tdb: tdb_allrecord_lock/tdb_allrecord_unlock/tdb_allrecord_upgrade
[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         /* old file size before transaction */
139         tdb_len_t old_map_size;
140
141         /* we should re-pack on commit */
142         bool need_repack;
143 };
144
145
146 /*
147   read while in a transaction. We need to check first if the data is in our list
148   of transaction elements, then if not do a real read
149 */
150 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf, 
151                             tdb_len_t len, int cv)
152 {
153         uint32_t blk;
154
155         /* break it down into block sized ops */
156         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
157                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
158                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
159                         return -1;
160                 }
161                 len -= len2;
162                 off += len2;
163                 buf = (void *)(len2 + (char *)buf);
164         }
165
166         if (len == 0) {
167                 return 0;
168         }
169
170         blk = off / tdb->transaction->block_size;
171
172         /* see if we have it in the block list */
173         if (tdb->transaction->num_blocks <= blk ||
174             tdb->transaction->blocks[blk] == NULL) {
175                 /* nope, do a real read */
176                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
177                         goto fail;
178                 }
179                 return 0;
180         }
181
182         /* it is in the block list. Now check for the last block */
183         if (blk == tdb->transaction->num_blocks-1) {
184                 if (len > tdb->transaction->last_block_size) {
185                         goto fail;
186                 }
187         }
188         
189         /* now copy it out of this block */
190         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
191         if (cv) {
192                 tdb_convert(buf, len);
193         }
194         return 0;
195
196 fail:
197         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
198         tdb->ecode = TDB_ERR_IO;
199         tdb->transaction->transaction_error = 1;
200         return -1;
201 }
202
203
204 /*
205   write while in a transaction
206 */
207 static int transaction_write(struct tdb_context *tdb, tdb_off_t off, 
208                              const void *buf, tdb_len_t len)
209 {
210         uint32_t blk;
211
212         /* Only a commit is allowed on a prepared transaction */
213         if (tdb->transaction->prepared) {
214                 tdb->ecode = TDB_ERR_EINVAL;
215                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
216                 tdb->transaction->transaction_error = 1;
217                 return -1;
218         }
219
220         /* if the write is to a hash head, then update the transaction
221            hash heads */
222         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
223             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
224                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
225                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
226         }
227
228         /* break it up into block sized chunks */
229         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
230                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
231                 if (transaction_write(tdb, off, buf, len2) != 0) {
232                         return -1;
233                 }
234                 len -= len2;
235                 off += len2;
236                 if (buf != NULL) {
237                         buf = (const void *)(len2 + (const char *)buf);
238                 }
239         }
240
241         if (len == 0) {
242                 return 0;
243         }
244
245         blk = off / tdb->transaction->block_size;
246         off = off % tdb->transaction->block_size;
247
248         if (tdb->transaction->num_blocks <= blk) {
249                 uint8_t **new_blocks;
250                 /* expand the blocks array */
251                 if (tdb->transaction->blocks == NULL) {
252                         new_blocks = (uint8_t **)malloc(
253                                 (blk+1)*sizeof(uint8_t *));
254                 } else {
255                         new_blocks = (uint8_t **)realloc(
256                                 tdb->transaction->blocks,
257                                 (blk+1)*sizeof(uint8_t *));
258                 }
259                 if (new_blocks == NULL) {
260                         tdb->ecode = TDB_ERR_OOM;
261                         goto fail;
262                 }
263                 memset(&new_blocks[tdb->transaction->num_blocks], 0, 
264                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
265                 tdb->transaction->blocks = new_blocks;
266                 tdb->transaction->num_blocks = blk+1;
267                 tdb->transaction->last_block_size = 0;
268         }
269
270         /* allocate and fill a block? */
271         if (tdb->transaction->blocks[blk] == NULL) {
272                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
273                 if (tdb->transaction->blocks[blk] == NULL) {
274                         tdb->ecode = TDB_ERR_OOM;
275                         tdb->transaction->transaction_error = 1;
276                         return -1;                      
277                 }
278                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
279                         tdb_len_t len2 = tdb->transaction->block_size;
280                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
281                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
282                         }
283                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size, 
284                                                                    tdb->transaction->blocks[blk], 
285                                                                    len2, 0) != 0) {
286                                 SAFE_FREE(tdb->transaction->blocks[blk]);                               
287                                 tdb->ecode = TDB_ERR_IO;
288                                 goto fail;
289                         }
290                         if (blk == tdb->transaction->num_blocks-1) {
291                                 tdb->transaction->last_block_size = len2;
292                         }                       
293                 }
294         }
295         
296         /* overwrite part of an existing block */
297         if (buf == NULL) {
298                 memset(tdb->transaction->blocks[blk] + off, 0, len);
299         } else {
300                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
301         }
302         if (blk == tdb->transaction->num_blocks-1) {
303                 if (len + off > tdb->transaction->last_block_size) {
304                         tdb->transaction->last_block_size = len + off;
305                 }
306         }
307
308         return 0;
309
310 fail:
311         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", 
312                  (blk*tdb->transaction->block_size) + off, len));
313         tdb->transaction->transaction_error = 1;
314         return -1;
315 }
316
317
318 /*
319   write while in a transaction - this varient never expands the transaction blocks, it only
320   updates existing blocks. This means it cannot change the recovery size
321 */
322 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off, 
323                                       const void *buf, tdb_len_t len)
324 {
325         uint32_t blk;
326
327         /* break it up into block sized chunks */
328         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
329                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
330                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
331                         return -1;
332                 }
333                 len -= len2;
334                 off += len2;
335                 if (buf != NULL) {
336                         buf = (const void *)(len2 + (const char *)buf);
337                 }
338         }
339
340         if (len == 0) {
341                 return 0;
342         }
343
344         blk = off / tdb->transaction->block_size;
345         off = off % tdb->transaction->block_size;
346
347         if (tdb->transaction->num_blocks <= blk ||
348             tdb->transaction->blocks[blk] == NULL) {
349                 return 0;
350         }
351
352         if (blk == tdb->transaction->num_blocks-1 &&
353             off + len > tdb->transaction->last_block_size) {
354                 if (off >= tdb->transaction->last_block_size) {
355                         return 0;
356                 }
357                 len = tdb->transaction->last_block_size - off;
358         }
359
360         /* overwrite part of an existing block */
361         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
362
363         return 0;
364 }
365
366
367 /*
368   accelerated hash chain head search, using the cached hash heads
369 */
370 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
371 {
372         uint32_t h = *chain;
373         for (;h < tdb->header.hash_size;h++) {
374                 /* the +1 takes account of the freelist */
375                 if (0 != tdb->transaction->hash_heads[h+1]) {
376                         break;
377                 }
378         }
379         (*chain) = h;
380 }
381
382 /*
383   out of bounds check during a transaction
384 */
385 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
386 {
387         if (len <= tdb->map_size) {
388                 return 0;
389         }
390         tdb->ecode = TDB_ERR_IO;
391         return -1;
392 }
393
394 /*
395   transaction version of tdb_expand().
396 */
397 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size, 
398                                    tdb_off_t addition)
399 {
400         /* add a write to the transaction elements, so subsequent
401            reads see the zero data */
402         if (transaction_write(tdb, size, NULL, addition) != 0) {
403                 return -1;
404         }
405
406         tdb->transaction->need_repack = true;
407
408         return 0;
409 }
410
411 /*
412   brlock during a transaction - ignore them
413 */
414 static int transaction_brlock(struct tdb_context *tdb,
415                               int rw_type, tdb_off_t offset, size_t len,
416                               enum tdb_lock_flags flags)
417 {
418         /* FIXME: We actually grab the open lock during a transaction. */
419         if (offset == OPEN_LOCK)
420                 return tdb_brlock(tdb, rw_type, offset, len, flags);
421         return 0;
422 }
423
424 static int transaction_brunlock(struct tdb_context *tdb,
425                                 int rw_type, tdb_off_t offset, size_t len)
426 {
427         return 0;
428 }
429
430 static const struct tdb_methods transaction_methods = {
431         transaction_read,
432         transaction_write,
433         transaction_next_hash_chain,
434         transaction_oob,
435         transaction_expand_file,
436         transaction_brlock,
437         transaction_brunlock
438 };
439
440
441 /*
442   start a tdb transaction. No token is returned, as only a single
443   transaction is allowed to be pending per tdb_context
444 */
445 int tdb_transaction_start(struct tdb_context *tdb)
446 {
447         /* some sanity checks */
448         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
449                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
450                 tdb->ecode = TDB_ERR_EINVAL;
451                 return -1;
452         }
453
454         /* cope with nested tdb_transaction_start() calls */
455         if (tdb->transaction != NULL) {
456                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
457                         tdb->ecode = TDB_ERR_NESTING;
458                         return -1;
459                 }
460                 tdb->transaction->nesting++;
461                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
462                          tdb->transaction->nesting));
463                 return 0;
464         }
465
466         if (tdb_have_extra_locks(tdb)) {
467                 /* the caller must not have any locks when starting a
468                    transaction as otherwise we'll be screwed by lack
469                    of nested locks in posix */
470                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
471                 tdb->ecode = TDB_ERR_LOCK;
472                 return -1;
473         }
474
475         if (tdb->travlocks.next != NULL) {
476                 /* you cannot use transactions inside a traverse (although you can use
477                    traverse inside a transaction) as otherwise you can end up with
478                    deadlock */
479                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
480                 tdb->ecode = TDB_ERR_LOCK;
481                 return -1;
482         }
483
484         tdb->transaction = (struct tdb_transaction *)
485                 calloc(sizeof(struct tdb_transaction), 1);
486         if (tdb->transaction == NULL) {
487                 tdb->ecode = TDB_ERR_OOM;
488                 return -1;
489         }
490
491         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
492         tdb->transaction->block_size = tdb->page_size;
493
494         /* get the transaction write lock. This is a blocking lock. As
495            discussed with Volker, there are a number of ways we could
496            make this async, which we will probably do in the future */
497         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
498                 SAFE_FREE(tdb->transaction->blocks);
499                 SAFE_FREE(tdb->transaction);
500                 return -1;
501         }
502         
503         /* get a read lock from the freelist to the end of file. This
504            is upgraded to a write lock during the commit */
505         if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
506                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
507                 goto fail_allrecord_lock;
508         }
509
510         /* setup a copy of the hash table heads so the hash scan in
511            traverse can be fast */
512         tdb->transaction->hash_heads = (uint32_t *)
513                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
514         if (tdb->transaction->hash_heads == NULL) {
515                 tdb->ecode = TDB_ERR_OOM;
516                 goto fail;
517         }
518         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
519                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
520                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
521                 tdb->ecode = TDB_ERR_IO;
522                 goto fail;
523         }
524
525         /* make sure we know about any file expansions already done by
526            anyone else */
527         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
528         tdb->transaction->old_map_size = tdb->map_size;
529
530         /* finally hook the io methods, replacing them with
531            transaction specific methods */
532         tdb->transaction->io_methods = tdb->methods;
533         tdb->methods = &transaction_methods;
534
535         /* Trace at the end, so we get sequence number correct. */
536         tdb_trace(tdb, "tdb_transaction_start");
537         return 0;
538         
539 fail:
540         tdb_allrecord_unlock(tdb, F_RDLCK, false);
541 fail_allrecord_lock:
542         tdb_transaction_unlock(tdb, F_WRLCK);
543         SAFE_FREE(tdb->transaction->blocks);
544         SAFE_FREE(tdb->transaction->hash_heads);
545         SAFE_FREE(tdb->transaction);
546         return -1;
547 }
548
549
550 /*
551   sync to disk
552 */
553 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
554 {       
555         if (tdb->flags & TDB_NOSYNC) {
556                 return 0;
557         }
558
559         if (fdatasync(tdb->fd) != 0) {
560                 tdb->ecode = TDB_ERR_IO;
561                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
562                 return -1;
563         }
564 #ifdef HAVE_MMAP
565         if (tdb->map_ptr) {
566                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
567                 if (msync(moffset + (char *)tdb->map_ptr, 
568                           length + (offset - moffset), MS_SYNC) != 0) {
569                         tdb->ecode = TDB_ERR_IO;
570                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
571                                  strerror(errno)));
572                         return -1;
573                 }
574         }
575 #endif
576         return 0;
577 }
578
579
580 /* 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         /* This also removes the OPEN_LOCK, if we have it. */
619         tdb_release_extra_locks(tdb);
620
621         /* restore the normal io methods */
622         tdb->methods = tdb->transaction->io_methods;
623
624         tdb_transaction_unlock(tdb, F_WRLCK);
625         SAFE_FREE(tdb->transaction->hash_heads);
626         SAFE_FREE(tdb->transaction);
627         
628         return ret;
629 }
630
631 /*
632   cancel the current transaction
633 */
634 int tdb_transaction_cancel(struct tdb_context *tdb)
635 {
636         int ltype = F_RDLCK;
637         tdb_trace(tdb, "tdb_transaction_cancel");
638         if (tdb->transaction && tdb->transaction->prepared)
639                 ltype = F_WRLCK;
640         return _tdb_transaction_cancel(tdb, ltype);
641 }
642
643 /*
644   work out how much space the linearised recovery data will consume
645 */
646 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
647 {
648         tdb_len_t recovery_size = 0;
649         int i;
650
651         recovery_size = sizeof(uint32_t);
652         for (i=0;i<tdb->transaction->num_blocks;i++) {
653                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
654                         break;
655                 }
656                 if (tdb->transaction->blocks[i] == NULL) {
657                         continue;
658                 }
659                 recovery_size += 2*sizeof(tdb_off_t);
660                 if (i == tdb->transaction->num_blocks-1) {
661                         recovery_size += tdb->transaction->last_block_size;
662                 } else {
663                         recovery_size += tdb->transaction->block_size;
664                 }
665         }       
666
667         return recovery_size;
668 }
669
670 /*
671   allocate the recovery area, or use an existing recovery area if it is
672   large enough
673 */
674 static int tdb_recovery_allocate(struct tdb_context *tdb, 
675                                  tdb_len_t *recovery_size,
676                                  tdb_off_t *recovery_offset,
677                                  tdb_len_t *recovery_max_size)
678 {
679         struct tdb_record rec;
680         const struct tdb_methods *methods = tdb->transaction->io_methods;
681         tdb_off_t recovery_head;
682
683         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
684                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
685                 return -1;
686         }
687
688         rec.rec_len = 0;
689
690         if (recovery_head != 0) {
691                 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
692                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
693                         return -1;
694                 }
695                 /* ignore invalid recovery regions: can happen in crash */
696                 if (rec.magic != TDB_RECOVERY_MAGIC &&
697                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
698                         recovery_head = 0;
699                 }
700         }
701
702         *recovery_size = tdb_recovery_size(tdb);
703
704         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
705                 /* it fits in the existing area */
706                 *recovery_max_size = rec.rec_len;
707                 *recovery_offset = recovery_head;
708                 return 0;
709         }
710
711         /* we need to free up the old recovery area, then allocate a
712            new one at the end of the file. Note that we cannot use
713            tdb_allocate() to allocate the new one as that might return
714            us an area that is being currently used (as of the start of
715            the transaction) */
716         if (recovery_head != 0) {
717                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
718                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
719                         return -1;
720                 }
721         }
722
723         /* the tdb_free() call might have increased the recovery size */
724         *recovery_size = tdb_recovery_size(tdb);
725
726         /* round up to a multiple of page size */
727         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
728         *recovery_offset = tdb->map_size;
729         recovery_head = *recovery_offset;
730
731         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
732                                      (tdb->map_size - tdb->transaction->old_map_size) +
733                                      sizeof(rec) + *recovery_max_size) == -1) {
734                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
735                 return -1;
736         }
737
738         /* remap the file (if using mmap) */
739         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
740
741         /* we have to reset the old map size so that we don't try to expand the file
742            again in the transaction commit, which would destroy the recovery area */
743         tdb->transaction->old_map_size = tdb->map_size;
744
745         /* write the recovery header offset and sync - we can sync without a race here
746            as the magic ptr in the recovery record has not been set */
747         CONVERT(recovery_head);
748         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
749                                &recovery_head, sizeof(tdb_off_t)) == -1) {
750                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
751                 return -1;
752         }
753         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
754                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
755                 return -1;
756         }
757
758         return 0;
759 }
760
761
762 /*
763   setup the recovery data that will be used on a crash during commit
764 */
765 static int transaction_setup_recovery(struct tdb_context *tdb, 
766                                       tdb_off_t *magic_offset)
767 {
768         tdb_len_t recovery_size;
769         unsigned char *data, *p;
770         const struct tdb_methods *methods = tdb->transaction->io_methods;
771         struct tdb_record *rec;
772         tdb_off_t recovery_offset, recovery_max_size;
773         tdb_off_t old_map_size = tdb->transaction->old_map_size;
774         uint32_t magic, tailer;
775         int i;
776
777         /*
778           check that the recovery area has enough space
779         */
780         if (tdb_recovery_allocate(tdb, &recovery_size, 
781                                   &recovery_offset, &recovery_max_size) == -1) {
782                 return -1;
783         }
784
785         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
786         if (data == NULL) {
787                 tdb->ecode = TDB_ERR_OOM;
788                 return -1;
789         }
790
791         rec = (struct tdb_record *)data;
792         memset(rec, 0, sizeof(*rec));
793
794         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
795         rec->data_len = recovery_size;
796         rec->rec_len  = recovery_max_size;
797         rec->key_len  = old_map_size;
798         CONVERT(rec);
799
800         /* build the recovery data into a single blob to allow us to do a single
801            large write, which should be more efficient */
802         p = data + sizeof(*rec);
803         for (i=0;i<tdb->transaction->num_blocks;i++) {
804                 tdb_off_t offset;
805                 tdb_len_t length;
806
807                 if (tdb->transaction->blocks[i] == NULL) {
808                         continue;
809                 }
810
811                 offset = i * tdb->transaction->block_size;
812                 length = tdb->transaction->block_size;
813                 if (i == tdb->transaction->num_blocks-1) {
814                         length = tdb->transaction->last_block_size;
815                 }
816                 
817                 if (offset >= old_map_size) {
818                         continue;
819                 }
820                 if (offset + length > tdb->transaction->old_map_size) {
821                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
822                         free(data);
823                         tdb->ecode = TDB_ERR_CORRUPT;
824                         return -1;
825                 }
826                 memcpy(p, &offset, 4);
827                 memcpy(p+4, &length, 4);
828                 if (DOCONV()) {
829                         tdb_convert(p, 8);
830                 }
831                 /* the recovery area contains the old data, not the
832                    new data, so we have to call the original tdb_read
833                    method to get it */
834                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
835                         free(data);
836                         tdb->ecode = TDB_ERR_IO;
837                         return -1;
838                 }
839                 p += 8 + length;
840         }
841
842         /* and the tailer */
843         tailer = sizeof(*rec) + recovery_max_size;
844         memcpy(p, &tailer, 4);
845         CONVERT(p);
846
847         /* write the recovery data to the recovery area */
848         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
849                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
850                 free(data);
851                 tdb->ecode = TDB_ERR_IO;
852                 return -1;
853         }
854         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
855                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
856                 free(data);
857                 tdb->ecode = TDB_ERR_IO;
858                 return -1;
859         }
860
861         /* as we don't have ordered writes, we have to sync the recovery
862            data before we update the magic to indicate that the recovery
863            data is present */
864         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
865                 free(data);
866                 return -1;
867         }
868
869         free(data);
870
871         magic = TDB_RECOVERY_MAGIC;
872         CONVERT(magic);
873
874         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
875
876         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
877                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
878                 tdb->ecode = TDB_ERR_IO;
879                 return -1;
880         }
881         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
882                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
883                 tdb->ecode = TDB_ERR_IO;
884                 return -1;
885         }
886
887         /* ensure the recovery magic marker is on disk */
888         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
889                 return -1;
890         }
891
892         return 0;
893 }
894
895 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
896 {       
897         const struct tdb_methods *methods;
898
899         if (tdb->transaction == NULL) {
900                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
901                 return -1;
902         }
903
904         if (tdb->transaction->prepared) {
905                 tdb->ecode = TDB_ERR_EINVAL;
906                 _tdb_transaction_cancel(tdb, F_WRLCK);
907                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
908                 return -1;
909         }
910
911         if (tdb->transaction->transaction_error) {
912                 tdb->ecode = TDB_ERR_IO;
913                 _tdb_transaction_cancel(tdb, F_RDLCK);
914                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
915                 return -1;
916         }
917
918
919         if (tdb->transaction->nesting != 0) {
920                 return 0;
921         }               
922
923         /* check for a null transaction */
924         if (tdb->transaction->blocks == NULL) {
925                 return 0;
926         }
927
928         methods = tdb->transaction->io_methods;
929         
930         /* if there are any locks pending then the caller has not
931            nested their locks properly, so fail the transaction */
932         if (tdb_have_extra_locks(tdb)) {
933                 tdb->ecode = TDB_ERR_LOCK;
934                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
935                 _tdb_transaction_cancel(tdb, F_RDLCK);
936                 return -1;
937         }
938
939         /* upgrade the main transaction lock region to a write lock */
940         if (tdb_allrecord_upgrade(tdb) == -1) {
941                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
942                 _tdb_transaction_cancel(tdb, F_RDLCK);
943                 return -1;
944         }
945
946         /* get the open lock - this prevents new users attaching to the database
947            during the commit */
948         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
949                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
950                 _tdb_transaction_cancel(tdb, F_WRLCK);
951                 return -1;
952         }
953
954         if (!(tdb->flags & TDB_NOSYNC)) {
955                 /* write the recovery data to the end of the file */
956                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
957                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
958                         _tdb_transaction_cancel(tdb, F_WRLCK);
959                         return -1;
960                 }
961         }
962
963         tdb->transaction->prepared = true;
964
965         /* expand the file to the new size if needed */
966         if (tdb->map_size != tdb->transaction->old_map_size) {
967                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
968                                              tdb->map_size - 
969                                              tdb->transaction->old_map_size) == -1) {
970                         tdb->ecode = TDB_ERR_IO;
971                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
972                         _tdb_transaction_cancel(tdb, F_WRLCK);
973                         return -1;
974                 }
975                 tdb->map_size = tdb->transaction->old_map_size;
976                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
977         }
978
979         /* Keep the open lock until the actual commit */
980
981         return 0;
982 }
983
984 /*
985    prepare to commit the current transaction
986 */
987 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
988 {       
989         tdb_trace(tdb, "tdb_transaction_prepare_commit");
990         return _tdb_transaction_prepare_commit(tdb);
991 }
992
993 /*
994   commit the current transaction
995 */
996 int tdb_transaction_commit(struct tdb_context *tdb)
997 {       
998         const struct tdb_methods *methods;
999         int i;
1000         bool need_repack;
1001
1002         if (tdb->transaction == NULL) {
1003                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1004                 return -1;
1005         }
1006
1007         tdb_trace(tdb, "tdb_transaction_commit");
1008
1009         if (tdb->transaction->transaction_error) {
1010                 tdb->ecode = TDB_ERR_IO;
1011                 _tdb_transaction_cancel(tdb, F_RDLCK);
1012                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1013                 return -1;
1014         }
1015
1016
1017         if (tdb->transaction->nesting != 0) {
1018                 tdb->transaction->nesting--;
1019                 return 0;
1020         }
1021
1022         /* check for a null transaction */
1023         if (tdb->transaction->blocks == NULL) {
1024                 _tdb_transaction_cancel(tdb, F_RDLCK);
1025                 return 0;
1026         }
1027
1028         if (!tdb->transaction->prepared) {
1029                 int ret = _tdb_transaction_prepare_commit(tdb);
1030                 if (ret)
1031                         return ret;
1032         }
1033
1034         methods = tdb->transaction->io_methods;
1035
1036         /* perform all the writes */
1037         for (i=0;i<tdb->transaction->num_blocks;i++) {
1038                 tdb_off_t offset;
1039                 tdb_len_t length;
1040
1041                 if (tdb->transaction->blocks[i] == NULL) {
1042                         continue;
1043                 }
1044
1045                 offset = i * tdb->transaction->block_size;
1046                 length = tdb->transaction->block_size;
1047                 if (i == tdb->transaction->num_blocks-1) {
1048                         length = tdb->transaction->last_block_size;
1049                 }
1050
1051                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1052                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1053                         
1054                         /* we've overwritten part of the data and
1055                            possibly expanded the file, so we need to
1056                            run the crash recovery code */
1057                         tdb->methods = methods;
1058                         tdb_transaction_recover(tdb); 
1059
1060                         _tdb_transaction_cancel(tdb, F_WRLCK);
1061
1062                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1063                         return -1;
1064                 }
1065                 SAFE_FREE(tdb->transaction->blocks[i]);
1066         } 
1067
1068         SAFE_FREE(tdb->transaction->blocks);
1069         tdb->transaction->num_blocks = 0;
1070
1071         /* ensure the new data is on disk */
1072         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1073                 return -1;
1074         }
1075
1076         /*
1077           TODO: maybe write to some dummy hdr field, or write to magic
1078           offset without mmap, before the last sync, instead of the
1079           utime() call
1080         */
1081
1082         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1083            don't change the mtime of the file, this means the file may
1084            not be backed up (as tdb rounding to block sizes means that
1085            file size changes are quite rare too). The following forces
1086            mtime changes when a transaction completes */
1087 #ifdef HAVE_UTIME
1088         utime(tdb->name, NULL);
1089 #endif
1090
1091         need_repack = tdb->transaction->need_repack;
1092
1093         /* use a transaction cancel to free memory and remove the
1094            transaction locks */
1095         _tdb_transaction_cancel(tdb, F_WRLCK);
1096
1097         if (need_repack) {
1098                 return tdb_repack(tdb);
1099         }
1100
1101         return 0;
1102 }
1103
1104
1105 /*
1106   recover from an aborted transaction. Must be called with exclusive
1107   database write access already established (including the open
1108   lock to prevent new processes attaching)
1109 */
1110 int tdb_transaction_recover(struct tdb_context *tdb)
1111 {
1112         tdb_off_t recovery_head, recovery_eof;
1113         unsigned char *data, *p;
1114         uint32_t zero = 0;
1115         struct tdb_record rec;
1116
1117         /* find the recovery area */
1118         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1119                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1120                 tdb->ecode = TDB_ERR_IO;
1121                 return -1;
1122         }
1123
1124         if (recovery_head == 0) {
1125                 /* we have never allocated a recovery record */
1126                 return 0;
1127         }
1128
1129         /* read the recovery record */
1130         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1131                                    sizeof(rec), DOCONV()) == -1) {
1132                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1133                 tdb->ecode = TDB_ERR_IO;
1134                 return -1;
1135         }
1136
1137         if (rec.magic != TDB_RECOVERY_MAGIC) {
1138                 /* there is no valid recovery data */
1139                 return 0;
1140         }
1141
1142         if (tdb->read_only) {
1143                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1144                 tdb->ecode = TDB_ERR_CORRUPT;
1145                 return -1;
1146         }
1147
1148         recovery_eof = rec.key_len;
1149
1150         data = (unsigned char *)malloc(rec.data_len);
1151         if (data == NULL) {
1152                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1153                 tdb->ecode = TDB_ERR_OOM;
1154                 return -1;
1155         }
1156
1157         /* read the full recovery data */
1158         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1159                                    rec.data_len, 0) == -1) {
1160                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1161                 tdb->ecode = TDB_ERR_IO;
1162                 return -1;
1163         }
1164
1165         /* recover the file data */
1166         p = data;
1167         while (p+8 < data + rec.data_len) {
1168                 uint32_t ofs, len;
1169                 if (DOCONV()) {
1170                         tdb_convert(p, 8);
1171                 }
1172                 memcpy(&ofs, p, 4);
1173                 memcpy(&len, p+4, 4);
1174
1175                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1176                         free(data);
1177                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1178                         tdb->ecode = TDB_ERR_IO;
1179                         return -1;
1180                 }
1181                 p += 8 + len;
1182         }
1183
1184         free(data);
1185
1186         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1187                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1188                 tdb->ecode = TDB_ERR_IO;
1189                 return -1;
1190         }
1191
1192         /* if the recovery area is after the recovered eof then remove it */
1193         if (recovery_eof <= recovery_head) {
1194                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1195                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1196                         tdb->ecode = TDB_ERR_IO;
1197                         return -1;                      
1198                 }
1199         }
1200
1201         /* remove the recovery magic */
1202         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1203                           &zero) == -1) {
1204                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1205                 tdb->ecode = TDB_ERR_IO;
1206                 return -1;                      
1207         }
1208         
1209         /* reduce the file size to the old size */
1210         tdb_munmap(tdb);
1211         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1212                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1213                 tdb->ecode = TDB_ERR_IO;
1214                 return -1;                      
1215         }
1216         tdb->map_size = recovery_eof;
1217         tdb_mmap(tdb);
1218
1219         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1220                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1221                 tdb->ecode = TDB_ERR_IO;
1222                 return -1;
1223         }
1224
1225         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1226                  recovery_eof));
1227
1228         /* all done */
1229         return 0;
1230 }