git.ozlabs.org Git - ccan/blob - ccan/tdb2/transaction.c

   1  /*
   2    Unix SMB/CIFS implementation.
   3
   4    trivial database library
   5
   6    Copyright (C) Andrew Tridgell              2005
   7    Copyright (C) Rusty Russell                2010
   8
   9      ** NOTE! The following LGPL license applies to the tdb
  10      ** library. This does NOT imply that all of Samba is released
  11      ** under the LGPL
  12
  13    This library is free software; you can redistribute it and/or
  14    modify it under the terms of the GNU Lesser General Public
  15    License as published by the Free Software Foundation; either
  16    version 3 of the License, or (at your option) any later version.
  17
  18    This library is distributed in the hope that it will be useful,
  19    but WITHOUT ANY WARRANTY; without even the implied warranty of
  20    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  21    Lesser General Public License for more details.
  22
  23    You should have received a copy of the GNU Lesser General Public
  24    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  25 */
  26
  27 #include "private.h"
  28 #define SAFE_FREE(x) do { if ((x) != NULL) {free(x); (x)=NULL;} } while(0)
  29
  30 /*
  31   transaction design:
  32
  33   - only allow a single transaction at a time per database. This makes
  34     using the transaction API simpler, as otherwise the caller would
  35     have to cope with temporary failures in transactions that conflict
  36     with other current transactions
  37
  38   - keep the transaction recovery information in the same file as the
  39     database, using a special 'transaction recovery' record pointed at
  40     by the header. This removes the need for extra journal files as
  41     used by some other databases
  42
  43   - dynamically allocated the transaction recover record, re-using it
  44     for subsequent transactions. If a larger record is needed then
  45     tdb_free() the old record to place it on the normal tdb freelist
  46     before allocating the new record
  47
  48   - during transactions, keep a linked list of writes all that have
  49     been performed by intercepting all tdb_write() calls. The hooked
  50     transaction versions of tdb_read() and tdb_write() check this
  51     linked list and try to use the elements of the list in preference
  52     to the real database.
  53
  54   - don't allow any locks to be held when a transaction starts,
  55     otherwise we can end up with deadlock (plus lack of lock nesting
  56     in POSIX locks would mean the lock is lost)
  57
  58   - if the caller gains a lock during the transaction but doesn't
  59     release it then fail the commit
  60
  61   - allow for nested calls to tdb_transaction_start(), re-using the
  62     existing transaction record. If the inner transaction is canceled
  63     then a subsequent commit will fail
  64
  65   - keep a mirrored copy of the tdb hash chain heads to allow for the
  66     fast hash heads scan on traverse, updating the mirrored copy in
  67     the transaction version of tdb_write
  68
  69   - allow callers to mix transaction and non-transaction use of tdb,
  70     although once a transaction is started then an exclusive lock is
  71     gained until the transaction is committed or canceled
  72
  73   - the commit stategy involves first saving away all modified data
  74     into a linearised buffer in the transaction recovery area, then
  75     marking the transaction recovery area with a magic value to
  76     indicate a valid recovery record. In total 4 fsync/msync calls are
  77     needed per commit to prevent race conditions. It might be possible
  78     to reduce this to 3 or even 2 with some more work.
  79
  80   - check for a valid recovery record on open of the tdb, while the
  81     open lock is held. Automatically recover from the transaction
  82     recovery area if needed, then continue with the open as
  83     usual. This allows for smooth crash recovery with no administrator
  84     intervention.
  85
  86   - if TDB_NOSYNC is passed to flags in tdb_open then transactions are
  87     still available, but no transaction recovery area is used and no
  88     fsync/msync calls are made.
  89 */
  90
  91
  92 /*
  93   hold the context of any current transaction
  94 */
  95 struct tdb_transaction {
  96         /* the original io methods - used to do IOs to the real db */
  97         const struct tdb_methods *io_methods;
  98
  99         /* the list of transaction blocks. When a block is first
 100            written to, it gets created in this list */
 101         uint8_t **blocks;
 102         size_t num_blocks;
 103         size_t last_block_size; /* number of valid bytes in the last block */
 104
 105         /* non-zero when an internal transaction error has
 106            occurred. All write operations will then fail until the
 107            transaction is ended */
 108         int transaction_error;
 109
 110         /* when inside a transaction we need to keep track of any
 111            nested tdb_transaction_start() calls, as these are allowed,
 112            but don't create a new transaction */
 113         unsigned int nesting;
 114
 115         /* set when a prepare has already occurred */
 116         bool prepared;
 117         tdb_off_t magic_offset;
 118
 119         /* old file size before transaction */
 120         tdb_len_t old_map_size;
 121 };
 122
 123 /* This doesn't really need to be pagesize, but we use it for similar reasons. */
 124 #define PAGESIZE 4096
 125
 126 /*
 127   read while in a transaction. We need to check first if the data is in our list
 128   of transaction elements, then if not do a real read
 129 */
 130 static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
 131                                        void *buf, tdb_len_t len)
 132 {
 133         size_t blk;
 134         enum TDB_ERROR ecode;
 135
 136         /* break it down into block sized ops */
 137         while (len + (off % PAGESIZE) > PAGESIZE) {
 138                 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
 139                 ecode = transaction_read(tdb, off, buf, len2);
 140                 if (ecode != TDB_SUCCESS) {
 141                         return ecode;
 142                 }
 143                 len -= len2;
 144                 off += len2;
 145                 buf = (void *)(len2 + (char *)buf);
 146         }
 147
 148         if (len == 0) {
 149                 return TDB_SUCCESS;
 150         }
 151
 152         blk = off / PAGESIZE;
 153
 154         /* see if we have it in the block list */
 155         if (tdb->transaction->num_blocks <= blk ||
 156             tdb->transaction->blocks[blk] == NULL) {
 157                 /* nope, do a real read */
 158                 ecode = tdb->transaction->io_methods->tread(tdb, off, buf, len);
 159                 if (ecode != TDB_SUCCESS) {
 160                         goto fail;
 161                 }
 162                 return 0;
 163         }
 164
 165         /* it is in the block list. Now check for the last block */
 166         if (blk == tdb->transaction->num_blocks-1) {
 167                 if (len > tdb->transaction->last_block_size) {
 168                         ecode = TDB_ERR_IO;
 169                         goto fail;
 170                 }
 171         }
 172
 173         /* now copy it out of this block */
 174         memcpy(buf, tdb->transaction->blocks[blk] + (off % PAGESIZE), len);
 175         return TDB_SUCCESS;
 176
 177 fail:
 178         tdb->transaction->transaction_error = 1;
 179         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 180                           "transaction_read: failed at off=%zu len=%zu",
 181                           (size_t)off, (size_t)len);
 182 }
 183
 184
 185 /*
 186   write while in a transaction
 187 */
 188 static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
 189                                         const void *buf, tdb_len_t len)
 190 {
 191         size_t blk;
 192         enum TDB_ERROR ecode;
 193
 194         /* Only a commit is allowed on a prepared transaction */
 195         if (tdb->transaction->prepared) {
 196                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
 197                                    "transaction_write: transaction already"
 198                                    " prepared, write not allowed");
 199                 goto fail;
 200         }
 201
 202         /* break it up into block sized chunks */
 203         while (len + (off % PAGESIZE) > PAGESIZE) {
 204                 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
 205                 ecode = transaction_write(tdb, off, buf, len2);
 206                 if (ecode != TDB_SUCCESS) {
 207                         return -1;
 208                 }
 209                 len -= len2;
 210                 off += len2;
 211                 if (buf != NULL) {
 212                         buf = (const void *)(len2 + (const char *)buf);
 213                 }
 214         }
 215
 216         if (len == 0) {
 217                 return TDB_SUCCESS;
 218         }
 219
 220         blk = off / PAGESIZE;
 221         off = off % PAGESIZE;
 222
 223         if (tdb->transaction->num_blocks <= blk) {
 224                 uint8_t **new_blocks;
 225                 /* expand the blocks array */
 226                 if (tdb->transaction->blocks == NULL) {
 227                         new_blocks = (uint8_t **)malloc(
 228                                 (blk+1)*sizeof(uint8_t *));
 229                 } else {
 230                         new_blocks = (uint8_t **)realloc(
 231                                 tdb->transaction->blocks,
 232                                 (blk+1)*sizeof(uint8_t *));
 233                 }
 234                 if (new_blocks == NULL) {
 235                         ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 236                                            "transaction_write:"
 237                                            " failed to allocate");
 238                         goto fail;
 239                 }
 240                 memset(&new_blocks[tdb->transaction->num_blocks], 0,
 241                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
 242                 tdb->transaction->blocks = new_blocks;
 243                 tdb->transaction->num_blocks = blk+1;
 244                 tdb->transaction->last_block_size = 0;
 245         }
 246
 247         /* allocate and fill a block? */
 248         if (tdb->transaction->blocks[blk] == NULL) {
 249                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(PAGESIZE, 1);
 250                 if (tdb->transaction->blocks[blk] == NULL) {
 251                         ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 252                                            "transaction_write:"
 253                                            " failed to allocate");
 254                         goto fail;
 255                 }
 256                 if (tdb->transaction->old_map_size > blk * PAGESIZE) {
 257                         tdb_len_t len2 = PAGESIZE;
 258                         if (len2 + (blk * PAGESIZE) > tdb->transaction->old_map_size) {
 259                                 len2 = tdb->transaction->old_map_size - (blk * PAGESIZE);
 260                         }
 261                         ecode = tdb->transaction->io_methods->tread(tdb,
 262                                         blk * PAGESIZE,
 263                                         tdb->transaction->blocks[blk],
 264                                         len2);
 265                         if (ecode != TDB_SUCCESS) {
 266                                 ecode = tdb_logerr(tdb, ecode,
 267                                                    TDB_LOG_ERROR,
 268                                                    "transaction_write:"
 269                                                    " failed to"
 270                                                    " read old block: %s",
 271                                                    strerror(errno));
 272                                 SAFE_FREE(tdb->transaction->blocks[blk]);
 273                                 goto fail;
 274                         }
 275                         if (blk == tdb->transaction->num_blocks-1) {
 276                                 tdb->transaction->last_block_size = len2;
 277                         }
 278                 }
 279         }
 280
 281         /* overwrite part of an existing block */
 282         if (buf == NULL) {
 283                 memset(tdb->transaction->blocks[blk] + off, 0, len);
 284         } else {
 285                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 286         }
 287         if (blk == tdb->transaction->num_blocks-1) {
 288                 if (len + off > tdb->transaction->last_block_size) {
 289                         tdb->transaction->last_block_size = len + off;
 290                 }
 291         }
 292
 293         return TDB_SUCCESS;
 294
 295 fail:
 296         tdb->transaction->transaction_error = 1;
 297         return ecode;
 298 }
 299
 300
 301 /*
 302   write while in a transaction - this variant never expands the transaction blocks, it only
 303   updates existing blocks. This means it cannot change the recovery size
 304 */
 305 static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
 306                                        const void *buf, tdb_len_t len)
 307 {
 308         size_t blk;
 309
 310         /* break it up into block sized chunks */
 311         while (len + (off % PAGESIZE) > PAGESIZE) {
 312                 tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
 313                 transaction_write_existing(tdb, off, buf, len2);
 314                 len -= len2;
 315                 off += len2;
 316                 if (buf != NULL) {
 317                         buf = (const void *)(len2 + (const char *)buf);
 318                 }
 319         }
 320
 321         if (len == 0) {
 322                 return;
 323         }
 324
 325         blk = off / PAGESIZE;
 326         off = off % PAGESIZE;
 327
 328         if (tdb->transaction->num_blocks <= blk ||
 329             tdb->transaction->blocks[blk] == NULL) {
 330                 return;
 331         }
 332
 333         if (blk == tdb->transaction->num_blocks-1 &&
 334             off + len > tdb->transaction->last_block_size) {
 335                 if (off >= tdb->transaction->last_block_size) {
 336                         return;
 337                 }
 338                 len = tdb->transaction->last_block_size - off;
 339         }
 340
 341         /* overwrite part of an existing block */
 342         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 343 }
 344
 345
 346 /*
 347   out of bounds check during a transaction
 348 */
 349 static enum TDB_ERROR transaction_oob(struct tdb_context *tdb, tdb_off_t len,
 350                                       bool probe)
 351 {
 352         if (len <= tdb->file->map_size) {
 353                 return TDB_SUCCESS;
 354         }
 355         if (!probe) {
 356                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 357                            "tdb_oob len %lld beyond transaction size %lld",
 358                            (long long)len,
 359                            (long long)tdb->file->map_size);
 360         }
 361         return TDB_ERR_IO;
 362 }
 363
 364 /*
 365   transaction version of tdb_expand().
 366 */
 367 static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
 368                                               tdb_off_t addition)
 369 {
 370         enum TDB_ERROR ecode;
 371
 372         /* add a write to the transaction elements, so subsequent
 373            reads see the zero data */
 374         ecode = transaction_write(tdb, tdb->file->map_size, NULL, addition);
 375         if (ecode == TDB_SUCCESS) {
 376                 tdb->file->map_size += addition;
 377         }
 378         return ecode;
 379 }
 380
 381 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
 382                                 size_t len, bool write_mode)
 383 {
 384         size_t blk = off / PAGESIZE, end_blk;
 385
 386         /* This is wrong for zero-length blocks, but will fail gracefully */
 387         end_blk = (off + len - 1) / PAGESIZE;
 388
 389         /* Can only do direct if in single block and we've already copied. */
 390         if (write_mode) {
 391                 if (blk != end_blk)
 392                         return NULL;
 393                 if (blk >= tdb->transaction->num_blocks)
 394                         return NULL;
 395                 if (tdb->transaction->blocks[blk] == NULL)
 396                         return NULL;
 397                 return tdb->transaction->blocks[blk] + off % PAGESIZE;
 398         }
 399
 400         /* Single which we have copied? */
 401         if (blk == end_blk
 402             && blk < tdb->transaction->num_blocks
 403             && tdb->transaction->blocks[blk])
 404                 return tdb->transaction->blocks[blk] + off % PAGESIZE;
 405
 406         /* Otherwise must be all not copied. */
 407         while (blk <= end_blk) {
 408                 if (blk >= tdb->transaction->num_blocks)
 409                         break;
 410                 if (tdb->transaction->blocks[blk])
 411                         return NULL;
 412                 blk++;
 413         }
 414         return tdb->transaction->io_methods->direct(tdb, off, len, false);
 415 }
 416
 417 static const struct tdb_methods transaction_methods = {
 418         transaction_read,
 419         transaction_write,
 420         transaction_oob,
 421         transaction_expand_file,
 422         transaction_direct,
 423 };
 424
 425 /*
 426   sync to disk
 427 */
 428 static enum TDB_ERROR transaction_sync(struct tdb_context *tdb,
 429                                        tdb_off_t offset, tdb_len_t length)
 430 {
 431         if (tdb->flags & TDB_NOSYNC) {
 432                 return TDB_SUCCESS;
 433         }
 434
 435         if (fsync(tdb->file->fd) != 0) {
 436                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 437                                   "tdb_transaction: fsync failed: %s",
 438                                   strerror(errno));
 439         }
 440 #ifdef MS_SYNC
 441         if (tdb->file->map_ptr) {
 442                 tdb_off_t moffset = offset & ~(PAGESIZE-1);
 443                 if (msync(moffset + (char *)tdb->file->map_ptr,
 444                           length + (offset - moffset), MS_SYNC) != 0) {
 445                         return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 446                                           "tdb_transaction: msync failed: %s",
 447                                           strerror(errno));
 448                 }
 449         }
 450 #endif
 451         return TDB_SUCCESS;
 452 }
 453
 454
 455 static void _tdb_transaction_cancel(struct tdb_context *tdb)
 456 {
 457         int i;
 458         enum TDB_ERROR ecode;
 459
 460         if (tdb->transaction == NULL) {
 461                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 462                            "tdb_transaction_cancel: no transaction");
 463                 return;
 464         }
 465
 466         if (tdb->transaction->nesting != 0) {
 467                 tdb->transaction->transaction_error = 1;
 468                 tdb->transaction->nesting--;
 469                 return;
 470         }
 471
 472         tdb->file->map_size = tdb->transaction->old_map_size;
 473
 474         /* free all the transaction blocks */
 475         for (i=0;i<tdb->transaction->num_blocks;i++) {
 476                 if (tdb->transaction->blocks[i] != NULL) {
 477                         free(tdb->transaction->blocks[i]);
 478                 }
 479         }
 480         SAFE_FREE(tdb->transaction->blocks);
 481
 482         if (tdb->transaction->magic_offset) {
 483                 const struct tdb_methods *methods = tdb->transaction->io_methods;
 484                 uint64_t invalid = TDB_RECOVERY_INVALID_MAGIC;
 485
 486                 /* remove the recovery marker */
 487                 ecode = methods->twrite(tdb, tdb->transaction->magic_offset,
 488                                         &invalid, sizeof(invalid));
 489                 if (ecode == TDB_SUCCESS)
 490                         ecode = transaction_sync(tdb,
 491                                                  tdb->transaction->magic_offset,
 492                                                  sizeof(invalid));
 493                 if (ecode != TDB_SUCCESS) {
 494                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 495                                    "tdb_transaction_cancel: failed to remove"
 496                                    " recovery magic");
 497                 }
 498         }
 499
 500         if (tdb->file->allrecord_lock.count)
 501                 tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
 502
 503         /* restore the normal io methods */
 504         tdb->methods = tdb->transaction->io_methods;
 505
 506         tdb_transaction_unlock(tdb, F_WRLCK);
 507
 508         if (tdb_has_open_lock(tdb))
 509                 tdb_unlock_open(tdb);
 510
 511         SAFE_FREE(tdb->transaction);
 512 }
 513
 514 /*
 515   start a tdb transaction. No token is returned, as only a single
 516   transaction is allowed to be pending per tdb_context
 517 */
 518 enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
 519 {
 520         enum TDB_ERROR ecode;
 521
 522         /* some sanity checks */
 523         if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
 524                 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 525                                                     TDB_LOG_USE_ERROR,
 526                                                     "tdb_transaction_start:"
 527                                                     " cannot start a"
 528                                                     " transaction on a "
 529                                                     "read-only or internal db");
 530         }
 531
 532         /* cope with nested tdb_transaction_start() calls */
 533         if (tdb->transaction != NULL) {
 534                 if (!(tdb->flags & TDB_ALLOW_NESTING)) {
 535                         return tdb->last_error
 536                                 = tdb_logerr(tdb, TDB_ERR_IO,
 537                                              TDB_LOG_USE_ERROR,
 538                                              "tdb_transaction_start:"
 539                                              " already inside transaction");
 540                 }
 541                 tdb->transaction->nesting++;
 542                 return 0;
 543         }
 544
 545         if (tdb_has_hash_locks(tdb)) {
 546                 /* the caller must not have any locks when starting a
 547                    transaction as otherwise we'll be screwed by lack
 548                    of nested locks in POSIX */
 549                 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
 550                                                     TDB_LOG_USE_ERROR,
 551                                                     "tdb_transaction_start:"
 552                                                     " cannot start a"
 553                                                     " transaction with locks"
 554                                                     " held");
 555         }
 556
 557         tdb->transaction = (struct tdb_transaction *)
 558                 calloc(sizeof(struct tdb_transaction), 1);
 559         if (tdb->transaction == NULL) {
 560                 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_OOM,
 561                                                     TDB_LOG_ERROR,
 562                                                     "tdb_transaction_start:"
 563                                                     " cannot allocate");
 564         }
 565
 566         /* get the transaction write lock. This is a blocking lock. As
 567            discussed with Volker, there are a number of ways we could
 568            make this async, which we will probably do in the future */
 569         ecode = tdb_transaction_lock(tdb, F_WRLCK);
 570         if (ecode != TDB_SUCCESS) {
 571                 SAFE_FREE(tdb->transaction->blocks);
 572                 SAFE_FREE(tdb->transaction);
 573                 return tdb->last_error = ecode;
 574         }
 575
 576         /* get a read lock over entire file. This is upgraded to a write
 577            lock during the commit */
 578         ecode = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true);
 579         if (ecode != TDB_SUCCESS) {
 580                 goto fail_allrecord_lock;
 581         }
 582
 583         /* make sure we know about any file expansions already done by
 584            anyone else */
 585         tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
 586         tdb->transaction->old_map_size = tdb->file->map_size;
 587
 588         /* finally hook the io methods, replacing them with
 589            transaction specific methods */
 590         tdb->transaction->io_methods = tdb->methods;
 591         tdb->methods = &transaction_methods;
 592         return tdb->last_error = TDB_SUCCESS;
 593
 594 fail_allrecord_lock:
 595         tdb_transaction_unlock(tdb, F_WRLCK);
 596         SAFE_FREE(tdb->transaction->blocks);
 597         SAFE_FREE(tdb->transaction);
 598         return tdb->last_error = ecode;
 599 }
 600
 601
 602 /*
 603   cancel the current transaction
 604 */
 605 void tdb_transaction_cancel(struct tdb_context *tdb)
 606 {
 607         _tdb_transaction_cancel(tdb);
 608 }
 609
 610 /*
 611   work out how much space the linearised recovery data will consume
 612 */
 613 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
 614 {
 615         tdb_len_t recovery_size = 0;
 616         int i;
 617
 618         recovery_size = sizeof(tdb_len_t);
 619         for (i=0;i<tdb->transaction->num_blocks;i++) {
 620                 if (i * PAGESIZE >= tdb->transaction->old_map_size) {
 621                         break;
 622                 }
 623                 if (tdb->transaction->blocks[i] == NULL) {
 624                         continue;
 625                 }
 626                 recovery_size += 2*sizeof(tdb_off_t);
 627                 if (i == tdb->transaction->num_blocks-1) {
 628                         recovery_size += tdb->transaction->last_block_size;
 629                 } else {
 630                         recovery_size += PAGESIZE;
 631                 }
 632         }
 633
 634         return recovery_size;
 635 }
 636
 637 /*
 638   allocate the recovery area, or use an existing recovery area if it is
 639   large enough
 640 */
 641 static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
 642                                             tdb_len_t *recovery_size,
 643                                             tdb_off_t *recovery_offset,
 644                                             tdb_len_t *recovery_max_size)
 645 {
 646         struct tdb_recovery_record rec;
 647         const struct tdb_methods *methods = tdb->transaction->io_methods;
 648         tdb_off_t recovery_head;
 649         size_t addition;
 650         enum TDB_ERROR ecode;
 651
 652         recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
 653         if (TDB_OFF_IS_ERR(recovery_head)) {
 654                 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
 655                                   "tdb_recovery_allocate:"
 656                                   " failed to read recovery head");
 657         }
 658
 659         if (recovery_head != 0) {
 660                 ecode = methods->tread(tdb, recovery_head, &rec, sizeof(rec));
 661                 if (ecode != TDB_SUCCESS) {
 662                         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 663                                           "tdb_recovery_allocate:"
 664                                           " failed to read recovery record");
 665                 }
 666                 tdb_convert(tdb, &rec, sizeof(rec));
 667                 /* ignore invalid recovery regions: can happen in crash */
 668                 if (rec.magic != TDB_RECOVERY_MAGIC &&
 669                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
 670                         recovery_head = 0;
 671                 }
 672         }
 673
 674         *recovery_size = tdb_recovery_size(tdb);
 675
 676         if (recovery_head != 0 && *recovery_size <= rec.max_len) {
 677                 /* it fits in the existing area */
 678                 *recovery_max_size = rec.max_len;
 679                 *recovery_offset = recovery_head;
 680                 return TDB_SUCCESS;
 681         }
 682
 683         /* we need to free up the old recovery area, then allocate a
 684            new one at the end of the file. Note that we cannot use
 685            normal allocation to allocate the new one as that might return
 686            us an area that is being currently used (as of the start of
 687            the transaction) */
 688         if (recovery_head != 0) {
 689                 tdb->stats.frees++;
 690                 ecode = add_free_record(tdb, recovery_head,
 691                                         sizeof(rec) + rec.max_len);
 692                 if (ecode != TDB_SUCCESS) {
 693                         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 694                                           "tdb_recovery_allocate:"
 695                                           " failed to free previous"
 696                                           " recovery area");
 697                 }
 698         }
 699
 700         /* the tdb_free() call might have increased the recovery size */
 701         *recovery_size = tdb_recovery_size(tdb);
 702
 703         /* round up to a multiple of page size */
 704         *recovery_max_size
 705                 = (((sizeof(rec) + *recovery_size) + PAGESIZE-1)
 706                    & ~(PAGESIZE-1))
 707                 - sizeof(rec);
 708         *recovery_offset = tdb->file->map_size;
 709         recovery_head = *recovery_offset;
 710
 711         /* Restore ->map_size before calling underlying expand_file.
 712            Also so that we don't try to expand the file again in the
 713            transaction commit, which would destroy the recovery
 714            area */
 715         addition = (tdb->file->map_size - tdb->transaction->old_map_size) +
 716                 sizeof(rec) + *recovery_max_size;
 717         tdb->file->map_size = tdb->transaction->old_map_size;
 718         ecode = methods->expand_file(tdb, addition);
 719         if (ecode != TDB_SUCCESS) {
 720                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 721                                   "tdb_recovery_allocate:"
 722                                   " failed to create recovery area");
 723         }
 724
 725         /* we have to reset the old map size so that we don't try to
 726            expand the file again in the transaction commit, which
 727            would destroy the recovery area */
 728         tdb->transaction->old_map_size = tdb->file->map_size;
 729
 730         /* write the recovery header offset and sync - we can sync without a race here
 731            as the magic ptr in the recovery record has not been set */
 732         tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
 733         ecode = methods->twrite(tdb, offsetof(struct tdb_header, recovery),
 734                                 &recovery_head, sizeof(tdb_off_t));
 735         if (ecode != TDB_SUCCESS) {
 736                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 737                                   "tdb_recovery_allocate:"
 738                                   " failed to write recovery head");
 739         }
 740         transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
 741                                    &recovery_head,
 742                                    sizeof(tdb_off_t));
 743         return TDB_SUCCESS;
 744 }
 745
 746 /* Set up header for the recovery record. */
 747 static void set_recovery_header(struct tdb_recovery_record *rec,
 748                                 uint64_t magic,
 749                                 uint64_t datalen, uint64_t actuallen,
 750                                 uint64_t oldsize)
 751 {
 752         rec->magic = magic;
 753         rec->max_len = actuallen;
 754         rec->len = datalen;
 755         rec->eof = oldsize;
 756 }
 757
 758 /*
 759   setup the recovery data that will be used on a crash during commit
 760 */
 761 static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
 762                                                  tdb_off_t *magic_offset)
 763 {
 764         /* Initialized for GCC's 4.4.5 overzealous uninitialized warnings. */
 765         tdb_len_t recovery_size = 0;
 766         tdb_off_t recovery_offset = 0, recovery_max_size = 0;
 767         unsigned char *data, *p;
 768         const struct tdb_methods *methods = tdb->transaction->io_methods;
 769         struct tdb_recovery_record *rec;
 770         tdb_off_t old_map_size = tdb->transaction->old_map_size;
 771         uint64_t magic, tailer;
 772         int i;
 773         enum TDB_ERROR ecode;
 774
 775         /*
 776           check that the recovery area has enough space
 777         */
 778         ecode = tdb_recovery_allocate(tdb, &recovery_size,
 779                                       &recovery_offset, &recovery_max_size);
 780         if (ecode != TDB_SUCCESS) {
 781                 return ecode;
 782         }
 783
 784         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
 785         if (data == NULL) {
 786                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 787                                   "transaction_setup_recovery:"
 788                                   " cannot allocate");
 789         }
 790
 791         rec = (struct tdb_recovery_record *)data;
 792         set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
 793                             recovery_size, recovery_max_size, old_map_size);
 794         tdb_convert(tdb, rec, sizeof(*rec));
 795
 796         /* build the recovery data into a single blob to allow us to do a single
 797            large write, which should be more efficient */
 798         p = data + sizeof(*rec);
 799         for (i=0;i<tdb->transaction->num_blocks;i++) {
 800                 tdb_off_t offset;
 801                 tdb_len_t length;
 802
 803                 if (tdb->transaction->blocks[i] == NULL) {
 804                         continue;
 805                 }
 806
 807                 offset = i * PAGESIZE;
 808                 length = PAGESIZE;
 809                 if (i == tdb->transaction->num_blocks-1) {
 810                         length = tdb->transaction->last_block_size;
 811                 }
 812
 813                 if (offset >= old_map_size) {
 814                         continue;
 815                 }
 816                 if (offset + length > tdb->file->map_size) {
 817                         free(data);
 818                         return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
 819                                           "tdb_transaction_setup_recovery:"
 820                                           " transaction data over new region"
 821                                           " boundary");
 822                 }
 823                 memcpy(p, &offset, sizeof(offset));
 824                 memcpy(p + sizeof(offset), &length, sizeof(length));
 825                 tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
 826
 827                 /* the recovery area contains the old data, not the
 828                    new data, so we have to call the original tdb_read
 829                    method to get it */
 830                 ecode = methods->tread(tdb, offset,
 831                                        p + sizeof(offset) + sizeof(length),
 832                                        length);
 833                 if (ecode != TDB_SUCCESS) {
 834                         free(data);
 835                         return ecode;
 836                 }
 837                 p += sizeof(offset) + sizeof(length) + length;
 838         }
 839
 840         /* and the tailer */
 841         tailer = sizeof(*rec) + recovery_max_size;
 842         memcpy(p, &tailer, sizeof(tailer));
 843         tdb_convert(tdb, p, sizeof(tailer));
 844
 845         /* write the recovery data to the recovery area */
 846         ecode = methods->twrite(tdb, recovery_offset, data,
 847                                 sizeof(*rec) + recovery_size);
 848         if (ecode != TDB_SUCCESS) {
 849                 free(data);
 850                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 851                                   "tdb_transaction_setup_recovery:"
 852                                   " failed to write recovery data");
 853         }
 854         transaction_write_existing(tdb, recovery_offset, data,
 855                                    sizeof(*rec) + recovery_size);
 856
 857         /* as we don't have ordered writes, we have to sync the recovery
 858            data before we update the magic to indicate that the recovery
 859            data is present */
 860         ecode = transaction_sync(tdb, recovery_offset,
 861                                  sizeof(*rec) + recovery_size);
 862         if (ecode != TDB_SUCCESS) {
 863                 free(data);
 864                 return ecode;
 865         }
 866
 867         free(data);
 868
 869         magic = TDB_RECOVERY_MAGIC;
 870         tdb_convert(tdb, &magic, sizeof(magic));
 871
 872         *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
 873                                                    magic);
 874
 875         ecode = methods->twrite(tdb, *magic_offset, &magic, sizeof(magic));
 876         if (ecode != TDB_SUCCESS) {
 877                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 878                                   "tdb_transaction_setup_recovery:"
 879                                   " failed to write recovery magic");
 880         }
 881         transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
 882
 883         /* ensure the recovery magic marker is on disk */
 884         return transaction_sync(tdb, *magic_offset, sizeof(magic));
 885 }
 886
 887 static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 888 {
 889         const struct tdb_methods *methods;
 890         enum TDB_ERROR ecode;
 891
 892         if (tdb->transaction == NULL) {
 893                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 894                                   "tdb_transaction_prepare_commit:"
 895                                   " no transaction");
 896         }
 897
 898         if (tdb->transaction->prepared) {
 899                 _tdb_transaction_cancel(tdb);
 900                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 901                                   "tdb_transaction_prepare_commit:"
 902                                   " transaction already prepared");
 903         }
 904
 905         if (tdb->transaction->transaction_error) {
 906                 _tdb_transaction_cancel(tdb);
 907                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
 908                                   "tdb_transaction_prepare_commit:"
 909                                   " transaction error pending");
 910         }
 911
 912
 913         if (tdb->transaction->nesting != 0) {
 914                 return TDB_SUCCESS;
 915         }
 916
 917         /* check for a null transaction */
 918         if (tdb->transaction->blocks == NULL) {
 919                 return TDB_SUCCESS;
 920         }
 921
 922         methods = tdb->transaction->io_methods;
 923
 924         /* upgrade the main transaction lock region to a write lock */
 925         ecode = tdb_allrecord_upgrade(tdb);
 926         if (ecode != TDB_SUCCESS) {
 927                 return ecode;
 928         }
 929
 930         /* get the open lock - this prevents new users attaching to the database
 931            during the commit */
 932         ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 933         if (ecode != TDB_SUCCESS) {
 934                 return ecode;
 935         }
 936
 937         /* Since we have whole db locked, we don't need the expansion lock. */
 938         if (!(tdb->flags & TDB_NOSYNC)) {
 939                 /* write the recovery data to the end of the file */
 940                 ecode = transaction_setup_recovery(tdb,
 941                                                    &tdb->transaction
 942                                                    ->magic_offset);
 943                 if (ecode != TDB_SUCCESS) {
 944                         return ecode;
 945                 }
 946         }
 947
 948         tdb->transaction->prepared = true;
 949
 950         /* expand the file to the new size if needed */
 951         if (tdb->file->map_size != tdb->transaction->old_map_size) {
 952                 tdb_len_t add;
 953
 954                 add = tdb->file->map_size - tdb->transaction->old_map_size;
 955                 /* Restore original map size for tdb_expand_file */
 956                 tdb->file->map_size = tdb->transaction->old_map_size;
 957                 ecode = methods->expand_file(tdb, add);
 958                 if (ecode != TDB_SUCCESS) {
 959                         return ecode;
 960                 }
 961         }
 962
 963         /* Keep the open lock until the actual commit */
 964         return TDB_SUCCESS;
 965 }
 966
 967 /*
 968    prepare to commit the current transaction
 969 */
 970 enum TDB_ERROR tdb_transaction_prepare_commit(struct tdb_context *tdb)
 971 {
 972         return _tdb_transaction_prepare_commit(tdb);
 973 }
 974
 975 /*
 976   commit the current transaction
 977 */
 978 enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
 979 {
 980         const struct tdb_methods *methods;
 981         int i;
 982         enum TDB_ERROR ecode;
 983
 984         if (tdb->transaction == NULL) {
 985                 return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 986                                                     TDB_LOG_USE_ERROR,
 987                                                     "tdb_transaction_commit:"
 988                                                     " no transaction");
 989         }
 990
 991         tdb_trace(tdb, "tdb_transaction_commit");
 992
 993         if (tdb->transaction->nesting != 0) {
 994                 tdb->transaction->nesting--;
 995                 return tdb->last_error = TDB_SUCCESS;
 996         }
 997
 998         /* check for a null transaction */
 999         if (tdb->transaction->blocks == NULL) {
1000                 _tdb_transaction_cancel(tdb);
1001                 return tdb->last_error = TDB_SUCCESS;
1002         }
1003
1004         if (!tdb->transaction->prepared) {
1005                 ecode = _tdb_transaction_prepare_commit(tdb);
1006                 if (ecode != TDB_SUCCESS) {
1007                         _tdb_transaction_cancel(tdb);
1008                         return tdb->last_error = ecode;
1009                 }
1010         }
1011
1012         methods = tdb->transaction->io_methods;
1013
1014         /* perform all the writes */
1015         for (i=0;i<tdb->transaction->num_blocks;i++) {
1016                 tdb_off_t offset;
1017                 tdb_len_t length;
1018
1019                 if (tdb->transaction->blocks[i] == NULL) {
1020                         continue;
1021                 }
1022
1023                 offset = i * PAGESIZE;
1024                 length = PAGESIZE;
1025                 if (i == tdb->transaction->num_blocks-1) {
1026                         length = tdb->transaction->last_block_size;
1027                 }
1028
1029                 ecode = methods->twrite(tdb, offset,
1030                                         tdb->transaction->blocks[i], length);
1031                 if (ecode != TDB_SUCCESS) {
1032                         /* we've overwritten part of the data and
1033                            possibly expanded the file, so we need to
1034                            run the crash recovery code */
1035                         tdb->methods = methods;
1036                         tdb_transaction_recover(tdb);
1037
1038                         _tdb_transaction_cancel(tdb);
1039
1040                         return tdb->last_error = ecode;
1041                 }
1042                 SAFE_FREE(tdb->transaction->blocks[i]);
1043         }
1044
1045         SAFE_FREE(tdb->transaction->blocks);
1046         tdb->transaction->num_blocks = 0;
1047
1048         /* ensure the new data is on disk */
1049         ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1050         if (ecode != TDB_SUCCESS) {
1051                 return tdb->last_error = ecode;
1052         }
1053
1054         /*
1055           TODO: maybe write to some dummy hdr field, or write to magic
1056           offset without mmap, before the last sync, instead of the
1057           utime() call
1058         */
1059
1060         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1061            don't change the mtime of the file, this means the file may
1062            not be backed up (as tdb rounding to block sizes means that
1063            file size changes are quite rare too). The following forces
1064            mtime changes when a transaction completes */
1065 #if HAVE_UTIME
1066         utime(tdb->name, NULL);
1067 #endif
1068
1069         /* use a transaction cancel to free memory and remove the
1070            transaction locks: it "restores" map_size, too. */
1071         tdb->transaction->old_map_size = tdb->file->map_size;
1072         _tdb_transaction_cancel(tdb);
1073
1074         return tdb->last_error = TDB_SUCCESS;
1075 }
1076
1077
1078 /*
1079   recover from an aborted transaction. Must be called with exclusive
1080   database write access already established (including the open
1081   lock to prevent new processes attaching)
1082 */
1083 enum TDB_ERROR tdb_transaction_recover(struct tdb_context *tdb)
1084 {
1085         tdb_off_t recovery_head, recovery_eof;
1086         unsigned char *data, *p;
1087         struct tdb_recovery_record rec;
1088         enum TDB_ERROR ecode;
1089
1090         /* find the recovery area */
1091         recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1092         if (TDB_OFF_IS_ERR(recovery_head)) {
1093                 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
1094                                   "tdb_transaction_recover:"
1095                                   " failed to read recovery head");
1096         }
1097
1098         if (recovery_head == 0) {
1099                 /* we have never allocated a recovery record */
1100                 return TDB_SUCCESS;
1101         }
1102
1103         /* read the recovery record */
1104         ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1105         if (ecode != TDB_SUCCESS) {
1106                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1107                                   "tdb_transaction_recover:"
1108                                   " failed to read recovery record");
1109         }
1110
1111         if (rec.magic != TDB_RECOVERY_MAGIC) {
1112                 /* there is no valid recovery data */
1113                 return TDB_SUCCESS;
1114         }
1115
1116         if (tdb->read_only) {
1117                 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
1118                                   "tdb_transaction_recover:"
1119                                   " attempt to recover read only database");
1120         }
1121
1122         recovery_eof = rec.eof;
1123
1124         data = (unsigned char *)malloc(rec.len);
1125         if (data == NULL) {
1126                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
1127                                   "tdb_transaction_recover:"
1128                                   " failed to allocate recovery data");
1129         }
1130
1131         /* read the full recovery data */
1132         ecode = tdb->methods->tread(tdb, recovery_head + sizeof(rec), data,
1133                                     rec.len);
1134         if (ecode != TDB_SUCCESS) {
1135                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1136                                   "tdb_transaction_recover:"
1137                                   " failed to read recovery data");
1138         }
1139
1140         /* recover the file data */
1141         p = data;
1142         while (p+sizeof(tdb_off_t)+sizeof(tdb_len_t) < data + rec.len) {
1143                 tdb_off_t ofs;
1144                 tdb_len_t len;
1145                 tdb_convert(tdb, p, sizeof(ofs) + sizeof(len));
1146                 memcpy(&ofs, p, sizeof(ofs));
1147                 memcpy(&len, p + sizeof(ofs), sizeof(len));
1148                 p += sizeof(ofs) + sizeof(len);
1149
1150                 ecode = tdb->methods->twrite(tdb, ofs, p, len);
1151                 if (ecode != TDB_SUCCESS) {
1152                         free(data);
1153                         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1154                                           "tdb_transaction_recover:"
1155                                           " failed to recover %zu bytes"
1156                                           " at offset %zu",
1157                                           (size_t)len, (size_t)ofs);
1158                 }
1159                 p += len;
1160         }
1161
1162         free(data);
1163
1164         ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1165         if (ecode != TDB_SUCCESS) {
1166                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1167                                   "tdb_transaction_recover:"
1168                                   " failed to sync recovery");
1169         }
1170
1171         /* if the recovery area is after the recovered eof then remove it */
1172         if (recovery_eof <= recovery_head) {
1173                 ecode = tdb_write_off(tdb, offsetof(struct tdb_header,
1174                                                     recovery),
1175                                       0);
1176                 if (ecode != TDB_SUCCESS) {
1177                         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1178                                           "tdb_transaction_recover:"
1179                                           " failed to remove recovery head");
1180                 }
1181         }
1182
1183         /* remove the recovery magic */
1184         ecode = tdb_write_off(tdb,
1185                               recovery_head
1186                               + offsetof(struct tdb_recovery_record, magic),
1187                               TDB_RECOVERY_INVALID_MAGIC);
1188         if (ecode != TDB_SUCCESS) {
1189                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1190                                   "tdb_transaction_recover:"
1191                                   " failed to remove recovery magic");
1192         }
1193
1194         ecode = transaction_sync(tdb, 0, recovery_eof);
1195         if (ecode != TDB_SUCCESS) {
1196                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1197                                   "tdb_transaction_recover:"
1198                                   " failed to sync2 recovery");
1199         }
1200
1201         tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
1202                    "tdb_transaction_recover: recovered %zu byte database",
1203                    (size_t)recovery_eof);
1204
1205         /* all done */
1206         return TDB_SUCCESS;
1207 }
1208
1209 tdb_bool_err tdb_needs_recovery(struct tdb_context *tdb)
1210 {
1211         tdb_off_t recovery_head;
1212         struct tdb_recovery_record rec;
1213         enum TDB_ERROR ecode;
1214
1215         /* find the recovery area */
1216         recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
1217         if (TDB_OFF_IS_ERR(recovery_head)) {
1218                 return recovery_head;
1219         }
1220
1221         if (recovery_head == 0) {
1222                 /* we have never allocated a recovery record */
1223                 return false;
1224         }
1225
1226         /* read the recovery record */
1227         ecode = tdb_read_convert(tdb, recovery_head, &rec, sizeof(rec));
1228         if (ecode != TDB_SUCCESS) {
1229                 return ecode;
1230         }
1231
1232         return (rec.magic == TDB_RECOVERY_MAGIC);
1233 }