02ef72a3574ae74939e2b2117799527495483a9d
[ira/wip.git] / lib / tdb / common / transaction.c
1  /* 
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              2005
7
8      ** NOTE! The following LGPL license applies to the tdb
9      ** library. This does NOT imply that all of Samba is released
10      ** under the LGPL
11    
12    This library is free software; you can redistribute it and/or
13    modify it under the terms of the GNU Lesser General Public
14    License as published by the Free Software Foundation; either
15    version 3 of the License, or (at your option) any later version.
16
17    This library is distributed in the hope that it will be useful,
18    but WITHOUT ANY WARRANTY; without even the implied warranty of
19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
20    Lesser General Public License for more details.
21
22    You should have received a copy of the GNU Lesser General Public
23    License along with this library; if not, see <http://www.gnu.org/licenses/>.
24 */
25
26 #include "tdb_private.h"
27
28 /*
29   transaction design:
30
31   - only allow a single transaction at a time per database. This makes
32     using the transaction API simpler, as otherwise the caller would
33     have to cope with temporary failures in transactions that conflict
34     with other current transactions
35
36   - keep the transaction recovery information in the same file as the
37     database, using a special 'transaction recovery' record pointed at
38     by the header. This removes the need for extra journal files as
39     used by some other databases
40
41   - dynamically allocated the transaction recover record, re-using it
42     for subsequent transactions. If a larger record is needed then
43     tdb_free() the old record to place it on the normal tdb freelist
44     before allocating the new record
45
46   - during transactions, keep a linked list of writes all that have
47     been performed by intercepting all tdb_write() calls. The hooked
48     transaction versions of tdb_read() and tdb_write() check this
49     linked list and try to use the elements of the list in preference
50     to the real database.
51
52   - don't allow any locks to be held when a transaction starts,
53     otherwise we can end up with deadlock (plus lack of lock nesting
54     in posix locks would mean the lock is lost)
55
56   - if the caller gains a lock during the transaction but doesn't
57     release it then fail the commit
58
59   - allow for nested calls to tdb_transaction_start(), re-using the
60     existing transaction record. If the inner transaction is cancelled
61     then a subsequent commit will fail
62  
63   - keep a mirrored copy of the tdb hash chain heads to allow for the
64     fast hash heads scan on traverse, updating the mirrored copy in
65     the transaction version of tdb_write
66
67   - allow callers to mix transaction and non-transaction use of tdb,
68     although once a transaction is started then an exclusive lock is
69     gained until the transaction is committed or cancelled
70
71   - the commit stategy involves first saving away all modified data
72     into a linearised buffer in the transaction recovery area, then
73     marking the transaction recovery area with a magic value to
74     indicate a valid recovery record. In total 4 fsync/msync calls are
75     needed per commit to prevent race conditions. It might be possible
76     to reduce this to 3 or even 2 with some more work.
77
78   - check for a valid recovery record on open of the tdb, while the
79     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         /* set when the OPEN_LOCK has been taken */
139         bool open_lock_taken;
140
141         /* old file size before transaction */
142         tdb_len_t old_map_size;
143
144         /* we should re-pack on commit */
145         bool need_repack;
146 };
147
148
149 /*
150   read while in a transaction. We need to check first if the data is in our list
151   of transaction elements, then if not do a real read
152 */
153 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
154                             tdb_len_t len, int cv)
155 {
156         uint32_t blk;
157
158         /* break it down into block sized ops */
159         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
160                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
161                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
162                         return -1;
163                 }
164                 len -= len2;
165                 off += len2;
166                 buf = (void *)(len2 + (char *)buf);
167         }
168
169         if (len == 0) {
170                 return 0;
171         }
172
173         blk = off / tdb->transaction->block_size;
174
175         /* see if we have it in the block list */
176         if (tdb->transaction->num_blocks <= blk ||
177             tdb->transaction->blocks[blk] == NULL) {
178                 /* nope, do a real read */
179                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
180                         goto fail;
181                 }
182                 return 0;
183         }
184
185         /* it is in the block list. Now check for the last block */
186         if (blk == tdb->transaction->num_blocks-1) {
187                 if (len > tdb->transaction->last_block_size) {
188                         goto fail;
189                 }
190         }
191         
192         /* now copy it out of this block */
193         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
194         if (cv) {
195                 tdb_convert(buf, len);
196         }
197         return 0;
198
199 fail:
200         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
201         tdb->ecode = TDB_ERR_IO;
202         tdb->transaction->transaction_error = 1;
203         return -1;
204 }
205
206
207 /*
208   write while in a transaction
209 */
210 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
211                              const void *buf, tdb_len_t len)
212 {
213         uint32_t blk;
214
215         /* Only a commit is allowed on a prepared transaction */
216         if (tdb->transaction->prepared) {
217                 tdb->ecode = TDB_ERR_EINVAL;
218                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
219                 tdb->transaction->transaction_error = 1;
220                 return -1;
221         }
222
223         /* if the write is to a hash head, then update the transaction
224            hash heads */
225         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
226             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
227                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
228                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
229         }
230
231         /* break it up into block sized chunks */
232         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
233                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
234                 if (transaction_write(tdb, off, buf, len2) != 0) {
235                         return -1;
236                 }
237                 len -= len2;
238                 off += len2;
239                 if (buf != NULL) {
240                         buf = (const void *)(len2 + (const char *)buf);
241                 }
242         }
243
244         if (len == 0) {
245                 return 0;
246         }
247
248         blk = off / tdb->transaction->block_size;
249         off = off % tdb->transaction->block_size;
250
251         if (tdb->transaction->num_blocks <= blk) {
252                 uint8_t **new_blocks;
253                 /* expand the blocks array */
254                 if (tdb->transaction->blocks == NULL) {
255                         new_blocks = (uint8_t **)malloc(
256                                 (blk+1)*sizeof(uint8_t *));
257                 } else {
258                         new_blocks = (uint8_t **)realloc(
259                                 tdb->transaction->blocks,
260                                 (blk+1)*sizeof(uint8_t *));
261                 }
262                 if (new_blocks == NULL) {
263                         tdb->ecode = TDB_ERR_OOM;
264                         goto fail;
265                 }
266                 memset(&new_blocks[tdb->transaction->num_blocks], 0, 
267                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
268                 tdb->transaction->blocks = new_blocks;
269                 tdb->transaction->num_blocks = blk+1;
270                 tdb->transaction->last_block_size = 0;
271         }
272
273         /* allocate and fill a block? */
274         if (tdb->transaction->blocks[blk] == NULL) {
275                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
276                 if (tdb->transaction->blocks[blk] == NULL) {
277                         tdb->ecode = TDB_ERR_OOM;
278                         tdb->transaction->transaction_error = 1;
279                         return -1;                      
280                 }
281                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
282                         tdb_len_t len2 = tdb->transaction->block_size;
283                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
284                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
285                         }
286                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
287                                                                    tdb->transaction->blocks[blk], 
288                                                                    len2, 0) != 0) {
289                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
290                                 tdb->ecode = TDB_ERR_IO;
291                                 goto fail;
292                         }
293                         if (blk == tdb->transaction->num_blocks-1) {
294                                 tdb->transaction->last_block_size = len2;
295                         }                       
296                 }
297         }
298         
299         /* overwrite part of an existing block */
300         if (buf == NULL) {
301                 memset(tdb->transaction->blocks[blk] + off, 0, len);
302         } else {
303                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
304         }
305         if (blk == tdb->transaction->num_blocks-1) {
306                 if (len + off > tdb->transaction->last_block_size) {
307                         tdb->transaction->last_block_size = len + off;
308                 }
309         }
310
311         return 0;
312
313 fail:
314         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", 
315                  (blk*tdb->transaction->block_size) + off, len));
316         tdb->transaction->transaction_error = 1;
317         return -1;
318 }
319
320
321 /*
322   write while in a transaction - this varient never expands the transaction blocks, it only
323   updates existing blocks. This means it cannot change the recovery size
324 */
325 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
326                                       const void *buf, tdb_len_t len)
327 {
328         uint32_t blk;
329
330         /* break it up into block sized chunks */
331         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
332                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
333                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
334                         return -1;
335                 }
336                 len -= len2;
337                 off += len2;
338                 if (buf != NULL) {
339                         buf = (const void *)(len2 + (const char *)buf);
340                 }
341         }
342
343         if (len == 0) {
344                 return 0;
345         }
346
347         blk = off / tdb->transaction->block_size;
348         off = off % tdb->transaction->block_size;
349
350         if (tdb->transaction->num_blocks <= blk ||
351             tdb->transaction->blocks[blk] == NULL) {
352                 return 0;
353         }
354
355         if (blk == tdb->transaction->num_blocks-1 &&
356             off + len > tdb->transaction->last_block_size) {
357                 if (off >= tdb->transaction->last_block_size) {
358                         return 0;
359                 }
360                 len = tdb->transaction->last_block_size - off;
361         }
362
363         /* overwrite part of an existing block */
364         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
365
366         return 0;
367 }
368
369
370 /*
371   accelerated hash chain head search, using the cached hash heads
372 */
373 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
374 {
375         uint32_t h = *chain;
376         for (;h < tdb->header.hash_size;h++) {
377                 /* the +1 takes account of the freelist */
378                 if (0 != tdb->transaction->hash_heads[h+1]) {
379                         break;
380                 }
381         }
382         (*chain) = h;
383 }
384
385 /*
386   out of bounds check during a transaction
387 */
388 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
389 {
390         if (len <= tdb->map_size) {
391                 return 0;
392         }
393         tdb->ecode = TDB_ERR_IO;
394         return -1;
395 }
396
397 /*
398   transaction version of tdb_expand().
399 */
400 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
401                                    tdb_off_t addition)
402 {
403         /* add a write to the transaction elements, so subsequent
404            reads see the zero data */
405         if (transaction_write(tdb, size, NULL, addition) != 0) {
406                 return -1;
407         }
408
409         tdb->transaction->need_repack = true;
410
411         return 0;
412 }
413
414 /*
415   brlock during a transaction - ignore them
416 */
417 static int transaction_brlock(struct tdb_context *tdb,
418                               int rw_type, tdb_off_t offset, size_t len,
419                               enum tdb_lock_flags flags)
420 {
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_brlock(tdb, F_RDLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT) == -1) {
506                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
507                 tdb->ecode = TDB_ERR_LOCK;
508                 goto fail;
509         }
510
511         /* setup a copy of the hash table heads so the hash scan in
512            traverse can be fast */
513         tdb->transaction->hash_heads = (uint32_t *)
514                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
515         if (tdb->transaction->hash_heads == NULL) {
516                 tdb->ecode = TDB_ERR_OOM;
517                 goto fail;
518         }
519         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
520                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
521                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
522                 tdb->ecode = TDB_ERR_IO;
523                 goto fail;
524         }
525
526         /* make sure we know about any file expansions already done by
527            anyone else */
528         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
529         tdb->transaction->old_map_size = tdb->map_size;
530
531         /* finally hook the io methods, replacing them with
532            transaction specific methods */
533         tdb->transaction->io_methods = tdb->methods;
534         tdb->methods = &transaction_methods;
535
536         /* Trace at the end, so we get sequence number correct. */
537         tdb_trace(tdb, "tdb_transaction_start");
538         return 0;
539         
540 fail:
541         tdb_brunlock(tdb, F_RDLCK, FREELIST_TOP, 0);
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 /* ltype is F_WRLCK after prepare. */
581 static int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype)
582 {       
583         int i, ret = 0;
584
585         if (tdb->transaction == NULL) {
586                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
587                 return -1;
588         }
589
590         if (tdb->transaction->nesting != 0) {
591                 tdb->transaction->transaction_error = 1;
592                 tdb->transaction->nesting--;
593                 return 0;
594         }               
595
596         tdb->map_size = tdb->transaction->old_map_size;
597
598         /* free all the transaction blocks */
599         for (i=0;i<tdb->transaction->num_blocks;i++) {
600                 if (tdb->transaction->blocks[i] != NULL) {
601                         free(tdb->transaction->blocks[i]);
602                 }
603         }
604         SAFE_FREE(tdb->transaction->blocks);
605
606         if (tdb->transaction->magic_offset) {
607                 const struct tdb_methods *methods = tdb->transaction->io_methods;
608                 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
609
610                 /* remove the recovery marker */
611                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
612                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
613                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
614                         ret = -1;
615                 }
616         }
617
618         if (tdb->transaction->open_lock_taken) {
619                 tdb_brunlock(tdb, F_WRLCK, OPEN_LOCK, 1);
620                 tdb->transaction->open_lock_taken = false;
621         }
622
623         tdb_release_extra_locks(tdb);
624
625         /* restore the normal io methods */
626         tdb->methods = tdb->transaction->io_methods;
627
628         tdb_brunlock(tdb, ltype, FREELIST_TOP, 0);
629         tdb_transaction_unlock(tdb, F_WRLCK);
630         SAFE_FREE(tdb->transaction->hash_heads);
631         SAFE_FREE(tdb->transaction);
632         
633         return ret;
634 }
635
636 /*
637   cancel the current transaction
638 */
639 int tdb_transaction_cancel(struct tdb_context *tdb)
640 {
641         int ltype = F_RDLCK;
642         tdb_trace(tdb, "tdb_transaction_cancel");
643         if (tdb->transaction && tdb->transaction->prepared)
644                 ltype = F_WRLCK;
645         return _tdb_transaction_cancel(tdb, ltype);
646 }
647
648 /*
649   work out how much space the linearised recovery data will consume
650 */
651 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
652 {
653         tdb_len_t recovery_size = 0;
654         int i;
655
656         recovery_size = sizeof(uint32_t);
657         for (i=0;i<tdb->transaction->num_blocks;i++) {
658                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
659                         break;
660                 }
661                 if (tdb->transaction->blocks[i] == NULL) {
662                         continue;
663                 }
664                 recovery_size += 2*sizeof(tdb_off_t);
665                 if (i == tdb->transaction->num_blocks-1) {
666                         recovery_size += tdb->transaction->last_block_size;
667                 } else {
668                         recovery_size += tdb->transaction->block_size;
669                 }
670         }       
671
672         return recovery_size;
673 }
674
675 /*
676   allocate the recovery area, or use an existing recovery area if it is
677   large enough
678 */
679 static int tdb_recovery_allocate(struct tdb_context *tdb, 
680                                  tdb_len_t *recovery_size,
681                                  tdb_off_t *recovery_offset,
682                                  tdb_len_t *recovery_max_size)
683 {
684         struct tdb_record rec;
685         const struct tdb_methods *methods = tdb->transaction->io_methods;
686         tdb_off_t recovery_head;
687
688         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
689                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
690                 return -1;
691         }
692
693         rec.rec_len = 0;
694
695         if (recovery_head != 0) {
696                 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
697                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
698                         return -1;
699                 }
700                 /* ignore invalid recovery regions: can happen in crash */
701                 if (rec.magic != TDB_RECOVERY_MAGIC &&
702                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
703                         recovery_head = 0;
704                 }
705         }
706
707         *recovery_size = tdb_recovery_size(tdb);
708
709         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
710                 /* it fits in the existing area */
711                 *recovery_max_size = rec.rec_len;
712                 *recovery_offset = recovery_head;
713                 return 0;
714         }
715
716         /* we need to free up the old recovery area, then allocate a
717            new one at the end of the file. Note that we cannot use
718            tdb_allocate() to allocate the new one as that might return
719            us an area that is being currently used (as of the start of
720            the transaction) */
721         if (recovery_head != 0) {
722                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
723                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
724                         return -1;
725                 }
726         }
727
728         /* the tdb_free() call might have increased the recovery size */
729         *recovery_size = tdb_recovery_size(tdb);
730
731         /* round up to a multiple of page size */
732         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
733         *recovery_offset = tdb->map_size;
734         recovery_head = *recovery_offset;
735
736         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
737                                      (tdb->map_size - tdb->transaction->old_map_size) +
738                                      sizeof(rec) + *recovery_max_size) == -1) {
739                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
740                 return -1;
741         }
742
743         /* remap the file (if using mmap) */
744         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
745
746         /* we have to reset the old map size so that we don't try to expand the file
747            again in the transaction commit, which would destroy the recovery area */
748         tdb->transaction->old_map_size = tdb->map_size;
749
750         /* write the recovery header offset and sync - we can sync without a race here
751            as the magic ptr in the recovery record has not been set */
752         CONVERT(recovery_head);
753         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
754                                &recovery_head, sizeof(tdb_off_t)) == -1) {
755                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
756                 return -1;
757         }
758         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
759                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
760                 return -1;
761         }
762
763         return 0;
764 }
765
766
767 /*
768   setup the recovery data that will be used on a crash during commit
769 */
770 static int transaction_setup_recovery(struct tdb_context *tdb, 
771                                       tdb_off_t *magic_offset)
772 {
773         tdb_len_t recovery_size;
774         unsigned char *data, *p;
775         const struct tdb_methods *methods = tdb->transaction->io_methods;
776         struct tdb_record *rec;
777         tdb_off_t recovery_offset, recovery_max_size;
778         tdb_off_t old_map_size = tdb->transaction->old_map_size;
779         uint32_t magic, tailer;
780         int i;
781
782         /*
783           check that the recovery area has enough space
784         */
785         if (tdb_recovery_allocate(tdb, &recovery_size, 
786                                   &recovery_offset, &recovery_max_size) == -1) {
787                 return -1;
788         }
789
790         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
791         if (data == NULL) {
792                 tdb->ecode = TDB_ERR_OOM;
793                 return -1;
794         }
795
796         rec = (struct tdb_record *)data;
797         memset(rec, 0, sizeof(*rec));
798
799         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
800         rec->data_len = recovery_size;
801         rec->rec_len  = recovery_max_size;
802         rec->key_len  = old_map_size;
803         CONVERT(rec);
804
805         /* build the recovery data into a single blob to allow us to do a single
806            large write, which should be more efficient */
807         p = data + sizeof(*rec);
808         for (i=0;i<tdb->transaction->num_blocks;i++) {
809                 tdb_off_t offset;
810                 tdb_len_t length;
811
812                 if (tdb->transaction->blocks[i] == NULL) {
813                         continue;
814                 }
815
816                 offset = i * tdb->transaction->block_size;
817                 length = tdb->transaction->block_size;
818                 if (i == tdb->transaction->num_blocks-1) {
819                         length = tdb->transaction->last_block_size;
820                 }
821                 
822                 if (offset >= old_map_size) {
823                         continue;
824                 }
825                 if (offset + length > tdb->transaction->old_map_size) {
826                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
827                         free(data);
828                         tdb->ecode = TDB_ERR_CORRUPT;
829                         return -1;
830                 }
831                 memcpy(p, &offset, 4);
832                 memcpy(p+4, &length, 4);
833                 if (DOCONV()) {
834                         tdb_convert(p, 8);
835                 }
836                 /* the recovery area contains the old data, not the
837                    new data, so we have to call the original tdb_read
838                    method to get it */
839                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
840                         free(data);
841                         tdb->ecode = TDB_ERR_IO;
842                         return -1;
843                 }
844                 p += 8 + length;
845         }
846
847         /* and the tailer */
848         tailer = sizeof(*rec) + recovery_max_size;
849         memcpy(p, &tailer, 4);
850         CONVERT(p);
851
852         /* write the recovery data to the recovery area */
853         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
854                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
855                 free(data);
856                 tdb->ecode = TDB_ERR_IO;
857                 return -1;
858         }
859         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
860                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
861                 free(data);
862                 tdb->ecode = TDB_ERR_IO;
863                 return -1;
864         }
865
866         /* as we don't have ordered writes, we have to sync the recovery
867            data before we update the magic to indicate that the recovery
868            data is present */
869         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
870                 free(data);
871                 return -1;
872         }
873
874         free(data);
875
876         magic = TDB_RECOVERY_MAGIC;
877         CONVERT(magic);
878
879         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
880
881         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
882                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
883                 tdb->ecode = TDB_ERR_IO;
884                 return -1;
885         }
886         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
887                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
888                 tdb->ecode = TDB_ERR_IO;
889                 return -1;
890         }
891
892         /* ensure the recovery magic marker is on disk */
893         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
894                 return -1;
895         }
896
897         return 0;
898 }
899
900 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
901 {       
902         const struct tdb_methods *methods;
903
904         if (tdb->transaction == NULL) {
905                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
906                 return -1;
907         }
908
909         if (tdb->transaction->prepared) {
910                 tdb->ecode = TDB_ERR_EINVAL;
911                 _tdb_transaction_cancel(tdb, F_WRLCK);
912                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
913                 return -1;
914         }
915
916         if (tdb->transaction->transaction_error) {
917                 tdb->ecode = TDB_ERR_IO;
918                 _tdb_transaction_cancel(tdb, F_RDLCK);
919                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
920                 return -1;
921         }
922
923
924         if (tdb->transaction->nesting != 0) {
925                 return 0;
926         }               
927
928         /* check for a null transaction */
929         if (tdb->transaction->blocks == NULL) {
930                 return 0;
931         }
932
933         methods = tdb->transaction->io_methods;
934         
935         /* if there are any locks pending then the caller has not
936            nested their locks properly, so fail the transaction */
937         if (tdb_have_extra_locks(tdb)) {
938                 tdb->ecode = TDB_ERR_LOCK;
939                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
940                 _tdb_transaction_cancel(tdb, F_RDLCK);
941                 return -1;
942         }
943
944         /* upgrade the main transaction lock region to a write lock */
945         if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
946                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
947                 tdb->ecode = TDB_ERR_LOCK;
948                 _tdb_transaction_cancel(tdb, F_RDLCK);
949                 return -1;
950         }
951
952         /* get the open lock - this prevents new users attaching to the database
953            during the commit */
954         if (tdb_brlock(tdb, F_WRLCK, OPEN_LOCK, 1, TDB_LOCK_WAIT) == -1) {
955                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
956                 tdb->ecode = TDB_ERR_LOCK;
957                 _tdb_transaction_cancel(tdb, F_WRLCK);
958                 return -1;
959         }
960
961         tdb->transaction->open_lock_taken = true;
962
963         if (!(tdb->flags & TDB_NOSYNC)) {
964                 /* write the recovery data to the end of the file */
965                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
966                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
967                         _tdb_transaction_cancel(tdb, F_WRLCK);
968                         return -1;
969                 }
970         }
971
972         tdb->transaction->prepared = true;
973
974         /* expand the file to the new size if needed */
975         if (tdb->map_size != tdb->transaction->old_map_size) {
976                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
977                                              tdb->map_size - 
978                                              tdb->transaction->old_map_size) == -1) {
979                         tdb->ecode = TDB_ERR_IO;
980                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
981                         _tdb_transaction_cancel(tdb, F_WRLCK);
982                         return -1;
983                 }
984                 tdb->map_size = tdb->transaction->old_map_size;
985                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
986         }
987
988         /* Keep the open lock until the actual commit */
989
990         return 0;
991 }
992
993 /*
994    prepare to commit the current transaction
995 */
996 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
997 {       
998         tdb_trace(tdb, "tdb_transaction_prepare_commit");
999         return _tdb_transaction_prepare_commit(tdb);
1000 }
1001
1002 /*
1003   commit the current transaction
1004 */
1005 int tdb_transaction_commit(struct tdb_context *tdb)
1006 {       
1007         const struct tdb_methods *methods;
1008         int i;
1009         bool need_repack;
1010
1011         if (tdb->transaction == NULL) {
1012                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1013                 return -1;
1014         }
1015
1016         tdb_trace(tdb, "tdb_transaction_commit");
1017
1018         if (tdb->transaction->transaction_error) {
1019                 tdb->ecode = TDB_ERR_IO;
1020                 _tdb_transaction_cancel(tdb, F_RDLCK);
1021                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1022                 return -1;
1023         }
1024
1025
1026         if (tdb->transaction->nesting != 0) {
1027                 tdb->transaction->nesting--;
1028                 return 0;
1029         }
1030
1031         /* check for a null transaction */
1032         if (tdb->transaction->blocks == NULL) {
1033                 _tdb_transaction_cancel(tdb, F_RDLCK);
1034                 return 0;
1035         }
1036
1037         if (!tdb->transaction->prepared) {
1038                 int ret = _tdb_transaction_prepare_commit(tdb);
1039                 if (ret)
1040                         return ret;
1041         }
1042
1043         methods = tdb->transaction->io_methods;
1044
1045         /* perform all the writes */
1046         for (i=0;i<tdb->transaction->num_blocks;i++) {
1047                 tdb_off_t offset;
1048                 tdb_len_t length;
1049
1050                 if (tdb->transaction->blocks[i] == NULL) {
1051                         continue;
1052                 }
1053
1054                 offset = i * tdb->transaction->block_size;
1055                 length = tdb->transaction->block_size;
1056                 if (i == tdb->transaction->num_blocks-1) {
1057                         length = tdb->transaction->last_block_size;
1058                 }
1059
1060                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1061                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1062                         
1063                         /* we've overwritten part of the data and
1064                            possibly expanded the file, so we need to
1065                            run the crash recovery code */
1066                         tdb->methods = methods;
1067                         tdb_transaction_recover(tdb); 
1068
1069                         _tdb_transaction_cancel(tdb, F_WRLCK);
1070
1071                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1072                         return -1;
1073                 }
1074                 SAFE_FREE(tdb->transaction->blocks[i]);
1075         } 
1076
1077         SAFE_FREE(tdb->transaction->blocks);
1078         tdb->transaction->num_blocks = 0;
1079
1080         /* ensure the new data is on disk */
1081         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1082                 return -1;
1083         }
1084
1085         /*
1086           TODO: maybe write to some dummy hdr field, or write to magic
1087           offset without mmap, before the last sync, instead of the
1088           utime() call
1089         */
1090
1091         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1092            don't change the mtime of the file, this means the file may
1093            not be backed up (as tdb rounding to block sizes means that
1094            file size changes are quite rare too). The following forces
1095            mtime changes when a transaction completes */
1096 #ifdef HAVE_UTIME
1097         utime(tdb->name, NULL);
1098 #endif
1099
1100         need_repack = tdb->transaction->need_repack;
1101
1102         /* use a transaction cancel to free memory and remove the
1103            transaction locks */
1104         _tdb_transaction_cancel(tdb, F_WRLCK);
1105
1106         if (need_repack) {
1107                 return tdb_repack(tdb);
1108         }
1109
1110         return 0;
1111 }
1112
1113
1114 /*
1115   recover from an aborted transaction. Must be called with exclusive
1116   database write access already established (including the open
1117   lock to prevent new processes attaching)
1118 */
1119 int tdb_transaction_recover(struct tdb_context *tdb)
1120 {
1121         tdb_off_t recovery_head, recovery_eof;
1122         unsigned char *data, *p;
1123         uint32_t zero = 0;
1124         struct tdb_record rec;
1125
1126         /* find the recovery area */
1127         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1128                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1129                 tdb->ecode = TDB_ERR_IO;
1130                 return -1;
1131         }
1132
1133         if (recovery_head == 0) {
1134                 /* we have never allocated a recovery record */
1135                 return 0;
1136         }
1137
1138         /* read the recovery record */
1139         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1140                                    sizeof(rec), DOCONV()) == -1) {
1141                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1142                 tdb->ecode = TDB_ERR_IO;
1143                 return -1;
1144         }
1145
1146         if (rec.magic != TDB_RECOVERY_MAGIC) {
1147                 /* there is no valid recovery data */
1148                 return 0;
1149         }
1150
1151         if (tdb->read_only) {
1152                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1153                 tdb->ecode = TDB_ERR_CORRUPT;
1154                 return -1;
1155         }
1156
1157         recovery_eof = rec.key_len;
1158
1159         data = (unsigned char *)malloc(rec.data_len);
1160         if (data == NULL) {
1161                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1162                 tdb->ecode = TDB_ERR_OOM;
1163                 return -1;
1164         }
1165
1166         /* read the full recovery data */
1167         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1168                                    rec.data_len, 0) == -1) {
1169                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1170                 tdb->ecode = TDB_ERR_IO;
1171                 return -1;
1172         }
1173
1174         /* recover the file data */
1175         p = data;
1176         while (p+8 < data + rec.data_len) {
1177                 uint32_t ofs, len;
1178                 if (DOCONV()) {
1179                         tdb_convert(p, 8);
1180                 }
1181                 memcpy(&ofs, p, 4);
1182                 memcpy(&len, p+4, 4);
1183
1184                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1185                         free(data);
1186                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1187                         tdb->ecode = TDB_ERR_IO;
1188                         return -1;
1189                 }
1190                 p += 8 + len;
1191         }
1192
1193         free(data);
1194
1195         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1196                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1197                 tdb->ecode = TDB_ERR_IO;
1198                 return -1;
1199         }
1200
1201         /* if the recovery area is after the recovered eof then remove it */
1202         if (recovery_eof <= recovery_head) {
1203                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1204                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1205                         tdb->ecode = TDB_ERR_IO;
1206                         return -1;                      
1207                 }
1208         }
1209
1210         /* remove the recovery magic */
1211         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1212                           &zero) == -1) {
1213                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1214                 tdb->ecode = TDB_ERR_IO;
1215                 return -1;                      
1216         }
1217         
1218         /* reduce the file size to the old size */
1219         tdb_munmap(tdb);
1220         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1221                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1222                 tdb->ecode = TDB_ERR_IO;
1223                 return -1;                      
1224         }
1225         tdb->map_size = recovery_eof;
1226         tdb_mmap(tdb);
1227
1228         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1229                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1230                 tdb->ecode = TDB_ERR_IO;
1231                 return -1;
1232         }
1233
1234         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1235                  recovery_eof));
1236
1237         /* all done */
1238         return 0;
1239 }