Import 898b5edfe757cb145960b8f3631029bfd5592119 from ctdb:
[ccan] / ccan / tdb / transaction.c
1  /* 
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              2005
7
8      ** NOTE! The following LGPL license applies to the tdb
9      ** library. This does NOT imply that all of Samba is released
10      ** under the LGPL
11    
12    This library is free software; you can redistribute it and/or
13    modify it under the terms of the GNU Lesser General Public
14    License as published by the Free Software Foundation; either
15    version 3 of the License, or (at your option) any later version.
16
17    This library is distributed in the hope that it will be useful,
18    but WITHOUT ANY WARRANTY; without even the implied warranty of
19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
20    Lesser General Public License for more details.
21
22    You should have received a copy of the GNU Lesser General Public
23    License along with this library; if not, see <http://www.gnu.org/licenses/>.
24 */
25
26 #include "tdb_private.h"
27
28 /*
29   transaction design:
30
31   - only allow a single transaction at a time per database. This makes
32     using the transaction API simpler, as otherwise the caller would
33     have to cope with temporary failures in transactions that conflict
34     with other current transactions
35
36   - keep the transaction recovery information in the same file as the
37     database, using a special 'transaction recovery' record pointed at
38     by the header. This removes the need for extra journal files as
39     used by some other databases
40
41   - dynamically allocated the transaction recover record, re-using it
42     for subsequent transactions. If a larger record is needed then
43     tdb_free() the old record to place it on the normal tdb freelist
44     before allocating the new record
45
46   - during transactions, keep a linked list of writes all that have
47     been performed by intercepting all tdb_write() calls. The hooked
48     transaction versions of tdb_read() and tdb_write() check this
49     linked list and try to use the elements of the list in preference
50     to the real database.
51
52   - don't allow any locks to be held when a transaction starts,
53     otherwise we can end up with deadlock (plus lack of lock nesting
54     in posix locks would mean the lock is lost)
55
56   - if the caller gains a lock during the transaction but doesn't
57     release it then fail the commit
58
59   - allow for nested calls to tdb_transaction_start(), re-using the
60     existing transaction record. If the inner transaction is cancelled
61     then a subsequent commit will fail
62  
63   - keep a mirrored copy of the tdb hash chain heads to allow for the
64     fast hash heads scan on traverse, updating the mirrored copy in
65     the transaction version of tdb_write
66
67   - allow callers to mix transaction and non-transaction use of tdb,
68     although once a transaction is started then an exclusive lock is
69     gained until the transaction is committed or cancelled
70
71   - the commit stategy involves first saving away all modified data
72     into a linearised buffer in the transaction recovery area, then
73     marking the transaction recovery area with a magic value to
74     indicate a valid recovery record. In total 4 fsync/msync calls are
75     needed per commit to prevent race conditions. It might be possible
76     to reduce this to 3 or even 2 with some more work.
77
78   - check for a valid recovery record on open of the tdb, while the
79     global lock is held. Automatically recover from the transaction
80     recovery area if needed, then continue with the open as
81     usual. This allows for smooth crash recovery with no administrator
82     intervention.
83
84   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
85     still available, but no transaction recovery area is used and no
86     fsync/msync calls are made.
87
88   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
89     tdb_add_flags() transaction 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 GLOBAL_LOCK has been taken */
139         bool global_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   sync to disk
442 */
443 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
444 {       
445         if (tdb->flags & TDB_NOSYNC) {
446                 return 0;
447         }
448
449         if (fsync(tdb->fd) != 0) {
450                 tdb->ecode = TDB_ERR_IO;
451                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
452                 return -1;
453         }
454 #ifdef MS_SYNC
455         if (tdb->map_ptr) {
456                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
457                 if (msync(moffset + (char *)tdb->map_ptr, 
458                           length + (offset - moffset), MS_SYNC) != 0) {
459                         tdb->ecode = TDB_ERR_IO;
460                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
461                                  strerror(errno)));
462                         return -1;
463                 }
464         }
465 #endif
466         return 0;
467 }
468
469 /* ltype is F_WRLCK after prepare. */
470 int _tdb_transaction_cancel(struct tdb_context *tdb, int ltype)
471 {
472         int i, ret = 0;
473
474         if (tdb->transaction == NULL) {
475                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
476                 return -1;
477         }
478
479         if (tdb->transaction->nesting != 0) {
480                 tdb->transaction->transaction_error = 1;
481                 tdb->transaction->nesting--;
482                 return 0;
483         }               
484
485         tdb->map_size = tdb->transaction->old_map_size;
486
487         /* free all the transaction blocks */
488         for (i=0;i<tdb->transaction->num_blocks;i++) {
489                 if (tdb->transaction->blocks[i] != NULL) {
490                         free(tdb->transaction->blocks[i]);
491                 }
492         }
493         SAFE_FREE(tdb->transaction->blocks);
494
495         if (tdb->transaction->magic_offset) {
496                 const struct tdb_methods *methods = tdb->transaction->io_methods;
497                 uint32_t zero = 0;
498
499                 /* remove the recovery marker */
500                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
501                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
502                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
503                         ret = -1;
504                 }
505         }
506
507         if (tdb->transaction->global_lock_taken) {
508                 tdb_brunlock(tdb, F_WRLCK, GLOBAL_LOCK, 1);
509                 tdb->transaction->global_lock_taken = false;
510         }
511
512         /* remove any global lock created during the transaction */
513         if (tdb->global_lock.count != 0) {
514                 tdb_brunlock(tdb, tdb->global_lock.ltype,
515                              FREELIST_TOP, 4*tdb->header.hash_size);
516                 tdb->global_lock.count = 0;
517         }
518
519         /* remove any locks created during the transaction */
520         if (tdb->num_locks != 0) {
521                 for (i=0;i<tdb->num_lockrecs;i++) {
522                         tdb_brunlock(tdb, tdb->lockrecs[i].ltype,
523                                      FREELIST_TOP+4*tdb->lockrecs[i].list, 1);
524                 }
525                 tdb->num_locks = 0;
526                 tdb->num_lockrecs = 0;
527                 SAFE_FREE(tdb->lockrecs);
528         }
529
530         /* restore the normal io methods */
531         tdb->methods = tdb->transaction->io_methods;
532
533         tdb_brunlock(tdb, ltype, FREELIST_TOP, 0);
534         tdb_transaction_unlock(tdb, F_WRLCK);
535         SAFE_FREE(tdb->transaction->hash_heads);
536         SAFE_FREE(tdb->transaction);
537         
538         return ret;
539 }
540
541 /*
542   start a tdb transaction. No token is returned, as only a single
543   transaction is allowed to be pending per tdb_context
544 */
545 int tdb_transaction_start(struct tdb_context *tdb)
546 {
547         /* some sanity checks */
548         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
549                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
550                 tdb->ecode = TDB_ERR_EINVAL;
551                 return -1;
552         }
553
554         /* cope with nested tdb_transaction_start() calls */
555         if (tdb->transaction != NULL) {
556                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
557                         tdb->ecode = TDB_ERR_NESTING;
558                         return -1;
559                 }
560                 tdb_trace(tdb, "tdb_transaction_start");
561                 tdb->transaction->nesting++;
562                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
563                          tdb->transaction->nesting));
564                 return 0;
565         }
566
567         if (tdb->num_locks != 0 || tdb->global_lock.count) {
568                 /* the caller must not have any locks when starting a
569                    transaction as otherwise we'll be screwed by lack
570                    of nested locks in posix */
571                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
572                 tdb->ecode = TDB_ERR_LOCK;
573                 return -1;
574         }
575
576         if (tdb->travlocks.next != NULL) {
577                 /* you cannot use transactions inside a traverse (although you can use
578                    traverse inside a transaction) as otherwise you can end up with
579                    deadlock */
580                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
581                 tdb->ecode = TDB_ERR_LOCK;
582                 return -1;
583         }
584
585         tdb->transaction = (struct tdb_transaction *)
586                 calloc(sizeof(struct tdb_transaction), 1);
587         if (tdb->transaction == NULL) {
588                 tdb->ecode = TDB_ERR_OOM;
589                 return -1;
590         }
591
592         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
593         tdb->transaction->block_size = tdb->page_size;
594
595         /* get the transaction write lock. This is a blocking lock. As
596            discussed with Volker, there are a number of ways we could
597            make this async, which we will probably do in the future */
598         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
599                 SAFE_FREE(tdb->transaction->blocks);
600                 SAFE_FREE(tdb->transaction);
601                 return -1;
602         }
603         
604         /* get a read lock from the freelist to the end of file. This
605            is upgraded to a write lock during the commit */
606         if (tdb_brlock(tdb, F_RDLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT) == -1) {
607                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
608                 tdb->ecode = TDB_ERR_LOCK;
609                 goto fail;
610         }
611
612         /* setup a copy of the hash table heads so the hash scan in
613            traverse can be fast */
614         tdb->transaction->hash_heads = (uint32_t *)
615                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
616         if (tdb->transaction->hash_heads == NULL) {
617                 tdb->ecode = TDB_ERR_OOM;
618                 goto fail;
619         }
620         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
621                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
622                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
623                 tdb->ecode = TDB_ERR_IO;
624                 goto fail;
625         }
626
627         /* make sure we know about any file expansions already done by
628            anyone else */
629         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
630         tdb->transaction->old_map_size = tdb->map_size;
631
632         /* finally hook the io methods, replacing them with
633            transaction specific methods */
634         tdb->transaction->io_methods = tdb->methods;
635         tdb->methods = &transaction_methods;
636
637         /* Trace at the end, so we get sequence number correct. */
638         tdb_trace(tdb, "tdb_transaction_start");
639         return 0;
640         
641 fail:
642         tdb_brunlock(tdb, F_RDLCK, FREELIST_TOP, 0);
643         tdb_transaction_unlock(tdb, F_WRLCK);
644         SAFE_FREE(tdb->transaction->blocks);
645         SAFE_FREE(tdb->transaction->hash_heads);
646         SAFE_FREE(tdb->transaction);
647         return -1;
648 }
649
650
651 /*
652   cancel the current transaction
653 */
654 int tdb_transaction_cancel(struct tdb_context *tdb)
655 {
656         int ltype = F_RDLCK;
657         tdb_trace(tdb, "tdb_transaction_cancel");
658         if (tdb->transaction && tdb->transaction->prepared)
659                 ltype = F_WRLCK;
660         return _tdb_transaction_cancel(tdb, ltype);
661 }
662
663 /*
664   work out how much space the linearised recovery data will consume
665 */
666 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
667 {
668         tdb_len_t recovery_size = 0;
669         int i;
670
671         recovery_size = sizeof(uint32_t);
672         for (i=0;i<tdb->transaction->num_blocks;i++) {
673                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
674                         break;
675                 }
676                 if (tdb->transaction->blocks[i] == NULL) {
677                         continue;
678                 }
679                 recovery_size += 2*sizeof(tdb_off_t);
680                 if (i == tdb->transaction->num_blocks-1) {
681                         recovery_size += tdb->transaction->last_block_size;
682                 } else {
683                         recovery_size += tdb->transaction->block_size;
684                 }
685         }       
686
687         return recovery_size;
688 }
689
690 /*
691   allocate the recovery area, or use an existing recovery area if it is
692   large enough
693 */
694 static int tdb_recovery_allocate(struct tdb_context *tdb, 
695                                  tdb_len_t *recovery_size,
696                                  tdb_off_t *recovery_offset,
697                                  tdb_len_t *recovery_max_size)
698 {
699         struct tdb_record rec;
700         const struct tdb_methods *methods = tdb->transaction->io_methods;
701         tdb_off_t recovery_head;
702
703         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
704                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
705                 return -1;
706         }
707
708         rec.rec_len = 0;
709
710         if (recovery_head != 0 && 
711             methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
712                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
713                 return -1;
714         }
715
716         *recovery_size = tdb_recovery_size(tdb);
717
718         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
719                 /* it fits in the existing area */
720                 *recovery_max_size = rec.rec_len;
721                 *recovery_offset = recovery_head;
722                 return 0;
723         }
724
725         /* we need to free up the old recovery area, then allocate a
726            new one at the end of the file. Note that we cannot use
727            tdb_allocate() to allocate the new one as that might return
728            us an area that is being currently used (as of the start of
729            the transaction) */
730         if (recovery_head != 0) {
731                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
732                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
733                         return -1;
734                 }
735         }
736
737         /* the tdb_free() call might have increased the recovery size */
738         *recovery_size = tdb_recovery_size(tdb);
739
740         /* round up to a multiple of page size */
741         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
742         *recovery_offset = tdb->map_size;
743         recovery_head = *recovery_offset;
744
745         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
746                                      (tdb->map_size - tdb->transaction->old_map_size) +
747                                      sizeof(rec) + *recovery_max_size) == -1) {
748                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
749                 return -1;
750         }
751
752         /* remap the file (if using mmap) */
753         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
754
755         /* we have to reset the old map size so that we don't try to expand the file
756            again in the transaction commit, which would destroy the recovery area */
757         tdb->transaction->old_map_size = tdb->map_size;
758
759         /* write the recovery header offset and sync - we can sync without a race here
760            as the magic ptr in the recovery record has not been set */
761         CONVERT(recovery_head);
762         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
763                                &recovery_head, sizeof(tdb_off_t)) == -1) {
764                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
765                 return -1;
766         }
767         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
768                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
769                 return -1;
770         }
771
772         return 0;
773 }
774
775
776 /*
777   setup the recovery data that will be used on a crash during commit
778 */
779 static int transaction_setup_recovery(struct tdb_context *tdb, 
780                                       tdb_off_t *magic_offset)
781 {
782         tdb_len_t recovery_size;
783         unsigned char *data, *p;
784         const struct tdb_methods *methods = tdb->transaction->io_methods;
785         struct tdb_record *rec;
786         tdb_off_t recovery_offset, recovery_max_size;
787         tdb_off_t old_map_size = tdb->transaction->old_map_size;
788         uint32_t magic, tailer;
789         int i;
790
791         /*
792           check that the recovery area has enough space
793         */
794         if (tdb_recovery_allocate(tdb, &recovery_size, 
795                                   &recovery_offset, &recovery_max_size) == -1) {
796                 return -1;
797         }
798
799         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
800         if (data == NULL) {
801                 tdb->ecode = TDB_ERR_OOM;
802                 return -1;
803         }
804
805         rec = (struct tdb_record *)data;
806         memset(rec, 0, sizeof(*rec));
807
808         rec->magic    = 0;
809         rec->data_len = recovery_size;
810         rec->rec_len  = recovery_max_size;
811         rec->key_len  = old_map_size;
812         CONVERT(rec);
813
814         /* build the recovery data into a single blob to allow us to do a single
815            large write, which should be more efficient */
816         p = data + sizeof(*rec);
817         for (i=0;i<tdb->transaction->num_blocks;i++) {
818                 tdb_off_t offset;
819                 tdb_len_t length;
820
821                 if (tdb->transaction->blocks[i] == NULL) {
822                         continue;
823                 }
824
825                 offset = i * tdb->transaction->block_size;
826                 length = tdb->transaction->block_size;
827                 if (i == tdb->transaction->num_blocks-1) {
828                         length = tdb->transaction->last_block_size;
829                 }
830                 
831                 if (offset >= old_map_size) {
832                         continue;
833                 }
834                 if (offset + length > tdb->transaction->old_map_size) {
835                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
836                         free(data);
837                         tdb->ecode = TDB_ERR_CORRUPT;
838                         return -1;
839                 }
840                 memcpy(p, &offset, 4);
841                 memcpy(p+4, &length, 4);
842                 if (DOCONV()) {
843                         tdb_convert(p, 8);
844                 }
845                 /* the recovery area contains the old data, not the
846                    new data, so we have to call the original tdb_read
847                    method to get it */
848                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
849                         free(data);
850                         tdb->ecode = TDB_ERR_IO;
851                         return -1;
852                 }
853                 p += 8 + length;
854         }
855
856         /* and the tailer */
857         tailer = sizeof(*rec) + recovery_max_size;
858         memcpy(p, &tailer, 4);
859         CONVERT(p);
860
861         /* write the recovery data to the recovery area */
862         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
863                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
864                 free(data);
865                 tdb->ecode = TDB_ERR_IO;
866                 return -1;
867         }
868         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
869                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
870                 free(data);
871                 tdb->ecode = TDB_ERR_IO;
872                 return -1;
873         }
874
875         /* as we don't have ordered writes, we have to sync the recovery
876            data before we update the magic to indicate that the recovery
877            data is present */
878         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
879                 free(data);
880                 return -1;
881         }
882
883         free(data);
884
885         magic = TDB_RECOVERY_MAGIC;
886         CONVERT(magic);
887
888         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
889
890         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
891                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
892                 tdb->ecode = TDB_ERR_IO;
893                 return -1;
894         }
895         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
896                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
897                 tdb->ecode = TDB_ERR_IO;
898                 return -1;
899         }
900
901         /* ensure the recovery magic marker is on disk */
902         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
903                 return -1;
904         }
905
906         return 0;
907 }
908
909 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
910 {       
911         const struct tdb_methods *methods;
912
913         if (tdb->transaction == NULL) {
914                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
915                 return -1;
916         }
917
918         if (tdb->transaction->prepared) {
919                 tdb->ecode = TDB_ERR_EINVAL;
920                 _tdb_transaction_cancel(tdb, F_WRLCK);
921                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
922                 return -1;
923         }
924
925         if (tdb->transaction->transaction_error) {
926                 tdb->ecode = TDB_ERR_IO;
927                 _tdb_transaction_cancel(tdb, F_RDLCK);
928                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
929                 return -1;
930         }
931
932
933         if (tdb->transaction->nesting != 0) {
934                 tdb->transaction->nesting--;
935                 return 0;
936         }               
937
938         /* check for a null transaction */
939         if (tdb->transaction->blocks == NULL) {
940                 return 0;
941         }
942
943         methods = tdb->transaction->io_methods;
944         
945         /* if there are any locks pending then the caller has not
946            nested their locks properly, so fail the transaction */
947         if (tdb->num_locks || tdb->global_lock.count) {
948                 tdb->ecode = TDB_ERR_LOCK;
949                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
950                 _tdb_transaction_cancel(tdb, F_RDLCK);
951                 return -1;
952         }
953
954         /* upgrade the main transaction lock region to a write lock */
955         if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
956                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
957                 tdb->ecode = TDB_ERR_LOCK;
958                 _tdb_transaction_cancel(tdb, F_RDLCK);
959                 return -1;
960         }
961
962         /* get the global lock - this prevents new users attaching to the database
963            during the commit */
964         if (tdb_brlock(tdb, F_WRLCK, GLOBAL_LOCK, 1, TDB_LOCK_WAIT) == -1) {
965                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
966                 tdb->ecode = TDB_ERR_LOCK;
967                 _tdb_transaction_cancel(tdb, F_WRLCK);
968                 return -1;
969         }
970
971         tdb->transaction->global_lock_taken = true;
972
973         if (!(tdb->flags & TDB_NOSYNC)) {
974                 /* write the recovery data to the end of the file */
975                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
976                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
977                         _tdb_transaction_cancel(tdb, F_WRLCK);
978                         return -1;
979                 }
980         }
981
982         tdb->transaction->prepared = true;
983
984         /* expand the file to the new size if needed */
985         if (tdb->map_size != tdb->transaction->old_map_size) {
986                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
987                                              tdb->map_size - 
988                                              tdb->transaction->old_map_size) == -1) {
989                         tdb->ecode = TDB_ERR_IO;
990                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
991                         _tdb_transaction_cancel(tdb, F_WRLCK);
992                         return -1;
993                 }
994                 tdb->map_size = tdb->transaction->old_map_size;
995                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
996         }
997
998         /* Keep the global lock until the actual commit */
999
1000         return 0;
1001 }
1002
1003 /*
1004    prepare to commit the current transaction
1005 */
1006 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1007 {       
1008         tdb_trace(tdb, "tdb_transaction_prepare_commit");
1009         return _tdb_transaction_prepare_commit(tdb);
1010 }
1011
1012 /*
1013   commit the current transaction
1014 */
1015 int tdb_transaction_commit(struct tdb_context *tdb)
1016 {       
1017         const struct tdb_methods *methods;
1018         int i;
1019         bool need_repack;
1020
1021         if (tdb->transaction == NULL) {
1022                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1023                 return -1;
1024         }
1025
1026         tdb_trace(tdb, "tdb_transaction_commit");
1027
1028         if (tdb->transaction->transaction_error) {
1029                 tdb->ecode = TDB_ERR_IO;
1030                 tdb_transaction_cancel(tdb);
1031                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1032                 return -1;
1033         }
1034
1035
1036         if (tdb->transaction->nesting != 0) {
1037                 tdb->transaction->nesting--;
1038                 return 0;
1039         }
1040
1041         /* check for a null transaction */
1042         if (tdb->transaction->blocks == NULL) {
1043                 _tdb_transaction_cancel(tdb, F_RDLCK);
1044                 return 0;
1045         }
1046
1047         if (!tdb->transaction->prepared) {
1048                 int ret = _tdb_transaction_prepare_commit(tdb);
1049                 if (ret)
1050                         return ret;
1051         }
1052
1053         methods = tdb->transaction->io_methods;
1054
1055         /* perform all the writes */
1056         for (i=0;i<tdb->transaction->num_blocks;i++) {
1057                 tdb_off_t offset;
1058                 tdb_len_t length;
1059
1060                 if (tdb->transaction->blocks[i] == NULL) {
1061                         continue;
1062                 }
1063
1064                 offset = i * tdb->transaction->block_size;
1065                 length = tdb->transaction->block_size;
1066                 if (i == tdb->transaction->num_blocks-1) {
1067                         length = tdb->transaction->last_block_size;
1068                 }
1069
1070                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1071                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1072                         
1073                         /* we've overwritten part of the data and
1074                            possibly expanded the file, so we need to
1075                            run the crash recovery code */
1076                         tdb->methods = methods;
1077                         tdb_transaction_recover(tdb); 
1078
1079                         _tdb_transaction_cancel(tdb, F_WRLCK);
1080
1081                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1082                         return -1;
1083                 }
1084                 SAFE_FREE(tdb->transaction->blocks[i]);
1085         } 
1086
1087         SAFE_FREE(tdb->transaction->blocks);
1088         tdb->transaction->num_blocks = 0;
1089
1090         /* ensure the new data is on disk */
1091         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1092                 return -1;
1093         }
1094
1095         /*
1096           TODO: maybe write to some dummy hdr field, or write to magic
1097           offset without mmap, before the last sync, instead of the
1098           utime() call
1099         */
1100
1101         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1102            don't change the mtime of the file, this means the file may
1103            not be backed up (as tdb rounding to block sizes means that
1104            file size changes are quite rare too). The following forces
1105            mtime changes when a transaction completes */
1106 #if HAVE_UTIME
1107         utime(tdb->name, NULL);
1108 #endif
1109
1110         need_repack = tdb->transaction->need_repack;
1111
1112         /* use a transaction cancel to free memory and remove the
1113            transaction locks */
1114         _tdb_transaction_cancel(tdb, F_WRLCK);
1115
1116         if (need_repack) {
1117                 return tdb_repack(tdb);
1118         }
1119
1120         return 0;
1121 }
1122
1123
1124 /*
1125   recover from an aborted transaction. Must be called with exclusive
1126   database write access already established (including the global
1127   lock to prevent new processes attaching)
1128 */
1129 int tdb_transaction_recover(struct tdb_context *tdb)
1130 {
1131         tdb_off_t recovery_head, recovery_eof;
1132         unsigned char *data, *p;
1133         uint32_t zero = 0;
1134         struct tdb_record rec;
1135
1136         /* find the recovery area */
1137         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1138                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1139                 tdb->ecode = TDB_ERR_IO;
1140                 return -1;
1141         }
1142
1143         if (recovery_head == 0) {
1144                 /* we have never allocated a recovery record */
1145                 return 0;
1146         }
1147
1148         /* read the recovery record */
1149         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1150                                    sizeof(rec), DOCONV()) == -1) {
1151                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1152                 tdb->ecode = TDB_ERR_IO;
1153                 return -1;
1154         }
1155
1156         if (rec.magic != TDB_RECOVERY_MAGIC) {
1157                 /* there is no valid recovery data */
1158                 return 0;
1159         }
1160
1161         if (tdb->read_only) {
1162                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1163                 tdb->ecode = TDB_ERR_CORRUPT;
1164                 return -1;
1165         }
1166
1167         recovery_eof = rec.key_len;
1168
1169         data = (unsigned char *)malloc(rec.data_len);
1170         if (data == NULL) {
1171                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1172                 tdb->ecode = TDB_ERR_OOM;
1173                 return -1;
1174         }
1175
1176         /* read the full recovery data */
1177         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1178                                    rec.data_len, 0) == -1) {
1179                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1180                 tdb->ecode = TDB_ERR_IO;
1181                 return -1;
1182         }
1183
1184         /* recover the file data */
1185         p = data;
1186         while (p+8 < data + rec.data_len) {
1187                 uint32_t ofs, len;
1188                 if (DOCONV()) {
1189                         tdb_convert(p, 8);
1190                 }
1191                 memcpy(&ofs, p, 4);
1192                 memcpy(&len, p+4, 4);
1193
1194                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1195                         free(data);
1196                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1197                         tdb->ecode = TDB_ERR_IO;
1198                         return -1;
1199                 }
1200                 p += 8 + len;
1201         }
1202
1203         free(data);
1204
1205         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1206                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1207                 tdb->ecode = TDB_ERR_IO;
1208                 return -1;
1209         }
1210
1211         /* if the recovery area is after the recovered eof then remove it */
1212         if (recovery_eof <= recovery_head) {
1213                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1214                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1215                         tdb->ecode = TDB_ERR_IO;
1216                         return -1;                      
1217                 }
1218         }
1219
1220         /* remove the recovery magic */
1221         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1222                           &zero) == -1) {
1223                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1224                 tdb->ecode = TDB_ERR_IO;
1225                 return -1;                      
1226         }
1227         
1228         /* reduce the file size to the old size */
1229         tdb_munmap(tdb);
1230         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1231                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1232                 tdb->ecode = TDB_ERR_IO;
1233                 return -1;                      
1234         }
1235         tdb->map_size = recovery_eof;
1236         tdb_mmap(tdb);
1237
1238         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1239                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1240                 tdb->ecode = TDB_ERR_IO;
1241                 return -1;
1242         }
1243
1244         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1245                  recovery_eof));
1246
1247         /* all done */
1248         return 0;
1249 }