304a03fa3836940197c1319eeae5ddbcb3fce823
[vlendec/samba-autobuild/.git] / ctdb / 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 static const struct tdb_methods transaction_methods = {
412         transaction_read,
413         transaction_write,
414         transaction_next_hash_chain,
415         transaction_oob,
416         transaction_expand_file,
417 };
418
419
420 /*
421   start a tdb transaction. No token is returned, as only a single
422   transaction is allowed to be pending per tdb_context
423 */
424 static int _tdb_transaction_start(struct tdb_context *tdb,
425                                   enum tdb_lock_flags lockflags)
426 {
427         /* some sanity checks */
428         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
429                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
430                 tdb->ecode = TDB_ERR_EINVAL;
431                 return -1;
432         }
433
434         /* cope with nested tdb_transaction_start() calls */
435         if (tdb->transaction != NULL) {
436                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
437                         tdb->ecode = TDB_ERR_NESTING;
438                         return -1;
439                 }
440                 tdb->transaction->nesting++;
441                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n", 
442                          tdb->transaction->nesting));
443                 return 0;
444         }
445
446         if (tdb_have_extra_locks(tdb)) {
447                 /* the caller must not have any locks when starting a
448                    transaction as otherwise we'll be screwed by lack
449                    of nested locks in posix */
450                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
451                 tdb->ecode = TDB_ERR_LOCK;
452                 return -1;
453         }
454
455         if (tdb->travlocks.next != NULL) {
456                 /* you cannot use transactions inside a traverse (although you can use
457                    traverse inside a transaction) as otherwise you can end up with
458                    deadlock */
459                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
460                 tdb->ecode = TDB_ERR_LOCK;
461                 return -1;
462         }
463
464         tdb->transaction = (struct tdb_transaction *)
465                 calloc(sizeof(struct tdb_transaction), 1);
466         if (tdb->transaction == NULL) {
467                 tdb->ecode = TDB_ERR_OOM;
468                 return -1;
469         }
470
471         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
472         tdb->transaction->block_size = tdb->page_size;
473
474         /* get the transaction write lock. This is a blocking lock. As
475            discussed with Volker, there are a number of ways we could
476            make this async, which we will probably do in the future */
477         if (tdb_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
478                 SAFE_FREE(tdb->transaction->blocks);
479                 SAFE_FREE(tdb->transaction);
480                 if ((lockflags & TDB_LOCK_WAIT) == 0) {
481                         tdb->ecode = TDB_ERR_NOLOCK;
482                 }
483                 return -1;
484         }
485
486         /* get a read lock from the freelist to the end of file. This
487            is upgraded to a write lock during the commit */
488         if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
489                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
490                 goto fail_allrecord_lock;
491         }
492
493         /* setup a copy of the hash table heads so the hash scan in
494            traverse can be fast */
495         tdb->transaction->hash_heads = (uint32_t *)
496                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
497         if (tdb->transaction->hash_heads == NULL) {
498                 tdb->ecode = TDB_ERR_OOM;
499                 goto fail;
500         }
501         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
502                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
503                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
504                 tdb->ecode = TDB_ERR_IO;
505                 goto fail;
506         }
507
508         /* make sure we know about any file expansions already done by
509            anyone else */
510         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
511         tdb->transaction->old_map_size = tdb->map_size;
512
513         /* finally hook the io methods, replacing them with
514            transaction specific methods */
515         tdb->transaction->io_methods = tdb->methods;
516         tdb->methods = &transaction_methods;
517
518         /* Trace at the end, so we get sequence number correct. */
519         tdb_trace(tdb, "tdb_transaction_start");
520         return 0;
521
522 fail:
523         tdb_allrecord_unlock(tdb, F_RDLCK, false);
524 fail_allrecord_lock:
525         tdb_transaction_unlock(tdb, F_WRLCK);
526         SAFE_FREE(tdb->transaction->blocks);
527         SAFE_FREE(tdb->transaction->hash_heads);
528         SAFE_FREE(tdb->transaction);
529         return -1;
530 }
531
532 int tdb_transaction_start(struct tdb_context *tdb)
533 {
534         return _tdb_transaction_start(tdb, TDB_LOCK_WAIT);
535 }
536
537 int tdb_transaction_start_nonblock(struct tdb_context *tdb)
538 {
539         return _tdb_transaction_start(tdb, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
540 }
541
542 /*
543   sync to disk
544 */
545 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
546 {       
547         if (tdb->flags & TDB_NOSYNC) {
548                 return 0;
549         }
550
551         if (fdatasync(tdb->fd) != 0) {
552                 tdb->ecode = TDB_ERR_IO;
553                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
554                 return -1;
555         }
556 #ifdef HAVE_MMAP
557         if (tdb->map_ptr) {
558                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
559                 if (msync(moffset + (char *)tdb->map_ptr, 
560                           length + (offset - moffset), MS_SYNC) != 0) {
561                         tdb->ecode = TDB_ERR_IO;
562                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
563                                  strerror(errno)));
564                         return -1;
565                 }
566         }
567 #endif
568         return 0;
569 }
570
571
572 static int _tdb_transaction_cancel(struct tdb_context *tdb)
573 {       
574         int i, ret = 0;
575
576         if (tdb->transaction == NULL) {
577                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
578                 return -1;
579         }
580
581         if (tdb->transaction->nesting != 0) {
582                 tdb->transaction->transaction_error = 1;
583                 tdb->transaction->nesting--;
584                 return 0;
585         }               
586
587         tdb->map_size = tdb->transaction->old_map_size;
588
589         /* free all the transaction blocks */
590         for (i=0;i<tdb->transaction->num_blocks;i++) {
591                 if (tdb->transaction->blocks[i] != NULL) {
592                         free(tdb->transaction->blocks[i]);
593                 }
594         }
595         SAFE_FREE(tdb->transaction->blocks);
596
597         if (tdb->transaction->magic_offset) {
598                 const struct tdb_methods *methods = tdb->transaction->io_methods;
599                 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
600
601                 /* remove the recovery marker */
602                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
603                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
604                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
605                         ret = -1;
606                 }
607         }
608
609         /* This also removes the OPEN_LOCK, if we have it. */
610         tdb_release_transaction_locks(tdb);
611
612         /* restore the normal io methods */
613         tdb->methods = tdb->transaction->io_methods;
614
615         SAFE_FREE(tdb->transaction->hash_heads);
616         SAFE_FREE(tdb->transaction);
617
618         return ret;
619 }
620
621 /*
622   cancel the current transaction
623 */
624 int tdb_transaction_cancel(struct tdb_context *tdb)
625 {
626         tdb_trace(tdb, "tdb_transaction_cancel");
627         return _tdb_transaction_cancel(tdb);
628 }
629
630 /*
631   work out how much space the linearised recovery data will consume
632 */
633 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
634 {
635         tdb_len_t recovery_size = 0;
636         int i;
637
638         recovery_size = sizeof(uint32_t);
639         for (i=0;i<tdb->transaction->num_blocks;i++) {
640                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
641                         break;
642                 }
643                 if (tdb->transaction->blocks[i] == NULL) {
644                         continue;
645                 }
646                 recovery_size += 2*sizeof(tdb_off_t);
647                 if (i == tdb->transaction->num_blocks-1) {
648                         recovery_size += tdb->transaction->last_block_size;
649                 } else {
650                         recovery_size += tdb->transaction->block_size;
651                 }
652         }       
653
654         return recovery_size;
655 }
656
657 /*
658   allocate the recovery area, or use an existing recovery area if it is
659   large enough
660 */
661 static int tdb_recovery_allocate(struct tdb_context *tdb, 
662                                  tdb_len_t *recovery_size,
663                                  tdb_off_t *recovery_offset,
664                                  tdb_len_t *recovery_max_size)
665 {
666         struct tdb_record rec;
667         const struct tdb_methods *methods = tdb->transaction->io_methods;
668         tdb_off_t recovery_head;
669
670         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
671                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
672                 return -1;
673         }
674
675         rec.rec_len = 0;
676
677         if (recovery_head != 0) {
678                 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
679                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
680                         return -1;
681                 }
682                 /* ignore invalid recovery regions: can happen in crash */
683                 if (rec.magic != TDB_RECOVERY_MAGIC &&
684                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
685                         recovery_head = 0;
686                 }
687         }
688
689         *recovery_size = tdb_recovery_size(tdb);
690
691         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
692                 /* it fits in the existing area */
693                 *recovery_max_size = rec.rec_len;
694                 *recovery_offset = recovery_head;
695                 return 0;
696         }
697
698         /* we need to free up the old recovery area, then allocate a
699            new one at the end of the file. Note that we cannot use
700            tdb_allocate() to allocate the new one as that might return
701            us an area that is being currently used (as of the start of
702            the transaction) */
703         if (recovery_head != 0) {
704                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
705                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
706                         return -1;
707                 }
708         }
709
710         /* the tdb_free() call might have increased the recovery size */
711         *recovery_size = tdb_recovery_size(tdb);
712
713         /* round up to a multiple of page size */
714         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
715         *recovery_offset = tdb->map_size;
716         recovery_head = *recovery_offset;
717
718         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
719                                      (tdb->map_size - tdb->transaction->old_map_size) +
720                                      sizeof(rec) + *recovery_max_size) == -1) {
721                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
722                 return -1;
723         }
724
725         /* remap the file (if using mmap) */
726         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
727
728         /* we have to reset the old map size so that we don't try to expand the file
729            again in the transaction commit, which would destroy the recovery area */
730         tdb->transaction->old_map_size = tdb->map_size;
731
732         /* write the recovery header offset and sync - we can sync without a race here
733            as the magic ptr in the recovery record has not been set */
734         CONVERT(recovery_head);
735         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD, 
736                                &recovery_head, sizeof(tdb_off_t)) == -1) {
737                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
738                 return -1;
739         }
740         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
741                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
742                 return -1;
743         }
744
745         return 0;
746 }
747
748
749 /*
750   setup the recovery data that will be used on a crash during commit
751 */
752 static int transaction_setup_recovery(struct tdb_context *tdb, 
753                                       tdb_off_t *magic_offset)
754 {
755         tdb_len_t recovery_size;
756         unsigned char *data, *p;
757         const struct tdb_methods *methods = tdb->transaction->io_methods;
758         struct tdb_record *rec;
759         tdb_off_t recovery_offset, recovery_max_size;
760         tdb_off_t old_map_size = tdb->transaction->old_map_size;
761         uint32_t magic, tailer;
762         int i;
763
764         /*
765           check that the recovery area has enough space
766         */
767         if (tdb_recovery_allocate(tdb, &recovery_size, 
768                                   &recovery_offset, &recovery_max_size) == -1) {
769                 return -1;
770         }
771
772         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
773         if (data == NULL) {
774                 tdb->ecode = TDB_ERR_OOM;
775                 return -1;
776         }
777
778         rec = (struct tdb_record *)data;
779         memset(rec, 0, sizeof(*rec));
780
781         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
782         rec->data_len = recovery_size;
783         rec->rec_len  = recovery_max_size;
784         rec->key_len  = old_map_size;
785         CONVERT(rec);
786
787         /* build the recovery data into a single blob to allow us to do a single
788            large write, which should be more efficient */
789         p = data + sizeof(*rec);
790         for (i=0;i<tdb->transaction->num_blocks;i++) {
791                 tdb_off_t offset;
792                 tdb_len_t length;
793
794                 if (tdb->transaction->blocks[i] == NULL) {
795                         continue;
796                 }
797
798                 offset = i * tdb->transaction->block_size;
799                 length = tdb->transaction->block_size;
800                 if (i == tdb->transaction->num_blocks-1) {
801                         length = tdb->transaction->last_block_size;
802                 }
803
804                 if (offset >= old_map_size) {
805                         continue;
806                 }
807                 if (offset + length > tdb->transaction->old_map_size) {
808                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
809                         free(data);
810                         tdb->ecode = TDB_ERR_CORRUPT;
811                         return -1;
812                 }
813                 memcpy(p, &offset, 4);
814                 memcpy(p+4, &length, 4);
815                 if (DOCONV()) {
816                         tdb_convert(p, 8);
817                 }
818                 /* the recovery area contains the old data, not the
819                    new data, so we have to call the original tdb_read
820                    method to get it */
821                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
822                         free(data);
823                         tdb->ecode = TDB_ERR_IO;
824                         return -1;
825                 }
826                 p += 8 + length;
827         }
828
829         /* and the tailer */
830         tailer = sizeof(*rec) + recovery_max_size;
831         memcpy(p, &tailer, 4);
832         CONVERT(p);
833
834         /* write the recovery data to the recovery area */
835         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
836                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
837                 free(data);
838                 tdb->ecode = TDB_ERR_IO;
839                 return -1;
840         }
841         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
842                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
843                 free(data);
844                 tdb->ecode = TDB_ERR_IO;
845                 return -1;
846         }
847
848         /* as we don't have ordered writes, we have to sync the recovery
849            data before we update the magic to indicate that the recovery
850            data is present */
851         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
852                 free(data);
853                 return -1;
854         }
855
856         free(data);
857
858         magic = TDB_RECOVERY_MAGIC;
859         CONVERT(magic);
860
861         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
862
863         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
864                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
865                 tdb->ecode = TDB_ERR_IO;
866                 return -1;
867         }
868         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
869                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
870                 tdb->ecode = TDB_ERR_IO;
871                 return -1;
872         }
873
874         /* ensure the recovery magic marker is on disk */
875         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
876                 return -1;
877         }
878
879         return 0;
880 }
881
882 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
883 {       
884         const struct tdb_methods *methods;
885
886         if (tdb->transaction == NULL) {
887                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
888                 return -1;
889         }
890
891         if (tdb->transaction->prepared) {
892                 tdb->ecode = TDB_ERR_EINVAL;
893                 _tdb_transaction_cancel(tdb);
894                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
895                 return -1;
896         }
897
898         if (tdb->transaction->transaction_error) {
899                 tdb->ecode = TDB_ERR_IO;
900                 _tdb_transaction_cancel(tdb);
901                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
902                 return -1;
903         }
904
905
906         if (tdb->transaction->nesting != 0) {
907                 return 0;
908         }               
909
910         /* check for a null transaction */
911         if (tdb->transaction->blocks == NULL) {
912                 return 0;
913         }
914
915         methods = tdb->transaction->io_methods;
916
917         /* if there are any locks pending then the caller has not
918            nested their locks properly, so fail the transaction */
919         if (tdb_have_extra_locks(tdb)) {
920                 tdb->ecode = TDB_ERR_LOCK;
921                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
922                 _tdb_transaction_cancel(tdb);
923                 return -1;
924         }
925
926         /* upgrade the main transaction lock region to a write lock */
927         if (tdb_allrecord_upgrade(tdb) == -1) {
928                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
929                 _tdb_transaction_cancel(tdb);
930                 return -1;
931         }
932
933         /* get the open lock - this prevents new users attaching to the database
934            during the commit */
935         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
936                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
937                 _tdb_transaction_cancel(tdb);
938                 return -1;
939         }
940
941         if (!(tdb->flags & TDB_NOSYNC)) {
942                 /* write the recovery data to the end of the file */
943                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
944                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
945                         _tdb_transaction_cancel(tdb);
946                         return -1;
947                 }
948         }
949
950         tdb->transaction->prepared = true;
951
952         /* expand the file to the new size if needed */
953         if (tdb->map_size != tdb->transaction->old_map_size) {
954                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size, 
955                                              tdb->map_size - 
956                                              tdb->transaction->old_map_size) == -1) {
957                         tdb->ecode = TDB_ERR_IO;
958                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
959                         _tdb_transaction_cancel(tdb);
960                         return -1;
961                 }
962                 tdb->map_size = tdb->transaction->old_map_size;
963                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
964         }
965
966         /* Keep the open lock until the actual commit */
967
968         return 0;
969 }
970
971 /*
972    prepare to commit the current transaction
973 */
974 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
975 {       
976         tdb_trace(tdb, "tdb_transaction_prepare_commit");
977         return _tdb_transaction_prepare_commit(tdb);
978 }
979
980 /*
981   commit the current transaction
982 */
983 int tdb_transaction_commit(struct tdb_context *tdb)
984 {       
985         const struct tdb_methods *methods;
986         int i;
987         bool need_repack;
988
989         if (tdb->transaction == NULL) {
990                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
991                 return -1;
992         }
993
994         tdb_trace(tdb, "tdb_transaction_commit");
995
996         if (tdb->transaction->transaction_error) {
997                 tdb->ecode = TDB_ERR_IO;
998                 _tdb_transaction_cancel(tdb);
999                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1000                 return -1;
1001         }
1002
1003
1004         if (tdb->transaction->nesting != 0) {
1005                 tdb->transaction->nesting--;
1006                 return 0;
1007         }
1008
1009         /* check for a null transaction */
1010         if (tdb->transaction->blocks == NULL) {
1011                 _tdb_transaction_cancel(tdb);
1012                 return 0;
1013         }
1014
1015         if (!tdb->transaction->prepared) {
1016                 int ret = _tdb_transaction_prepare_commit(tdb);
1017                 if (ret)
1018                         return ret;
1019         }
1020
1021         methods = tdb->transaction->io_methods;
1022
1023         /* perform all the writes */
1024         for (i=0;i<tdb->transaction->num_blocks;i++) {
1025                 tdb_off_t offset;
1026                 tdb_len_t length;
1027
1028                 if (tdb->transaction->blocks[i] == NULL) {
1029                         continue;
1030                 }
1031
1032                 offset = i * tdb->transaction->block_size;
1033                 length = tdb->transaction->block_size;
1034                 if (i == tdb->transaction->num_blocks-1) {
1035                         length = tdb->transaction->last_block_size;
1036                 }
1037
1038                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1039                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1040
1041                         /* we've overwritten part of the data and
1042                            possibly expanded the file, so we need to
1043                            run the crash recovery code */
1044                         tdb->methods = methods;
1045                         tdb_transaction_recover(tdb); 
1046
1047                         _tdb_transaction_cancel(tdb);
1048
1049                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1050                         return -1;
1051                 }
1052                 SAFE_FREE(tdb->transaction->blocks[i]);
1053         } 
1054
1055         SAFE_FREE(tdb->transaction->blocks);
1056         tdb->transaction->num_blocks = 0;
1057
1058         /* ensure the new data is on disk */
1059         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1060                 return -1;
1061         }
1062
1063         /*
1064           TODO: maybe write to some dummy hdr field, or write to magic
1065           offset without mmap, before the last sync, instead of the
1066           utime() call
1067         */
1068
1069         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1070            don't change the mtime of the file, this means the file may
1071            not be backed up (as tdb rounding to block sizes means that
1072            file size changes are quite rare too). The following forces
1073            mtime changes when a transaction completes */
1074 #ifdef HAVE_UTIME
1075         utime(tdb->name, NULL);
1076 #endif
1077
1078         need_repack = tdb->transaction->need_repack;
1079
1080         /* use a transaction cancel to free memory and remove the
1081            transaction locks */
1082         _tdb_transaction_cancel(tdb);
1083
1084         if (need_repack) {
1085                 return tdb_repack(tdb);
1086         }
1087
1088         return 0;
1089 }
1090
1091
1092 /*
1093   recover from an aborted transaction. Must be called with exclusive
1094   database write access already established (including the open
1095   lock to prevent new processes attaching)
1096 */
1097 int tdb_transaction_recover(struct tdb_context *tdb)
1098 {
1099         tdb_off_t recovery_head, recovery_eof;
1100         unsigned char *data, *p;
1101         uint32_t zero = 0;
1102         struct tdb_record rec;
1103
1104         /* find the recovery area */
1105         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1106                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1107                 tdb->ecode = TDB_ERR_IO;
1108                 return -1;
1109         }
1110
1111         if (recovery_head == 0) {
1112                 /* we have never allocated a recovery record */
1113                 return 0;
1114         }
1115
1116         /* read the recovery record */
1117         if (tdb->methods->tdb_read(tdb, recovery_head, &rec, 
1118                                    sizeof(rec), DOCONV()) == -1) {
1119                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));           
1120                 tdb->ecode = TDB_ERR_IO;
1121                 return -1;
1122         }
1123
1124         if (rec.magic != TDB_RECOVERY_MAGIC) {
1125                 /* there is no valid recovery data */
1126                 return 0;
1127         }
1128
1129         if (tdb->read_only) {
1130                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1131                 tdb->ecode = TDB_ERR_CORRUPT;
1132                 return -1;
1133         }
1134
1135         recovery_eof = rec.key_len;
1136
1137         data = (unsigned char *)malloc(rec.data_len);
1138         if (data == NULL) {
1139                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));         
1140                 tdb->ecode = TDB_ERR_OOM;
1141                 return -1;
1142         }
1143
1144         /* read the full recovery data */
1145         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1146                                    rec.data_len, 0) == -1) {
1147                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));             
1148                 tdb->ecode = TDB_ERR_IO;
1149                 return -1;
1150         }
1151
1152         /* recover the file data */
1153         p = data;
1154         while (p+8 < data + rec.data_len) {
1155                 uint32_t ofs, len;
1156                 if (DOCONV()) {
1157                         tdb_convert(p, 8);
1158                 }
1159                 memcpy(&ofs, p, 4);
1160                 memcpy(&len, p+4, 4);
1161
1162                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1163                         free(data);
1164                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1165                         tdb->ecode = TDB_ERR_IO;
1166                         return -1;
1167                 }
1168                 p += 8 + len;
1169         }
1170
1171         free(data);
1172
1173         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1174                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1175                 tdb->ecode = TDB_ERR_IO;
1176                 return -1;
1177         }
1178
1179         /* if the recovery area is after the recovered eof then remove it */
1180         if (recovery_eof <= recovery_head) {
1181                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1182                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1183                         tdb->ecode = TDB_ERR_IO;
1184                         return -1;                      
1185                 }
1186         }
1187
1188         /* remove the recovery magic */
1189         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1190                           &zero) == -1) {
1191                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1192                 tdb->ecode = TDB_ERR_IO;
1193                 return -1;                      
1194         }
1195
1196         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1197                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1198                 tdb->ecode = TDB_ERR_IO;
1199                 return -1;
1200         }
1201
1202         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n", 
1203                  recovery_eof));
1204
1205         /* all done */
1206         return 0;
1207 }
1208
1209 /* Any I/O failures we say "needs recovery". */
1210 bool tdb_needs_recovery(struct tdb_context *tdb)
1211 {
1212         tdb_off_t recovery_head;
1213         struct tdb_record rec;
1214
1215         /* find the recovery area */
1216         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1217                 return true;
1218         }
1219
1220         if (recovery_head == 0) {
1221                 /* we have never allocated a recovery record */
1222                 return false;
1223         }
1224
1225         /* read the recovery record */
1226         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1227                                    sizeof(rec), DOCONV()) == -1) {
1228                 return true;
1229         }
1230
1231         return (rec.magic == TDB_RECOVERY_MAGIC);
1232 }