git.ozlabs.org Git - ccan/blob - ccan/tdb/transaction.c

   1  /*
   2    Unix SMB/CIFS implementation.
   3
   4    trivial database library
   5
   6    Copyright (C) Andrew Tridgell              2005
   7
   8      ** NOTE! The following LGPL license applies to the tdb
   9      ** library. This does NOT imply that all of Samba is released
  10      ** under the LGPL
  11
  12    This library is free software; you can redistribute it and/or
  13    modify it under the terms of the GNU Lesser General Public
  14    License as published by the Free Software Foundation; either
  15    version 3 of the License, or (at your option) any later version.
  16
  17    This library is distributed in the hope that it will be useful,
  18    but WITHOUT ANY WARRANTY; without even the implied warranty of
  19    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20    Lesser General Public License for more details.
  21
  22    You should have received a copy of the GNU Lesser General Public
  23    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  24 */
  25
  26 #include "tdb_private.h"
  27
  28 /*
  29   transaction design:
  30
  31   - only allow a single transaction at a time per database. This makes
  32     using the transaction API simpler, as otherwise the caller would
  33     have to cope with temporary failures in transactions that conflict
  34     with other current transactions
  35
  36   - keep the transaction recovery information in the same file as the
  37     database, using a special 'transaction recovery' record pointed at
  38     by the header. This removes the need for extra journal files as
  39     used by some other databases
  40
  41   - dynamically allocated the transaction recover record, re-using it
  42     for subsequent transactions. If a larger record is needed then
  43     tdb_free() the old record to place it on the normal tdb freelist
  44     before allocating the new record
  45
  46   - during transactions, keep a linked list of writes all that have
  47     been performed by intercepting all tdb_write() calls. The hooked
  48     transaction versions of tdb_read() and tdb_write() check this
  49     linked list and try to use the elements of the list in preference
  50     to the real database.
  51
  52   - don't allow any locks to be held when a transaction starts,
  53     otherwise we can end up with deadlock (plus lack of lock nesting
  54     in posix locks would mean the lock is lost)
  55
  56   - if the caller gains a lock during the transaction but doesn't
  57     release it then fail the commit
  58
  59   - allow for nested calls to tdb_transaction_start(), re-using the
  60     existing transaction record. If the inner transaction is cancelled
  61     then a subsequent commit will fail
  62
  63   - keep a mirrored copy of the tdb hash chain heads to allow for the
  64     fast hash heads scan on traverse, updating the mirrored copy in
  65     the transaction version of tdb_write
  66
  67   - allow callers to mix transaction and non-transaction use of tdb,
  68     although once a transaction is started then an exclusive lock is
  69     gained until the transaction is committed or cancelled
  70
  71   - the commit stategy involves first saving away all modified data
  72     into a linearised buffer in the transaction recovery area, then
  73     marking the transaction recovery area with a magic value to
  74     indicate a valid recovery record. In total 4 fsync/msync calls are
  75     needed per commit to prevent race conditions. It might be possible
  76     to reduce this to 3 or even 2 with some more work.
  77
  78   - check for a valid recovery record on open of the tdb, while the
  79     open lock is held. Automatically recover from the transaction
  80     recovery area if needed, then continue with the open as
  81     usual. This allows for smooth crash recovery with no administrator
  82     intervention.
  83
  84   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
  85     still available, but no transaction recovery area is used and no
  86     fsync/msync calls are made.
  87
  88   - if TDB_ALLOW_NESTING is passed to flags in tdb open, or added using
  89     tdb_add_flags() transaction nesting is enabled.
  90     It resets the TDB_DISALLOW_NESTING flag, as both cannot be used together.
  91     The default is that transaction nesting is allowed.
  92     Note: this default may change in future versions of tdb.
  93
  94     Beware. when transactions are nested a transaction successfully
  95     completed with tdb_transaction_commit() can be silently unrolled later.
  96
  97   - if TDB_DISALLOW_NESTING is passed to flags in tdb open, or added using
  98     tdb_add_flags() transaction nesting is disabled.
  99     It resets the TDB_ALLOW_NESTING flag, as both cannot be used together.
 100     An attempt create a nested transaction will fail with TDB_ERR_NESTING.
 101     The default is that transaction nesting is allowed.
 102     Note: this default may change in future versions of tdb.
 103 */
 104
 105
 106 /*
 107   hold the context of any current transaction
 108 */
 109 struct tdb_transaction {
 110         /* we keep a mirrored copy of the tdb hash heads here so
 111            tdb_next_hash_chain() can operate efficiently */
 112         uint32_t *hash_heads;
 113
 114         /* the original io methods - used to do IOs to the real db */
 115         const struct tdb_methods *io_methods;
 116
 117         /* the list of transaction blocks. When a block is first
 118            written to, it gets created in this list */
 119         uint8_t **blocks;
 120         uint32_t num_blocks;
 121         uint32_t block_size;      /* bytes in each block */
 122         uint32_t last_block_size; /* number of valid bytes in the last block */
 123
 124         /* non-zero when an internal transaction error has
 125            occurred. All write operations will then fail until the
 126            transaction is ended */
 127         int transaction_error;
 128
 129         /* when inside a transaction we need to keep track of any
 130            nested tdb_transaction_start() calls, as these are allowed,
 131            but don't create a new transaction */
 132         int nesting;
 133
 134         /* set when a prepare has already occurred */
 135         bool prepared;
 136         tdb_off_t magic_offset;
 137
 138         /* old file size before transaction */
 139         tdb_len_t old_map_size;
 140
 141         /* we should re-pack on commit */
 142         bool need_repack;
 143 };
 144
 145
 146 /*
 147   read while in a transaction. We need to check first if the data is in our list
 148   of transaction elements, then if not do a real read
 149 */
 150 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
 151                             tdb_len_t len, int cv)
 152 {
 153         uint32_t blk;
 154
 155         /* break it down into block sized ops */
 156         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 157                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 158                 if (transaction_read(tdb, off, buf, len2, cv) != 0) {
 159                         return -1;
 160                 }
 161                 len -= len2;
 162                 off += len2;
 163                 buf = (void *)(len2 + (char *)buf);
 164         }
 165
 166         if (len == 0) {
 167                 return 0;
 168         }
 169
 170         blk = off / tdb->transaction->block_size;
 171
 172         /* see if we have it in the block list */
 173         if (tdb->transaction->num_blocks <= blk ||
 174             tdb->transaction->blocks[blk] == NULL) {
 175                 /* nope, do a real read */
 176                 if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
 177                         goto fail;
 178                 }
 179                 return 0;
 180         }
 181
 182         /* it is in the block list. Now check for the last block */
 183         if (blk == tdb->transaction->num_blocks-1) {
 184                 if (len > tdb->transaction->last_block_size) {
 185                         goto fail;
 186                 }
 187         }
 188
 189         /* now copy it out of this block */
 190         memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
 191         if (cv) {
 192                 tdb_convert(buf, len);
 193         }
 194         return 0;
 195
 196 fail:
 197         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
 198         tdb->ecode = TDB_ERR_IO;
 199         tdb->transaction->transaction_error = 1;
 200         return -1;
 201 }
 202
 203
 204 /*
 205   write while in a transaction
 206 */
 207 static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
 208                              const void *buf, tdb_len_t len)
 209 {
 210         uint32_t blk;
 211
 212         /* Only a commit is allowed on a prepared transaction */
 213         if (tdb->transaction->prepared) {
 214                 tdb->ecode = TDB_ERR_EINVAL;
 215                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: transaction already prepared, write not allowed\n"));
 216                 tdb->transaction->transaction_error = 1;
 217                 return -1;
 218         }
 219
 220         /* if the write is to a hash head, then update the transaction
 221            hash heads */
 222         if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
 223             off < FREELIST_TOP+TDB_HASHTABLE_SIZE(tdb)) {
 224                 uint32_t chain = (off-FREELIST_TOP) / sizeof(tdb_off_t);
 225                 memcpy(&tdb->transaction->hash_heads[chain], buf, len);
 226         }
 227
 228         /* break it up into block sized chunks */
 229         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 230                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 231                 if (transaction_write(tdb, off, buf, len2) != 0) {
 232                         return -1;
 233                 }
 234                 len -= len2;
 235                 off += len2;
 236                 if (buf != NULL) {
 237                         buf = (const void *)(len2 + (const char *)buf);
 238                 }
 239         }
 240
 241         if (len == 0) {
 242                 return 0;
 243         }
 244
 245         blk = off / tdb->transaction->block_size;
 246         off = off % tdb->transaction->block_size;
 247
 248         if (tdb->transaction->num_blocks <= blk) {
 249                 uint8_t **new_blocks;
 250                 /* expand the blocks array */
 251                 if (tdb->transaction->blocks == NULL) {
 252                         new_blocks = (uint8_t **)malloc(
 253                                 (blk+1)*sizeof(uint8_t *));
 254                 } else {
 255                         new_blocks = (uint8_t **)realloc(
 256                                 tdb->transaction->blocks,
 257                                 (blk+1)*sizeof(uint8_t *));
 258                 }
 259                 if (new_blocks == NULL) {
 260                         tdb->ecode = TDB_ERR_OOM;
 261                         goto fail;
 262                 }
 263                 memset(&new_blocks[tdb->transaction->num_blocks], 0,
 264                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
 265                 tdb->transaction->blocks = new_blocks;
 266                 tdb->transaction->num_blocks = blk+1;
 267                 tdb->transaction->last_block_size = 0;
 268         }
 269
 270         /* allocate and fill a block? */
 271         if (tdb->transaction->blocks[blk] == NULL) {
 272                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
 273                 if (tdb->transaction->blocks[blk] == NULL) {
 274                         tdb->ecode = TDB_ERR_OOM;
 275                         tdb->transaction->transaction_error = 1;
 276                         return -1;
 277                 }
 278                 if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
 279                         tdb_len_t len2 = tdb->transaction->block_size;
 280                         if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
 281                                 len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
 282                         }
 283                         if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
 284                                                                    tdb->transaction->blocks[blk],
 285                                                                    len2, 0) != 0) {
 286                                 SAFE_FREE(tdb->transaction->blocks[blk]);
 287                                 tdb->ecode = TDB_ERR_IO;
 288                                 goto fail;
 289                         }
 290                         if (blk == tdb->transaction->num_blocks-1) {
 291                                 tdb->transaction->last_block_size = len2;
 292                         }
 293                 }
 294         }
 295
 296         /* overwrite part of an existing block */
 297         if (buf == NULL) {
 298                 memset(tdb->transaction->blocks[blk] + off, 0, len);
 299         } else {
 300                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 301         }
 302         if (blk == tdb->transaction->num_blocks-1) {
 303                 if (len + off > tdb->transaction->last_block_size) {
 304                         tdb->transaction->last_block_size = len + off;
 305                 }
 306         }
 307
 308         return 0;
 309
 310 fail:
 311         TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
 312                  (blk*tdb->transaction->block_size) + off, len));
 313         tdb->transaction->transaction_error = 1;
 314         return -1;
 315 }
 316
 317
 318 /*
 319   write while in a transaction - this varient never expands the transaction blocks, it only
 320   updates existing blocks. This means it cannot change the recovery size
 321 */
 322 static int transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
 323                                       const void *buf, tdb_len_t len)
 324 {
 325         uint32_t blk;
 326
 327         /* break it up into block sized chunks */
 328         while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
 329                 tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
 330                 if (transaction_write_existing(tdb, off, buf, len2) != 0) {
 331                         return -1;
 332                 }
 333                 len -= len2;
 334                 off += len2;
 335                 if (buf != NULL) {
 336                         buf = (const void *)(len2 + (const char *)buf);
 337                 }
 338         }
 339
 340         if (len == 0) {
 341                 return 0;
 342         }
 343
 344         blk = off / tdb->transaction->block_size;
 345         off = off % tdb->transaction->block_size;
 346
 347         if (tdb->transaction->num_blocks <= blk ||
 348             tdb->transaction->blocks[blk] == NULL) {
 349                 return 0;
 350         }
 351
 352         if (blk == tdb->transaction->num_blocks-1 &&
 353             off + len > tdb->transaction->last_block_size) {
 354                 if (off >= tdb->transaction->last_block_size) {
 355                         return 0;
 356                 }
 357                 len = tdb->transaction->last_block_size - off;
 358         }
 359
 360         /* overwrite part of an existing block */
 361         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 362
 363         return 0;
 364 }
 365
 366
 367 /*
 368   accelerated hash chain head search, using the cached hash heads
 369 */
 370 static void transaction_next_hash_chain(struct tdb_context *tdb, uint32_t *chain)
 371 {
 372         uint32_t h = *chain;
 373         for (;h < tdb->header.hash_size;h++) {
 374                 /* the +1 takes account of the freelist */
 375                 if (0 != tdb->transaction->hash_heads[h+1]) {
 376                         break;
 377                 }
 378         }
 379         (*chain) = h;
 380 }
 381
 382 /*
 383   out of bounds check during a transaction
 384 */
 385 static int transaction_oob(struct tdb_context *tdb, tdb_off_t len, int probe)
 386 {
 387         if (len <= tdb->map_size) {
 388                 return 0;
 389         }
 390         tdb->ecode = TDB_ERR_IO;
 391         return -1;
 392 }
 393
 394 /*
 395   transaction version of tdb_expand().
 396 */
 397 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
 398                                    tdb_off_t addition)
 399 {
 400         /* add a write to the transaction elements, so subsequent
 401            reads see the zero data */
 402         if (transaction_write(tdb, size, NULL, addition) != 0) {
 403                 return -1;
 404         }
 405
 406         tdb->transaction->need_repack = true;
 407
 408         return 0;
 409 }
 410
 411 static const struct tdb_methods transaction_methods = {
 412         transaction_read,
 413         transaction_write,
 414         transaction_next_hash_chain,
 415         transaction_oob,
 416         transaction_expand_file,
 417 };
 418
 419 /*
 420   sync to disk
 421 */
 422 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
 423 {
 424         if (tdb->flags & TDB_NOSYNC) {
 425                 return 0;
 426         }
 427
 428         if (fsync(tdb->fd) != 0) {
 429                 tdb->ecode = TDB_ERR_IO;
 430                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
 431                 return -1;
 432         }
 433 #ifdef MS_SYNC
 434         if (tdb->map_ptr) {
 435                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
 436                 if (msync(moffset + (char *)tdb->map_ptr,
 437                           length + (offset - moffset), MS_SYNC) != 0) {
 438                         tdb->ecode = TDB_ERR_IO;
 439                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
 440                                  strerror(errno)));
 441                         return -1;
 442                 }
 443         }
 444 #endif
 445         return 0;
 446 }
 447
 448
 449 static int _tdb_transaction_cancel(struct tdb_context *tdb)
 450 {
 451         int i, ret = 0;
 452
 453         if (tdb->transaction == NULL) {
 454                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
 455                 return -1;
 456         }
 457
 458         if (tdb->transaction->nesting != 0) {
 459                 tdb->transaction->transaction_error = 1;
 460                 tdb->transaction->nesting--;
 461                 return 0;
 462         }
 463
 464         tdb->map_size = tdb->transaction->old_map_size;
 465
 466         /* free all the transaction blocks */
 467         for (i=0;i<tdb->transaction->num_blocks;i++) {
 468                 if (tdb->transaction->blocks[i] != NULL) {
 469                         free(tdb->transaction->blocks[i]);
 470                 }
 471         }
 472         SAFE_FREE(tdb->transaction->blocks);
 473
 474         if (tdb->transaction->magic_offset) {
 475                 const struct tdb_methods *methods = tdb->transaction->io_methods;
 476                 uint32_t invalid = TDB_RECOVERY_INVALID_MAGIC;
 477
 478                 /* remove the recovery marker */
 479                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &invalid, 4) == -1 ||
 480                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
 481                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
 482                         ret = -1;
 483                 }
 484         }
 485
 486         /* This also removes the OPEN_LOCK, if we have it. */
 487         tdb_release_extra_locks(tdb);
 488
 489         /* restore the normal io methods */
 490         tdb->methods = tdb->transaction->io_methods;
 491
 492         tdb_transaction_unlock(tdb, F_WRLCK);
 493         SAFE_FREE(tdb->transaction->hash_heads);
 494         SAFE_FREE(tdb->transaction);
 495
 496         return ret;
 497 }
 498
 499 /*
 500   start a tdb transaction. No token is returned, as only a single
 501   transaction is allowed to be pending per tdb_context
 502 */
 503 int tdb_transaction_start(struct tdb_context *tdb)
 504 {
 505         /* some sanity checks */
 506         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
 507                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
 508                 tdb->ecode = TDB_ERR_EINVAL;
 509                 return -1;
 510         }
 511
 512         /* cope with nested tdb_transaction_start() calls */
 513         if (tdb->transaction != NULL) {
 514                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
 515                         tdb->ecode = TDB_ERR_NESTING;
 516                         return -1;
 517                 }
 518                 tdb_trace(tdb, "tdb_transaction_start");
 519                 tdb->transaction->nesting++;
 520                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
 521                          tdb->transaction->nesting));
 522                 return 0;
 523         }
 524
 525         if (tdb_have_extra_locks(tdb)) {
 526                 /* the caller must not have any locks when starting a
 527                    transaction as otherwise we'll be screwed by lack
 528                    of nested locks in posix */
 529                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
 530                 tdb->ecode = TDB_ERR_LOCK;
 531                 return -1;
 532         }
 533
 534         if (tdb->travlocks.next != NULL) {
 535                 /* you cannot use transactions inside a traverse (although you can use
 536                    traverse inside a transaction) as otherwise you can end up with
 537                    deadlock */
 538                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
 539                 tdb->ecode = TDB_ERR_LOCK;
 540                 return -1;
 541         }
 542
 543         tdb->transaction = (struct tdb_transaction *)
 544                 calloc(sizeof(struct tdb_transaction), 1);
 545         if (tdb->transaction == NULL) {
 546                 tdb->ecode = TDB_ERR_OOM;
 547                 return -1;
 548         }
 549
 550         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
 551         tdb->transaction->block_size = tdb->page_size;
 552
 553         /* get the transaction write lock. This is a blocking lock. As
 554            discussed with Volker, there are a number of ways we could
 555            make this async, which we will probably do in the future */
 556         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
 557                 SAFE_FREE(tdb->transaction->blocks);
 558                 SAFE_FREE(tdb->transaction);
 559                 return -1;
 560         }
 561
 562         /* get a read lock from the freelist to the end of file. This
 563            is upgraded to a write lock during the commit */
 564         if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true) == -1) {
 565                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
 566                 goto fail_allrecord_lock;
 567         }
 568
 569         /* setup a copy of the hash table heads so the hash scan in
 570            traverse can be fast */
 571         tdb->transaction->hash_heads = (uint32_t *)
 572                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
 573         if (tdb->transaction->hash_heads == NULL) {
 574                 tdb->ecode = TDB_ERR_OOM;
 575                 goto fail;
 576         }
 577         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
 578                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
 579                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
 580                 tdb->ecode = TDB_ERR_IO;
 581                 goto fail;
 582         }
 583
 584         /* make sure we know about any file expansions already done by
 585            anyone else */
 586         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
 587         tdb->transaction->old_map_size = tdb->map_size;
 588
 589         /* finally hook the io methods, replacing them with
 590            transaction specific methods */
 591         tdb->transaction->io_methods = tdb->methods;
 592         tdb->methods = &transaction_methods;
 593
 594         /* Trace at the end, so we get sequence number correct. */
 595         tdb_trace(tdb, "tdb_transaction_start");
 596         return 0;
 597
 598 fail:
 599         tdb_allrecord_unlock(tdb, F_RDLCK, false);
 600 fail_allrecord_lock:
 601         tdb_transaction_unlock(tdb, F_WRLCK);
 602         SAFE_FREE(tdb->transaction->blocks);
 603         SAFE_FREE(tdb->transaction->hash_heads);
 604         SAFE_FREE(tdb->transaction);
 605         return -1;
 606 }
 607
 608
 609 /*
 610   cancel the current transaction
 611 */
 612 int tdb_transaction_cancel(struct tdb_context *tdb)
 613 {
 614         tdb_trace(tdb, "tdb_transaction_cancel");
 615         return _tdb_transaction_cancel(tdb);
 616 }
 617
 618 /*
 619   work out how much space the linearised recovery data will consume
 620 */
 621 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
 622 {
 623         tdb_len_t recovery_size = 0;
 624         int i;
 625
 626         recovery_size = sizeof(uint32_t);
 627         for (i=0;i<tdb->transaction->num_blocks;i++) {
 628                 if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
 629                         break;
 630                 }
 631                 if (tdb->transaction->blocks[i] == NULL) {
 632                         continue;
 633                 }
 634                 recovery_size += 2*sizeof(tdb_off_t);
 635                 if (i == tdb->transaction->num_blocks-1) {
 636                         recovery_size += tdb->transaction->last_block_size;
 637                 } else {
 638                         recovery_size += tdb->transaction->block_size;
 639                 }
 640         }
 641
 642         return recovery_size;
 643 }
 644
 645 /*
 646   allocate the recovery area, or use an existing recovery area if it is
 647   large enough
 648 */
 649 static int tdb_recovery_allocate(struct tdb_context *tdb,
 650                                  tdb_len_t *recovery_size,
 651                                  tdb_off_t *recovery_offset,
 652                                  tdb_len_t *recovery_max_size)
 653 {
 654         struct tdb_record rec;
 655         const struct tdb_methods *methods = tdb->transaction->io_methods;
 656         tdb_off_t recovery_head;
 657
 658         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
 659                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
 660                 return -1;
 661         }
 662
 663         rec.rec_len = 0;
 664
 665         if (recovery_head != 0) {
 666                 if (methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
 667                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
 668                         return -1;
 669                 }
 670                 /* ignore invalid recovery regions: can happen in crash */
 671                 if (rec.magic != TDB_RECOVERY_MAGIC &&
 672                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
 673                         recovery_head = 0;
 674                 }
 675         }
 676
 677         *recovery_size = tdb_recovery_size(tdb);
 678
 679         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
 680                 /* it fits in the existing area */
 681                 *recovery_max_size = rec.rec_len;
 682                 *recovery_offset = recovery_head;
 683                 return 0;
 684         }
 685
 686         /* we need to free up the old recovery area, then allocate a
 687            new one at the end of the file. Note that we cannot use
 688            tdb_allocate() to allocate the new one as that might return
 689            us an area that is being currently used (as of the start of
 690            the transaction) */
 691         if (recovery_head != 0) {
 692                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
 693                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
 694                         return -1;
 695                 }
 696         }
 697
 698         /* the tdb_free() call might have increased the recovery size */
 699         *recovery_size = tdb_recovery_size(tdb);
 700
 701         /* round up to a multiple of page size */
 702         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
 703         *recovery_offset = tdb->map_size;
 704         recovery_head = *recovery_offset;
 705
 706         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
 707                                      (tdb->map_size - tdb->transaction->old_map_size) +
 708                                      sizeof(rec) + *recovery_max_size) == -1) {
 709                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
 710                 return -1;
 711         }
 712
 713         /* remap the file (if using mmap) */
 714         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
 715
 716         /* we have to reset the old map size so that we don't try to expand the file
 717            again in the transaction commit, which would destroy the recovery area */
 718         tdb->transaction->old_map_size = tdb->map_size;
 719
 720         /* write the recovery header offset and sync - we can sync without a race here
 721            as the magic ptr in the recovery record has not been set */
 722         CONVERT(recovery_head);
 723         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
 724                                &recovery_head, sizeof(tdb_off_t)) == -1) {
 725                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 726                 return -1;
 727         }
 728         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
 729                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 730                 return -1;
 731         }
 732
 733         return 0;
 734 }
 735
 736
 737 /*
 738   setup the recovery data that will be used on a crash during commit
 739 */
 740 static int transaction_setup_recovery(struct tdb_context *tdb,
 741                                       tdb_off_t *magic_offset)
 742 {
 743         tdb_len_t recovery_size;
 744         unsigned char *data, *p;
 745         const struct tdb_methods *methods = tdb->transaction->io_methods;
 746         struct tdb_record *rec;
 747         tdb_off_t recovery_offset, recovery_max_size;
 748         tdb_off_t old_map_size = tdb->transaction->old_map_size;
 749         uint32_t magic, tailer;
 750         int i;
 751
 752         /*
 753           check that the recovery area has enough space
 754         */
 755         if (tdb_recovery_allocate(tdb, &recovery_size,
 756                                   &recovery_offset, &recovery_max_size) == -1) {
 757                 return -1;
 758         }
 759
 760         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
 761         if (data == NULL) {
 762                 tdb->ecode = TDB_ERR_OOM;
 763                 return -1;
 764         }
 765
 766         rec = (struct tdb_record *)data;
 767         memset(rec, 0, sizeof(*rec));
 768
 769         rec->magic    = TDB_RECOVERY_INVALID_MAGIC;
 770         rec->data_len = recovery_size;
 771         rec->rec_len  = recovery_max_size;
 772         rec->key_len  = old_map_size;
 773         CONVERT(rec);
 774
 775         /* build the recovery data into a single blob to allow us to do a single
 776            large write, which should be more efficient */
 777         p = data + sizeof(*rec);
 778         for (i=0;i<tdb->transaction->num_blocks;i++) {
 779                 tdb_off_t offset;
 780                 tdb_len_t length;
 781
 782                 if (tdb->transaction->blocks[i] == NULL) {
 783                         continue;
 784                 }
 785
 786                 offset = i * tdb->transaction->block_size;
 787                 length = tdb->transaction->block_size;
 788                 if (i == tdb->transaction->num_blocks-1) {
 789                         length = tdb->transaction->last_block_size;
 790                 }
 791
 792                 if (offset >= old_map_size) {
 793                         continue;
 794                 }
 795                 if (offset + length > tdb->transaction->old_map_size) {
 796                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
 797                         free(data);
 798                         tdb->ecode = TDB_ERR_CORRUPT;
 799                         return -1;
 800                 }
 801                 memcpy(p, &offset, 4);
 802                 memcpy(p+4, &length, 4);
 803                 if (DOCONV()) {
 804                         tdb_convert(p, 8);
 805                 }
 806                 /* the recovery area contains the old data, not the
 807                    new data, so we have to call the original tdb_read
 808                    method to get it */
 809                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
 810                         free(data);
 811                         tdb->ecode = TDB_ERR_IO;
 812                         return -1;
 813                 }
 814                 p += 8 + length;
 815         }
 816
 817         /* and the tailer */
 818         tailer = sizeof(*rec) + recovery_max_size;
 819         memcpy(p, &tailer, 4);
 820         CONVERT(p);
 821
 822         /* write the recovery data to the recovery area */
 823         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 824                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
 825                 free(data);
 826                 tdb->ecode = TDB_ERR_IO;
 827                 return -1;
 828         }
 829         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 830                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
 831                 free(data);
 832                 tdb->ecode = TDB_ERR_IO;
 833                 return -1;
 834         }
 835
 836         /* as we don't have ordered writes, we have to sync the recovery
 837            data before we update the magic to indicate that the recovery
 838            data is present */
 839         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
 840                 free(data);
 841                 return -1;
 842         }
 843
 844         free(data);
 845
 846         magic = TDB_RECOVERY_MAGIC;
 847         CONVERT(magic);
 848
 849         *magic_offset = recovery_offset + offsetof(struct tdb_record, magic);
 850
 851         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 852                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
 853                 tdb->ecode = TDB_ERR_IO;
 854                 return -1;
 855         }
 856         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 857                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
 858                 tdb->ecode = TDB_ERR_IO;
 859                 return -1;
 860         }
 861
 862         /* ensure the recovery magic marker is on disk */
 863         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
 864                 return -1;
 865         }
 866
 867         return 0;
 868 }
 869
 870 static int _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 871 {
 872         const struct tdb_methods *methods;
 873
 874         if (tdb->transaction == NULL) {
 875                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
 876                 return -1;
 877         }
 878
 879         if (tdb->transaction->prepared) {
 880                 tdb->ecode = TDB_ERR_EINVAL;
 881                 _tdb_transaction_cancel(tdb);
 882                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
 883                 return -1;
 884         }
 885
 886         if (tdb->transaction->transaction_error) {
 887                 tdb->ecode = TDB_ERR_IO;
 888                 _tdb_transaction_cancel(tdb);
 889                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
 890                 return -1;
 891         }
 892
 893
 894         if (tdb->transaction->nesting != 0) {
 895                 tdb->transaction->nesting--;
 896                 return 0;
 897         }
 898
 899         /* check for a null transaction */
 900         if (tdb->transaction->blocks == NULL) {
 901                 return 0;
 902         }
 903
 904         methods = tdb->transaction->io_methods;
 905
 906         /* if there are any locks pending then the caller has not
 907            nested their locks properly, so fail the transaction */
 908         if (tdb_have_extra_locks(tdb)) {
 909                 tdb->ecode = TDB_ERR_LOCK;
 910                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
 911                 _tdb_transaction_cancel(tdb);
 912                 return -1;
 913         }
 914
 915         /* upgrade the main transaction lock region to a write lock */
 916         if (tdb_allrecord_upgrade(tdb) == -1) {
 917                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
 918                 _tdb_transaction_cancel(tdb);
 919                 return -1;
 920         }
 921
 922         /* get the open lock - this prevents new users attaching to the database
 923            during the commit */
 924         if (tdb_nest_lock(tdb, OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT) == -1) {
 925                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get open lock\n"));
 926                 _tdb_transaction_cancel(tdb);
 927                 return -1;
 928         }
 929
 930         if (!(tdb->flags & TDB_NOSYNC)) {
 931                 /* write the recovery data to the end of the file */
 932                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
 933                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
 934                         _tdb_transaction_cancel(tdb);
 935                         return -1;
 936                 }
 937         }
 938
 939         tdb->transaction->prepared = true;
 940
 941         /* expand the file to the new size if needed */
 942         if (tdb->map_size != tdb->transaction->old_map_size) {
 943                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
 944                                              tdb->map_size -
 945                                              tdb->transaction->old_map_size) == -1) {
 946                         tdb->ecode = TDB_ERR_IO;
 947                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
 948                         _tdb_transaction_cancel(tdb);
 949                         return -1;
 950                 }
 951                 tdb->map_size = tdb->transaction->old_map_size;
 952                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
 953         }
 954
 955         /* Keep the open lock until the actual commit */
 956
 957         return 0;
 958 }
 959
 960 /*
 961    prepare to commit the current transaction
 962 */
 963 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
 964 {
 965         tdb_trace(tdb, "tdb_transaction_prepare_commit");
 966         return _tdb_transaction_prepare_commit(tdb);
 967 }
 968
 969 /*
 970   commit the current transaction
 971 */
 972 int tdb_transaction_commit(struct tdb_context *tdb)
 973 {
 974         const struct tdb_methods *methods;
 975         int i;
 976         bool need_repack;
 977
 978         if (tdb->transaction == NULL) {
 979                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
 980                 return -1;
 981         }
 982
 983         tdb_trace(tdb, "tdb_transaction_commit");
 984
 985         if (tdb->transaction->transaction_error) {
 986                 tdb->ecode = TDB_ERR_IO;
 987                 tdb_transaction_cancel(tdb);
 988                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
 989                 return -1;
 990         }
 991
 992
 993         if (tdb->transaction->nesting != 0) {
 994                 tdb->transaction->nesting--;
 995                 return 0;
 996         }
 997
 998         /* check for a null transaction */
 999         if (tdb->transaction->blocks == NULL) {
1000                 _tdb_transaction_cancel(tdb);
1001                 return 0;
1002         }
1003
1004         if (!tdb->transaction->prepared) {
1005                 int ret = _tdb_transaction_prepare_commit(tdb);
1006                 if (ret)
1007                         return ret;
1008         }
1009
1010         methods = tdb->transaction->io_methods;
1011
1012         /* perform all the writes */
1013         for (i=0;i<tdb->transaction->num_blocks;i++) {
1014                 tdb_off_t offset;
1015                 tdb_len_t length;
1016
1017                 if (tdb->transaction->blocks[i] == NULL) {
1018                         continue;
1019                 }
1020
1021                 offset = i * tdb->transaction->block_size;
1022                 length = tdb->transaction->block_size;
1023                 if (i == tdb->transaction->num_blocks-1) {
1024                         length = tdb->transaction->last_block_size;
1025                 }
1026
1027                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1028                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1029
1030                         /* we've overwritten part of the data and
1031                            possibly expanded the file, so we need to
1032                            run the crash recovery code */
1033                         tdb->methods = methods;
1034                         tdb_transaction_recover(tdb);
1035
1036                         _tdb_transaction_cancel(tdb);
1037
1038                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1039                         return -1;
1040                 }
1041                 SAFE_FREE(tdb->transaction->blocks[i]);
1042         }
1043
1044         SAFE_FREE(tdb->transaction->blocks);
1045         tdb->transaction->num_blocks = 0;
1046
1047         /* ensure the new data is on disk */
1048         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1049                 return -1;
1050         }
1051
1052         /*
1053           TODO: maybe write to some dummy hdr field, or write to magic
1054           offset without mmap, before the last sync, instead of the
1055           utime() call
1056         */
1057
1058         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1059            don't change the mtime of the file, this means the file may
1060            not be backed up (as tdb rounding to block sizes means that
1061            file size changes are quite rare too). The following forces
1062            mtime changes when a transaction completes */
1063 #if HAVE_UTIME
1064         utime(tdb->name, NULL);
1065 #endif
1066
1067         need_repack = tdb->transaction->need_repack;
1068
1069         /* use a transaction cancel to free memory and remove the
1070            transaction locks */
1071         _tdb_transaction_cancel(tdb);
1072
1073         if (need_repack) {
1074                 return tdb_repack(tdb);
1075         }
1076
1077         return 0;
1078 }
1079
1080
1081 /*
1082   recover from an aborted transaction. Must be called with exclusive
1083   database write access already established (including the open
1084   lock to prevent new processes attaching)
1085 */
1086 int tdb_transaction_recover(struct tdb_context *tdb)
1087 {
1088         tdb_off_t recovery_head, recovery_eof;
1089         unsigned char *data, *p;
1090         uint32_t zero = 0;
1091         struct tdb_record rec;
1092
1093         /* find the recovery area */
1094         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1095                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1096                 tdb->ecode = TDB_ERR_IO;
1097                 return -1;
1098         }
1099
1100         if (recovery_head == 0) {
1101                 /* we have never allocated a recovery record */
1102                 return 0;
1103         }
1104
1105         /* read the recovery record */
1106         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1107                                    sizeof(rec), DOCONV()) == -1) {
1108                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1109                 tdb->ecode = TDB_ERR_IO;
1110                 return -1;
1111         }
1112
1113         if (rec.magic != TDB_RECOVERY_MAGIC) {
1114                 /* there is no valid recovery data */
1115                 return 0;
1116         }
1117
1118         if (tdb->read_only) {
1119                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1120                 tdb->ecode = TDB_ERR_CORRUPT;
1121                 return -1;
1122         }
1123
1124         recovery_eof = rec.key_len;
1125
1126         data = (unsigned char *)malloc(rec.data_len);
1127         if (data == NULL) {
1128                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1129                 tdb->ecode = TDB_ERR_OOM;
1130                 return -1;
1131         }
1132
1133         /* read the full recovery data */
1134         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1135                                    rec.data_len, 0) == -1) {
1136                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1137                 tdb->ecode = TDB_ERR_IO;
1138                 return -1;
1139         }
1140
1141         /* recover the file data */
1142         p = data;
1143         while (p+8 < data + rec.data_len) {
1144                 uint32_t ofs, len;
1145                 if (DOCONV()) {
1146                         tdb_convert(p, 8);
1147                 }
1148                 memcpy(&ofs, p, 4);
1149                 memcpy(&len, p+4, 4);
1150
1151                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1152                         free(data);
1153                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1154                         tdb->ecode = TDB_ERR_IO;
1155                         return -1;
1156                 }
1157                 p += 8 + len;
1158         }
1159
1160         free(data);
1161
1162         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1163                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1164                 tdb->ecode = TDB_ERR_IO;
1165                 return -1;
1166         }
1167
1168         /* if the recovery area is after the recovered eof then remove it */
1169         if (recovery_eof <= recovery_head) {
1170                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1171                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1172                         tdb->ecode = TDB_ERR_IO;
1173                         return -1;
1174                 }
1175         }
1176
1177         /* remove the recovery magic */
1178         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct tdb_record, magic),
1179                           &zero) == -1) {
1180                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1181                 tdb->ecode = TDB_ERR_IO;
1182                 return -1;
1183         }
1184
1185         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1186                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1187                 tdb->ecode = TDB_ERR_IO;
1188                 return -1;
1189         }
1190
1191         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",
1192                  recovery_eof));
1193
1194         /* all done */
1195         return 0;
1196 }
1197
1198 /* Any I/O failures we say "needs recovery". */
1199 bool tdb_needs_recovery(struct tdb_context *tdb)
1200 {
1201         tdb_off_t recovery_head;
1202         struct tdb_record rec;
1203
1204         /* find the recovery area */
1205         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1206                 return true;
1207         }
1208
1209         if (recovery_head == 0) {
1210                 /* we have never allocated a recovery record */
1211                 return false;
1212         }
1213
1214         /* read the recovery record */
1215         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1216                                    sizeof(rec), DOCONV()) == -1) {
1217                 return true;
1218         }
1219
1220         return (rec.magic == TDB_RECOVERY_MAGIC);
1221 }