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     global 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 is enabled.
  90     The default is that transaction nesting is not allowed and an attempt
  91     to create a nested transaction will fail with TDB_ERR_NESTING.
  92
  93     Beware. when transactions are nested a transaction successfully
  94     completed with tdb_transaction_commit() can be silently unrolled later.
  95 */
  96
  97
  98 /*
  99   hold the context of any current transaction
 100 */
 101 struct tdb_transaction {
 102         /* we keep a mirrored copy of the tdb hash heads here so
 103            tdb_next_hash_chain() can operate efficiently */
 104         uint32_t *hash_heads;
 105
 106         /* the original io methods - used to do IOs to the real db */
 107         const struct tdb_methods *io_methods;
 108
 109         /* the list of transaction blocks. When a block is first
 110            written to, it gets created in this list */
 111         uint8_t **blocks;
 112         uint32_t num_blocks;
 113         uint32_t block_size;      /* bytes in each block */
 114         uint32_t last_block_size; /* number of valid bytes in the last block */
 115
 116         /* non-zero when an internal transaction error has
 117            occurred. All write operations will then fail until the
 118            transaction is ended */
 119         int transaction_error;
 120
 121         /* when inside a transaction we need to keep track of any
 122            nested tdb_transaction_start() calls, as these are allowed,
 123            but don't create a new transaction */
 124         int nesting;
 125
 126         /* set when a prepare has already occurred */
 127         bool prepared;
 128         tdb_off_t magic_offset;
 129
 130         /* old file size before transaction */
 131         tdb_len_t old_map_size;
 132
 133         /* we should re-pack on commit */
 134         bool need_repack;
 135 };
 136
 137
 138 /*
 139   read while in a transaction. We need to check first if the data is in our list
 140   of transaction elements, then if not do a real read
 141 */
 142 static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
 143                             tdb_len_t len, int cv)
 144 {
 145         uint32_t blk;
 146
 147         /* Only a commit is allowed on a prepared transaction */
 148         if (tdb->transaction->prepared) {
 149                 tdb->ecode = TDB_ERR_EINVAL;
 150                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: transaction already prepared, read not allowed\n"));
 151                 tdb->transaction->transaction_error = 1;
 152                 return -1;
 153         }
 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         return TDB_ERRCODE(TDB_ERR_IO, -1);
 391 }
 392
 393 /*
 394   transaction version of tdb_expand().
 395 */
 396 static int transaction_expand_file(struct tdb_context *tdb, tdb_off_t size,
 397                                    tdb_off_t addition)
 398 {
 399         /* add a write to the transaction elements, so subsequent
 400            reads see the zero data */
 401         if (transaction_write(tdb, size, NULL, addition) != 0) {
 402                 return -1;
 403         }
 404
 405         tdb->transaction->need_repack = true;
 406
 407         return 0;
 408 }
 409
 410 /*
 411   brlock during a transaction - ignore them
 412 */
 413 static int transaction_brlock(struct tdb_context *tdb, tdb_off_t offset,
 414                               int rw_type, int lck_type, int probe, size_t len)
 415 {
 416         return 0;
 417 }
 418
 419 static const struct tdb_methods transaction_methods = {
 420         transaction_read,
 421         transaction_write,
 422         transaction_next_hash_chain,
 423         transaction_oob,
 424         transaction_expand_file,
 425         transaction_brlock
 426 };
 427
 428
 429 /*
 430   start a tdb transaction. No token is returned, as only a single
 431   transaction is allowed to be pending per tdb_context
 432 */
 433 int tdb_transaction_start(struct tdb_context *tdb)
 434 {
 435         /* some sanity checks */
 436         if (tdb->read_only || (tdb->flags & TDB_INTERNAL) || tdb->traverse_read) {
 437                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction on a read-only or internal db\n"));
 438                 tdb->ecode = TDB_ERR_EINVAL;
 439                 return -1;
 440         }
 441
 442         /* cope with nested tdb_transaction_start() calls */
 443         if (tdb->transaction != NULL) {
 444                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
 445                         tdb->ecode = TDB_ERR_NESTING;
 446                         return -1;
 447                 }
 448                 tdb->transaction->nesting++;
 449                 TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_start: nesting %d\n",
 450                          tdb->transaction->nesting));
 451                 return 0;
 452         }
 453
 454         if (tdb->num_locks != 0 || tdb->global_lock.count) {
 455                 /* the caller must not have any locks when starting a
 456                    transaction as otherwise we'll be screwed by lack
 457                    of nested locks in posix */
 458                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction with locks held\n"));
 459                 tdb->ecode = TDB_ERR_LOCK;
 460                 return -1;
 461         }
 462
 463         if (tdb->travlocks.next != NULL) {
 464                 /* you cannot use transactions inside a traverse (although you can use
 465                    traverse inside a transaction) as otherwise you can end up with
 466                    deadlock */
 467                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: cannot start a transaction within a traverse\n"));
 468                 tdb->ecode = TDB_ERR_LOCK;
 469                 return -1;
 470         }
 471
 472         tdb->transaction = (struct tdb_transaction *)
 473                 calloc(sizeof(struct tdb_transaction), 1);
 474         if (tdb->transaction == NULL) {
 475                 tdb->ecode = TDB_ERR_OOM;
 476                 return -1;
 477         }
 478
 479         /* a page at a time seems like a reasonable compromise between compactness and efficiency */
 480         tdb->transaction->block_size = tdb->page_size;
 481
 482         /* get the transaction write lock. This is a blocking lock. As
 483            discussed with Volker, there are a number of ways we could
 484            make this async, which we will probably do in the future */
 485         if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
 486                 SAFE_FREE(tdb->transaction->blocks);
 487                 SAFE_FREE(tdb->transaction);
 488                 return -1;
 489         }
 490
 491         /* get a read lock from the freelist to the end of file. This
 492            is upgraded to a write lock during the commit */
 493         if (tdb_brlock(tdb, FREELIST_TOP, F_RDLCK, F_SETLKW, 0, 0) == -1) {
 494                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_start: failed to get hash locks\n"));
 495                 tdb->ecode = TDB_ERR_LOCK;
 496                 goto fail;
 497         }
 498
 499         /* setup a copy of the hash table heads so the hash scan in
 500            traverse can be fast */
 501         tdb->transaction->hash_heads = (uint32_t *)
 502                 calloc(tdb->header.hash_size+1, sizeof(uint32_t));
 503         if (tdb->transaction->hash_heads == NULL) {
 504                 tdb->ecode = TDB_ERR_OOM;
 505                 goto fail;
 506         }
 507         if (tdb->methods->tdb_read(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
 508                                    TDB_HASHTABLE_SIZE(tdb), 0) != 0) {
 509                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to read hash heads\n"));
 510                 tdb->ecode = TDB_ERR_IO;
 511                 goto fail;
 512         }
 513
 514         /* make sure we know about any file expansions already done by
 515            anyone else */
 516         tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
 517         tdb->transaction->old_map_size = tdb->map_size;
 518
 519         /* finally hook the io methods, replacing them with
 520            transaction specific methods */
 521         tdb->transaction->io_methods = tdb->methods;
 522         tdb->methods = &transaction_methods;
 523
 524         return 0;
 525
 526 fail:
 527         tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
 528         tdb_transaction_unlock(tdb);
 529         SAFE_FREE(tdb->transaction->blocks);
 530         SAFE_FREE(tdb->transaction->hash_heads);
 531         SAFE_FREE(tdb->transaction);
 532         return -1;
 533 }
 534
 535
 536 /*
 537   sync to disk
 538 */
 539 static int transaction_sync(struct tdb_context *tdb, tdb_off_t offset, tdb_len_t length)
 540 {
 541         if (tdb->flags & TDB_NOSYNC) {
 542                 return 0;
 543         }
 544
 545         if (fsync(tdb->fd) != 0) {
 546                 tdb->ecode = TDB_ERR_IO;
 547                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: fsync failed\n"));
 548                 return -1;
 549         }
 550 #ifdef MS_SYNC
 551         if (tdb->map_ptr) {
 552                 tdb_off_t moffset = offset & ~(tdb->page_size-1);
 553                 if (msync(moffset + (char *)tdb->map_ptr,
 554                           length + (offset - moffset), MS_SYNC) != 0) {
 555                         tdb->ecode = TDB_ERR_IO;
 556                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction: msync failed - %s\n",
 557                                  strerror(errno)));
 558                         return -1;
 559                 }
 560         }
 561 #endif
 562         return 0;
 563 }
 564
 565
 566 /*
 567   cancel the current transaction
 568 */
 569 int tdb_transaction_cancel(struct tdb_context *tdb)
 570 {
 571         int i, ret = 0;
 572
 573         if (tdb->transaction == NULL) {
 574                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
 575                 return -1;
 576         }
 577
 578         if (tdb->transaction->nesting != 0) {
 579                 tdb->transaction->transaction_error = 1;
 580                 tdb->transaction->nesting--;
 581                 return 0;
 582         }
 583
 584         tdb->map_size = tdb->transaction->old_map_size;
 585
 586         /* free all the transaction blocks */
 587         for (i=0;i<tdb->transaction->num_blocks;i++) {
 588                 if (tdb->transaction->blocks[i] != NULL) {
 589                         free(tdb->transaction->blocks[i]);
 590                 }
 591         }
 592         SAFE_FREE(tdb->transaction->blocks);
 593
 594         if (tdb->transaction->magic_offset) {
 595                 const struct tdb_methods *methods = tdb->transaction->io_methods;
 596                 uint32_t zero = 0;
 597
 598                 /* remove the recovery marker */
 599                 if (methods->tdb_write(tdb, tdb->transaction->magic_offset, &zero, 4) == -1 ||
 600                 transaction_sync(tdb, tdb->transaction->magic_offset, 4) == -1) {
 601                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_cancel: failed to remove recovery magic\n"));
 602                         ret = -1;
 603                 }
 604         }
 605
 606         /* remove any global lock created during the transaction */
 607         if (tdb->global_lock.count != 0) {
 608                 tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 4*tdb->header.hash_size);
 609                 tdb->global_lock.count = 0;
 610         }
 611
 612         /* remove any locks created during the transaction */
 613         if (tdb->num_locks != 0) {
 614                 for (i=0;i<tdb->num_lockrecs;i++) {
 615                         tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
 616                                    F_UNLCK,F_SETLKW, 0, 1);
 617                 }
 618                 tdb->num_locks = 0;
 619                 tdb->num_lockrecs = 0;
 620                 SAFE_FREE(tdb->lockrecs);
 621         }
 622
 623         /* restore the normal io methods */
 624         tdb->methods = tdb->transaction->io_methods;
 625
 626         tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
 627         tdb_transaction_unlock(tdb);
 628         SAFE_FREE(tdb->transaction->hash_heads);
 629         SAFE_FREE(tdb->transaction);
 630
 631         return ret;
 632 }
 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 /*
 663   allocate the recovery area, or use an existing recovery area if it is
 664   large enough
 665 */
 666 static int tdb_recovery_allocate(struct tdb_context *tdb,
 667                                  tdb_len_t *recovery_size,
 668                                  tdb_off_t *recovery_offset,
 669                                  tdb_len_t *recovery_max_size)
 670 {
 671         struct list_struct rec;
 672         const struct tdb_methods *methods = tdb->transaction->io_methods;
 673         tdb_off_t recovery_head;
 674
 675         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
 676                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery head\n"));
 677                 return -1;
 678         }
 679
 680         rec.rec_len = 0;
 681
 682         if (recovery_head != 0 &&
 683             methods->tdb_read(tdb, recovery_head, &rec, sizeof(rec), DOCONV()) == -1) {
 684                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to read recovery record\n"));
 685                 return -1;
 686         }
 687
 688         *recovery_size = tdb_recovery_size(tdb);
 689
 690         if (recovery_head != 0 && *recovery_size <= rec.rec_len) {
 691                 /* it fits in the existing area */
 692                 *recovery_max_size = rec.rec_len;
 693                 *recovery_offset = recovery_head;
 694                 return 0;
 695         }
 696
 697         /* we need to free up the old recovery area, then allocate a
 698            new one at the end of the file. Note that we cannot use
 699            tdb_allocate() to allocate the new one as that might return
 700            us an area that is being currently used (as of the start of
 701            the transaction) */
 702         if (recovery_head != 0) {
 703                 if (tdb_free(tdb, recovery_head, &rec) == -1) {
 704                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to free previous recovery area\n"));
 705                         return -1;
 706                 }
 707         }
 708
 709         /* the tdb_free() call might have increased the recovery size */
 710         *recovery_size = tdb_recovery_size(tdb);
 711
 712         /* round up to a multiple of page size */
 713         *recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
 714         *recovery_offset = tdb->map_size;
 715         recovery_head = *recovery_offset;
 716
 717         if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
 718                                      (tdb->map_size - tdb->transaction->old_map_size) +
 719                                      sizeof(rec) + *recovery_max_size) == -1) {
 720                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to create recovery area\n"));
 721                 return -1;
 722         }
 723
 724         /* remap the file (if using mmap) */
 725         methods->tdb_oob(tdb, tdb->map_size + 1, 1);
 726
 727         /* we have to reset the old map size so that we don't try to expand the file
 728            again in the transaction commit, which would destroy the recovery area */
 729         tdb->transaction->old_map_size = tdb->map_size;
 730
 731         /* write the recovery header offset and sync - we can sync without a race here
 732            as the magic ptr in the recovery record has not been set */
 733         CONVERT(recovery_head);
 734         if (methods->tdb_write(tdb, TDB_RECOVERY_HEAD,
 735                                &recovery_head, sizeof(tdb_off_t)) == -1) {
 736                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 737                 return -1;
 738         }
 739         if (transaction_write_existing(tdb, TDB_RECOVERY_HEAD, &recovery_head, sizeof(tdb_off_t)) == -1) {
 740                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_recovery_allocate: failed to write recovery head\n"));
 741                 return -1;
 742         }
 743
 744         return 0;
 745 }
 746
 747
 748 /*
 749   setup the recovery data that will be used on a crash during commit
 750 */
 751 static int transaction_setup_recovery(struct tdb_context *tdb,
 752                                       tdb_off_t *magic_offset)
 753 {
 754         tdb_len_t recovery_size;
 755         unsigned char *data, *p;
 756         const struct tdb_methods *methods = tdb->transaction->io_methods;
 757         struct list_struct *rec;
 758         tdb_off_t recovery_offset, recovery_max_size;
 759         tdb_off_t old_map_size = tdb->transaction->old_map_size;
 760         uint32_t magic, tailer;
 761         int i;
 762
 763         /*
 764           check that the recovery area has enough space
 765         */
 766         if (tdb_recovery_allocate(tdb, &recovery_size,
 767                                   &recovery_offset, &recovery_max_size) == -1) {
 768                 return -1;
 769         }
 770
 771         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
 772         if (data == NULL) {
 773                 tdb->ecode = TDB_ERR_OOM;
 774                 return -1;
 775         }
 776
 777         rec = (struct list_struct *)data;
 778         memset(rec, 0, sizeof(*rec));
 779
 780         rec->magic    = 0;
 781         rec->data_len = recovery_size;
 782         rec->rec_len  = recovery_max_size;
 783         rec->key_len  = old_map_size;
 784         CONVERT(rec);
 785
 786         /* build the recovery data into a single blob to allow us to do a single
 787            large write, which should be more efficient */
 788         p = data + sizeof(*rec);
 789         for (i=0;i<tdb->transaction->num_blocks;i++) {
 790                 tdb_off_t offset;
 791                 tdb_len_t length;
 792
 793                 if (tdb->transaction->blocks[i] == NULL) {
 794                         continue;
 795                 }
 796
 797                 offset = i * tdb->transaction->block_size;
 798                 length = tdb->transaction->block_size;
 799                 if (i == tdb->transaction->num_blocks-1) {
 800                         length = tdb->transaction->last_block_size;
 801                 }
 802
 803                 if (offset >= old_map_size) {
 804                         continue;
 805                 }
 806                 if (offset + length > tdb->transaction->old_map_size) {
 807                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
 808                         free(data);
 809                         tdb->ecode = TDB_ERR_CORRUPT;
 810                         return -1;
 811                 }
 812                 memcpy(p, &offset, 4);
 813                 memcpy(p+4, &length, 4);
 814                 if (DOCONV()) {
 815                         tdb_convert(p, 8);
 816                 }
 817                 /* the recovery area contains the old data, not the
 818                    new data, so we have to call the original tdb_read
 819                    method to get it */
 820                 if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
 821                         free(data);
 822                         tdb->ecode = TDB_ERR_IO;
 823                         return -1;
 824                 }
 825                 p += 8 + length;
 826         }
 827
 828         /* and the tailer */
 829         tailer = sizeof(*rec) + recovery_max_size;
 830         memcpy(p, &tailer, 4);
 831         CONVERT(p);
 832
 833         /* write the recovery data to the recovery area */
 834         if (methods->tdb_write(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 835                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery data\n"));
 836                 free(data);
 837                 tdb->ecode = TDB_ERR_IO;
 838                 return -1;
 839         }
 840         if (transaction_write_existing(tdb, recovery_offset, data, sizeof(*rec) + recovery_size) == -1) {
 841                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery data\n"));
 842                 free(data);
 843                 tdb->ecode = TDB_ERR_IO;
 844                 return -1;
 845         }
 846
 847         /* as we don't have ordered writes, we have to sync the recovery
 848            data before we update the magic to indicate that the recovery
 849            data is present */
 850         if (transaction_sync(tdb, recovery_offset, sizeof(*rec) + recovery_size) == -1) {
 851                 free(data);
 852                 return -1;
 853         }
 854
 855         free(data);
 856
 857         magic = TDB_RECOVERY_MAGIC;
 858         CONVERT(magic);
 859
 860         *magic_offset = recovery_offset + offsetof(struct list_struct, magic);
 861
 862         if (methods->tdb_write(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 863                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write recovery magic\n"));
 864                 tdb->ecode = TDB_ERR_IO;
 865                 return -1;
 866         }
 867         if (transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic)) == -1) {
 868                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: failed to write secondary recovery magic\n"));
 869                 tdb->ecode = TDB_ERR_IO;
 870                 return -1;
 871         }
 872
 873         /* ensure the recovery magic marker is on disk */
 874         if (transaction_sync(tdb, *magic_offset, sizeof(magic)) == -1) {
 875                 return -1;
 876         }
 877
 878         return 0;
 879 }
 880
 881 /*
 882   prepare to commit the current transaction
 883 */
 884 int tdb_transaction_prepare_commit(struct tdb_context *tdb)
 885 {
 886         const struct tdb_methods *methods;
 887
 888         if (tdb->transaction == NULL) {
 889                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: no transaction\n"));
 890                 return -1;
 891         }
 892
 893         if (tdb->transaction->prepared) {
 894                 tdb->ecode = TDB_ERR_EINVAL;
 895                 tdb_transaction_cancel(tdb);
 896                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction already prepared\n"));
 897                 return -1;
 898         }
 899
 900         if (tdb->transaction->transaction_error) {
 901                 tdb->ecode = TDB_ERR_IO;
 902                 tdb_transaction_cancel(tdb);
 903                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: transaction error pending\n"));
 904                 return -1;
 905         }
 906
 907
 908         if (tdb->transaction->nesting != 0) {
 909                 return 0;
 910         }
 911
 912         /* check for a null transaction */
 913         if (tdb->transaction->blocks == NULL) {
 914                 return 0;
 915         }
 916
 917         methods = tdb->transaction->io_methods;
 918
 919         /* if there are any locks pending then the caller has not
 920            nested their locks properly, so fail the transaction */
 921         if (tdb->num_locks || tdb->global_lock.count) {
 922                 tdb->ecode = TDB_ERR_LOCK;
 923                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: locks pending on commit\n"));
 924                 tdb_transaction_cancel(tdb);
 925                 return -1;
 926         }
 927
 928         /* upgrade the main transaction lock region to a write lock */
 929         if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) {
 930                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to upgrade hash locks\n"));
 931                 tdb->ecode = TDB_ERR_LOCK;
 932                 tdb_transaction_cancel(tdb);
 933                 return -1;
 934         }
 935
 936         /* get the global lock - this prevents new users attaching to the database
 937            during the commit */
 938         if (tdb_brlock(tdb, GLOBAL_LOCK, F_WRLCK, F_SETLKW, 0, 1) == -1) {
 939                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_prepare_commit: failed to get global lock\n"));
 940                 tdb->ecode = TDB_ERR_LOCK;
 941                 tdb_transaction_cancel(tdb);
 942                 return -1;
 943         }
 944
 945         if (!(tdb->flags & TDB_NOSYNC)) {
 946                 /* write the recovery data to the end of the file */
 947                 if (transaction_setup_recovery(tdb, &tdb->transaction->magic_offset) == -1) {
 948                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: failed to setup recovery data\n"));
 949                         tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
 950                         tdb_transaction_cancel(tdb);
 951                         return -1;
 952                 }
 953         }
 954
 955         tdb->transaction->prepared = true;
 956
 957         /* expand the file to the new size if needed */
 958         if (tdb->map_size != tdb->transaction->old_map_size) {
 959                 if (methods->tdb_expand_file(tdb, tdb->transaction->old_map_size,
 960                                              tdb->map_size -
 961                                              tdb->transaction->old_map_size) == -1) {
 962                         tdb->ecode = TDB_ERR_IO;
 963                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_prepare_commit: expansion failed\n"));
 964                         tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
 965                         tdb_transaction_cancel(tdb);
 966                         return -1;
 967                 }
 968                 tdb->map_size = tdb->transaction->old_map_size;
 969                 methods->tdb_oob(tdb, tdb->map_size + 1, 1);
 970         }
 971
 972         /* Keep the global lock until the actual commit */
 973
 974         return 0;
 975 }
 976
 977 /*
 978   commit the current transaction
 979 */
 980 int tdb_transaction_commit(struct tdb_context *tdb)
 981 {
 982         const struct tdb_methods *methods;
 983         int i;
 984         bool need_repack;
 985
 986         if (tdb->transaction == NULL) {
 987                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
 988                 return -1;
 989         }
 990
 991         if (tdb->transaction->transaction_error) {
 992                 tdb->ecode = TDB_ERR_IO;
 993                 tdb_transaction_cancel(tdb);
 994                 TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: transaction error pending\n"));
 995                 return -1;
 996         }
 997
 998
 999         if (tdb->transaction->nesting != 0) {
1000                 tdb->transaction->nesting--;
1001                 return 0;
1002         }
1003
1004         /* check for a null transaction */
1005         if (tdb->transaction->blocks == NULL) {
1006                 tdb_transaction_cancel(tdb);
1007                 return 0;
1008         }
1009
1010         if (!tdb->transaction->prepared) {
1011                 int ret = tdb_transaction_prepare_commit(tdb);
1012                 if (ret)
1013                         return ret;
1014         }
1015
1016         methods = tdb->transaction->io_methods;
1017
1018         /* perform all the writes */
1019         for (i=0;i<tdb->transaction->num_blocks;i++) {
1020                 tdb_off_t offset;
1021                 tdb_len_t length;
1022
1023                 if (tdb->transaction->blocks[i] == NULL) {
1024                         continue;
1025                 }
1026
1027                 offset = i * tdb->transaction->block_size;
1028                 length = tdb->transaction->block_size;
1029                 if (i == tdb->transaction->num_blocks-1) {
1030                         length = tdb->transaction->last_block_size;
1031                 }
1032
1033                 if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
1034                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
1035
1036                         /* we've overwritten part of the data and
1037                            possibly expanded the file, so we need to
1038                            run the crash recovery code */
1039                         tdb->methods = methods;
1040                         tdb_transaction_recover(tdb);
1041
1042                         tdb_transaction_cancel(tdb);
1043                         tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1044
1045                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
1046                         return -1;
1047                 }
1048                 SAFE_FREE(tdb->transaction->blocks[i]);
1049         }
1050
1051         SAFE_FREE(tdb->transaction->blocks);
1052         tdb->transaction->num_blocks = 0;
1053
1054         /* ensure the new data is on disk */
1055         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1056                 return -1;
1057         }
1058
1059         tdb_brlock(tdb, GLOBAL_LOCK, F_UNLCK, F_SETLKW, 0, 1);
1060
1061         /*
1062           TODO: maybe write to some dummy hdr field, or write to magic
1063           offset without mmap, before the last sync, instead of the
1064           utime() call
1065         */
1066
1067         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1068            don't change the mtime of the file, this means the file may
1069            not be backed up (as tdb rounding to block sizes means that
1070            file size changes are quite rare too). The following forces
1071            mtime changes when a transaction completes */
1072 #ifdef HAVE_UTIME
1073         utime(tdb->name, NULL);
1074 #endif
1075
1076         need_repack = tdb->transaction->need_repack;
1077
1078         /* use a transaction cancel to free memory and remove the
1079            transaction locks */
1080         tdb_transaction_cancel(tdb);
1081
1082         if (need_repack) {
1083                 return tdb_repack(tdb);
1084         }
1085
1086         return 0;
1087 }
1088
1089
1090 /*
1091   recover from an aborted transaction. Must be called with exclusive
1092   database write access already established (including the global
1093   lock to prevent new processes attaching)
1094 */
1095 int tdb_transaction_recover(struct tdb_context *tdb)
1096 {
1097         tdb_off_t recovery_head, recovery_eof;
1098         unsigned char *data, *p;
1099         uint32_t zero = 0;
1100         struct list_struct rec;
1101
1102         /* find the recovery area */
1103         if (tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &recovery_head) == -1) {
1104                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery head\n"));
1105                 tdb->ecode = TDB_ERR_IO;
1106                 return -1;
1107         }
1108
1109         if (recovery_head == 0) {
1110                 /* we have never allocated a recovery record */
1111                 return 0;
1112         }
1113
1114         /* read the recovery record */
1115         if (tdb->methods->tdb_read(tdb, recovery_head, &rec,
1116                                    sizeof(rec), DOCONV()) == -1) {
1117                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery record\n"));
1118                 tdb->ecode = TDB_ERR_IO;
1119                 return -1;
1120         }
1121
1122         if (rec.magic != TDB_RECOVERY_MAGIC) {
1123                 /* there is no valid recovery data */
1124                 return 0;
1125         }
1126
1127         if (tdb->read_only) {
1128                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: attempt to recover read only database\n"));
1129                 tdb->ecode = TDB_ERR_CORRUPT;
1130                 return -1;
1131         }
1132
1133         recovery_eof = rec.key_len;
1134
1135         data = (unsigned char *)malloc(rec.data_len);
1136         if (data == NULL) {
1137                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to allocate recovery data\n"));
1138                 tdb->ecode = TDB_ERR_OOM;
1139                 return -1;
1140         }
1141
1142         /* read the full recovery data */
1143         if (tdb->methods->tdb_read(tdb, recovery_head + sizeof(rec), data,
1144                                    rec.data_len, 0) == -1) {
1145                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to read recovery data\n"));
1146                 tdb->ecode = TDB_ERR_IO;
1147                 return -1;
1148         }
1149
1150         /* recover the file data */
1151         p = data;
1152         while (p+8 < data + rec.data_len) {
1153                 uint32_t ofs, len;
1154                 if (DOCONV()) {
1155                         tdb_convert(p, 8);
1156                 }
1157                 memcpy(&ofs, p, 4);
1158                 memcpy(&len, p+4, 4);
1159
1160                 if (tdb->methods->tdb_write(tdb, ofs, p+8, len) == -1) {
1161                         free(data);
1162                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to recover %d bytes at offset %d\n", len, ofs));
1163                         tdb->ecode = TDB_ERR_IO;
1164                         return -1;
1165                 }
1166                 p += 8 + len;
1167         }
1168
1169         free(data);
1170
1171         if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
1172                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync recovery\n"));
1173                 tdb->ecode = TDB_ERR_IO;
1174                 return -1;
1175         }
1176
1177         /* if the recovery area is after the recovered eof then remove it */
1178         if (recovery_eof <= recovery_head) {
1179                 if (tdb_ofs_write(tdb, TDB_RECOVERY_HEAD, &zero) == -1) {
1180                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery head\n"));
1181                         tdb->ecode = TDB_ERR_IO;
1182                         return -1;
1183                 }
1184         }
1185
1186         /* remove the recovery magic */
1187         if (tdb_ofs_write(tdb, recovery_head + offsetof(struct list_struct, magic),
1188                           &zero) == -1) {
1189                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to remove recovery magic\n"));
1190                 tdb->ecode = TDB_ERR_IO;
1191                 return -1;
1192         }
1193
1194         /* reduce the file size to the old size */
1195         tdb_munmap(tdb);
1196         if (ftruncate(tdb->fd, recovery_eof) != 0) {
1197                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to reduce to recovery size\n"));
1198                 tdb->ecode = TDB_ERR_IO;
1199                 return -1;
1200         }
1201         tdb->map_size = recovery_eof;
1202         tdb_mmap(tdb);
1203
1204         if (transaction_sync(tdb, 0, recovery_eof) == -1) {
1205                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_recover: failed to sync2 recovery\n"));
1206                 tdb->ecode = TDB_ERR_IO;
1207                 return -1;
1208         }
1209
1210         TDB_LOG((tdb, TDB_DEBUG_TRACE, "tdb_transaction_recover: recovered %d byte database\n",
1211                  recovery_eof));
1212
1213         /* all done */
1214         return 0;
1215 }