tdb2: remove _PUBLIC_ in tdb1 functions.
[ccan] / ccan / tdb2 / tdb1_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 "tdb1_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     tdb1_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 tdb1_write() calls. The hooked
48     transaction versions of tdb1_read() and tdb1_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 tdb1_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 tdb1_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 TDB1_NOSYNC is passed to flags in tdb1_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 TDB1_ALLOW_NESTING is passed to flags in tdb open, or added using
89     tdb1_add_flags() transaction nesting is enabled.
90     It resets the TDB1_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 tdb1_transaction_commit() can be silently unrolled later.
96
97   - if TDB1_DISALLOW_NESTING is passed to flags in tdb open, or added using
98     tdb1_add_flags() transaction nesting is disabled.
99     It resets the TDB1_ALLOW_NESTING flag, as both cannot be used together.
100     An attempt create a nested transaction will fail with TDB1_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 tdb1_transaction {
110         /* we keep a mirrored copy of the tdb hash heads here so
111            tdb1_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 tdb1_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 tdb1_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         tdb1_off_t magic_offset;
137
138         /* old file size before transaction */
139         tdb1_len_t old_map_size;
140
141         /* did we expand in this transaction */
142         bool expanded;
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 transaction1_read(struct tdb1_context *tdb, tdb1_off_t off, void *buf,
151                              tdb1_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                 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
158                 if (transaction1_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->tdb1_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                 tdb1_convert(buf, len);
193         }
194         return 0;
195
196 fail:
197         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
198         tdb->ecode = TDB1_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 transaction1_write(struct tdb1_context *tdb, tdb1_off_t off,
208                              const void *buf, tdb1_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 = TDB1_ERR_EINVAL;
215                 TDB1_LOG((tdb, TDB1_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(tdb1_off_t) && off >= TDB1_FREELIST_TOP &&
223             off < TDB1_FREELIST_TOP+TDB1_HASHTABLE_SIZE(tdb)) {
224                 uint32_t chain = (off-TDB1_FREELIST_TOP) / sizeof(tdb1_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                 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
231                 if (transaction1_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 = TDB1_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 = TDB1_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                         tdb1_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->tdb1_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 = TDB1_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         TDB1_LOG((tdb, TDB1_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 transaction1_write_existing(struct tdb1_context *tdb, tdb1_off_t off,
323                                       const void *buf, tdb1_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                 tdb1_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
330                 if (transaction1_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 transaction1_next_hash_chain(struct tdb1_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 transaction1_oob(struct tdb1_context *tdb, tdb1_off_t len, int probe)
386 {
387         if (len <= tdb->map_size) {
388                 return 0;
389         }
390         tdb->ecode = TDB1_ERR_IO;
391         return -1;
392 }
393
394 /*
395   transaction version of tdb1_expand().
396 */
397 static int transaction1_expand_file(struct tdb1_context *tdb, tdb1_off_t size,
398                                     tdb1_off_t addition)
399 {
400         /* add a write to the transaction elements, so subsequent
401            reads see the zero data */
402         if (transaction1_write(tdb, size, NULL, addition) != 0) {
403                 return -1;
404         }
405
406         tdb->transaction->expanded = true;
407
408         return 0;
409 }
410
411 static const struct tdb1_methods transaction1_methods = {
412         transaction1_read,
413         transaction1_write,
414         transaction1_next_hash_chain,
415         transaction1_oob,
416         transaction1_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 tdb1_context
423 */
424 static int _tdb1_transaction_start(struct tdb1_context *tdb,
425                                   enum tdb1_lock_flags lockflags)
426 {
427         /* some sanity checks */
428         if (tdb->read_only || (tdb->flags & TDB1_INTERNAL) || tdb->traverse_read) {
429                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction on a read-only or internal db\n"));
430                 tdb->ecode = TDB1_ERR_EINVAL;
431                 return -1;
432         }
433
434         /* cope with nested tdb1_transaction_start() calls */
435         if (tdb->transaction != NULL) {
436                 if (!(tdb->flags & TDB1_ALLOW_NESTING)) {
437                         tdb->ecode = TDB1_ERR_NESTING;
438                         return -1;
439                 }
440                 tdb->transaction->nesting++;
441                 TDB1_LOG((tdb, TDB1_DEBUG_TRACE, "tdb1_transaction_start: nesting %d\n",
442                          tdb->transaction->nesting));
443                 return 0;
444         }
445
446         if (tdb1_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                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction with locks held\n"));
451                 tdb->ecode = TDB1_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                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_start: cannot start a transaction within a traverse\n"));
460                 tdb->ecode = TDB1_ERR_LOCK;
461                 return -1;
462         }
463
464         tdb->transaction = (struct tdb1_transaction *)
465                 calloc(sizeof(struct tdb1_transaction), 1);
466         if (tdb->transaction == NULL) {
467                 tdb->ecode = TDB1_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 (tdb1_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
478                 SAFE_FREE(tdb->transaction->blocks);
479                 SAFE_FREE(tdb->transaction);
480                 if ((lockflags & TDB1_LOCK_WAIT) == 0) {
481                         tdb->ecode = TDB1_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 (tdb1_allrecord_lock(tdb, F_RDLCK, TDB1_LOCK_WAIT, true) == -1) {
489                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_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 = TDB1_ERR_OOM;
499                 goto fail;
500         }
501         if (tdb->methods->tdb1_read(tdb, TDB1_FREELIST_TOP, tdb->transaction->hash_heads,
502                                    TDB1_HASHTABLE_SIZE(tdb), 0) != 0) {
503                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_start: failed to read hash heads\n"));
504                 tdb->ecode = TDB1_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->tdb1_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 = &transaction1_methods;
517
518         return 0;
519
520 fail:
521         tdb1_allrecord_unlock(tdb, F_RDLCK, false);
522 fail_allrecord_lock:
523         tdb1_transaction_unlock(tdb, F_WRLCK);
524         SAFE_FREE(tdb->transaction->blocks);
525         SAFE_FREE(tdb->transaction->hash_heads);
526         SAFE_FREE(tdb->transaction);
527         return -1;
528 }
529
530 int tdb1_transaction_start(struct tdb1_context *tdb)
531 {
532         return _tdb1_transaction_start(tdb, TDB1_LOCK_WAIT);
533 }
534
535 /*
536   sync to disk
537 */
538 static int transaction1_sync(struct tdb1_context *tdb, tdb1_off_t offset, tdb1_len_t length)
539 {
540         if (tdb->flags & TDB1_NOSYNC) {
541                 return 0;
542         }
543
544 #if HAVE_FDATASYNC
545         if (fdatasync(tdb->fd) != 0) {
546 #else
547         if (fsync(tdb->fd) != 0) {
548 #endif
549                 tdb->ecode = TDB1_ERR_IO;
550                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction: fsync failed\n"));
551                 return -1;
552         }
553 #if HAVE_MMAP
554         if (tdb->map_ptr) {
555                 tdb1_off_t moffset = offset & ~(tdb->page_size-1);
556                 if (msync(moffset + (char *)tdb->map_ptr,
557                           length + (offset - moffset), MS_SYNC) != 0) {
558                         tdb->ecode = TDB1_ERR_IO;
559                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction: msync failed - %s\n",
560                                  strerror(errno)));
561                         return -1;
562                 }
563         }
564 #endif
565         return 0;
566 }
567
568
569 static int _tdb1_transaction_cancel(struct tdb1_context *tdb)
570 {
571         int i, ret = 0;
572
573         if (tdb->transaction == NULL) {
574                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_cancel: no transaction\n"));
575                 return -1;
576         }
577
578         if (tdb->transaction->nesting != 0) {
579                 tdb->transaction->transaction_error = 1;
580                 tdb->transaction->nesting--;
581                 return 0;
582         }
583
584         tdb->map_size = tdb->transaction->old_map_size;
585
586         /* free all the transaction blocks */
587         for (i=0;i<tdb->transaction->num_blocks;i++) {
588                 if (tdb->transaction->blocks[i] != NULL) {
589                         free(tdb->transaction->blocks[i]);
590                 }
591         }
592         SAFE_FREE(tdb->transaction->blocks);
593
594         if (tdb->transaction->magic_offset) {
595                 const struct tdb1_methods *methods = tdb->transaction->io_methods;
596                 const uint32_t invalid = TDB1_RECOVERY_INVALID_MAGIC;
597
598                 /* remove the recovery marker */
599                 if (methods->tdb1_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
600                 transaction1_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
601                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_cancel: failed to remove recovery magic\n"));
602                         ret = -1;
603                 }
604         }
605
606         /* This also removes the OPEN_LOCK, if we have it. */
607         tdb1_release_transaction_locks(tdb);
608
609         /* restore the normal io methods */
610         tdb->methods = tdb->transaction->io_methods;
611
612         SAFE_FREE(tdb->transaction->hash_heads);
613         SAFE_FREE(tdb->transaction);
614
615         return ret;
616 }
617
618 /*
619   cancel the current transaction
620 */
621 int tdb1_transaction_cancel(struct tdb1_context *tdb)
622 {
623         return _tdb1_transaction_cancel(tdb);
624 }
625
626 /*
627   work out how much space the linearised recovery data will consume
628 */
629 static tdb1_len_t tdb1_recovery_size(struct tdb1_context *tdb)
630 {
631         tdb1_len_t recovery_size = 0;
632         int i;
633
634         recovery_size = sizeof(uint32_t);
635         for (i=0;i<tdb->transaction->num_blocks;i++) {
636                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
637                         break;
638                 }
639                 if (tdb->transaction->blocks[i] == NULL) {
640                         continue;
641                 }
642                 recovery_size += 2*sizeof(tdb1_off_t);
643                 if (i == tdb->transaction->num_blocks-1) {
644                         recovery_size += tdb->transaction->last_block_size;
645                 } else {
646                         recovery_size += tdb->transaction->block_size;
647                 }
648         }
649
650         return recovery_size;
651 }
652
653 int tdb1_recovery_area(struct tdb1_context *tdb,
654                       const struct tdb1_methods *methods,
655                       tdb1_off_t *recovery_offset,
656                       struct tdb1_record *rec)
657 {
658         if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, recovery_offset) == -1) {
659                 return -1;
660         }
661
662         if (*recovery_offset == 0) {
663                 rec->rec_len = 0;
664                 return 0;
665         }
666
667         if (methods->tdb1_read(tdb, *recovery_offset, rec, sizeof(*rec),
668                               TDB1_DOCONV()) == -1) {
669                 return -1;
670         }
671
672         /* ignore invalid recovery regions: can happen in crash */
673         if (rec->magic != TDB1_RECOVERY_MAGIC &&
674             rec->magic != TDB1_RECOVERY_INVALID_MAGIC) {
675                 *recovery_offset = 0;
676                 rec->rec_len = 0;
677         }
678         return 0;
679 }
680
681 /*
682   allocate the recovery area, or use an existing recovery area if it is
683   large enough
684 */
685 static int tdb1_recovery_allocate(struct tdb1_context *tdb,
686                                  tdb1_len_t *recovery_size,
687                                  tdb1_off_t *recovery_offset,
688                                  tdb1_len_t *recovery_max_size)
689 {
690         struct tdb1_record rec;
691         const struct tdb1_methods *methods = tdb->transaction->io_methods;
692         tdb1_off_t recovery_head;
693
694         if (tdb1_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
695                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to read recovery head\n"));
696                 return -1;
697         }
698
699         *recovery_size = tdb1_recovery_size(tdb);
700
701         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
702                 /* it fits in the existing area */
703                 *recovery_max_size = rec.rec_len;
704                 *recovery_offset = recovery_head;
705                 return 0;
706         }
707
708         /* we need to free up the old recovery area, then allocate a
709            new one at the end of the file. Note that we cannot use
710            tdb1_allocate() to allocate the new one as that might return
711            us an area that is being currently used (as of the start of
712            the transaction) */
713         if (recovery_head != 0) {
714                 if (tdb1_free(tdb, recovery_head, &rec) == -1) {
715                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to free previous recovery area\n"));
716                         return -1;
717                 }
718         }
719
720         /* the tdb1_free() call might have increased the recovery size */
721         *recovery_size = tdb1_recovery_size(tdb);
722
723         /* round up to a multiple of page size */
724         *recovery_max_size = TDB1_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
725         *recovery_offset = tdb->map_size;
726         recovery_head = *recovery_offset;
727
728         if (methods->tdb1_expand_file(tdb, tdb->transaction->old_map_size,
729                                      (tdb->map_size - tdb->transaction->old_map_size) +
730                                      sizeof(rec) + *recovery_max_size) == -1) {
731                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to create recovery area\n"));
732                 return -1;
733         }
734
735         /* remap the file (if using mmap) */
736         methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
737
738         /* we have to reset the old map size so that we don't try to expand the file
739            again in the transaction commit, which would destroy the recovery area */
740         tdb->transaction->old_map_size = tdb->map_size;
741
742         /* write the recovery header offset and sync - we can sync without a race here
743            as the magic ptr in the recovery record has not been set */
744         TDB1_CONV(recovery_head);
745         if (methods->tdb1_write(tdb, TDB1_RECOVERY_HEAD,
746                                &recovery_head, sizeof(tdb1_off_t)) == -1) {
747                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to write recovery head\n"));
748                 return -1;
749         }
750         if (transaction1_write_existing(tdb, TDB1_RECOVERY_HEAD, &recovery_head, sizeof(tdb1_off_t)) == -1) {
751                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_recovery_allocate: failed to write recovery head\n"));
752                 return -1;
753         }
754
755         return 0;
756 }
757
758
759 /*
760   setup the recovery data that will be used on a crash during commit
761 */
762 static int transaction1_setup_recovery(struct tdb1_context *tdb,
763                                        tdb1_off_t *magic_offset)
764 {
765         tdb1_len_t recovery_size;
766         unsigned char *data, *p;
767         const struct tdb1_methods *methods = tdb->transaction->io_methods;
768         struct tdb1_record *rec;
769         tdb1_off_t recovery_offset, recovery_max_size;
770         tdb1_off_t old_map_size = tdb->transaction->old_map_size;
771         uint32_t magic, tailer;
772         int i;
773
774         /*
775           check that the recovery area has enough space
776         */
777         if (tdb1_recovery_allocate(tdb, &recovery_size,
778                                   &recovery_offset, &recovery_max_size) == -1) {
779                 return -1;
780         }
781
782         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
783         if (data == NULL) {
784                 tdb->ecode = TDB1_ERR_OOM;
785                 return -1;
786         }
787
788         rec = (struct tdb1_record *)data;
789         memset(rec, 0, sizeof(*rec));
790
791         rec->magic    = TDB1_RECOVERY_INVALID_MAGIC;
792         rec->data_len = recovery_size;
793         rec->rec_len  = recovery_max_size;
794         rec->key_len  = old_map_size;
795         TDB1_CONV(*rec);
796
797         /* build the recovery data into a single blob to allow us to do a single
798            large write, which should be more efficient */
799         p = data + sizeof(*rec);
800         for (i=0;i<tdb->transaction->num_blocks;i++) {
801                 tdb1_off_t offset;
802                 tdb1_len_t length;
803
804                 if (tdb->transaction->blocks[i] == NULL) {
805                         continue;
806                 }
807
808                 offset = i * tdb->transaction->block_size;
809                 length = tdb->transaction->block_size;
810                 if (i == tdb->transaction->num_blocks-1) {
811                         length = tdb->transaction->last_block_size;
812                 }
813
814                 if (offset >= old_map_size) {
815                         continue;
816                 }
817                 if (offset + length > tdb->transaction->old_map_size) {
818                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: transaction data over new region boundary\n"));
819                         free(data);
820                         tdb->ecode = TDB1_ERR_CORRUPT;
821                         return -1;
822                 }
823                 memcpy(p, &offset, 4);
824                 memcpy(p+4, &length, 4);
825                 if (TDB1_DOCONV()) {
826                         tdb1_convert(p, 8);
827                 }
828                 /* the recovery area contains the old data, not the
829                    new data, so we have to call the original tdb1_read
830                    method to get it */
831                 if (methods->tdb1_read(tdb, offset, p + 8, length, 0) != 0) {
832                         free(data);
833                         tdb->ecode = TDB1_ERR_IO;
834                         return -1;
835                 }
836                 p += 8 + length;
837         }
838
839         /* and the tailer */
840         tailer = sizeof(*rec) + recovery_max_size;
841         memcpy(p, &tailer, 4);
842         if (TDB1_DOCONV()) {
843                 tdb1_convert(p, 4);
844         }
845
846         /* write the recovery data to the recovery area */
847         if (methods->tdb1_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
848                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write recovery data\n"));
849                 free(data);
850                 tdb->ecode = TDB1_ERR_IO;
851                 return -1;
852         }
853         if (transaction1_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
854                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write secondary recovery data\n"));
855                 free(data);
856                 tdb->ecode = TDB1_ERR_IO;
857                 return -1;
858         }
859
860         /* as we don't have ordered writes, we have to sync the recovery
861            data before we update the magic to indicate that the recovery
862            data is present */
863         if (transaction1_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
864                 free(data);
865                 return -1;
866         }
867
868         free(data);
869
870         magic = TDB1_RECOVERY_MAGIC;
871         TDB1_CONV(magic);
872
873         *magic_offset = recovery_offset + offsetof(struct tdb1_record, magic);
874
875         if (methods->tdb1_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
876                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write recovery magic\n"));
877                 tdb->ecode = TDB1_ERR_IO;
878                 return -1;
879         }
880         if (transaction1_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
881                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_setup_recovery: failed to write secondary recovery magic\n"));
882                 tdb->ecode = TDB1_ERR_IO;
883                 return -1;
884         }
885
886         /* ensure the recovery magic marker is on disk */
887         if (transaction1_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
888                 return -1;
889         }
890
891         return 0;
892 }
893
894 static int _tdb1_transaction_prepare_commit(struct tdb1_context *tdb)
895 {
896         const struct tdb1_methods *methods;
897
898         if (tdb->transaction == NULL) {
899                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: no transaction\n"));
900                 return -1;
901         }
902
903         if (tdb->transaction->prepared) {
904                 tdb->ecode = TDB1_ERR_EINVAL;
905                 _tdb1_transaction_cancel(tdb);
906                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: transaction already prepared\n"));
907                 return -1;
908         }
909
910         if (tdb->transaction->transaction_error) {
911                 tdb->ecode = TDB1_ERR_IO;
912                 _tdb1_transaction_cancel(tdb);
913                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: transaction error pending\n"));
914                 return -1;
915         }
916
917
918         if (tdb->transaction->nesting != 0) {
919                 return 0;
920         }
921
922         /* check for a null transaction */
923         if (tdb->transaction->blocks == NULL) {
924                 return 0;
925         }
926
927         methods = tdb->transaction->io_methods;
928
929         /* if there are any locks pending then the caller has not
930            nested their locks properly, so fail the transaction */
931         if (tdb1_have_extra_locks(tdb)) {
932                 tdb->ecode = TDB1_ERR_LOCK;
933                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: locks pending on commit\n"));
934                 _tdb1_transaction_cancel(tdb);
935                 return -1;
936         }
937
938         /* upgrade the main transaction lock region to a write lock */
939         if (tdb1_allrecord_upgrade(tdb) == -1) {
940                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: failed to upgrade hash locks\n"));
941                 _tdb1_transaction_cancel(tdb);
942                 return -1;
943         }
944
945         /* get the open lock - this prevents new users attaching to the database
946            during the commit */
947         if (tdb1_nest_lock(tdb, TDB1_OPEN_LOCK, F_WRLCK, TDB1_LOCK_WAIT) == -1) {
948                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_prepare_commit: failed to get open lock\n"));
949                 _tdb1_transaction_cancel(tdb);
950                 return -1;
951         }
952
953         if (!(tdb->flags & TDB1_NOSYNC)) {
954                 /* write the recovery data to the end of the file */
955                 if (transaction1_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
956                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_prepare_commit: failed to setup recovery data\n"));
957                         _tdb1_transaction_cancel(tdb);
958                         return -1;
959                 }
960         }
961
962         tdb->transaction->prepared = true;
963
964         /* expand the file to the new size if needed */
965         if (tdb->map_size != tdb->transaction->old_map_size) {
966                 if (methods->tdb1_expand_file(tdb, tdb->transaction->old_map_size,
967                                              tdb->map_size -
968                                              tdb->transaction->old_map_size) == -1) {
969                         tdb->ecode = TDB1_ERR_IO;
970                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_prepare_commit: expansion failed\n"));
971                         _tdb1_transaction_cancel(tdb);
972                         return -1;
973                 }
974                 tdb->map_size = tdb->transaction->old_map_size;
975                 methods->tdb1_oob(tdb, tdb->map_size + 1, 1);
976         }
977
978         /* Keep the open lock until the actual commit */
979
980         return 0;
981 }
982
983 /*
984    prepare to commit the current transaction
985 */
986 int tdb1_transaction_prepare_commit(struct tdb1_context *tdb)
987 {
988         return _tdb1_transaction_prepare_commit(tdb);
989 }
990
991 /* A repack is worthwhile if the largest is less than half total free. */
992 static bool repack_worthwhile(struct tdb1_context *tdb)
993 {
994         tdb1_off_t ptr;
995         struct tdb1_record rec;
996         tdb1_len_t total = 0, largest = 0;
997
998         if (tdb1_ofs_read(tdb, TDB1_FREELIST_TOP, &ptr) == -1) {
999                 return false;
1000         }
1001
1002         while (ptr != 0 && tdb1_rec_free_read(tdb, ptr, &rec) == 0) {
1003                 total += rec.rec_len;
1004                 if (rec.rec_len > largest) {
1005                         largest = rec.rec_len;
1006                 }
1007                 ptr = rec.next;
1008         }
1009
1010         return total > largest * 2;
1011 }
1012
1013 /*
1014   commit the current transaction
1015 */
1016 int tdb1_transaction_commit(struct tdb1_context *tdb)
1017 {
1018         const struct tdb1_methods *methods;
1019         int i;
1020         bool need_repack = false;
1021
1022         if (tdb->transaction == NULL) {
1023                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_commit: no transaction\n"));
1024                 return -1;
1025         }
1026
1027         if (tdb->transaction->transaction_error) {
1028                 tdb->ecode = TDB1_ERR_IO;
1029                 _tdb1_transaction_cancel(tdb);
1030                 TDB1_LOG((tdb, TDB1_DEBUG_ERROR, "tdb1_transaction_commit: transaction error pending\n"));
1031                 return -1;
1032         }
1033
1034
1035         if (tdb->transaction->nesting != 0) {
1036                 tdb->transaction->nesting--;
1037                 return 0;
1038         }
1039
1040         /* check for a null transaction */
1041         if (tdb->transaction->blocks == NULL) {
1042                 _tdb1_transaction_cancel(tdb);
1043                 return 0;
1044         }
1045
1046         if (!tdb->transaction->prepared) {
1047                 int ret = _tdb1_transaction_prepare_commit(tdb);
1048                 if (ret)
1049                         return ret;
1050         }
1051
1052         methods = tdb->transaction->io_methods;
1053
1054         /* perform all the writes */
1055         for (i=0;i<tdb->transaction->num_blocks;i++) {
1056                 tdb1_off_t offset;
1057                 tdb1_len_t length;
1058
1059                 if (tdb->transaction->blocks[i] == NULL) {
1060                         continue;
1061                 }
1062
1063                 offset = i * tdb->transaction->block_size;
1064                 length = tdb->transaction->block_size;
1065                 if (i == tdb->transaction->num_blocks-1) {
1066                         length = tdb->transaction->last_block_size;
1067                 }
1068
1069                 if (methods->tdb1_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1070                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_commit: write failed during commit\n"));
1071
1072                         /* we've overwritten part of the data and
1073                            possibly expanded the file, so we need to
1074                            run the crash recovery code */
1075                         tdb->methods = methods;
1076                         tdb1_transaction_recover(tdb);
1077
1078                         _tdb1_transaction_cancel(tdb);
1079
1080                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_commit: write failed\n"));
1081                         return -1;
1082                 }
1083                 SAFE_FREE(tdb->transaction->blocks[i]);
1084         }
1085
1086         /* Do this before we drop lock or blocks. */
1087         if (tdb->transaction->expanded) {
1088                 need_repack = repack_worthwhile(tdb);
1089         }
1090
1091         SAFE_FREE(tdb->transaction->blocks);
1092         tdb->transaction->num_blocks = 0;
1093
1094         /* ensure the new data is on disk */
1095         if (transaction1_sync(tdb, 0, tdb->map_size) == -1) {
1096                 return -1;
1097         }
1098
1099         /*
1100           TODO: maybe write to some dummy hdr field, or write to magic
1101           offset without mmap, before the last sync, instead of the
1102           utime() call
1103         */
1104
1105         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1106            don't change the mtime of the file, this means the file may
1107            not be backed up (as tdb rounding to block sizes means that
1108            file size changes are quite rare too). The following forces
1109            mtime changes when a transaction completes */
1110 #if HAVE_UTIME
1111         utime(tdb->name, NULL);
1112 #endif
1113
1114         /* use a transaction cancel to free memory and remove the
1115            transaction locks */
1116         _tdb1_transaction_cancel(tdb);
1117
1118         if (need_repack) {
1119                 return tdb1_repack(tdb);
1120         }
1121
1122         return 0;
1123 }
1124
1125
1126 /*
1127   recover from an aborted transaction. Must be called with exclusive
1128   database write access already established (including the open
1129   lock to prevent new processes attaching)
1130 */
1131 int tdb1_transaction_recover(struct tdb1_context *tdb)
1132 {
1133         tdb1_off_t recovery_head, recovery_eof;
1134         unsigned char *data, *p;
1135         uint32_t zero = 0;
1136         struct tdb1_record rec;
1137
1138         /* find the recovery area */
1139         if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1140                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery head\n"));
1141                 tdb->ecode = TDB1_ERR_IO;
1142                 return -1;
1143         }
1144
1145         if (recovery_head == 0) {
1146                 /* we have never allocated a recovery record */
1147                 return 0;
1148         }
1149
1150         /* read the recovery record */
1151         if (tdb->methods->tdb1_read(tdb, recovery_head, &rec,
1152                                    sizeof(rec), TDB1_DOCONV()) == -1) {
1153                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery record\n"));
1154                 tdb->ecode = TDB1_ERR_IO;
1155                 return -1;
1156         }
1157
1158         if (rec.magic != TDB1_RECOVERY_MAGIC) {
1159                 /* there is no valid recovery data */
1160                 return 0;
1161         }
1162
1163         if (tdb->read_only) {
1164                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: attempt to recover read only database\n"));
1165                 tdb->ecode = TDB1_ERR_CORRUPT;
1166                 return -1;
1167         }
1168
1169         recovery_eof = rec.key_len;
1170
1171         data = (unsigned char *)malloc(rec.data_len);
1172         if (data == NULL) {
1173                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to allocate recovery data\n"));
1174                 tdb->ecode = TDB1_ERR_OOM;
1175                 return -1;
1176         }
1177
1178         /* read the full recovery data */
1179         if (tdb->methods->tdb1_read(tdb, recovery_head + sizeof(rec), data,
1180                                    rec.data_len, 0) == -1) {
1181                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to read recovery data\n"));
1182                 tdb->ecode = TDB1_ERR_IO;
1183                 return -1;
1184         }
1185
1186         /* recover the file data */
1187         p = data;
1188         while (p+8 < data + rec.data_len) {
1189                 uint32_t ofs, len;
1190                 if (TDB1_DOCONV()) {
1191                         tdb1_convert(p, 8);
1192                 }
1193                 memcpy(&ofs, p, 4);
1194                 memcpy(&len, p+4, 4);
1195
1196                 if (tdb->methods->tdb1_write(tdb, ofs, p+8, len) == -1) {
1197                         free(data);
1198                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1199                         tdb->ecode = TDB1_ERR_IO;
1200                         return -1;
1201                 }
1202                 p += 8 + len;
1203         }
1204
1205         free(data);
1206
1207         if (transaction1_sync(tdb, 0, tdb->map_size) == -1) {
1208                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to sync recovery\n"));
1209                 tdb->ecode = TDB1_ERR_IO;
1210                 return -1;
1211         }
1212
1213         /* if the recovery area is after the recovered eof then remove it */
1214         if (recovery_eof <= recovery_head) {
1215                 if (tdb1_ofs_write(tdb, TDB1_RECOVERY_HEAD, &zero) == -1) {
1216                         TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to remove recovery head\n"));
1217                         tdb->ecode = TDB1_ERR_IO;
1218                         return -1;
1219                 }
1220         }
1221
1222         /* remove the recovery magic */
1223         if (tdb1_ofs_write(tdb, recovery_head + offsetof(struct tdb1_record, magic),
1224                           &zero) == -1) {
1225                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to remove recovery magic\n"));
1226                 tdb->ecode = TDB1_ERR_IO;
1227                 return -1;
1228         }
1229
1230         if (transaction1_sync(tdb, 0, recovery_eof) == -1) {
1231                 TDB1_LOG((tdb, TDB1_DEBUG_FATAL, "tdb1_transaction_recover: failed to sync2 recovery\n"));
1232                 tdb->ecode = TDB1_ERR_IO;
1233                 return -1;
1234         }
1235
1236         TDB1_LOG((tdb, TDB1_DEBUG_TRACE, "tdb1_transaction_recover: recovered %d byte database\n",
1237                  recovery_eof));
1238
1239         /* all done */
1240         return 0;
1241 }
1242
1243 /* Any I/O failures we say "needs recovery". */
1244 bool tdb1_needs_recovery(struct tdb1_context *tdb)
1245 {
1246         tdb1_off_t recovery_head;
1247         struct tdb1_record rec;
1248
1249         /* find the recovery area */
1250         if (tdb1_ofs_read(tdb, TDB1_RECOVERY_HEAD, &recovery_head) == -1) {
1251                 return true;
1252         }
1253
1254         if (recovery_head == 0) {
1255                 /* we have never allocated a recovery record */
1256                 return false;
1257         }
1258
1259         /* read the recovery record */
1260         if (tdb->methods->tdb1_read(tdb, recovery_head, &rec,
1261                                    sizeof(rec), TDB1_DOCONV()) == -1) {
1262                 return true;
1263         }
1264
1265         return (rec.magic == TDB1_RECOVERY_MAGIC);
1266 }