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