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