tdb: rename tdb_release_extra_locks() to tdb_release_transaction_locks()
[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     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 /*
442   start a tdb transaction. No token is returned, as only a single
443   transaction is allowed to be pending per tdb_context
444 */
445 int tdb_transaction_start(struct tdb_context *tdb)
446 {
447         /* some sanity checks */
448         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
449                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
450                 tdb->ecode = TDB_ERR_EINVAL;
451                 return -1;
452         }
453
454         /* cope with nested tdb_transaction_start() calls */
455         if (tdb->transaction != NULL) {
456                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
457                         tdb->ecode = TDB_ERR_NESTING;
458                         return -1;
459                 }
460                 tdb->transaction->nesting++;
461                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
462                          tdb->transaction->nesting));
463                 return 0;
464         }
465
466         if (tdb_have_extra_locks(tdb)) {
467                 /* the caller must not have any locks when starting a
468                    transaction as otherwise we'll be screwed by lack
469                    of nested locks in posix */
470                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
471                 tdb->ecode = TDB_ERR_LOCK;
472                 return -1;
473         }
474
475         if (tdb->travlocks.next != NULL) {
476                 /* you cannot use transactions inside a traverse (although you can use
477                    traverse inside a transaction) as otherwise you can end up with
478                    deadlock */
479                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
480                 tdb->ecode = TDB_ERR_LOCK;
481                 return -1;
482         }
483
484         tdb->transaction = (struct tdb_transaction *)
485                 calloc(sizeof(struct tdb_transaction), 1);
486         if (tdb->transaction == NULL) {
487                 tdb->ecode = TDB_ERR_OOM;
488                 return -1;
489         }
490
491         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
492         tdb->transaction->block_size = tdb->page_size;
493
494         /* get the transaction write lock. This is a blocking lock. As
495            discussed with Volker, there are a number of ways we could
496            make this async, which we will probably do in the future */
497         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
498                 SAFE_FREE(tdb->transaction->blocks);
499                 SAFE_FREE(tdb->transaction);
500                 return -1;
501         }
502         
503         /* get a read lock from the freelist to the end of file. This
504            is upgraded to a write lock during the commit */
505         if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
506                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
507                 goto fail_allrecord_lock;
508         }
509
510         /* setup a copy of the hash table heads so the hash scan in
511            traverse can be fast */
512         tdb->transaction->hash_heads = (uint32_t *)
513                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
514         if (tdb->transaction->hash_heads == NULL) {
515                 tdb->ecode = TDB_ERR_OOM;
516                 goto fail;
517         }
518         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
519                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
520                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
521                 tdb->ecode = TDB_ERR_IO;
522                 goto fail;
523         }
524
525         /* make sure we know about any file expansions already done by
526            anyone else */
527         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
528         tdb->transaction->old_map_size = tdb->map_size;
529
530         /* finally hook the io methods, replacing them with
531            transaction specific methods */
532         tdb->transaction->io_methods = tdb->methods;
533         tdb->methods = &transaction_methods;
534
535         /* Trace at the end, so we get sequence number correct. */
536         tdb_trace(tdb, "tdb_transaction_start");
537         return 0;
538         
539 fail:
540         tdb_allrecord_unlock(tdb, F_RDLCK, false);
541 fail_allrecord_lock:
542         tdb_transaction_unlock(tdb, F_WRLCK);
543         SAFE_FREE(tdb->transaction->blocks);
544         SAFE_FREE(tdb->transaction->hash_heads);
545         SAFE_FREE(tdb->transaction);
546         return -1;
547 }
548
549
550 /*
551   sync to disk
552 */
553 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
554 {       
555         if (tdb->flags & TDB_NOSYNC) {
556                 return 0;
557         }
558
559         if (fdatasync(tdb->fd) != 0) {
560                 tdb->ecode = TDB_ERR_IO;
561                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
562                 return -1;
563         }
564 #ifdef HAVE_MMAP
565         if (tdb->map_ptr) {
566                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
567                 if (msync(moffset + (char *)tdb->map_ptr, 
568                           length + (offset - moffset), MS_SYNC) != 0) {
569                         tdb->ecode = TDB_ERR_IO;
570                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
571                                  strerror(errno)));
572                         return -1;
573                 }
574         }
575 #endif
576         return 0;
577 }
578
579
580 static int _tdb_transaction_cancel(struct tdb_context *tdb)
581 {       
582         int i, ret = 0;
583
584         if (tdb->transaction == NULL) {
585                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
586                 return -1;
587         }
588
589         if (tdb->transaction->nesting != 0) {
590                 tdb->transaction->transaction_error = 1;
591                 tdb->transaction->nesting--;
592                 return 0;
593         }               
594
595         tdb->map_size = tdb->transaction->old_map_size;
596
597         /* free all the transaction blocks */
598         for (i=0;i<tdb->transaction->num_blocks;i++) {
599                 if (tdb->transaction->blocks[i] != NULL) {
600                         free(tdb->transaction->blocks[i]);
601                 }
602         }
603         SAFE_FREE(tdb->transaction->blocks);
604
605         if (tdb->transaction->magic_offset) {
606                 const struct tdb_methods *methods = tdb->transaction->io_methods;
607                 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
608
609                 /* remove the recovery marker */
610                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
611                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
612                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
613                         ret = -1;
614                 }
615         }
616
617         /* This also removes the OPEN_LOCK, if we have it. */
618         tdb_release_transaction_locks(tdb);
619
620         /* restore the normal io methods */
621         tdb->methods = tdb->transaction->io_methods;
622
623         SAFE_FREE(tdb->transaction->hash_heads);
624         SAFE_FREE(tdb->transaction);
625         
626         return ret;
627 }
628
629 /*
630   cancel the current transaction
631 */
632 int tdb_transaction_cancel(struct tdb_context *tdb)
633 {
634         tdb_trace(tdb, "tdb_transaction_cancel");
635         return _tdb_transaction_cancel(tdb);
636 }
637
638 /*
639   work out how much space the linearised recovery data will consume
640 */
641 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
642 {
643         tdb_len_t recovery_size = 0;
644         int i;
645
646         recovery_size = sizeof(uint32_t);
647         for (i=0;i<tdb->transaction->num_blocks;i++) {
648                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
649                         break;
650                 }
651                 if (tdb->transaction->blocks[i] == NULL) {
652                         continue;
653                 }
654                 recovery_size += 2*sizeof(tdb_off_t);
655                 if (i == tdb->transaction->num_blocks-1) {
656                         recovery_size += tdb->transaction->last_block_size;
657                 } else {
658                         recovery_size += tdb->transaction->block_size;
659                 }
660         }       
661
662         return recovery_size;
663 }
664
665 /*
666   allocate the recovery area, or use an existing recovery area if it is
667   large enough
668 */
669 static int tdb_recovery_allocate(struct tdb_context *tdb, 
670                                  tdb_len_t *recovery_size,
671                                  tdb_off_t *recovery_offset,
672                                  tdb_len_t *recovery_max_size)
673 {
674         struct tdb_record rec;
675         const struct tdb_methods *methods = tdb->transaction->io_methods;
676         tdb_off_t recovery_head;
677
678         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
679                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
680                 return -1;
681         }
682
683         rec.rec_len = 0;
684
685         if (recovery_head != 0) {
686                 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
687                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
688                         return -1;
689                 }
690                 /* ignore invalid recovery regions: can happen in crash */
691                 if (rec.magic != TDB_RECOVERY_MAGIC &&
692                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
693                         recovery_head = 0;
694                 }
695         }
696
697         *recovery_size = tdb_recovery_size(tdb);
698
699         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
700                 /* it fits in the existing area */
701                 *recovery_max_size = rec.rec_len;
702                 *recovery_offset = recovery_head;
703                 return 0;
704         }
705
706         /* we need to free up the old recovery area, then allocate a
707            new one at the end of the file. Note that we cannot use
708            tdb_allocate() to allocate the new one as that might return
709            us an area that is being currently used (as of the start of
710            the transaction) */
711         if (recovery_head != 0) {
712                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
713                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
714                         return -1;
715                 }
716         }
717
718         /* the tdb_free() call might have increased the recovery size */
719         *recovery_size = tdb_recovery_size(tdb);
720
721         /* round up to a multiple of page size */
722         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
723         *recovery_offset = tdb->map_size;
724         recovery_head = *recovery_offset;
725
726         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
727                                      (tdb->map_size - tdb->transaction->old_map_size) +
728                                      sizeof(rec) + *recovery_max_size) == -1) {
729                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
730                 return -1;
731         }
732
733         /* remap the file (if using mmap) */
734         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
735
736         /* we have to reset the old map size so that we don't try to expand the file
737            again in the transaction commit, which would destroy the recovery area */
738         tdb->transaction->old_map_size = tdb->map_size;
739
740         /* write the recovery header offset and sync - we can sync without a race here
741            as the magic ptr in the recovery record has not been set */
742         CONVERT(recovery_head);
743         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
744                                &recovery_head, sizeof(tdb_off_t)) == -1) {
745                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
746                 return -1;
747         }
748         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
749                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
750                 return -1;
751         }
752
753         return 0;
754 }
755
756
757 /*
758   setup the recovery data that will be used on a crash during commit
759 */
760 static int transaction_setup_recovery(struct tdb_context *tdb, 
761                                       tdb_off_t *magic_offset)
762 {
763         tdb_len_t recovery_size;
764         unsigned char *data, *p;
765         const struct tdb_methods *methods = tdb->transaction->io_methods;
766         struct tdb_record *rec;
767         tdb_off_t recovery_offset, recovery_max_size;
768         tdb_off_t old_map_size = tdb->transaction->old_map_size;
769         uint32_t magic, tailer;
770         int i;
771
772         /*
773           check that the recovery area has enough space
774         */
775         if (tdb_recovery_allocate(tdb, &recovery_size, 
776                                   &recovery_offset, &recovery_max_size) == -1) {
777                 return -1;
778         }
779
780         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
781         if (data == NULL) {
782                 tdb->ecode = TDB_ERR_OOM;
783                 return -1;
784         }
785
786         rec = (struct tdb_record *)data;
787         memset(rec, 0, sizeof(*rec));
788
789         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
790         rec->data_len = recovery_size;
791         rec->rec_len  = recovery_max_size;
792         rec->key_len  = old_map_size;
793         CONVERT(rec);
794
795         /* build the recovery data into a single blob to allow us to do a single
796            large write, which should be more efficient */
797         p = data + sizeof(*rec);
798         for (i=0;i<tdb->transaction->num_blocks;i++) {
799                 tdb_off_t offset;
800                 tdb_len_t length;
801
802                 if (tdb->transaction->blocks[i] == NULL) {
803                         continue;
804                 }
805
806                 offset = i * tdb->transaction->block_size;
807                 length = tdb->transaction->block_size;
808                 if (i == tdb->transaction->num_blocks-1) {
809                         length = tdb->transaction->last_block_size;
810                 }
811                 
812                 if (offset >= old_map_size) {
813                         continue;
814                 }
815                 if (offset + length > tdb->transaction->old_map_size) {
816                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
817                         free(data);
818                         tdb->ecode = TDB_ERR_CORRUPT;
819                         return -1;
820                 }
821                 memcpy(p, &offset, 4);
822                 memcpy(p+4, &length, 4);
823                 if (DOCONV()) {
824                         tdb_convert(p, 8);
825                 }
826                 /* the recovery area contains the old data, not the
827                    new data, so we have to call the original tdb_read
828                    method to get it */
829                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
830                         free(data);
831                         tdb->ecode = TDB_ERR_IO;
832                         return -1;
833                 }
834                 p += 8 + length;
835         }
836
837         /* and the tailer */
838         tailer = sizeof(*rec) + recovery_max_size;
839         memcpy(p, &tailer, 4);
840         CONVERT(p);
841
842         /* write the recovery data to the recovery area */
843         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
844                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
845                 free(data);
846                 tdb->ecode = TDB_ERR_IO;
847                 return -1;
848         }
849         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
850                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
851                 free(data);
852                 tdb->ecode = TDB_ERR_IO;
853                 return -1;
854         }
855
856         /* as we don't have ordered writes, we have to sync the recovery
857            data before we update the magic to indicate that the recovery
858            data is present */
859         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
860                 free(data);
861                 return -1;
862         }
863
864         free(data);
865
866         magic = TDB_RECOVERY_MAGIC;
867         CONVERT(magic);
868
869         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
870
871         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
872                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
873                 tdb->ecode = TDB_ERR_IO;
874                 return -1;
875         }
876         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
877                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
878                 tdb->ecode = TDB_ERR_IO;
879                 return -1;
880         }
881
882         /* ensure the recovery magic marker is on disk */
883         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
884                 return -1;
885         }
886
887         return 0;
888 }
889
890 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
891 {       
892         const struct tdb_methods *methods;
893
894         if (tdb->transaction == NULL) {
895                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
896                 return -1;
897         }
898
899         if (tdb->transaction->prepared) {
900                 tdb->ecode = TDB_ERR_EINVAL;
901                 _tdb_transaction_cancel(tdb);
902                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
903                 return -1;
904         }
905
906         if (tdb->transaction->transaction_error) {
907                 tdb->ecode = TDB_ERR_IO;
908                 _tdb_transaction_cancel(tdb);
909                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
910                 return -1;
911         }
912
913
914         if (tdb->transaction->nesting != 0) {
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);
931                 return -1;
932         }
933
934         /* upgrade the main transaction lock region to a write lock */
935         if (tdb_allrecord_upgrade(tdb) == -1) {
936                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
937                 _tdb_transaction_cancel(tdb);
938                 return -1;
939         }
940
941         /* get the open lock - this prevents new users attaching to the database
942            during the commit */
943         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
944                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
945                 _tdb_transaction_cancel(tdb);
946                 return -1;
947         }
948
949         if (!(tdb->flags & TDB_NOSYNC)) {
950                 /* write the recovery data to the end of the file */
951                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
952                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
953                         _tdb_transaction_cancel(tdb);
954                         return -1;
955                 }
956         }
957
958         tdb->transaction->prepared = true;
959
960         /* expand the file to the new size if needed */
961         if (tdb->map_size != tdb->transaction->old_map_size) {
962                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
963                                              tdb->map_size - 
964                                              tdb->transaction->old_map_size) == -1) {
965                         tdb->ecode = TDB_ERR_IO;
966                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
967                         _tdb_transaction_cancel(tdb);
968                         return -1;
969                 }
970                 tdb->map_size = tdb->transaction->old_map_size;
971                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
972         }
973
974         /* Keep the open lock until the actual commit */
975
976         return 0;
977 }
978
979 /*
980    prepare to commit the current transaction
981 */
982 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
983 {       
984         tdb_trace(tdb, "tdb_transaction_prepare_commit");
985         return _tdb_transaction_prepare_commit(tdb);
986 }
987
988 /*
989   commit the current transaction
990 */
991 int tdb_transaction_commit(struct tdb_context *tdb)
992 {       
993         const struct tdb_methods *methods;
994         int i;
995         bool need_repack;
996
997         if (tdb->transaction == NULL) {
998                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
999                 return -1;
1000         }
1001
1002         tdb_trace(tdb, "tdb_transaction_commit");
1003
1004         if (tdb->transaction->transaction_error) {
1005                 tdb->ecode = TDB_ERR_IO;
1006                 _tdb_transaction_cancel(tdb);
1007                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1008                 return -1;
1009         }
1010
1011
1012         if (tdb->transaction->nesting != 0) {
1013                 tdb->transaction->nesting--;
1014                 return 0;
1015         }
1016
1017         /* check for a null transaction */
1018         if (tdb->transaction->blocks == NULL) {
1019                 _tdb_transaction_cancel(tdb);
1020                 return 0;
1021         }
1022
1023         if (!tdb->transaction->prepared) {
1024                 int ret = _tdb_transaction_prepare_commit(tdb);
1025                 if (ret)
1026                         return ret;
1027         }
1028
1029         methods = tdb->transaction->io_methods;
1030
1031         /* perform all the writes */
1032         for (i=0;i<tdb->transaction->num_blocks;i++) {
1033                 tdb_off_t offset;
1034                 tdb_len_t length;
1035
1036                 if (tdb->transaction->blocks[i] == NULL) {
1037                         continue;
1038                 }
1039
1040                 offset = i * tdb->transaction->block_size;
1041                 length = tdb->transaction->block_size;
1042                 if (i == tdb->transaction->num_blocks-1) {
1043                         length = tdb->transaction->last_block_size;
1044                 }
1045
1046                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1047                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1048                         
1049                         /* we've overwritten part of the data and
1050                            possibly expanded the file, so we need to
1051                            run the crash recovery code */
1052                         tdb->methods = methods;
1053                         tdb_transaction_recover(tdb); 
1054
1055                         _tdb_transaction_cancel(tdb);
1056
1057                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1058                         return -1;
1059                 }
1060                 SAFE_FREE(tdb->transaction->blocks[i]);
1061         } 
1062
1063         SAFE_FREE(tdb->transaction->blocks);
1064         tdb->transaction->num_blocks = 0;
1065
1066         /* ensure the new data is on disk */
1067         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1068                 return -1;
1069         }
1070
1071         /*
1072           TODO: maybe write to some dummy hdr field, or write to magic
1073           offset without mmap, before the last sync, instead of the
1074           utime() call
1075         */
1076
1077         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1078            don't change the mtime of the file, this means the file may
1079            not be backed up (as tdb rounding to block sizes means that
1080            file size changes are quite rare too). The following forces
1081            mtime changes when a transaction completes */
1082 #ifdef HAVE_UTIME
1083         utime(tdb->name, NULL);
1084 #endif
1085
1086         need_repack = tdb->transaction->need_repack;
1087
1088         /* use a transaction cancel to free memory and remove the
1089            transaction locks */
1090         _tdb_transaction_cancel(tdb);
1091
1092         if (need_repack) {
1093                 return tdb_repack(tdb);
1094         }
1095
1096         return 0;
1097 }
1098
1099
1100 /*
1101   recover from an aborted transaction. Must be called with exclusive
1102   database write access already established (including the open
1103   lock to prevent new processes attaching)
1104 */
1105 int tdb_transaction_recover(struct tdb_context *tdb)
1106 {
1107         tdb_off_t recovery_head, recovery_eof;
1108         unsigned char *data, *p;
1109         uint32_t zero = 0;
1110         struct tdb_record rec;
1111
1112         /* find the recovery area */
1113         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1114                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1115                 tdb->ecode = TDB_ERR_IO;
1116                 return -1;
1117         }
1118
1119         if (recovery_head == 0) {
1120                 /* we have never allocated a recovery record */
1121                 return 0;
1122         }
1123
1124         /* read the recovery record */
1125         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1126                                    sizeof(rec), DOCONV()) == -1) {
1127                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1128                 tdb->ecode = TDB_ERR_IO;
1129                 return -1;
1130         }
1131
1132         if (rec.magic != TDB_RECOVERY_MAGIC) {
1133                 /* there is no valid recovery data */
1134                 return 0;
1135         }
1136
1137         if (tdb->read_only) {
1138                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1139                 tdb->ecode = TDB_ERR_CORRUPT;
1140                 return -1;
1141         }
1142
1143         recovery_eof = rec.key_len;
1144
1145         data = (unsigned char *)malloc(rec.data_len);
1146         if (data == NULL) {
1147                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1148                 tdb->ecode = TDB_ERR_OOM;
1149                 return -1;
1150         }
1151
1152         /* read the full recovery data */
1153         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1154                                    rec.data_len, 0) == -1) {
1155                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1156                 tdb->ecode = TDB_ERR_IO;
1157                 return -1;
1158         }
1159
1160         /* recover the file data */
1161         p = data;
1162         while (p+8 < data + rec.data_len) {
1163                 uint32_t ofs, len;
1164                 if (DOCONV()) {
1165                         tdb_convert(p, 8);
1166                 }
1167                 memcpy(&ofs, p, 4);
1168                 memcpy(&len, p+4, 4);
1169
1170                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1171                         free(data);
1172                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1173                         tdb->ecode = TDB_ERR_IO;
1174                         return -1;
1175                 }
1176                 p += 8 + len;
1177         }
1178
1179         free(data);
1180
1181         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1182                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1183                 tdb->ecode = TDB_ERR_IO;
1184                 return -1;
1185         }
1186
1187         /* if the recovery area is after the recovered eof then remove it */
1188         if (recovery_eof <= recovery_head) {
1189                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1190                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1191                         tdb->ecode = TDB_ERR_IO;
1192                         return -1;                      
1193                 }
1194         }
1195
1196         /* remove the recovery magic */
1197         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1198                           &zero) == -1) {
1199                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1200                 tdb->ecode = TDB_ERR_IO;
1201                 return -1;                      
1202         }
1203         
1204         /* reduce the file size to the old size */
1205         tdb_munmap(tdb);
1206         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1207                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1208                 tdb->ecode = TDB_ERR_IO;
1209                 return -1;                      
1210         }
1211         tdb->map_size = recovery_eof;
1212         tdb_mmap(tdb);
1213
1214         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1215                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1216                 tdb->ecode = TDB_ERR_IO;
1217                 return -1;
1218         }
1219
1220         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1221                  recovery_eof));
1222
1223         /* all done */
1224         return 0;
1225 }