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 fsync/msync calls are made.  This means we
  86     are still proof against a process dying during transaction commit,
  87     but not against machine reboot.
  88
  89   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
  90     tdb_add_flags() transaction nesting is enabled.
  91     It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
  92     The default is that transaction nesting is allowed.
  93     Note: this default may change in future versions of tdb.
  94
  95     Beware. when transactions are nested a transaction successfully
  96     completed with tdb_transaction_commit() can be silently unrolled later.
  97
  98   - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
  99     tdb_add_flags() transaction nesting is disabled.
 100     It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
 101     An attempt create a nested transaction will fail with TDB_ERR_NESTING.
 102     The default is that transaction nesting is allowed.
 103     Note: this default may change in future versions of tdb.
 104 */
 105
 106
 107 /*
 108   hold the context of any current transaction
 109 */
 110 struct tdb_transaction {
 111         /* we keep a mirrored copy of the tdb hash heads here so
 112            tdb_next_hash_chain() can operate efficiently */
 113         uint32_t *hash_heads;
 114
 115         /* the original io methods - used to do IOs to the real db */
 116         const struct tdb_methods *io_methods;
 117
 118         /* the list of transaction blocks. When a block is first
 119            written to, it gets created in this list */
 120         uint8_t **blocks;
 121         uint32_t num_blocks;
 122         uint32_t block_size;      /* bytes in each block */
 123         uint32_t last_block_size; /* number of valid bytes in the last block */
 124
 125         /* non-zero when an internal transaction error has
 126            occurred. All write operations will then fail until the
 127            transaction is ended */
 128         int transaction_error;
 129
 130         /* when inside a transaction we need to keep track of any
 131            nested tdb_transaction_start() calls, as these are allowed,
 132            but don't create a new transaction */
 133         int nesting;
 134
 135         /* set when a prepare has already occurred */
 136         bool prepared;
 137         tdb_off_t magic_offset;
 138
 139         /* old file size before transaction */
 140         tdb_len_t old_map_size;
 141
 142         /* did we expand in this transaction */
 143         bool expanded;
 144 };
 145
 146
 147 /*
 148   read while in a transaction. We need to check first if the data is in our list
 149   of transaction elements, then if not do a real read
 150 */
 151 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
 152                             tdb_len_t len, int cv)
 153 {
 154         uint32_t blk;
 155
 156         /* break it down into block sized ops */
 157         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 158                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 159                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
 160                         return -1;
 161                 }
 162                 len -= len2;
 163                 off += len2;
 164                 buf = (void *)(len2 + (char *)buf);
 165         }
 166
 167         if (len == 0) {
 168                 return 0;
 169         }
 170
 171         blk = off / tdb->transaction->block_size;
 172
 173         /* see if we have it in the block list */
 174         if (tdb->transaction->num_blocks <= blk ||
 175             tdb->transaction->blocks[blk] == NULL) {
 176                 /* nope, do a real read */
 177                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
 178                         goto fail;
 179                 }
 180                 return 0;
 181         }
 182
 183         /* it is in the block list. Now check for the last block */
 184         if (blk == tdb->transaction->num_blocks-1) {
 185                 if (len > tdb->transaction->last_block_size) {
 186                         goto fail;
 187                 }
 188         }
 189
 190         /* now copy it out of this block */
 191         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
 192         if (cv) {
 193                 tdb_convert(buf, len);
 194         }
 195         return 0;
 196
 197 fail:
 198         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%u len=%u\n", off, len));
 199         tdb->ecode = TDB_ERR_IO;
 200         tdb->transaction->transaction_error = 1;
 201         return -1;
 202 }
 203
 204
 205 /*
 206   write while in a transaction
 207 */
 208 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
 209                              const void *buf, tdb_len_t len)
 210 {
 211         uint32_t blk;
 212
 213         /* Only a commit is allowed on a prepared transaction */
 214         if (tdb->transaction->prepared) {
 215                 tdb->ecode = TDB_ERR_EINVAL;
 216                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
 217                 tdb->transaction->transaction_error = 1;
 218                 return -1;
 219         }
 220
 221         /* if the write is to a hash head, then update the transaction
 222            hash heads */
 223         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
 224             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
 225                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
 226                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
 227         }
 228
 229         /* break it up into block sized chunks */
 230         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 231                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 232                 if (transaction_write(tdb, off, buf, len2) != 0) {
 233                         return -1;
 234                 }
 235                 len -= len2;
 236                 off += len2;
 237                 if (buf != NULL) {
 238                         buf = (const void *)(len2 + (const char *)buf);
 239                 }
 240         }
 241
 242         if (len == 0) {
 243                 return 0;
 244         }
 245
 246         blk = off / tdb->transaction->block_size;
 247         off = off % tdb->transaction->block_size;
 248
 249         if (tdb->transaction->num_blocks <= blk) {
 250                 uint8_t **new_blocks;
 251                 /* expand the blocks array */
 252                 new_blocks = (uint8_t **)realloc(tdb->transaction->blocks,
 253                                                  (blk+1)*sizeof(uint8_t *));
 254                 if (new_blocks == NULL) {
 255                         tdb->ecode = TDB_ERR_OOM;
 256                         goto fail;
 257                 }
 258                 memset(&new_blocks[tdb->transaction->num_blocks], 0,
 259                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
 260                 tdb->transaction->blocks = new_blocks;
 261                 tdb->transaction->num_blocks = blk+1;
 262                 tdb->transaction->last_block_size = 0;
 263         }
 264
 265         /* allocate and fill a block? */
 266         if (tdb->transaction->blocks[blk] == NULL) {
 267                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
 268                 if (tdb->transaction->blocks[blk] == NULL) {
 269                         tdb->ecode = TDB_ERR_OOM;
 270                         tdb->transaction->transaction_error = 1;
 271                         return -1;
 272                 }
 273                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
 274                         tdb_len_t len2 = tdb->transaction->block_size;
 275                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
 276                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
 277                         }
 278                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
 279                                                                    tdb->transaction->blocks[blk],
 280                                                                    len2, 0) != 0) {
 281                                 SAFE_FREE(tdb->transaction->blocks[blk]);
 282                                 tdb->ecode = TDB_ERR_IO;
 283                                 goto fail;
 284                         }
 285                         if (blk == tdb->transaction->num_blocks-1) {
 286                                 tdb->transaction->last_block_size = len2;
 287                         }
 288                 }
 289         }
 290
 291         /* overwrite part of an existing block */
 292         if (buf == NULL) {
 293                 memset(tdb->transaction->blocks[blk] + off, 0, len);
 294         } else {
 295                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 296         }
 297         if (blk == tdb->transaction->num_blocks-1) {
 298                 if (len + off > tdb->transaction->last_block_size) {
 299                         tdb->transaction->last_block_size = len + off;
 300                 }
 301         }
 302
 303         return 0;
 304
 305 fail:
 306         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%u len=%u\n",
 307                  (blk*tdb->transaction->block_size) + off, len));
 308         tdb->transaction->transaction_error = 1;
 309         return -1;
 310 }
 311
 312
 313 /*
 314   write while in a transaction - this variant never expands the transaction blocks, it only
 315   updates existing blocks. This means it cannot change the recovery size
 316 */
 317 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
 318                                       const void *buf, tdb_len_t len)
 319 {
 320         uint32_t blk;
 321
 322         /* break it up into block sized chunks */
 323         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 324                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 325                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
 326                         return -1;
 327                 }
 328                 len -= len2;
 329                 off += len2;
 330                 if (buf != NULL) {
 331                         buf = (const void *)(len2 + (const char *)buf);
 332                 }
 333         }
 334
 335         if (len == 0) {
 336                 return 0;
 337         }
 338
 339         blk = off / tdb->transaction->block_size;
 340         off = off % tdb->transaction->block_size;
 341
 342         if (tdb->transaction->num_blocks <= blk ||
 343             tdb->transaction->blocks[blk] == NULL) {
 344                 return 0;
 345         }
 346
 347         if (blk == tdb->transaction->num_blocks-1 &&
 348             off + len > tdb->transaction->last_block_size) {
 349                 if (off >= tdb->transaction->last_block_size) {
 350                         return 0;
 351                 }
 352                 len = tdb->transaction->last_block_size - off;
 353         }
 354
 355         /* overwrite part of an existing block */
 356         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 357
 358         return 0;
 359 }
 360
 361
 362 /*
 363   accelerated hash chain head search, using the cached hash heads
 364 */
 365 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
 366 {
 367         uint32_t h = *chain;
 368         for (;h < tdb->hash_size;h++) {
 369                 /* the +1 takes account of the freelist */
 370                 if (0 != tdb->transaction->hash_heads[h+1]) {
 371                         break;
 372                 }
 373         }
 374         (*chain) = h;
 375 }
 376
 377 /*
 378   out of bounds check during a transaction
 379 */
 380 static int transaction_oob(struct tdb_context *tdb, tdb_off_t off,
 381                            tdb_len_t len, int probe)
 382 {
 383         if (off + len >= off && off + len <= tdb->map_size) {
 384                 return 0;
 385         }
 386         tdb->ecode = TDB_ERR_IO;
 387         return -1;
 388 }
 389
 390 /*
 391   transaction version of tdb_expand().
 392 */
 393 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
 394                                    tdb_off_t addition)
 395 {
 396         /* add a write to the transaction elements, so subsequent
 397            reads see the zero data */
 398         if (transaction_write(tdb, size, NULL, addition) != 0) {
 399                 return -1;
 400         }
 401
 402         tdb->transaction->expanded = true;
 403
 404         return 0;
 405 }
 406
 407 static const struct tdb_methods transaction_methods = {
 408         transaction_read,
 409         transaction_write,
 410         transaction_next_hash_chain,
 411         transaction_oob,
 412         transaction_expand_file,
 413 };
 414
 415
 416 /*
 417   start a tdb transaction. No token is returned, as only a single
 418   transaction is allowed to be pending per tdb_context
 419 */
 420 static int _tdb_transaction_start(struct tdb_context *tdb,
 421                                   enum tdb_lock_flags lockflags)
 422 {
 423         /* some sanity checks */
 424         if (tdb->read_only || (tdb->flags & TDB_INTERNAL)
 425             || tdb->traverse_read) {
 426                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
 427                 tdb->ecode = TDB_ERR_EINVAL;
 428                 return -1;
 429         }
 430
 431         /* cope with nested tdb_transaction_start() calls */
 432         if (tdb->transaction != NULL) {
 433                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
 434                         tdb->ecode = TDB_ERR_NESTING;
 435                         return -1;
 436                 }
 437                 tdb->transaction->nesting++;
 438                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
 439                          tdb->transaction->nesting));
 440                 return 0;
 441         }
 442
 443         if (tdb_have_extra_locks(tdb)) {
 444                 /* the caller must not have any locks when starting a
 445                    transaction as otherwise we'll be screwed by lack
 446                    of nested locks in posix */
 447                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
 448                 tdb->ecode = TDB_ERR_LOCK;
 449                 return -1;
 450         }
 451
 452         if (tdb->travlocks.next != NULL) {
 453                 /* you cannot use transactions inside a traverse (although you can use
 454                    traverse inside a transaction) as otherwise you can end up with
 455                    deadlock */
 456                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
 457                 tdb->ecode = TDB_ERR_LOCK;
 458                 return -1;
 459         }
 460
 461         tdb->transaction = (struct tdb_transaction *)
 462                 calloc(sizeof(struct tdb_transaction), 1);
 463         if (tdb->transaction == NULL) {
 464                 tdb->ecode = TDB_ERR_OOM;
 465                 return -1;
 466         }
 467
 468         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
 469         tdb->transaction->block_size = tdb->page_size;
 470
 471         /* get the transaction write lock. This is a blocking lock. As
 472            discussed with Volker, there are a number of ways we could
 473            make this async, which we will probably do in the future */
 474         if (tdb_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
 475                 SAFE_FREE(tdb->transaction->blocks);
 476                 SAFE_FREE(tdb->transaction);
 477                 if ((lockflags & TDB_LOCK_WAIT) == 0) {
 478                         tdb->ecode = TDB_ERR_NOLOCK;
 479                 } else {
 480                         TDB_LOG((tdb, TDB_DEBUG_ERROR,
 481                                  "tdb_transaction_start: "
 482                                  "failed to get transaction lock\n"));
 483                 }
 484                 return -1;
 485         }
 486
 487         /* get a read lock from the freelist to the end of file. This
 488            is upgraded to a write lock during the commit */
 489         if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
 490                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
 491                 goto fail_allrecord_lock;
 492         }
 493
 494         /* setup a copy of the hash table heads so the hash scan in
 495            traverse can be fast */
 496         tdb->transaction->hash_heads = (uint32_t *)
 497                 calloc(tdb->hash_size+1, sizeof(uint32_t));
 498         if (tdb->transaction->hash_heads == NULL) {
 499                 tdb->ecode = TDB_ERR_OOM;
 500                 goto fail;
 501         }
 502         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
 503                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
 504                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
 505                 tdb->ecode = TDB_ERR_IO;
 506                 goto fail;
 507         }
 508
 509         /* make sure we know about any file expansions already done by
 510            anyone else */
 511         tdb->methods->tdb_oob(tdb, tdb->map_size, 1, 1);
 512         tdb->transaction->old_map_size = tdb->map_size;
 513
 514         /* finally hook the io methods, replacing them with
 515            transaction specific methods */
 516         tdb->transaction->io_methods = tdb->methods;
 517         tdb->methods = &transaction_methods;
 518
 519         /* Trace at the end, so we get sequence number correct. */
 520         tdb_trace(tdb, "tdb_transaction_start");
 521         return 0;
 522
 523 fail:
 524         tdb_allrecord_unlock(tdb, F_RDLCK, false);
 525 fail_allrecord_lock:
 526         tdb_transaction_unlock(tdb, F_WRLCK);
 527         SAFE_FREE(tdb->transaction->blocks);
 528         SAFE_FREE(tdb->transaction->hash_heads);
 529         SAFE_FREE(tdb->transaction);
 530         return -1;
 531 }
 532
 533 _PUBLIC_ int tdb_transaction_start(struct tdb_context *tdb)
 534 {
 535         return _tdb_transaction_start(tdb, TDB_LOCK_WAIT);
 536 }
 537
 538 _PUBLIC_ int tdb_transaction_start_nonblock(struct tdb_context *tdb)
 539 {
 540         return _tdb_transaction_start(tdb, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
 541 }
 542
 543 /*
 544   sync to disk
 545 */
 546 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
 547 {
 548         if (tdb->flags & TDB_NOSYNC) {
 549                 return 0;
 550         }
 551
 552 #ifdef HAVE_FDATASYNC
 553         if (fdatasync(tdb->fd) != 0) {
 554 #else
 555         if (fsync(tdb->fd) != 0) {
 556 #endif
 557                 tdb->ecode = TDB_ERR_IO;
 558                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
 559                 return -1;
 560         }
 561 #ifdef HAVE_MMAP
 562         if (tdb->map_ptr) {
 563                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
 564                 if (msync(moffset + (char *)tdb->map_ptr,
 565                           length + (offset - moffset), MS_SYNC) != 0) {
 566                         tdb->ecode = TDB_ERR_IO;
 567                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
 568                                  strerror(errno)));
 569                         return -1;
 570                 }
 571         }
 572 #endif
 573         return 0;
 574 }
 575
 576
 577 static int _tdb_transaction_cancel(struct tdb_context *tdb)
 578 {
 579         int i, ret = 0;
 580
 581         if (tdb->transaction == NULL) {
 582                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
 583                 return -1;
 584         }
 585
 586         if (tdb->transaction->nesting != 0) {
 587                 tdb->transaction->transaction_error = 1;
 588                 tdb->transaction->nesting--;
 589                 return 0;
 590         }
 591
 592         tdb->map_size = tdb->transaction->old_map_size;
 593
 594         /* free all the transaction blocks */
 595         for (i=0;i<tdb->transaction->num_blocks;i++) {
 596                 if (tdb->transaction->blocks[i] != NULL) {
 597                         free(tdb->transaction->blocks[i]);
 598                 }
 599         }
 600         SAFE_FREE(tdb->transaction->blocks);
 601
 602         if (tdb->transaction->magic_offset) {
 603                 const struct tdb_methods *methods = tdb->transaction->io_methods;
 604                 const uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
 605
 606                 /* remove the recovery marker */
 607                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
 608                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
 609                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
 610                         ret = -1;
 611                 }
 612         }
 613
 614         /* This also removes the OPEN_LOCK, if we have it. */
 615         tdb_release_transaction_locks(tdb);
 616
 617         /* restore the normal io methods */
 618         tdb->methods = tdb->transaction->io_methods;
 619
 620         SAFE_FREE(tdb->transaction->hash_heads);
 621         SAFE_FREE(tdb->transaction);
 622
 623         return ret;
 624 }
 625
 626 /*
 627   cancel the current transaction
 628 */
 629 _PUBLIC_ int tdb_transaction_cancel(struct tdb_context *tdb)
 630 {
 631         tdb_trace(tdb, "tdb_transaction_cancel");
 632         return _tdb_transaction_cancel(tdb);
 633 }
 634
 635 /*
 636   work out how much space the linearised recovery data will consume
 637 */
 638 static bool tdb_recovery_size(struct tdb_context *tdb, tdb_len_t *result)
 639 {
 640         tdb_len_t recovery_size = 0;
 641         int i;
 642
 643         recovery_size = sizeof(uint32_t);
 644         for (i=0;i<tdb->transaction->num_blocks;i++) {
 645                 tdb_len_t block_size;
 646                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
 647                         break;
 648                 }
 649                 if (tdb->transaction->blocks[i] == NULL) {
 650                         continue;
 651                 }
 652                 if (!tdb_add_len_t(recovery_size, 2*sizeof(tdb_off_t),
 653                                    &recovery_size)) {
 654                         return false;
 655                 }
 656                 if (i == tdb->transaction->num_blocks-1) {
 657                         block_size = tdb->transaction->last_block_size;
 658                 } else {
 659                         block_size =  tdb->transaction->block_size;
 660                 }
 661                 if (!tdb_add_len_t(recovery_size, block_size,
 662                                    &recovery_size)) {
 663                         return false;
 664                 }
 665         }
 666
 667         *result = recovery_size;
 668         return true;
 669 }
 670
 671 int tdb_recovery_area(struct tdb_context *tdb,
 672                       const struct tdb_methods *methods,
 673                       tdb_off_t *recovery_offset,
 674                       struct tdb_record *rec)
 675 {
 676         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, recovery_offset) == -1) {
 677                 return -1;
 678         }
 679
 680         if (*recovery_offset == 0) {
 681                 rec->rec_len = 0;
 682                 return 0;
 683         }
 684
 685         if (methods->tdb_read(tdb, *recovery_offset, rec, sizeof(*rec),
 686                               DOCONV()) == -1) {
 687                 return -1;
 688         }
 689
 690         /* ignore invalid recovery regions: can happen in crash */
 691         if (rec->magic != TDB_RECOVERY_MAGIC &&
 692             rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
 693                 *recovery_offset = 0;
 694                 rec->rec_len = 0;
 695         }
 696         return 0;
 697 }
 698
 699 /*
 700   allocate the recovery area, or use an existing recovery area if it is
 701   large enough
 702 */
 703 static int tdb_recovery_allocate(struct tdb_context *tdb,
 704                                  tdb_len_t *recovery_size,
 705                                  tdb_off_t *recovery_offset,
 706                                  tdb_len_t *recovery_max_size)
 707 {
 708         struct tdb_record rec;
 709         const struct tdb_methods *methods = tdb->transaction->io_methods;
 710         tdb_off_t recovery_head, new_end;
 711
 712         if (tdb_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
 713                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
 714                 return -1;
 715         }
 716
 717         if (!tdb_recovery_size(tdb, recovery_size)) {
 718                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
 719                          "overflow recovery size\n"));
 720                 return -1;
 721         }
 722
 723         /* Existing recovery area? */
 724         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
 725                 /* it fits in the existing area */
 726                 *recovery_max_size = rec.rec_len;
 727                 *recovery_offset = recovery_head;
 728                 return 0;
 729         }
 730
 731         /* If recovery area in middle of file, we need a new one. */
 732         if (recovery_head == 0
 733             || recovery_head + sizeof(rec) + rec.rec_len != tdb->map_size) {
 734                 /* we need to free up the old recovery area, then allocate a
 735                    new one at the end of the file. Note that we cannot use
 736                    tdb_allocate() to allocate the new one as that might return
 737                    us an area that is being currently used (as of the start of
 738                    the transaction) */
 739                 if (recovery_head) {
 740                         if (tdb_free(tdb, recovery_head, &rec) == -1) {
 741                                 TDB_LOG((tdb, TDB_DEBUG_FATAL,
 742                                          "tdb_recovery_allocate: failed to"
 743                                          " free previous recovery area\n"));
 744                                 return -1;
 745                         }
 746
 747                         /* the tdb_free() call might have increased
 748                          * the recovery size */
 749                         if (!tdb_recovery_size(tdb, recovery_size)) {
 750                                 TDB_LOG((tdb, TDB_DEBUG_FATAL,
 751                                          "tdb_recovery_allocate: "
 752                                          "overflow recovery size\n"));
 753                                 return -1;
 754                         }
 755                 }
 756
 757                 /* New head will be at end of file. */
 758                 recovery_head = tdb->map_size;
 759         }
 760
 761         /* Now we know where it will be. */
 762         *recovery_offset = recovery_head;
 763
 764         /* Expand by more than we need, so we don't do it often. */
 765         *recovery_max_size = tdb_expand_adjust(tdb->map_size,
 766                                                *recovery_size,
 767                                                tdb->page_size)
 768                 - sizeof(rec);
 769
 770         if (!tdb_add_off_t(recovery_head, sizeof(rec), &new_end) ||
 771             !tdb_add_off_t(new_end, *recovery_max_size, &new_end)) {
 772                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: "
 773                          "overflow recovery area\n"));
 774                 return -1;
 775         }
 776
 777         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
 778                                      new_end - tdb->transaction->old_map_size)
 779             == -1) {
 780                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
 781                 return -1;
 782         }
 783
 784         /* remap the file (if using mmap) */
 785         methods->tdb_oob(tdb, tdb->map_size, 1, 1);
 786
 787         /* we have to reset the old map size so that we don't try to expand the file
 788            again in the transaction commit, which would destroy the recovery area */
 789         tdb->transaction->old_map_size = tdb->map_size;
 790
 791         /* write the recovery header offset and sync - we can sync without a race here
 792            as the magic ptr in the recovery record has not been set */
 793         CONVERT(recovery_head);
 794         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
 795                                &recovery_head, sizeof(tdb_off_t)) == -1) {
 796                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 797                 return -1;
 798         }
 799         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
 800                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 801                 return -1;
 802         }
 803
 804         return 0;
 805 }
 806
 807
 808 /*
 809   setup the recovery data that will be used on a crash during commit
 810 */
 811 static int transaction_setup_recovery(struct tdb_context *tdb,
 812                                       tdb_off_t *magic_offset)
 813 {
 814         tdb_len_t recovery_size;
 815         unsigned char *data, *p;
 816         const struct tdb_methods *methods = tdb->transaction->io_methods;
 817         struct tdb_record *rec;
 818         tdb_off_t recovery_offset, recovery_max_size;
 819         tdb_off_t old_map_size = tdb->transaction->old_map_size;
 820         uint32_t magic, tailer;
 821         int i;
 822
 823         /*
 824           check that the recovery area has enough space
 825         */
 826         if (tdb_recovery_allocate(tdb, &recovery_size,
 827                                   &recovery_offset, &recovery_max_size) == -1) {
 828                 return -1;
 829         }
 830
 831         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
 832         if (data == NULL) {
 833                 tdb->ecode = TDB_ERR_OOM;
 834                 return -1;
 835         }
 836
 837         rec = (struct tdb_record *)data;
 838         memset(rec, 0, sizeof(*rec));
 839
 840         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
 841         rec->data_len = recovery_size;
 842         rec->rec_len  = recovery_max_size;
 843         rec->key_len  = old_map_size;
 844         CONVERT(*rec);
 845
 846         /* build the recovery data into a single blob to allow us to do a single
 847            large write, which should be more efficient */
 848         p = data + sizeof(*rec);
 849         for (i=0;i<tdb->transaction->num_blocks;i++) {
 850                 tdb_off_t offset;
 851                 tdb_len_t length;
 852
 853                 if (tdb->transaction->blocks[i] == NULL) {
 854                         continue;
 855                 }
 856
 857                 offset = i * tdb->transaction->block_size;
 858                 length = tdb->transaction->block_size;
 859                 if (i == tdb->transaction->num_blocks-1) {
 860                         length = tdb->transaction->last_block_size;
 861                 }
 862
 863                 if (offset >= old_map_size) {
 864                         continue;
 865                 }
 866                 if (offset + length > tdb->transaction->old_map_size) {
 867                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
 868                         free(data);
 869                         tdb->ecode = TDB_ERR_CORRUPT;
 870                         return -1;
 871                 }
 872                 memcpy(p, &offset, 4);
 873                 memcpy(p+4, &length, 4);
 874                 if (DOCONV()) {
 875                         tdb_convert(p, 8);
 876                 }
 877                 /* the recovery area contains the old data, not the
 878                    new data, so we have to call the original tdb_read
 879                    method to get it */
 880                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
 881                         free(data);
 882                         tdb->ecode = TDB_ERR_IO;
 883                         return -1;
 884                 }
 885                 p += 8 + length;
 886         }
 887
 888         /* and the tailer */
 889         tailer = sizeof(*rec) + recovery_max_size;
 890         memcpy(p, &tailer, 4);
 891         if (DOCONV()) {
 892                 tdb_convert(p, 4);
 893         }
 894
 895         /* write the recovery data to the recovery area */
 896         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 897                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
 898                 free(data);
 899                 tdb->ecode = TDB_ERR_IO;
 900                 return -1;
 901         }
 902         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 903                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
 904                 free(data);
 905                 tdb->ecode = TDB_ERR_IO;
 906                 return -1;
 907         }
 908
 909         /* as we don't have ordered writes, we have to sync the recovery
 910            data before we update the magic to indicate that the recovery
 911            data is present */
 912         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
 913                 free(data);
 914                 return -1;
 915         }
 916
 917         free(data);
 918
 919         magic = TDB_RECOVERY_MAGIC;
 920         CONVERT(magic);
 921
 922         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
 923
 924         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 925                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
 926                 tdb->ecode = TDB_ERR_IO;
 927                 return -1;
 928         }
 929         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 930                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
 931                 tdb->ecode = TDB_ERR_IO;
 932                 return -1;
 933         }
 934
 935         /* ensure the recovery magic marker is on disk */
 936         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
 937                 return -1;
 938         }
 939
 940         return 0;
 941 }
 942
 943 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 944 {
 945         const struct tdb_methods *methods;
 946
 947         if (tdb->transaction == NULL) {
 948                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
 949                 return -1;
 950         }
 951
 952         if (tdb->transaction->prepared) {
 953                 tdb->ecode = TDB_ERR_EINVAL;
 954                 _tdb_transaction_cancel(tdb);
 955                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
 956                 return -1;
 957         }
 958
 959         if (tdb->transaction->transaction_error) {
 960                 tdb->ecode = TDB_ERR_IO;
 961                 _tdb_transaction_cancel(tdb);
 962                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
 963                 return -1;
 964         }
 965
 966
 967         if (tdb->transaction->nesting != 0) {
 968                 return 0;
 969         }
 970
 971         /* check for a null transaction */
 972         if (tdb->transaction->blocks == NULL) {
 973                 return 0;
 974         }
 975
 976         methods = tdb->transaction->io_methods;
 977
 978         /* if there are any locks pending then the caller has not
 979            nested their locks properly, so fail the transaction */
 980         if (tdb_have_extra_locks(tdb)) {
 981                 tdb->ecode = TDB_ERR_LOCK;
 982                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
 983                 _tdb_transaction_cancel(tdb);
 984                 return -1;
 985         }
 986
 987         /* upgrade the main transaction lock region to a write lock */
 988         if (tdb_allrecord_upgrade(tdb) == -1) {
 989                 if (tdb->ecode == TDB_ERR_RDONLY && tdb->read_only) {
 990                         TDB_LOG((tdb, TDB_DEBUG_ERROR,
 991                                  "tdb_transaction_prepare_commit: "
 992                                  "failed to upgrade hash locks: "
 993                                  "database is read only\n"));
 994                 } else if (tdb->ecode == TDB_ERR_RDONLY
 995                            && tdb->traverse_read) {
 996                         TDB_LOG((tdb, TDB_DEBUG_ERROR,
 997                                  "tdb_transaction_prepare_commit: "
 998                                  "failed to upgrade hash locks: "
 999                                  "a database traverse is in progress\n"));
1000                 } else {
1001                         TDB_LOG((tdb, TDB_DEBUG_ERROR,
1002                                  "tdb_transaction_prepare_commit: "
1003                                  "failed to upgrade hash locks: %s\n",
1004                                  tdb_errorstr(tdb)));
1005                 }
1006                 _tdb_transaction_cancel(tdb);
1007                 return -1;
1008         }
1009
1010         /* get the open lock - this prevents new users attaching to the database
1011            during the commit */
1012         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
1013                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
1014                 _tdb_transaction_cancel(tdb);
1015                 return -1;
1016         }
1017
1018         /* write the recovery data to the end of the file */
1019         if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
1020                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
1021                 _tdb_transaction_cancel(tdb);
1022                 return -1;
1023         }
1024
1025         tdb->transaction->prepared = true;
1026
1027         /* expand the file to the new size if needed */
1028         if (tdb->map_size != tdb->transaction->old_map_size) {
1029                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
1030                                              tdb->map_size -
1031                                              tdb->transaction->old_map_size) == -1) {
1032                         tdb->ecode = TDB_ERR_IO;
1033                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
1034                         _tdb_transaction_cancel(tdb);
1035                         return -1;
1036                 }
1037                 tdb->map_size = tdb->transaction->old_map_size;
1038                 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
1039         }
1040
1041         /* Keep the open lock until the actual commit */
1042
1043         return 0;
1044 }
1045
1046 /*
1047    prepare to commit the current transaction
1048 */
1049 _PUBLIC_ int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1050 {
1051         tdb_trace(tdb, "tdb_transaction_prepare_commit");
1052         return _tdb_transaction_prepare_commit(tdb);
1053 }
1054
1055 /* A repack is worthwhile if the largest is less than half total free. */
1056 static bool repack_worthwhile(struct tdb_context *tdb)
1057 {
1058         tdb_off_t ptr;
1059         struct tdb_record rec;
1060         tdb_len_t total = 0, largest = 0;
1061
1062         if (tdb_ofs_read(tdb, FREELIST_TOP, &ptr) == -1) {
1063                 return false;
1064         }
1065
1066         while (ptr != 0 && tdb_rec_free_read(tdb, ptr, &rec) == 0) {
1067                 total += rec.rec_len;
1068                 if (rec.rec_len > largest) {
1069                         largest = rec.rec_len;
1070                 }
1071                 ptr = rec.next;
1072         }
1073
1074         return total > largest * 2;
1075 }
1076
1077 /*
1078   commit the current transaction
1079 */
1080 _PUBLIC_ int tdb_transaction_commit(struct tdb_context *tdb)
1081 {
1082         const struct tdb_methods *methods;
1083         int i;
1084         bool need_repack = false;
1085
1086         if (tdb->transaction == NULL) {
1087                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1088                 return -1;
1089         }
1090
1091         tdb_trace(tdb, "tdb_transaction_commit");
1092
1093         if (tdb->transaction->transaction_error) {
1094                 tdb->ecode = TDB_ERR_IO;
1095                 _tdb_transaction_cancel(tdb);
1096                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1097                 return -1;
1098         }
1099
1100
1101         if (tdb->transaction->nesting != 0) {
1102                 tdb->transaction->nesting--;
1103                 return 0;
1104         }
1105
1106         /* check for a null transaction */
1107         if (tdb->transaction->blocks == NULL) {
1108                 _tdb_transaction_cancel(tdb);
1109                 return 0;
1110         }
1111
1112         if (!tdb->transaction->prepared) {
1113                 int ret = _tdb_transaction_prepare_commit(tdb);
1114                 if (ret)
1115                         return ret;
1116         }
1117
1118         methods = tdb->transaction->io_methods;
1119
1120         /* perform all the writes */
1121         for (i=0;i<tdb->transaction->num_blocks;i++) {
1122                 tdb_off_t offset;
1123                 tdb_len_t length;
1124
1125                 if (tdb->transaction->blocks[i] == NULL) {
1126                         continue;
1127                 }
1128
1129                 offset = i * tdb->transaction->block_size;
1130                 length = tdb->transaction->block_size;
1131                 if (i == tdb->transaction->num_blocks-1) {
1132                         length = tdb->transaction->last_block_size;
1133                 }
1134
1135                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1136                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1137
1138                         /* we've overwritten part of the data and
1139                            possibly expanded the file, so we need to
1140                            run the crash recovery code */
1141                         tdb->methods = methods;
1142                         tdb_transaction_recover(tdb);
1143
1144                         _tdb_transaction_cancel(tdb);
1145
1146                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1147                         return -1;
1148                 }
1149                 SAFE_FREE(tdb->transaction->blocks[i]);
1150         }
1151
1152         /* Do this before we drop lock or blocks. */
1153         if (tdb->transaction->expanded) {
1154                 need_repack = repack_worthwhile(tdb);
1155         }
1156
1157         SAFE_FREE(tdb->transaction->blocks);
1158         tdb->transaction->num_blocks = 0;
1159
1160         /* ensure the new data is on disk */
1161         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1162                 return -1;
1163         }
1164
1165         /*
1166           TODO: maybe write to some dummy hdr field, or write to magic
1167           offset without mmap, before the last sync, instead of the
1168           utime() call
1169         */
1170
1171         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1172            don't change the mtime of the file, this means the file may
1173            not be backed up (as tdb rounding to block sizes means that
1174            file size changes are quite rare too). The following forces
1175            mtime changes when a transaction completes */
1176 #ifdef HAVE_UTIME
1177         utime(tdb->name, NULL);
1178 #endif
1179
1180         /* use a transaction cancel to free memory and remove the
1181            transaction locks */
1182         _tdb_transaction_cancel(tdb);
1183
1184         if (need_repack) {
1185                 return tdb_repack(tdb);
1186         }
1187
1188         return 0;
1189 }
1190
1191
1192 /*
1193   recover from an aborted transaction. Must be called with exclusive
1194   database write access already established (including the open
1195   lock to prevent new processes attaching)
1196 */
1197 int tdb_transaction_recover(struct tdb_context *tdb)
1198 {
1199         tdb_off_t recovery_head, recovery_eof;
1200         unsigned char *data, *p;
1201         uint32_t zero = 0;
1202         struct tdb_record rec;
1203
1204         /* find the recovery area */
1205         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1206                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1207                 tdb->ecode = TDB_ERR_IO;
1208                 return -1;
1209         }
1210
1211         if (recovery_head == 0) {
1212                 /* we have never allocated a recovery record */
1213                 return 0;
1214         }
1215
1216         /* read the recovery record */
1217         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1218                                    sizeof(rec), DOCONV()) == -1) {
1219                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1220                 tdb->ecode = TDB_ERR_IO;
1221                 return -1;
1222         }
1223
1224         if (rec.magic != TDB_RECOVERY_MAGIC) {
1225                 /* there is no valid recovery data */
1226                 return 0;
1227         }
1228
1229         if (tdb->read_only) {
1230                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1231                 tdb->ecode = TDB_ERR_CORRUPT;
1232                 return -1;
1233         }
1234
1235         recovery_eof = rec.key_len;
1236
1237         data = (unsigned char *)malloc(rec.data_len);
1238         if (data == NULL) {
1239                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1240                 tdb->ecode = TDB_ERR_OOM;
1241                 return -1;
1242         }
1243
1244         /* read the full recovery data */
1245         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1246                                    rec.data_len, 0) == -1) {
1247                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1248                 tdb->ecode = TDB_ERR_IO;
1249                 return -1;
1250         }
1251
1252         /* recover the file data */
1253         p = data;
1254         while (p+8 < data + rec.data_len) {
1255                 uint32_t ofs, len;
1256                 if (DOCONV()) {
1257                         tdb_convert(p, 8);
1258                 }
1259                 memcpy(&ofs, p, 4);
1260                 memcpy(&len, p+4, 4);
1261
1262                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1263                         free(data);
1264                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %u bytes at offset %u\n", len, ofs));
1265                         tdb->ecode = TDB_ERR_IO;
1266                         return -1;
1267                 }
1268                 p += 8 + len;
1269         }
1270
1271         free(data);
1272
1273         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1274                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1275                 tdb->ecode = TDB_ERR_IO;
1276                 return -1;
1277         }
1278
1279         /* if the recovery area is after the recovered eof then remove it */
1280         if (recovery_eof <= recovery_head) {
1281                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1282                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1283                         tdb->ecode = TDB_ERR_IO;
1284                         return -1;
1285                 }
1286         }
1287
1288         /* remove the recovery magic */
1289         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1290                           &zero) == -1) {
1291                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1292                 tdb->ecode = TDB_ERR_IO;
1293                 return -1;
1294         }
1295
1296         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1297                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1298                 tdb->ecode = TDB_ERR_IO;
1299                 return -1;
1300         }
1301
1302         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %u byte database\n",
1303                  recovery_eof));
1304
1305         /* all done */
1306         return 0;
1307 }
1308
1309 /* Any I/O failures we say "needs recovery". */
1310 bool tdb_needs_recovery(struct tdb_context *tdb)
1311 {
1312         tdb_off_t recovery_head;
1313         struct tdb_record rec;
1314
1315         /* find the recovery area */
1316         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1317                 return true;
1318         }
1319
1320         if (recovery_head == 0) {
1321                 /* we have never allocated a recovery record */
1322                 return false;
1323         }
1324
1325         /* read the recovery record */
1326         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1327                                    sizeof(rec), DOCONV()) == -1) {
1328                 return true;
1329         }
1330
1331         return (rec.magic == TDB_RECOVERY_MAGIC);
1332 }