lib/tdb/common/transaction.c

   1  /*
   2    Unix SMB/CIFS implementation.
   3
   4    trivial database library
   5
   6    Copyright (C) Andrew Tridgell              2005
   7
   8      ** NOTE! The following LGPL license applies to the tdb
   9      ** library. This does NOT imply that all of Samba is released
  10      ** under the LGPL
  11
  12    This library is free software; you can redistribute it and/or
  13    modify it under the terms of the GNU Lesser General Public
  14    License as published by the Free Software Foundation; either
  15    version 3 of the License, or (at your option) any later version.
  16
  17    This library is distributed in the hope that it will be useful,
  18    but WITHOUT ANY WARRANTY; without even the implied warranty of
  19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20    Lesser General Public License for more details.
  21
  22    You should have received a copy of the GNU Lesser General Public
  23    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  24 */
  25
  26 #include "tdb_private.h"
  27
  28 /*
  29   transaction design:
  30
  31   - only allow a single transaction at a time per database. This makes
  32     using the transaction API simpler, as otherwise the caller would
  33     have to cope with temporary failures in transactions that conflict
  34     with other current transactions
  35
  36   - keep the transaction recovery information in the same file as the
  37     database, using a special 'transaction recovery' record pointed at
  38     by the header. This removes the need for extra journal files as
  39     used by some other databases
  40
  41   - dynamically allocated the transaction recover record, re-using it
  42     for subsequent transactions. If a larger record is needed then
  43     tdb_free() the old record to place it on the normal tdb freelist
  44     before allocating the new record
  45
  46   - during transactions, keep a linked list of writes all that have
  47     been performed by intercepting all tdb_write() calls. The hooked
  48     transaction versions of tdb_read() and tdb_write() check this
  49     linked list and try to use the elements of the list in preference
  50     to the real database.
  51
  52   - don't allow any locks to be held when a transaction starts,
  53     otherwise we can end up with deadlock (plus lack of lock nesting
  54     in posix locks would mean the lock is lost)
  55
  56   - if the caller gains a lock during the transaction but doesn't
  57     release it then fail the commit
  58
  59   - allow for nested calls to tdb_transaction_start(), re-using the
  60     existing transaction record. If the inner transaction is cancelled
  61     then a subsequent commit will fail
  62
  63   - keep a mirrored copy of the tdb hash chain heads to allow for the
  64     fast hash heads scan on traverse, updating the mirrored copy in
  65     the transaction version of tdb_write
  66
  67   - allow callers to mix transaction and non-transaction use of tdb,
  68     although once a transaction is started then an exclusive lock is
  69     gained until the transaction is committed or cancelled
  70
  71   - the commit stategy involves first saving away all modified data
  72     into a linearised buffer in the transaction recovery area, then
  73     marking the transaction recovery area with a magic value to
  74     indicate a valid recovery record. In total 4 fsync/msync calls are
  75     needed per commit to prevent race conditions. It might be possible
  76     to reduce this to 3 or even 2 with some more work.
  77
  78   - check for a valid recovery record on open of the tdb, while the
  79     open lock is held. Automatically recover from the transaction
  80     recovery area if needed, then continue with the open as
  81     usual. This allows for smooth crash recovery with no administrator
  82     intervention.
  83
  84   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
  85     still available, but no transaction recovery area is used and no
  86     fsync/msync calls are made.
  87
  88   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
  89     tdb_add_flags() transaction nesting is enabled.
  90     It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
  91     The default is that transaction nesting is allowed.
  92     Note: this default may change in future versions of tdb.
  93
  94     Beware. when transactions are nested a transaction successfully
  95     completed with tdb_transaction_commit() can be silently unrolled later.
  96
  97   - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
  98     tdb_add_flags() transaction nesting is disabled.
  99     It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
 100     An attempt create a nested transaction will fail with TDB_ERR_NESTING.
 101     The default is that transaction nesting is allowed.
 102     Note: this default may change in future versions of tdb.
 103 */
 104
 105
 106 /*
 107   hold the context of any current transaction
 108 */
 109 struct tdb_transaction {
 110         /* we keep a mirrored copy of the tdb hash heads here so
 111            tdb_next_hash_chain() can operate efficiently */
 112         uint32_t *hash_heads;
 113
 114         /* the original io methods - used to do IOs to the real db */
 115         const struct tdb_methods *io_methods;
 116
 117         /* the list of transaction blocks. When a block is first
 118            written to, it gets created in this list */
 119         uint8_t **blocks;
 120         uint32_t num_blocks;
 121         uint32_t block_size;      /* bytes in each block */
 122         uint32_t last_block_size; /* number of valid bytes in the last block */
 123
 124         /* non-zero when an internal transaction error has
 125            occurred. All write operations will then fail until the
 126            transaction is ended */
 127         int transaction_error;
 128
 129         /* when inside a transaction we need to keep track of any
 130            nested tdb_transaction_start() calls, as these are allowed,
 131            but don't create a new transaction */
 132         int nesting;
 133
 134         /* set when a prepare has already occurred */
 135         bool prepared;
 136         tdb_off_t magic_offset;
 137
 138         /* old file size before transaction */
 139         tdb_len_t old_map_size;
 140
 141         /* 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 transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
 151                             tdb_len_t len, int cv)
 152 {
 153         uint32_t blk;
 154
 155         /* break it down into block sized ops */
 156         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 157                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 158                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
 159                         return -1;
 160                 }
 161                 len -= len2;
 162                 off += len2;
 163                 buf = (void *)(len2 + (char *)buf);
 164         }
 165
 166         if (len == 0) {
 167                 return 0;
 168         }
 169
 170         blk = off / tdb->transaction->block_size;
 171
 172         /* see if we have it in the block list */
 173         if (tdb->transaction->num_blocks <= blk ||
 174             tdb->transaction->blocks[blk] == NULL) {
 175                 /* nope, do a real read */
 176                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
 177                         goto fail;
 178                 }
 179                 return 0;
 180         }
 181
 182         /* it is in the block list. Now check for the last block */
 183         if (blk == tdb->transaction->num_blocks-1) {
 184                 if (len > tdb->transaction->last_block_size) {
 185                         goto fail;
 186                 }
 187         }
 188
 189         /* now copy it out of this block */
 190         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
 191         if (cv) {
 192                 tdb_convert(buf, len);
 193         }
 194         return 0;
 195
 196 fail:
 197         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
 198         tdb->ecode = TDB_ERR_IO;
 199         tdb->transaction->transaction_error = 1;
 200         return -1;
 201 }
 202
 203
 204 /*
 205   write while in a transaction
 206 */
 207 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
 208                              const void *buf, tdb_len_t len)
 209 {
 210         uint32_t blk;
 211
 212         /* Only a commit is allowed on a prepared transaction */
 213         if (tdb->transaction->prepared) {
 214                 tdb->ecode = TDB_ERR_EINVAL;
 215                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
 216                 tdb->transaction->transaction_error = 1;
 217                 return -1;
 218         }
 219
 220         /* if the write is to a hash head, then update the transaction
 221            hash heads */
 222         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
 223             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
 224                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
 225                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
 226         }
 227
 228         /* break it up into block sized chunks */
 229         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 230                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 231                 if (transaction_write(tdb, off, buf, len2) != 0) {
 232                         return -1;
 233                 }
 234                 len -= len2;
 235                 off += len2;
 236                 if (buf != NULL) {
 237                         buf = (const void *)(len2 + (const char *)buf);
 238                 }
 239         }
 240
 241         if (len == 0) {
 242                 return 0;
 243         }
 244
 245         blk = off / tdb->transaction->block_size;
 246         off = off % tdb->transaction->block_size;
 247
 248         if (tdb->transaction->num_blocks <= blk) {
 249                 uint8_t **new_blocks;
 250                 /* expand the blocks array */
 251                 if (tdb->transaction->blocks == NULL) {
 252                         new_blocks = (uint8_t **)malloc(
 253                                 (blk+1)*sizeof(uint8_t *));
 254                 } else {
 255                         new_blocks = (uint8_t **)realloc(
 256                                 tdb->transaction->blocks,
 257                                 (blk+1)*sizeof(uint8_t *));
 258                 }
 259                 if (new_blocks == NULL) {
 260                         tdb->ecode = TDB_ERR_OOM;
 261                         goto fail;
 262                 }
 263                 memset(&new_blocks[tdb->transaction->num_blocks], 0,
 264                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
 265                 tdb->transaction->blocks = new_blocks;
 266                 tdb->transaction->num_blocks = blk+1;
 267                 tdb->transaction->last_block_size = 0;
 268         }
 269
 270         /* allocate and fill a block? */
 271         if (tdb->transaction->blocks[blk] == NULL) {
 272                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
 273                 if (tdb->transaction->blocks[blk] == NULL) {
 274                         tdb->ecode = TDB_ERR_OOM;
 275                         tdb->transaction->transaction_error = 1;
 276                         return -1;
 277                 }
 278                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
 279                         tdb_len_t len2 = tdb->transaction->block_size;
 280                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
 281                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
 282                         }
 283                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
 284                                                                    tdb->transaction->blocks[blk],
 285                                                                    len2, 0) != 0) {
 286                                 SAFE_FREE(tdb->transaction->blocks[blk]);
 287                                 tdb->ecode = TDB_ERR_IO;
 288                                 goto fail;
 289                         }
 290                         if (blk == tdb->transaction->num_blocks-1) {
 291                                 tdb->transaction->last_block_size = len2;
 292                         }
 293                 }
 294         }
 295
 296         /* overwrite part of an existing block */
 297         if (buf == NULL) {
 298                 memset(tdb->transaction->blocks[blk] + off, 0, len);
 299         } else {
 300                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 301         }
 302         if (blk == tdb->transaction->num_blocks-1) {
 303                 if (len + off > tdb->transaction->last_block_size) {
 304                         tdb->transaction->last_block_size = len + off;
 305                 }
 306         }
 307
 308         return 0;
 309
 310 fail:
 311         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
 312                  (blk*tdb->transaction->block_size) + off, len));
 313         tdb->transaction->transaction_error = 1;
 314         return -1;
 315 }
 316
 317
 318 /*
 319   write while in a transaction - this varient never expands the transaction blocks, it only
 320   updates existing blocks. This means it cannot change the recovery size
 321 */
 322 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
 323                                       const void *buf, tdb_len_t len)
 324 {
 325         uint32_t blk;
 326
 327         /* break it up into block sized chunks */
 328         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 329                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 330                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
 331                         return -1;
 332                 }
 333                 len -= len2;
 334                 off += len2;
 335                 if (buf != NULL) {
 336                         buf = (const void *)(len2 + (const char *)buf);
 337                 }
 338         }
 339
 340         if (len == 0) {
 341                 return 0;
 342         }
 343
 344         blk = off / tdb->transaction->block_size;
 345         off = off % tdb->transaction->block_size;
 346
 347         if (tdb->transaction->num_blocks <= blk ||
 348             tdb->transaction->blocks[blk] == NULL) {
 349                 return 0;
 350         }
 351
 352         if (blk == tdb->transaction->num_blocks-1 &&
 353             off + len > tdb->transaction->last_block_size) {
 354                 if (off >= tdb->transaction->last_block_size) {
 355                         return 0;
 356                 }
 357                 len = tdb->transaction->last_block_size - off;
 358         }
 359
 360         /* overwrite part of an existing block */
 361         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 362
 363         return 0;
 364 }
 365
 366
 367 /*
 368   accelerated hash chain head search, using the cached hash heads
 369 */
 370 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
 371 {
 372         uint32_t h = *chain;
 373         for (;h < tdb->header.hash_size;h++) {
 374                 /* the +1 takes account of the freelist */
 375                 if (0 != tdb->transaction->hash_heads[h+1]) {
 376                         break;
 377                 }
 378         }
 379         (*chain) = h;
 380 }
 381
 382 /*
 383   out of bounds check during a transaction
 384 */
 385 static int transaction_oob(struct tdb_context *tdb, tdb_off_t off,
 386                            tdb_len_t len, int probe)
 387 {
 388         if (off + len >= off && off + len <= tdb->map_size) {
 389                 return 0;
 390         }
 391         tdb->ecode = TDB_ERR_IO;
 392         return -1;
 393 }
 394
 395 /*
 396   transaction version of tdb_expand().
 397 */
 398 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
 399                                    tdb_off_t addition)
 400 {
 401         /* add a write to the transaction elements, so subsequent
 402            reads see the zero data */
 403         if (transaction_write(tdb, size, NULL, addition) != 0) {
 404                 return -1;
 405         }
 406
 407         tdb->transaction->expanded = true;
 408
 409         return 0;
 410 }
 411
 412 static const struct tdb_methods transaction_methods = {
 413         transaction_read,
 414         transaction_write,
 415         transaction_next_hash_chain,
 416         transaction_oob,
 417         transaction_expand_file,
 418 };
 419
 420
 421 /*
 422   start a tdb transaction. No token is returned, as only a single
 423   transaction is allowed to be pending per tdb_context
 424 */
 425 static int _tdb_transaction_start(struct tdb_context *tdb,
 426                                   enum tdb_lock_flags lockflags)
 427 {
 428         /* some sanity checks */
 429         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
 430                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
 431                 tdb->ecode = TDB_ERR_EINVAL;
 432                 return -1;
 433         }
 434
 435         /* cope with nested tdb_transaction_start() calls */
 436         if (tdb->transaction != NULL) {
 437                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
 438                         tdb->ecode = TDB_ERR_NESTING;
 439                         return -1;
 440                 }
 441                 tdb->transaction->nesting++;
 442                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
 443                          tdb->transaction->nesting));
 444                 return 0;
 445         }
 446
 447         if (tdb_have_extra_locks(tdb)) {
 448                 /* the caller must not have any locks when starting a
 449                    transaction as otherwise we'll be screwed by lack
 450                    of nested locks in posix */
 451                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
 452                 tdb->ecode = TDB_ERR_LOCK;
 453                 return -1;
 454         }
 455
 456         if (tdb->travlocks.next != NULL) {
 457                 /* you cannot use transactions inside a traverse (although you can use
 458                    traverse inside a transaction) as otherwise you can end up with
 459                    deadlock */
 460                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
 461                 tdb->ecode = TDB_ERR_LOCK;
 462                 return -1;
 463         }
 464
 465         tdb->transaction = (struct tdb_transaction *)
 466                 calloc(sizeof(struct tdb_transaction), 1);
 467         if (tdb->transaction == NULL) {
 468                 tdb->ecode = TDB_ERR_OOM;
 469                 return -1;
 470         }
 471
 472         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
 473         tdb->transaction->block_size = tdb->page_size;
 474
 475         /* get the transaction write lock. This is a blocking lock. As
 476            discussed with Volker, there are a number of ways we could
 477            make this async, which we will probably do in the future */
 478         if (tdb_transaction_lock(tdb, F_WRLCK, lockflags) == -1) {
 479                 SAFE_FREE(tdb->transaction->blocks);
 480                 SAFE_FREE(tdb->transaction);
 481                 if ((lockflags & TDB_LOCK_WAIT) == 0) {
 482                         tdb->ecode = TDB_ERR_NOLOCK;
 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->header.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 tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
 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                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
 646                         break;
 647                 }
 648                 if (tdb->transaction->blocks[i] == NULL) {
 649                         continue;
 650                 }
 651                 recovery_size += 2*sizeof(tdb_off_t);
 652                 if (i == tdb->transaction->num_blocks-1) {
 653                         recovery_size += tdb->transaction->last_block_size;
 654                 } else {
 655                         recovery_size += tdb->transaction->block_size;
 656                 }
 657         }
 658
 659         return recovery_size;
 660 }
 661
 662 int tdb_recovery_area(struct tdb_context *tdb,
 663                       const struct tdb_methods *methods,
 664                       tdb_off_t *recovery_offset,
 665                       struct tdb_record *rec)
 666 {
 667         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, recovery_offset) == -1) {
 668                 return -1;
 669         }
 670
 671         if (*recovery_offset == 0) {
 672                 rec->rec_len = 0;
 673                 return 0;
 674         }
 675
 676         if (methods->tdb_read(tdb, *recovery_offset, rec, sizeof(*rec),
 677                               DOCONV()) == -1) {
 678                 return -1;
 679         }
 680
 681         /* ignore invalid recovery regions: can happen in crash */
 682         if (rec->magic != TDB_RECOVERY_MAGIC &&
 683             rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
 684                 *recovery_offset = 0;
 685                 rec->rec_len = 0;
 686         }
 687         return 0;
 688 }
 689
 690 /*
 691   allocate the recovery area, or use an existing recovery area if it is
 692   large enough
 693 */
 694 static int tdb_recovery_allocate(struct tdb_context *tdb,
 695                                  tdb_len_t *recovery_size,
 696                                  tdb_off_t *recovery_offset,
 697                                  tdb_len_t *recovery_max_size)
 698 {
 699         struct tdb_record rec;
 700         const struct tdb_methods *methods = tdb->transaction->io_methods;
 701         tdb_off_t recovery_head, new_end;
 702
 703         if (tdb_recovery_area(tdb, methods, &recovery_head, &rec) == -1) {
 704                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
 705                 return -1;
 706         }
 707
 708         *recovery_size = tdb_recovery_size(tdb);
 709
 710         /* Existing recovery area? */
 711         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
 712                 /* it fits in the existing area */
 713                 *recovery_max_size = rec.rec_len;
 714                 *recovery_offset = recovery_head;
 715                 return 0;
 716         }
 717
 718         /* If recovery area in middle of file, we need a new one. */
 719         if (recovery_head == 0
 720             || recovery_head + sizeof(rec) + rec.rec_len != tdb->map_size) {
 721                 /* we need to free up the old recovery area, then allocate a
 722                    new one at the end of the file. Note that we cannot use
 723                    tdb_allocate() to allocate the new one as that might return
 724                    us an area that is being currently used (as of the start of
 725                    the transaction) */
 726                 if (recovery_head) {
 727                         if (tdb_free(tdb, recovery_head, &rec) == -1) {
 728                                 TDB_LOG((tdb, TDB_DEBUG_FATAL,
 729                                          "tdb_recovery_allocate: failed to"
 730                                          " free previous recovery area\n"));
 731                                 return -1;
 732                         }
 733
 734                         /* the tdb_free() call might have increased
 735                          * the recovery size */
 736                         *recovery_size = tdb_recovery_size(tdb);
 737                 }
 738
 739                 /* New head will be at end of file. */
 740                 recovery_head = tdb->map_size;
 741         }
 742
 743         /* Now we know where it will be. */
 744         *recovery_offset = recovery_head;
 745
 746         /* Expand by more than we need, so we don't do it often. */
 747         *recovery_max_size = tdb_expand_adjust(tdb->map_size,
 748                                                *recovery_size,
 749                                                tdb->page_size)
 750                 - sizeof(rec);
 751
 752         new_end = recovery_head + sizeof(rec) + *recovery_max_size;
 753
 754         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
 755                                      new_end - tdb->transaction->old_map_size)
 756             == -1) {
 757                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
 758                 return -1;
 759         }
 760
 761         /* remap the file (if using mmap) */
 762         methods->tdb_oob(tdb, tdb->map_size, 1, 1);
 763
 764         /* we have to reset the old map size so that we don't try to expand the file
 765            again in the transaction commit, which would destroy the recovery area */
 766         tdb->transaction->old_map_size = tdb->map_size;
 767
 768         /* write the recovery header offset and sync - we can sync without a race here
 769            as the magic ptr in the recovery record has not been set */
 770         CONVERT(recovery_head);
 771         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
 772                                &recovery_head, sizeof(tdb_off_t)) == -1) {
 773                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 774                 return -1;
 775         }
 776         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
 777                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 778                 return -1;
 779         }
 780
 781         return 0;
 782 }
 783
 784
 785 /*
 786   setup the recovery data that will be used on a crash during commit
 787 */
 788 static int transaction_setup_recovery(struct tdb_context *tdb,
 789                                       tdb_off_t *magic_offset)
 790 {
 791         tdb_len_t recovery_size;
 792         unsigned char *data, *p;
 793         const struct tdb_methods *methods = tdb->transaction->io_methods;
 794         struct tdb_record *rec;
 795         tdb_off_t recovery_offset, recovery_max_size;
 796         tdb_off_t old_map_size = tdb->transaction->old_map_size;
 797         uint32_t magic, tailer;
 798         int i;
 799
 800         /*
 801           check that the recovery area has enough space
 802         */
 803         if (tdb_recovery_allocate(tdb, &recovery_size,
 804                                   &recovery_offset, &recovery_max_size) == -1) {
 805                 return -1;
 806         }
 807
 808         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
 809         if (data == NULL) {
 810                 tdb->ecode = TDB_ERR_OOM;
 811                 return -1;
 812         }
 813
 814         rec = (struct tdb_record *)data;
 815         memset(rec, 0, sizeof(*rec));
 816
 817         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
 818         rec->data_len = recovery_size;
 819         rec->rec_len  = recovery_max_size;
 820         rec->key_len  = old_map_size;
 821         CONVERT(*rec);
 822
 823         /* build the recovery data into a single blob to allow us to do a single
 824            large write, which should be more efficient */
 825         p = data + sizeof(*rec);
 826         for (i=0;i<tdb->transaction->num_blocks;i++) {
 827                 tdb_off_t offset;
 828                 tdb_len_t length;
 829
 830                 if (tdb->transaction->blocks[i] == NULL) {
 831                         continue;
 832                 }
 833
 834                 offset = i * tdb->transaction->block_size;
 835                 length = tdb->transaction->block_size;
 836                 if (i == tdb->transaction->num_blocks-1) {
 837                         length = tdb->transaction->last_block_size;
 838                 }
 839
 840                 if (offset >= old_map_size) {
 841                         continue;
 842                 }
 843                 if (offset + length > tdb->transaction->old_map_size) {
 844                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
 845                         free(data);
 846                         tdb->ecode = TDB_ERR_CORRUPT;
 847                         return -1;
 848                 }
 849                 memcpy(p, &offset, 4);
 850                 memcpy(p+4, &length, 4);
 851                 if (DOCONV()) {
 852                         tdb_convert(p, 8);
 853                 }
 854                 /* the recovery area contains the old data, not the
 855                    new data, so we have to call the original tdb_read
 856                    method to get it */
 857                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
 858                         free(data);
 859                         tdb->ecode = TDB_ERR_IO;
 860                         return -1;
 861                 }
 862                 p += 8 + length;
 863         }
 864
 865         /* and the tailer */
 866         tailer = sizeof(*rec) + recovery_max_size;
 867         memcpy(p, &tailer, 4);
 868         if (DOCONV()) {
 869                 tdb_convert(p, 4);
 870         }
 871
 872         /* write the recovery data to the recovery area */
 873         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 874                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
 875                 free(data);
 876                 tdb->ecode = TDB_ERR_IO;
 877                 return -1;
 878         }
 879         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 880                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
 881                 free(data);
 882                 tdb->ecode = TDB_ERR_IO;
 883                 return -1;
 884         }
 885
 886         /* as we don't have ordered writes, we have to sync the recovery
 887            data before we update the magic to indicate that the recovery
 888            data is present */
 889         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
 890                 free(data);
 891                 return -1;
 892         }
 893
 894         free(data);
 895
 896         magic = TDB_RECOVERY_MAGIC;
 897         CONVERT(magic);
 898
 899         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
 900
 901         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 902                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
 903                 tdb->ecode = TDB_ERR_IO;
 904                 return -1;
 905         }
 906         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 907                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
 908                 tdb->ecode = TDB_ERR_IO;
 909                 return -1;
 910         }
 911
 912         /* ensure the recovery magic marker is on disk */
 913         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
 914                 return -1;
 915         }
 916
 917         return 0;
 918 }
 919
 920 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 921 {
 922         const struct tdb_methods *methods;
 923
 924         if (tdb->transaction == NULL) {
 925                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
 926                 return -1;
 927         }
 928
 929         if (tdb->transaction->prepared) {
 930                 tdb->ecode = TDB_ERR_EINVAL;
 931                 _tdb_transaction_cancel(tdb);
 932                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
 933                 return -1;
 934         }
 935
 936         if (tdb->transaction->transaction_error) {
 937                 tdb->ecode = TDB_ERR_IO;
 938                 _tdb_transaction_cancel(tdb);
 939                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
 940                 return -1;
 941         }
 942
 943
 944         if (tdb->transaction->nesting != 0) {
 945                 return 0;
 946         }
 947
 948         /* check for a null transaction */
 949         if (tdb->transaction->blocks == NULL) {
 950                 return 0;
 951         }
 952
 953         methods = tdb->transaction->io_methods;
 954
 955         /* if there are any locks pending then the caller has not
 956            nested their locks properly, so fail the transaction */
 957         if (tdb_have_extra_locks(tdb)) {
 958                 tdb->ecode = TDB_ERR_LOCK;
 959                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
 960                 _tdb_transaction_cancel(tdb);
 961                 return -1;
 962         }
 963
 964         /* upgrade the main transaction lock region to a write lock */
 965         if (tdb_allrecord_upgrade(tdb) == -1) {
 966                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
 967                 _tdb_transaction_cancel(tdb);
 968                 return -1;
 969         }
 970
 971         /* get the open lock - this prevents new users attaching to the database
 972            during the commit */
 973         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
 974                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
 975                 _tdb_transaction_cancel(tdb);
 976                 return -1;
 977         }
 978
 979         if (!(tdb->flags & TDB_NOSYNC)) {
 980                 /* write the recovery data to the end of the file */
 981                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
 982                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
 983                         _tdb_transaction_cancel(tdb);
 984                         return -1;
 985                 }
 986         }
 987
 988         tdb->transaction->prepared = true;
 989
 990         /* expand the file to the new size if needed */
 991         if (tdb->map_size != tdb->transaction->old_map_size) {
 992                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
 993                                              tdb->map_size -
 994                                              tdb->transaction->old_map_size) == -1) {
 995                         tdb->ecode = TDB_ERR_IO;
 996                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
 997                         _tdb_transaction_cancel(tdb);
 998                         return -1;
 999                 }
1000                 tdb->map_size = tdb->transaction->old_map_size;
1001                 methods->tdb_oob(tdb, tdb->map_size, 1, 1);
1002         }
1003
1004         /* Keep the open lock until the actual commit */
1005
1006         return 0;
1007 }
1008
1009 /*
1010    prepare to commit the current transaction
1011 */
1012 _PUBLIC_ int tdb_transaction_prepare_commit(struct tdb_context *tdb)
1013 {
1014         tdb_trace(tdb, "tdb_transaction_prepare_commit");
1015         return _tdb_transaction_prepare_commit(tdb);
1016 }
1017
1018 /* A repack is worthwhile if the largest is less than half total free. */
1019 static bool repack_worthwhile(struct tdb_context *tdb)
1020 {
1021         tdb_off_t ptr;
1022         struct tdb_record rec;
1023         tdb_len_t total = 0, largest = 0;
1024
1025         if (tdb_ofs_read(tdb, FREELIST_TOP, &ptr) == -1) {
1026                 return false;
1027         }
1028
1029         while (ptr != 0 && tdb_rec_free_read(tdb, ptr, &rec) == 0) {
1030                 total += rec.rec_len;
1031                 if (rec.rec_len > largest) {
1032                         largest = rec.rec_len;
1033                 }
1034                 ptr = rec.next;
1035         }
1036
1037         return total > largest * 2;
1038 }
1039
1040 /*
1041   commit the current transaction
1042 */
1043 _PUBLIC_ int tdb_transaction_commit(struct tdb_context *tdb)
1044 {
1045         const struct tdb_methods *methods;
1046         int i;
1047         bool need_repack = false;
1048
1049         if (tdb->transaction == NULL) {
1050                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
1051                 return -1;
1052         }
1053
1054         tdb_trace(tdb, "tdb_transaction_commit");
1055
1056         if (tdb->transaction->transaction_error) {
1057                 tdb->ecode = TDB_ERR_IO;
1058                 _tdb_transaction_cancel(tdb);
1059                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
1060                 return -1;
1061         }
1062
1063
1064         if (tdb->transaction->nesting != 0) {
1065                 tdb->transaction->nesting--;
1066                 return 0;
1067         }
1068
1069         /* check for a null transaction */
1070         if (tdb->transaction->blocks == NULL) {
1071                 _tdb_transaction_cancel(tdb);
1072                 return 0;
1073         }
1074
1075         if (!tdb->transaction->prepared) {
1076                 int ret = _tdb_transaction_prepare_commit(tdb);
1077                 if (ret)
1078                         return ret;
1079         }
1080
1081         methods = tdb->transaction->io_methods;
1082
1083         /* perform all the writes */
1084         for (i=0;i<tdb->transaction->num_blocks;i++) {
1085                 tdb_off_t offset;
1086                 tdb_len_t length;
1087
1088                 if (tdb->transaction->blocks[i] == NULL) {
1089                         continue;
1090                 }
1091
1092                 offset = i * tdb->transaction->block_size;
1093                 length = tdb->transaction->block_size;
1094                 if (i == tdb->transaction->num_blocks-1) {
1095                         length = tdb->transaction->last_block_size;
1096                 }
1097
1098                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1099                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1100
1101                         /* we've overwritten part of the data and
1102                            possibly expanded the file, so we need to
1103                            run the crash recovery code */
1104                         tdb->methods = methods;
1105                         tdb_transaction_recover(tdb);
1106
1107                         _tdb_transaction_cancel(tdb);
1108
1109                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1110                         return -1;
1111                 }
1112                 SAFE_FREE(tdb->transaction->blocks[i]);
1113         }
1114
1115         /* Do this before we drop lock or blocks. */
1116         if (tdb->transaction->expanded) {
1117                 need_repack = repack_worthwhile(tdb);
1118         }
1119
1120         SAFE_FREE(tdb->transaction->blocks);
1121         tdb->transaction->num_blocks = 0;
1122
1123         /* ensure the new data is on disk */
1124         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1125                 return -1;
1126         }
1127
1128         /*
1129           TODO: maybe write to some dummy hdr field, or write to magic
1130           offset without mmap, before the last sync, instead of the
1131           utime() call
1132         */
1133
1134         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1135            don't change the mtime of the file, this means the file may
1136            not be backed up (as tdb rounding to block sizes means that
1137            file size changes are quite rare too). The following forces
1138            mtime changes when a transaction completes */
1139 #ifdef HAVE_UTIME
1140         utime(tdb->name, NULL);
1141 #endif
1142
1143         /* use a transaction cancel to free memory and remove the
1144            transaction locks */
1145         _tdb_transaction_cancel(tdb);
1146
1147         if (need_repack) {
1148                 return tdb_repack(tdb);
1149         }
1150
1151         return 0;
1152 }
1153
1154
1155 /*
1156   recover from an aborted transaction. Must be called with exclusive
1157   database write access already established (including the open
1158   lock to prevent new processes attaching)
1159 */
1160 int tdb_transaction_recover(struct tdb_context *tdb)
1161 {
1162         tdb_off_t recovery_head, recovery_eof;
1163         unsigned char *data, *p;
1164         uint32_t zero = 0;
1165         struct tdb_record rec;
1166
1167         /* find the recovery area */
1168         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1169                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1170                 tdb->ecode = TDB_ERR_IO;
1171                 return -1;
1172         }
1173
1174         if (recovery_head == 0) {
1175                 /* we have never allocated a recovery record */
1176                 return 0;
1177         }
1178
1179         /* read the recovery record */
1180         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1181                                    sizeof(rec), DOCONV()) == -1) {
1182                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1183                 tdb->ecode = TDB_ERR_IO;
1184                 return -1;
1185         }
1186
1187         if (rec.magic != TDB_RECOVERY_MAGIC) {
1188                 /* there is no valid recovery data */
1189                 return 0;
1190         }
1191
1192         if (tdb->read_only) {
1193                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1194                 tdb->ecode = TDB_ERR_CORRUPT;
1195                 return -1;
1196         }
1197
1198         recovery_eof = rec.key_len;
1199
1200         data = (unsigned char *)malloc(rec.data_len);
1201         if (data == NULL) {
1202                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1203                 tdb->ecode = TDB_ERR_OOM;
1204                 return -1;
1205         }
1206
1207         /* read the full recovery data */
1208         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1209                                    rec.data_len, 0) == -1) {
1210                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1211                 tdb->ecode = TDB_ERR_IO;
1212                 return -1;
1213         }
1214
1215         /* recover the file data */
1216         p = data;
1217         while (p+8 < data + rec.data_len) {
1218                 uint32_t ofs, len;
1219                 if (DOCONV()) {
1220                         tdb_convert(p, 8);
1221                 }
1222                 memcpy(&ofs, p, 4);
1223                 memcpy(&len, p+4, 4);
1224
1225                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1226                         free(data);
1227                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1228                         tdb->ecode = TDB_ERR_IO;
1229                         return -1;
1230                 }
1231                 p += 8 + len;
1232         }
1233
1234         free(data);
1235
1236         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1237                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1238                 tdb->ecode = TDB_ERR_IO;
1239                 return -1;
1240         }
1241
1242         /* if the recovery area is after the recovered eof then remove it */
1243         if (recovery_eof <= recovery_head) {
1244                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1245                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1246                         tdb->ecode = TDB_ERR_IO;
1247                         return -1;
1248                 }
1249         }
1250
1251         /* remove the recovery magic */
1252         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1253                           &zero) == -1) {
1254                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1255                 tdb->ecode = TDB_ERR_IO;
1256                 return -1;
1257         }
1258
1259         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1260                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1261                 tdb->ecode = TDB_ERR_IO;
1262                 return -1;
1263         }
1264
1265         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",
1266                  recovery_eof));
1267
1268         /* all done */
1269         return 0;
1270 }
1271
1272 /* Any I/O failures we say "needs recovery". */
1273 bool tdb_needs_recovery(struct tdb_context *tdb)
1274 {
1275         tdb_off_t recovery_head;
1276         struct tdb_record rec;
1277
1278         /* find the recovery area */
1279         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1280                 return true;
1281         }
1282
1283         if (recovery_head == 0) {
1284                 /* we have never allocated a recovery record */
1285                 return false;
1286         }
1287
1288         /* read the recovery record */
1289         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1290                                    sizeof(rec), DOCONV()) == -1) {
1291                 return true;
1292         }
1293
1294         return (rec.magic == TDB_RECOVERY_MAGIC);
1295 }