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1224                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1225                 tdb->ecode = TDB_ERR_IO;
1226                 return -1;                      
1227         }
1228         tdb->map_size = recovery_eof;
1229         tdb_mmap(tdb);
1230
1231         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1232                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1233                 tdb->ecode = TDB_ERR_IO;
1234                 return -1;
1235         }
1236
1237         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1238                  recovery_eof));
1239
1240         /* all done */
1241         return 0;
1242 }