tdb: locking cleanup
[ccan] / ccan / tdb / transaction.c
1  /* 
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              2005
7
8      ** NOTE! The following LGPL license applies to the tdb
9      ** library. This does NOT imply that all of Samba is released
10      ** under the LGPL
11    
12    This library is free software; you can redistribute it and/or
13    modify it under the terms of the GNU Lesser General Public
14    License as published by the Free Software Foundation; either
15    version 3 of the License, or (at your option) any later version.
16
17    This library is distributed in the hope that it will be useful,
18    but WITHOUT ANY WARRANTY; without even the implied warranty of
19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
20    Lesser General Public License for more details.
21
22    You should have received a copy of the GNU Lesser General Public
23    License along with this library; if not, see <http://www.gnu.org/licenses/>.
24 */
25
26 #include "tdb_private.h"
27
28 /*
29   transaction design:
30
31   - only allow a single transaction at a time per database. This makes
32     using the transaction API simpler, as otherwise the caller would
33     have to cope with temporary failures in transactions that conflict
34     with other current transactions
35
36   - keep the transaction recovery information in the same file as the
37     database, using a special 'transaction recovery' record pointed at
38     by the header. This removes the need for extra journal files as
39     used by some other databases
40
41   - dynamically allocated the transaction recover record, re-using it
42     for subsequent transactions. If a larger record is needed then
43     tdb_free() the old record to place it on the normal tdb freelist
44     before allocating the new record
45
46   - during transactions, keep a linked list of writes all that have
47     been performed by intercepting all tdb_write() calls. The hooked
48     transaction versions of tdb_read() and tdb_write() check this
49     linked list and try to use the elements of the list in preference
50     to the real database.
51
52   - don't allow any locks to be held when a transaction starts,
53     otherwise we can end up with deadlock (plus lack of lock nesting
54     in posix locks would mean the lock is lost)
55
56   - if the caller gains a lock during the transaction but doesn't
57     release it then fail the commit
58
59   - allow for nested calls to tdb_transaction_start(), re-using the
60     existing transaction record. If the inner transaction is cancelled
61     then a subsequent commit will fail
62  
63   - keep a mirrored copy of the tdb hash chain heads to allow for the
64     fast hash heads scan on traverse, updating the mirrored copy in
65     the transaction version of tdb_write
66
67   - allow callers to mix transaction and non-transaction use of tdb,
68     although once a transaction is started then an exclusive lock is
69     gained until the transaction is committed or cancelled
70
71   - the commit stategy involves first saving away all modified data
72     into a linearised buffer in the transaction recovery area, then
73     marking the transaction recovery area with a magic value to
74     indicate a valid recovery record. In total 4 fsync/msync calls are
75     needed per commit to prevent race conditions. It might be possible
76     to reduce this to 3 or even 2 with some more work.
77
78   - check for a valid recovery record on open of the tdb, while the
79     global lock is held. Automatically recover from the transaction
80     recovery area if needed, then continue with the open as
81     usual. This allows for smooth crash recovery with no administrator
82     intervention.
83
84   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85     still available, but no transaction recovery area is used and no
86     fsync/msync calls are made.
87
88   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89     tdb_add_flags() transaction is enabled.
90     The default is that transaction nesting is not allowed and an attempt
91     to create a nested transaction will fail with TDB_ERR_NESTING.
92
93     Beware. when transactions are nested a transaction successfully
94     completed with tdb_transaction_commit() can be silently unrolled later.
95 */
96
97
98 /*
99   hold the context of any current transaction
100 */
101 struct tdb_transaction {
102         /* we keep a mirrored copy of the tdb hash heads here so
103            tdb_next_hash_chain() can operate efficiently */
104         uint32_t *hash_heads;
105
106         /* the original io methods - used to do IOs to the real db */
107         const struct tdb_methods *io_methods;
108
109         /* the list of transaction blocks. When a block is first
110            written to, it gets created in this list */
111         uint8_t **blocks;
112         uint32_t num_blocks;
113         uint32_t block_size;      /* bytes in each block */
114         uint32_t last_block_size; /* number of valid bytes in the last block */
115
116         /* non-zero when an internal transaction error has
117            occurred. All write operations will then fail until the
118            transaction is ended */
119         int transaction_error;
120
121         /* when inside a transaction we need to keep track of any
122            nested tdb_transaction_start() calls, as these are allowed,
123            but don't create a new transaction */
124         int nesting;
125
126         /* set when a prepare has already occurred */
127         bool prepared;
128         tdb_off_t magic_offset;
129
130         /* old file size before transaction */
131         tdb_len_t old_map_size;
132
133         /* we should re-pack on commit */
134         bool need_repack;
135 };
136
137
138 /*
139   read while in a transaction. We need to check first if the data is in our list
140   of transaction elements, then if not do a real read
141 */
142 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
143                             tdb_len_t len, int cv)
144 {
145         uint32_t blk;
146
147         /* Only a commit is allowed on a prepared transaction */
148         if (tdb->transaction->prepared) {
149                 tdb->ecode = TDB_ERR_EINVAL;
150                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: transaction already prepared, read not allowed\n"));
151                 tdb->transaction->transaction_error = 1;
152                 return -1;
153         }
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         return 0;
419 }
420
421 static int transaction_brunlock(struct tdb_context *tdb,
422                                 int rw_type, tdb_off_t offset, size_t len)
423 {
424         return 0;
425 }
426
427 static const struct tdb_methods transaction_methods = {
428         transaction_read,
429         transaction_write,
430         transaction_next_hash_chain,
431         transaction_oob,
432         transaction_expand_file,
433         transaction_brlock,
434         transaction_brunlock
435 };
436
437 /*
438   sync to disk
439 */
440 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
441 {       
442         if (tdb->flags & TDB_NOSYNC) {
443                 return 0;
444         }
445
446         if (fsync(tdb->fd) != 0) {
447                 tdb->ecode = TDB_ERR_IO;
448                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
449                 return -1;
450         }
451 #ifdef MS_SYNC
452         if (tdb->map_ptr) {
453                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
454                 if (msync(moffset + (char *)tdb->map_ptr, 
455                           length + (offset - moffset), MS_SYNC) != 0) {
456                         tdb->ecode = TDB_ERR_IO;
457                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
458                                  strerror(errno)));
459                         return -1;
460                 }
461         }
462 #endif
463         return 0;
464 }
465
466 /* ltype is F_WRLCK after prepare. */
467 int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype)
468 {
469         int i, ret = 0;
470
471         if (tdb->transaction == NULL) {
472                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
473                 return -1;
474         }
475
476         if (tdb->transaction->nesting != 0) {
477                 tdb->transaction->transaction_error = 1;
478                 tdb->transaction->nesting--;
479                 return 0;
480         }               
481
482         tdb->map_size = tdb->transaction->old_map_size;
483
484         /* free all the transaction blocks */
485         for (i=0;i<tdb->transaction->num_blocks;i++) {
486                 if (tdb->transaction->blocks[i] != NULL) {
487                         free(tdb->transaction->blocks[i]);
488                 }
489         }
490         SAFE_FREE(tdb->transaction->blocks);
491
492         if (tdb->transaction->magic_offset) {
493                 const struct tdb_methods *methods = tdb->transaction->io_methods;
494                 uint32_t zero = 0;
495
496                 /* remove the recovery marker */
497                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
498                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
499                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
500                         ret = -1;
501                 }
502         }
503
504         /* remove any global lock created during the transaction */
505         if (tdb->global_lock.count != 0) {
506                 tdb_brunlock(tdb, tdb->global_lock.ltype,
507                              FREELIST_TOP, 4*tdb->header.hash_size);
508                 tdb->global_lock.count = 0;
509         }
510
511         /* remove any locks created during the transaction */
512         if (tdb->num_locks != 0) {
513                 for (i=0;i<tdb->num_lockrecs;i++) {
514                         tdb_brunlock(tdb, tdb->lockrecs[i].ltype,
515                                      FREELIST_TOP+4*tdb->lockrecs[i].list, 1);
516                 }
517                 tdb->num_locks = 0;
518                 tdb->num_lockrecs = 0;
519                 SAFE_FREE(tdb->lockrecs);
520         }
521
522         /* restore the normal io methods */
523         tdb->methods = tdb->transaction->io_methods;
524
525         tdb_brunlock(tdb, ltype, FREELIST_TOP, 0);
526         tdb_transaction_unlock(tdb, F_WRLCK);
527         SAFE_FREE(tdb->transaction->hash_heads);
528         SAFE_FREE(tdb->transaction);
529         
530         return ret;
531 }
532
533 /*
534   start a tdb transaction. No token is returned, as only a single
535   transaction is allowed to be pending per tdb_context
536 */
537 int tdb_transaction_start(struct tdb_context *tdb)
538 {
539         /* some sanity checks */
540         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
541                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
542                 tdb->ecode = TDB_ERR_EINVAL;
543                 return -1;
544         }
545
546         /* cope with nested tdb_transaction_start() calls */
547         if (tdb->transaction != NULL) {
548                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
549                         tdb->ecode = TDB_ERR_NESTING;
550                         return -1;
551                 }
552                 tdb_trace(tdb, "tdb_transaction_start");
553                 tdb->transaction->nesting++;
554                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
555                          tdb->transaction->nesting));
556                 return 0;
557         }
558
559         if (tdb->num_locks != 0 || tdb->global_lock.count) {
560                 /* the caller must not have any locks when starting a
561                    transaction as otherwise we'll be screwed by lack
562                    of nested locks in posix */
563                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
564                 tdb->ecode = TDB_ERR_LOCK;
565                 return -1;
566         }
567
568         if (tdb->travlocks.next != NULL) {
569                 /* you cannot use transactions inside a traverse (although you can use
570                    traverse inside a transaction) as otherwise you can end up with
571                    deadlock */
572                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
573                 tdb->ecode = TDB_ERR_LOCK;
574                 return -1;
575         }
576
577         tdb->transaction = (struct tdb_transaction *)
578                 calloc(sizeof(struct tdb_transaction), 1);
579         if (tdb->transaction == NULL) {
580                 tdb->ecode = TDB_ERR_OOM;
581                 return -1;
582         }
583
584         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
585         tdb->transaction->block_size = tdb->page_size;
586
587         /* get the transaction write lock. This is a blocking lock. As
588            discussed with Volker, there are a number of ways we could
589            make this async, which we will probably do in the future */
590         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
591                 SAFE_FREE(tdb->transaction->blocks);
592                 SAFE_FREE(tdb->transaction);
593                 return -1;
594         }
595         
596         /* get a read lock from the freelist to the end of file. This
597            is upgraded to a write lock during the commit */
598         if (tdb_brlock(tdb, F_RDLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT) == -1) {
599                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
600                 tdb->ecode = TDB_ERR_LOCK;
601                 goto fail;
602         }
603
604         /* setup a copy of the hash table heads so the hash scan in
605            traverse can be fast */
606         tdb->transaction->hash_heads = (uint32_t *)
607                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
608         if (tdb->transaction->hash_heads == NULL) {
609                 tdb->ecode = TDB_ERR_OOM;
610                 goto fail;
611         }
612         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
613                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
614                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
615                 tdb->ecode = TDB_ERR_IO;
616                 goto fail;
617         }
618
619         /* make sure we know about any file expansions already done by
620            anyone else */
621         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
622         tdb->transaction->old_map_size = tdb->map_size;
623
624         /* finally hook the io methods, replacing them with
625            transaction specific methods */
626         tdb->transaction->io_methods = tdb->methods;
627         tdb->methods = &transaction_methods;
628
629         /* Trace at the end, so we get sequence number correct. */
630         tdb_trace(tdb, "tdb_transaction_start");
631         return 0;
632         
633 fail:
634         tdb_brunlock(tdb, F_RDLCK, FREELIST_TOP, 0);
635         tdb_transaction_unlock(tdb, F_WRLCK);
636         SAFE_FREE(tdb->transaction->blocks);
637         SAFE_FREE(tdb->transaction->hash_heads);
638         SAFE_FREE(tdb->transaction);
639         return -1;
640 }
641
642
643 /*
644   cancel the current transaction
645 */
646 int tdb_transaction_cancel(struct tdb_context *tdb)
647 {
648         int ltype = F_RDLCK;
649         tdb_trace(tdb, "tdb_transaction_cancel");
650         if (tdb->transaction && tdb->transaction->prepared)
651                 ltype = F_WRLCK;
652         return _tdb_transaction_cancel(tdb, ltype);
653 }
654
655 /*
656   work out how much space the linearised recovery data will consume
657 */
658 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
659 {
660         tdb_len_t recovery_size = 0;
661         int i;
662
663         recovery_size = sizeof(uint32_t);
664         for (i=0;i<tdb->transaction->num_blocks;i++) {
665                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
666                         break;
667                 }
668                 if (tdb->transaction->blocks[i] == NULL) {
669                         continue;
670                 }
671                 recovery_size += 2*sizeof(tdb_off_t);
672                 if (i == tdb->transaction->num_blocks-1) {
673                         recovery_size += tdb->transaction->last_block_size;
674                 } else {
675                         recovery_size += tdb->transaction->block_size;
676                 }
677         }       
678
679         return recovery_size;
680 }
681
682 /*
683   allocate the recovery area, or use an existing recovery area if it is
684   large enough
685 */
686 static int tdb_recovery_allocate(struct tdb_context *tdb, 
687                                  tdb_len_t *recovery_size,
688                                  tdb_off_t *recovery_offset,
689                                  tdb_len_t *recovery_max_size)
690 {
691         struct list_struct rec;
692         const struct tdb_methods *methods = tdb->transaction->io_methods;
693         tdb_off_t recovery_head;
694
695         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
696                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
697                 return -1;
698         }
699
700         rec.rec_len = 0;
701
702         if (recovery_head != 0 && 
703             methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
704                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
705                 return -1;
706         }
707
708         *recovery_size = tdb_recovery_size(tdb);
709
710         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
711                 /* it fits in the existing area */
712                 *recovery_max_size = rec.rec_len;
713                 *recovery_offset = recovery_head;
714                 return 0;
715         }
716
717         /* we need to free up the old recovery area, then allocate a
718            new one at the end of the file. Note that we cannot use
719            tdb_allocate() to allocate the new one as that might return
720            us an area that is being currently used (as of the start of
721            the transaction) */
722         if (recovery_head != 0) {
723                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
724                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
725                         return -1;
726                 }
727         }
728
729         /* the tdb_free() call might have increased the recovery size */
730         *recovery_size = tdb_recovery_size(tdb);
731
732         /* round up to a multiple of page size */
733         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
734         *recovery_offset = tdb->map_size;
735         recovery_head = *recovery_offset;
736
737         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
738                                      (tdb->map_size - tdb->transaction->old_map_size) +
739                                      sizeof(rec) + *recovery_max_size) == -1) {
740                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
741                 return -1;
742         }
743
744         /* remap the file (if using mmap) */
745         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
746
747         /* we have to reset the old map size so that we don't try to expand the file
748            again in the transaction commit, which would destroy the recovery area */
749         tdb->transaction->old_map_size = tdb->map_size;
750
751         /* write the recovery header offset and sync - we can sync without a race here
752            as the magic ptr in the recovery record has not been set */
753         CONVERT(recovery_head);
754         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
755                                &recovery_head, sizeof(tdb_off_t)) == -1) {
756                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
757                 return -1;
758         }
759         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
760                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
761                 return -1;
762         }
763
764         return 0;
765 }
766
767
768 /*
769   setup the recovery data that will be used on a crash during commit
770 */
771 static int transaction_setup_recovery(struct tdb_context *tdb, 
772                                       tdb_off_t *magic_offset)
773 {
774         tdb_len_t recovery_size;
775         unsigned char *data, *p;
776         const struct tdb_methods *methods = tdb->transaction->io_methods;
777         struct list_struct *rec;
778         tdb_off_t recovery_offset, recovery_max_size;
779         tdb_off_t old_map_size = tdb->transaction->old_map_size;
780         uint32_t magic, tailer;
781         int i;
782
783         /*
784           check that the recovery area has enough space
785         */
786         if (tdb_recovery_allocate(tdb, &recovery_size, 
787                                   &recovery_offset, &recovery_max_size) == -1) {
788                 return -1;
789         }
790
791         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
792         if (data == NULL) {
793                 tdb->ecode = TDB_ERR_OOM;
794                 return -1;
795         }
796
797         rec = (struct list_struct *)data;
798         memset(rec, 0, sizeof(*rec));
799
800         rec->magic    = 0;
801         rec->data_len = recovery_size;
802         rec->rec_len  = recovery_max_size;
803         rec->key_len  = old_map_size;
804         CONVERT(rec);
805
806         /* build the recovery data into a single blob to allow us to do a single
807            large write, which should be more efficient */
808         p = data + sizeof(*rec);
809         for (i=0;i<tdb->transaction->num_blocks;i++) {
810                 tdb_off_t offset;
811                 tdb_len_t length;
812
813                 if (tdb->transaction->blocks[i] == NULL) {
814                         continue;
815                 }
816
817                 offset = i * tdb->transaction->block_size;
818                 length = tdb->transaction->block_size;
819                 if (i == tdb->transaction->num_blocks-1) {
820                         length = tdb->transaction->last_block_size;
821                 }
822                 
823                 if (offset >= old_map_size) {
824                         continue;
825                 }
826                 if (offset + length > tdb->transaction->old_map_size) {
827                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
828                         free(data);
829                         tdb->ecode = TDB_ERR_CORRUPT;
830                         return -1;
831                 }
832                 memcpy(p, &offset, 4);
833                 memcpy(p+4, &length, 4);
834                 if (DOCONV()) {
835                         tdb_convert(p, 8);
836                 }
837                 /* the recovery area contains the old data, not the
838                    new data, so we have to call the original tdb_read
839                    method to get it */
840                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
841                         free(data);
842                         tdb->ecode = TDB_ERR_IO;
843                         return -1;
844                 }
845                 p += 8 + length;
846         }
847
848         /* and the tailer */
849         tailer = sizeof(*rec) + recovery_max_size;
850         memcpy(p, &tailer, 4);
851         CONVERT(p);
852
853         /* write the recovery data to the recovery area */
854         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
855                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
856                 free(data);
857                 tdb->ecode = TDB_ERR_IO;
858                 return -1;
859         }
860         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
861                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
862                 free(data);
863                 tdb->ecode = TDB_ERR_IO;
864                 return -1;
865         }
866
867         /* as we don't have ordered writes, we have to sync the recovery
868            data before we update the magic to indicate that the recovery
869            data is present */
870         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
871                 free(data);
872                 return -1;
873         }
874
875         free(data);
876
877         magic = TDB_RECOVERY_MAGIC;
878         CONVERT(magic);
879
880         *magic_offset = recovery_offset + offsetof(struct list_struct, magic);
881
882         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
883                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
884                 tdb->ecode = TDB_ERR_IO;
885                 return -1;
886         }
887         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
888                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
889                 tdb->ecode = TDB_ERR_IO;
890                 return -1;
891         }
892
893         /* ensure the recovery magic marker is on disk */
894         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
895                 return -1;
896         }
897
898         return 0;
899 }
900
901 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
902 {       
903         const struct tdb_methods *methods;
904
905         if (tdb->transaction == NULL) {
906                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
907                 return -1;
908         }
909
910         if (tdb->transaction->prepared) {
911                 tdb->ecode = TDB_ERR_EINVAL;
912                 _tdb_transaction_cancel(tdb, F_WRLCK);
913                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
914                 return -1;
915         }
916
917         if (tdb->transaction->transaction_error) {
918                 tdb->ecode = TDB_ERR_IO;
919                 _tdb_transaction_cancel(tdb, F_RDLCK);
920                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
921                 return -1;
922         }
923
924
925         if (tdb->transaction->nesting != 0) {
926                 tdb->transaction->nesting--;
927                 return 0;
928         }               
929
930 #ifdef TDB_TRACE
931         /* store seqnum now, before reading becomes illegal. */
932         tdb_ofs_read(tdb, TDB_SEQNUM_OFS, &tdb->transaction_prepare_seqnum);
933 #endif
934
935         /* check for a null transaction */
936         if (tdb->transaction->blocks == NULL) {
937                 return 0;
938         }
939
940         methods = tdb->transaction->io_methods;
941         
942         /* if there are any locks pending then the caller has not
943            nested their locks properly, so fail the transaction */
944         if (tdb->num_locks || tdb->global_lock.count) {
945                 tdb->ecode = TDB_ERR_LOCK;
946                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
947                 _tdb_transaction_cancel(tdb, F_RDLCK);
948                 return -1;
949         }
950
951         /* upgrade the main transaction lock region to a write lock */
952         if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
953                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
954                 tdb->ecode = TDB_ERR_LOCK;
955                 _tdb_transaction_cancel(tdb, F_RDLCK);
956                 return -1;
957         }
958
959         /* get the global lock - this prevents new users attaching to the database
960            during the commit */
961         if (tdb_brlock(tdb, F_WRLCK, GLOBAL_LOCK, 1, TDB_LOCK_WAIT) == -1) {
962                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
963                 tdb->ecode = TDB_ERR_LOCK;
964                 _tdb_transaction_cancel(tdb, F_WRLCK);
965                 return -1;
966         }
967
968         if (!(tdb->flags & TDB_NOSYNC)) {
969                 /* write the recovery data to the end of the file */
970                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
971                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
972                         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 0);
973                         _tdb_transaction_cancel(tdb, F_WRLCK);
974                         return -1;
975                 }
976         }
977
978         tdb->transaction->prepared = true;
979
980         /* expand the file to the new size if needed */
981         if (tdb->map_size != tdb->transaction->old_map_size) {
982                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
983                                              tdb->map_size - 
984                                              tdb->transaction->old_map_size) == -1) {
985                         tdb->ecode = TDB_ERR_IO;
986                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
987                         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 0);
988                         _tdb_transaction_cancel(tdb, F_WRLCK);
989                         return -1;
990                 }
991                 tdb->map_size = tdb->transaction->old_map_size;
992                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
993         }
994
995         /* Keep the global lock until the actual commit */
996
997         return 0;
998 }
999
1000 /*
1001    prepare to commit the current transaction
1002 */
1003 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1004 {       
1005         tdb_trace(tdb, "tdb_transaction_prepare_commit");
1006         return _tdb_transaction_prepare_commit(tdb);
1007 }
1008
1009 /*
1010   commit the current transaction
1011 */
1012 int tdb_transaction_commit(struct tdb_context *tdb)
1013 {       
1014         const struct tdb_methods *methods;
1015         int i;
1016         bool need_repack;
1017
1018         if (tdb->transaction == NULL) {
1019                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1020                 return -1;
1021         }
1022
1023         /* If we've prepared, can't read seqnum. */
1024         if (tdb->transaction->prepared) {
1025                 tdb_trace_seqnum(tdb, tdb->transaction_prepare_seqnum,
1026                                  "tdb_transaction_commit");
1027         } else {
1028                 tdb_trace(tdb, "tdb_transaction_commit");
1029         }
1030
1031         if (tdb->transaction->transaction_error) {
1032                 tdb->ecode = TDB_ERR_IO;
1033                 tdb_transaction_cancel(tdb);
1034                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1035                 return -1;
1036         }
1037
1038
1039         if (tdb->transaction->nesting != 0) {
1040                 tdb->transaction->nesting--;
1041                 return 0;
1042         }
1043
1044         /* check for a null transaction */
1045         if (tdb->transaction->blocks == NULL) {
1046                 _tdb_transaction_cancel(tdb, F_RDLCK);
1047                 return 0;
1048         }
1049
1050         if (!tdb->transaction->prepared) {
1051                 int ret = _tdb_transaction_prepare_commit(tdb);
1052                 if (ret)
1053                         return ret;
1054         }
1055
1056         methods = tdb->transaction->io_methods;
1057
1058         /* perform all the writes */
1059         for (i=0;i<tdb->transaction->num_blocks;i++) {
1060                 tdb_off_t offset;
1061                 tdb_len_t length;
1062
1063                 if (tdb->transaction->blocks[i] == NULL) {
1064                         continue;
1065                 }
1066
1067                 offset = i * tdb->transaction->block_size;
1068                 length = tdb->transaction->block_size;
1069                 if (i == tdb->transaction->num_blocks-1) {
1070                         length = tdb->transaction->last_block_size;
1071                 }
1072
1073                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1074                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1075                         
1076                         /* we've overwritten part of the data and
1077                            possibly expanded the file, so we need to
1078                            run the crash recovery code */
1079                         tdb->methods = methods;
1080                         tdb_transaction_recover(tdb); 
1081
1082                         _tdb_transaction_cancel(tdb, F_WRLCK);
1083                         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 0);
1084
1085                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1086                         return -1;
1087                 }
1088                 SAFE_FREE(tdb->transaction->blocks[i]);
1089         } 
1090
1091         SAFE_FREE(tdb->transaction->blocks);
1092         tdb->transaction->num_blocks = 0;
1093
1094         /* ensure the new data is on disk */
1095         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1096                 return -1;
1097         }
1098
1099         tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
1100
1101         /*
1102           TODO: maybe write to some dummy hdr field, or write to magic
1103           offset without mmap, before the last sync, instead of the
1104           utime() call
1105         */
1106
1107         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1108            don't change the mtime of the file, this means the file may
1109            not be backed up (as tdb rounding to block sizes means that
1110            file size changes are quite rare too). The following forces
1111            mtime changes when a transaction completes */
1112 #if HAVE_UTIME
1113         utime(tdb->name, NULL);
1114 #endif
1115
1116         need_repack = tdb->transaction->need_repack;
1117
1118         /* use a transaction cancel to free memory and remove the
1119            transaction locks */
1120         _tdb_transaction_cancel(tdb, F_WRLCK);
1121
1122         if (need_repack) {
1123                 return tdb_repack(tdb);
1124         }
1125
1126         return 0;
1127 }
1128
1129
1130 /*
1131   recover from an aborted transaction. Must be called with exclusive
1132   database write access already established (including the global
1133   lock to prevent new processes attaching)
1134 */
1135 int tdb_transaction_recover(struct tdb_context *tdb)
1136 {
1137         tdb_off_t recovery_head, recovery_eof;
1138         unsigned char *data, *p;
1139         uint32_t zero = 0;
1140         struct list_struct rec;
1141
1142         /* find the recovery area */
1143         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1144                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1145                 tdb->ecode = TDB_ERR_IO;
1146                 return -1;
1147         }
1148
1149         if (recovery_head == 0) {
1150                 /* we have never allocated a recovery record */
1151                 return 0;
1152         }
1153
1154         /* read the recovery record */
1155         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1156                                    sizeof(rec), DOCONV()) == -1) {
1157                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1158                 tdb->ecode = TDB_ERR_IO;
1159                 return -1;
1160         }
1161
1162         if (rec.magic != TDB_RECOVERY_MAGIC) {
1163                 /* there is no valid recovery data */
1164                 return 0;
1165         }
1166
1167         if (tdb->read_only) {
1168                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1169                 tdb->ecode = TDB_ERR_CORRUPT;
1170                 return -1;
1171         }
1172
1173         recovery_eof = rec.key_len;
1174
1175         data = (unsigned char *)malloc(rec.data_len);
1176         if (data == NULL) {
1177                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1178                 tdb->ecode = TDB_ERR_OOM;
1179                 return -1;
1180         }
1181
1182         /* read the full recovery data */
1183         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1184                                    rec.data_len, 0) == -1) {
1185                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1186                 tdb->ecode = TDB_ERR_IO;
1187                 return -1;
1188         }
1189
1190         /* recover the file data */
1191         p = data;
1192         while (p+8 < data + rec.data_len) {
1193                 uint32_t ofs, len;
1194                 if (DOCONV()) {
1195                         tdb_convert(p, 8);
1196                 }
1197                 memcpy(&ofs, p, 4);
1198                 memcpy(&len, p+4, 4);
1199
1200                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1201                         free(data);
1202                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1203                         tdb->ecode = TDB_ERR_IO;
1204                         return -1;
1205                 }
1206                 p += 8 + len;
1207         }
1208
1209         free(data);
1210
1211         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1212                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1213                 tdb->ecode = TDB_ERR_IO;
1214                 return -1;
1215         }
1216
1217         /* if the recovery area is after the recovered eof then remove it */
1218         if (recovery_eof <= recovery_head) {
1219                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1220                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1221                         tdb->ecode = TDB_ERR_IO;
1222                         return -1;                      
1223                 }
1224         }
1225
1226         /* remove the recovery magic */
1227         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic), 
1228                           &zero) == -1) {
1229                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1230                 tdb->ecode = TDB_ERR_IO;
1231                 return -1;                      
1232         }
1233         
1234         /* reduce the file size to the old size */
1235         tdb_munmap(tdb);
1236         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1237                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1238                 tdb->ecode = TDB_ERR_IO;
1239                 return -1;                      
1240         }
1241         tdb->map_size = recovery_eof;
1242         tdb_mmap(tdb);
1243
1244         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1245                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1246                 tdb->ecode = TDB_ERR_IO;
1247                 return -1;
1248         }
1249
1250         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1251                  recovery_eof));
1252
1253         /* all done */
1254         return 0;
1255 }