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         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
 124 /*
 125   read while in a transaction. We need to check first if the data is in our list
 126   of transaction elements, then if not do a real read
 127 */
 128 static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
 129                                        void *buf, tdb_len_t len)
 130 {
 131         size_t blk;
 132         enum TDB_ERROR ecode;
 133
 134         /* break it down into block sized ops */
 135         while (len + (off % getpagesize()) > getpagesize()) {
 136                 tdb_len_t len2 = getpagesize() - (off % getpagesize());
 137                 ecode = transaction_read(tdb, off, buf, len2);
 138                 if (ecode != TDB_SUCCESS) {
 139                         return ecode;
 140                 }
 141                 len -= len2;
 142                 off += len2;
 143                 buf = (void *)(len2 + (char *)buf);
 144         }
 145
 146         if (len == 0) {
 147                 return TDB_SUCCESS;
 148         }
 149
 150         blk = off / getpagesize();
 151
 152         /* see if we have it in the block list */
 153         if (tdb->transaction->num_blocks <= blk ||
 154             tdb->transaction->blocks[blk] == NULL) {
 155                 /* nope, do a real read */
 156                 ecode = tdb->transaction->io_methods->tread(tdb, off, buf, len);
 157                 if (ecode != TDB_SUCCESS) {
 158                         goto fail;
 159                 }
 160                 return 0;
 161         }
 162
 163         /* it is in the block list. Now check for the last block */
 164         if (blk == tdb->transaction->num_blocks-1) {
 165                 if (len > tdb->transaction->last_block_size) {
 166                         ecode = TDB_ERR_IO;
 167                         goto fail;
 168                 }
 169         }
 170
 171         /* now copy it out of this block */
 172         memcpy(buf, tdb->transaction->blocks[blk] + (off % getpagesize()), len);
 173         return TDB_SUCCESS;
 174
 175 fail:
 176         tdb->transaction->transaction_error = 1;
 177         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 178                           "transaction_read: failed at off=%zu len=%zu",
 179                           (size_t)off, (size_t)len);
 180 }
 181
 182
 183 /*
 184   write while in a transaction
 185 */
 186 static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
 187                                         const void *buf, tdb_len_t len)
 188 {
 189         size_t blk;
 190         enum TDB_ERROR ecode;
 191
 192         /* Only a commit is allowed on a prepared transaction */
 193         if (tdb->transaction->prepared) {
 194                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
 195                                    "transaction_write: transaction already"
 196                                    " prepared, write not allowed");
 197                 goto fail;
 198         }
 199
 200         /* break it up into block sized chunks */
 201         while (len + (off % getpagesize()) > getpagesize()) {
 202                 tdb_len_t len2 = getpagesize() - (off % getpagesize());
 203                 ecode = transaction_write(tdb, off, buf, len2);
 204                 if (ecode != TDB_SUCCESS) {
 205                         return -1;
 206                 }
 207                 len -= len2;
 208                 off += len2;
 209                 if (buf != NULL) {
 210                         buf = (const void *)(len2 + (const char *)buf);
 211                 }
 212         }
 213
 214         if (len == 0) {
 215                 return TDB_SUCCESS;
 216         }
 217
 218         blk = off / getpagesize();
 219         off = off % getpagesize();
 220
 221         if (tdb->transaction->num_blocks <= blk) {
 222                 uint8_t **new_blocks;
 223                 /* expand the blocks array */
 224                 if (tdb->transaction->blocks == NULL) {
 225                         new_blocks = (uint8_t **)malloc(
 226                                 (blk+1)*sizeof(uint8_t *));
 227                 } else {
 228                         new_blocks = (uint8_t **)realloc(
 229                                 tdb->transaction->blocks,
 230                                 (blk+1)*sizeof(uint8_t *));
 231                 }
 232                 if (new_blocks == NULL) {
 233                         ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 234                                            "transaction_write:"
 235                                            " failed to allocate");
 236                         goto fail;
 237                 }
 238                 memset(&new_blocks[tdb->transaction->num_blocks], 0,
 239                        (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
 240                 tdb->transaction->blocks = new_blocks;
 241                 tdb->transaction->num_blocks = blk+1;
 242                 tdb->transaction->last_block_size = 0;
 243         }
 244
 245         /* allocate and fill a block? */
 246         if (tdb->transaction->blocks[blk] == NULL) {
 247                 tdb->transaction->blocks[blk] = (uint8_t *)calloc(getpagesize(), 1);
 248                 if (tdb->transaction->blocks[blk] == NULL) {
 249                         ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 250                                            "transaction_write:"
 251                                            " failed to allocate");
 252                         goto fail;
 253                 }
 254                 if (tdb->transaction->old_map_size > blk * getpagesize()) {
 255                         tdb_len_t len2 = getpagesize();
 256                         if (len2 + (blk * getpagesize()) > tdb->transaction->old_map_size) {
 257                                 len2 = tdb->transaction->old_map_size - (blk * getpagesize());
 258                         }
 259                         ecode = tdb->transaction->io_methods->tread(tdb,
 260                                         blk * getpagesize(),
 261                                         tdb->transaction->blocks[blk],
 262                                         len2);
 263                         if (ecode != TDB_SUCCESS) {
 264                                 ecode = tdb_logerr(tdb, ecode,
 265                                                    TDB_LOG_ERROR,
 266                                                    "transaction_write:"
 267                                                    " failed to"
 268                                                    " read old block: %s",
 269                                                    strerror(errno));
 270                                 SAFE_FREE(tdb->transaction->blocks[blk]);
 271                                 goto fail;
 272                         }
 273                         if (blk == tdb->transaction->num_blocks-1) {
 274                                 tdb->transaction->last_block_size = len2;
 275                         }
 276                 }
 277         }
 278
 279         /* overwrite part of an existing block */
 280         if (buf == NULL) {
 281                 memset(tdb->transaction->blocks[blk] + off, 0, len);
 282         } else {
 283                 memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 284         }
 285         if (blk == tdb->transaction->num_blocks-1) {
 286                 if (len + off > tdb->transaction->last_block_size) {
 287                         tdb->transaction->last_block_size = len + off;
 288                 }
 289         }
 290
 291         return TDB_SUCCESS;
 292
 293 fail:
 294         tdb->transaction->transaction_error = 1;
 295         return ecode;
 296 }
 297
 298
 299 /*
 300   write while in a transaction - this variant never expands the transaction blocks, it only
 301   updates existing blocks. This means it cannot change the recovery size
 302 */
 303 static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
 304                                        const void *buf, tdb_len_t len)
 305 {
 306         size_t blk;
 307
 308         /* break it up into block sized chunks */
 309         while (len + (off % getpagesize()) > getpagesize()) {
 310                 tdb_len_t len2 = getpagesize() - (off % getpagesize());
 311                 transaction_write_existing(tdb, off, buf, len2);
 312                 len -= len2;
 313                 off += len2;
 314                 if (buf != NULL) {
 315                         buf = (const void *)(len2 + (const char *)buf);
 316                 }
 317         }
 318
 319         if (len == 0) {
 320                 return;
 321         }
 322
 323         blk = off / getpagesize();
 324         off = off % getpagesize();
 325
 326         if (tdb->transaction->num_blocks <= blk ||
 327             tdb->transaction->blocks[blk] == NULL) {
 328                 return;
 329         }
 330
 331         if (blk == tdb->transaction->num_blocks-1 &&
 332             off + len > tdb->transaction->last_block_size) {
 333                 if (off >= tdb->transaction->last_block_size) {
 334                         return;
 335                 }
 336                 len = tdb->transaction->last_block_size - off;
 337         }
 338
 339         /* overwrite part of an existing block */
 340         memcpy(tdb->transaction->blocks[blk] + off, buf, len);
 341 }
 342
 343
 344 /*
 345   out of bounds check during a transaction
 346 */
 347 static enum TDB_ERROR transaction_oob(struct tdb_context *tdb, tdb_off_t len,
 348                                       bool probe)
 349 {
 350         if (len <= tdb->file->map_size) {
 351                 return TDB_SUCCESS;
 352         }
 353         if (!probe) {
 354                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 355                            "tdb_oob len %lld beyond transaction size %lld",
 356                            (long long)len,
 357                            (long long)tdb->file->map_size);
 358         }
 359         return TDB_ERR_IO;
 360 }
 361
 362 /*
 363   transaction version of tdb_expand().
 364 */
 365 static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
 366                                               tdb_off_t addition)
 367 {
 368         enum TDB_ERROR ecode;
 369
 370         /* add a write to the transaction elements, so subsequent
 371            reads see the zero data */
 372         ecode = transaction_write(tdb, tdb->file->map_size, NULL, addition);
 373         if (ecode == TDB_SUCCESS) {
 374                 tdb->file->map_size += addition;
 375         }
 376         return ecode;
 377 }
 378
 379 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
 380                                 size_t len, bool write_mode)
 381 {
 382         size_t blk = off / getpagesize(), end_blk;
 383
 384         /* This is wrong for zero-length blocks, but will fail gracefully */
 385         end_blk = (off + len - 1) / getpagesize();
 386
 387         /* Can only do direct if in single block and we've already copied. */
 388         if (write_mode) {
 389                 if (blk != end_blk)
 390                         return NULL;
 391                 if (blk >= tdb->transaction->num_blocks)
 392                         return NULL;
 393                 if (tdb->transaction->blocks[blk] == NULL)
 394                         return NULL;
 395                 return tdb->transaction->blocks[blk] + off % getpagesize();
 396         }
 397
 398         /* Single which we have copied? */
 399         if (blk == end_blk
 400             && blk < tdb->transaction->num_blocks
 401             && tdb->transaction->blocks[blk])
 402                 return tdb->transaction->blocks[blk] + off % getpagesize();
 403
 404         /* Otherwise must be all not copied. */
 405         while (blk < end_blk) {
 406                 if (blk >= tdb->transaction->num_blocks)
 407                         break;
 408                 if (tdb->transaction->blocks[blk])
 409                         return NULL;
 410                 blk++;
 411         }
 412         return tdb->transaction->io_methods->direct(tdb, off, len, false);
 413 }
 414
 415 static const struct tdb_methods transaction_methods = {
 416         transaction_read,
 417         transaction_write,
 418         transaction_oob,
 419         transaction_expand_file,
 420         transaction_direct,
 421 };
 422
 423 /*
 424   sync to disk
 425 */
 426 static enum TDB_ERROR transaction_sync(struct tdb_context *tdb,
 427                                        tdb_off_t offset, tdb_len_t length)
 428 {
 429         if (tdb->flags & TDB_NOSYNC) {
 430                 return TDB_SUCCESS;
 431         }
 432
 433         if (fsync(tdb->file->fd) != 0) {
 434                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 435                                   "tdb_transaction: fsync failed: %s",
 436                                   strerror(errno));
 437         }
 438 #ifdef MS_SYNC
 439         if (tdb->file->map_ptr) {
 440                 tdb_off_t moffset = offset & ~(getpagesize()-1);
 441                 if (msync(moffset + (char *)tdb->file->map_ptr,
 442                           length + (offset - moffset), MS_SYNC) != 0) {
 443                         return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 444                                           "tdb_transaction: msync failed: %s",
 445                                           strerror(errno));
 446                 }
 447         }
 448 #endif
 449         return TDB_SUCCESS;
 450 }
 451
 452
 453 static void _tdb_transaction_cancel(struct tdb_context *tdb)
 454 {
 455         int i;
 456         enum TDB_ERROR ecode;
 457
 458         if (tdb->transaction == NULL) {
 459                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 460                            "tdb_transaction_cancel: no transaction");
 461                 return;
 462         }
 463
 464         if (tdb->transaction->nesting != 0) {
 465                 tdb->transaction->transaction_error = 1;
 466                 tdb->transaction->nesting--;
 467                 return;
 468         }
 469
 470         tdb->file->map_size = tdb->transaction->old_map_size;
 471
 472         /* free all the transaction blocks */
 473         for (i=0;i<tdb->transaction->num_blocks;i++) {
 474                 if (tdb->transaction->blocks[i] != NULL) {
 475                         free(tdb->transaction->blocks[i]);
 476                 }
 477         }
 478         SAFE_FREE(tdb->transaction->blocks);
 479
 480         if (tdb->transaction->magic_offset) {
 481                 const struct tdb_methods *methods = tdb->transaction->io_methods;
 482                 uint64_t invalid = TDB_RECOVERY_INVALID_MAGIC;
 483
 484                 /* remove the recovery marker */
 485                 ecode = methods->twrite(tdb, tdb->transaction->magic_offset,
 486                                         &invalid, sizeof(invalid));
 487                 if (ecode == TDB_SUCCESS)
 488                         ecode = transaction_sync(tdb,
 489                                                  tdb->transaction->magic_offset,
 490                                                  sizeof(invalid));
 491                 if (ecode != TDB_SUCCESS) {
 492                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 493                                    "tdb_transaction_cancel: failed to remove"
 494                                    " recovery magic");
 495                 }
 496         }
 497
 498         if (tdb->file->allrecord_lock.count)
 499                 tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
 500
 501         /* restore the normal io methods */
 502         tdb->methods = tdb->transaction->io_methods;
 503
 504         tdb_transaction_unlock(tdb, F_WRLCK);
 505
 506         if (tdb_has_open_lock(tdb))
 507                 tdb_unlock_open(tdb);
 508
 509         SAFE_FREE(tdb->transaction);
 510 }
 511
 512 /*
 513   start a tdb transaction. No token is returned, as only a single
 514   transaction is allowed to be pending per tdb_context
 515 */
 516 enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
 517 {
 518         enum TDB_ERROR ecode;
 519
 520         /* some sanity checks */
 521         if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
 522                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 523                                   "tdb_transaction_start: cannot start a"
 524                                   " transaction on a read-only or internal db");
 525         }
 526
 527         /* cope with nested tdb_transaction_start() calls */
 528         if (tdb->transaction != NULL) {
 529                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_USE_ERROR,
 530                                   "tdb_transaction_start:"
 531                                   " already inside transaction");
 532         }
 533
 534         if (tdb_has_hash_locks(tdb)) {
 535                 /* the caller must not have any locks when starting a
 536                    transaction as otherwise we'll be screwed by lack
 537                    of nested locks in POSIX */
 538                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 539                                   "tdb_transaction_start: cannot start a"
 540                                   " transaction with locks held");
 541         }
 542
 543         tdb->transaction = (struct tdb_transaction *)
 544                 calloc(sizeof(struct tdb_transaction), 1);
 545         if (tdb->transaction == NULL) {
 546                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 547                                   "tdb_transaction_start: cannot allocate");
 548         }
 549
 550         /* get the transaction write lock. This is a blocking lock. As
 551            discussed with Volker, there are a number of ways we could
 552            make this async, which we will probably do in the future */
 553         ecode = tdb_transaction_lock(tdb, F_WRLCK);
 554         if (ecode != TDB_SUCCESS) {
 555                 SAFE_FREE(tdb->transaction->blocks);
 556                 SAFE_FREE(tdb->transaction);
 557                 return ecode;
 558         }
 559
 560         /* get a read lock over entire file. This is upgraded to a write
 561            lock during the commit */
 562         ecode = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, true);
 563         if (ecode != TDB_SUCCESS) {
 564                 goto fail_allrecord_lock;
 565         }
 566
 567         /* make sure we know about any file expansions already done by
 568            anyone else */
 569         tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
 570         tdb->transaction->old_map_size = tdb->file->map_size;
 571
 572         /* finally hook the io methods, replacing them with
 573            transaction specific methods */
 574         tdb->transaction->io_methods = tdb->methods;
 575         tdb->methods = &transaction_methods;
 576         return TDB_SUCCESS;
 577
 578 fail_allrecord_lock:
 579         tdb_transaction_unlock(tdb, F_WRLCK);
 580         SAFE_FREE(tdb->transaction->blocks);
 581         SAFE_FREE(tdb->transaction);
 582         return ecode;
 583 }
 584
 585
 586 /*
 587   cancel the current transaction
 588 */
 589 void tdb_transaction_cancel(struct tdb_context *tdb)
 590 {
 591         _tdb_transaction_cancel(tdb);
 592 }
 593
 594 /*
 595   work out how much space the linearised recovery data will consume
 596 */
 597 static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
 598 {
 599         tdb_len_t recovery_size = 0;
 600         int i;
 601
 602         recovery_size = sizeof(tdb_len_t);
 603         for (i=0;i<tdb->transaction->num_blocks;i++) {
 604                 if (i * getpagesize() >= tdb->transaction->old_map_size) {
 605                         break;
 606                 }
 607                 if (tdb->transaction->blocks[i] == NULL) {
 608                         continue;
 609                 }
 610                 recovery_size += 2*sizeof(tdb_off_t);
 611                 if (i == tdb->transaction->num_blocks-1) {
 612                         recovery_size += tdb->transaction->last_block_size;
 613                 } else {
 614                         recovery_size += getpagesize();
 615                 }
 616         }
 617
 618         return recovery_size;
 619 }
 620
 621 /*
 622   allocate the recovery area, or use an existing recovery area if it is
 623   large enough
 624 */
 625 static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
 626                                             tdb_len_t *recovery_size,
 627                                             tdb_off_t *recovery_offset,
 628                                             tdb_len_t *recovery_max_size)
 629 {
 630         struct tdb_recovery_record rec;
 631         const struct tdb_methods *methods = tdb->transaction->io_methods;
 632         tdb_off_t recovery_head;
 633         size_t addition;
 634         enum TDB_ERROR ecode;
 635
 636         recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
 637         if (TDB_OFF_IS_ERR(recovery_head)) {
 638                 return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
 639                                   "tdb_recovery_allocate:"
 640                                   " failed to read recovery head");
 641         }
 642
 643         if (recovery_head != 0) {
 644                 ecode = methods->tread(tdb, recovery_head, &rec, sizeof(rec));
 645                 if (ecode != TDB_SUCCESS) {
 646                         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 647                                           "tdb_recovery_allocate:"
 648                                           " failed to read recovery record");
 649                 }
 650                 tdb_convert(tdb, &rec, sizeof(rec));
 651                 /* ignore invalid recovery regions: can happen in crash */
 652                 if (rec.magic != TDB_RECOVERY_MAGIC &&
 653                     rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
 654                         recovery_head = 0;
 655                 }
 656         }
 657
 658         *recovery_size = tdb_recovery_size(tdb);
 659
 660         if (recovery_head != 0 && *recovery_size <= rec.max_len) {
 661                 /* it fits in the existing area */
 662                 *recovery_max_size = rec.max_len;
 663                 *recovery_offset = recovery_head;
 664                 return TDB_SUCCESS;
 665         }
 666
 667         /* we need to free up the old recovery area, then allocate a
 668            new one at the end of the file. Note that we cannot use
 669            normal allocation to allocate the new one as that might return
 670            us an area that is being currently used (as of the start of
 671            the transaction) */
 672         if (recovery_head != 0) {
 673                 add_stat(tdb, frees, 1);
 674                 ecode = add_free_record(tdb, recovery_head,
 675                                         sizeof(rec) + rec.max_len);
 676                 if (ecode != TDB_SUCCESS) {
 677                         return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 678                                           "tdb_recovery_allocate:"
 679                                           " failed to free previous"
 680                                           " recovery area");
 681                 }
 682         }
 683
 684         /* the tdb_free() call might have increased the recovery size */
 685         *recovery_size = tdb_recovery_size(tdb);
 686
 687         /* round up to a multiple of page size */
 688         *recovery_max_size
 689                 = (((sizeof(rec) + *recovery_size) + getpagesize()-1)
 690                    & ~(getpagesize()-1))
 691                 - sizeof(rec);
 692         *recovery_offset = tdb->file->map_size;
 693         recovery_head = *recovery_offset;
 694
 695         /* Restore ->map_size before calling underlying expand_file.
 696            Also so that we don't try to expand the file again in the
 697            transaction commit, which would destroy the recovery
 698            area */
 699         addition = (tdb->file->map_size - tdb->transaction->old_map_size) +
 700                 sizeof(rec) + *recovery_max_size;
 701         tdb->file->map_size = tdb->transaction->old_map_size;
 702         ecode = methods->expand_file(tdb, addition);
 703         if (ecode != TDB_SUCCESS) {
 704                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 705                                   "tdb_recovery_allocate:"
 706                                   " failed to create recovery area");
 707         }
 708
 709         /* we have to reset the old map size so that we don't try to
 710            expand the file again in the transaction commit, which
 711            would destroy the recovery area */
 712         tdb->transaction->old_map_size = tdb->file->map_size;
 713
 714         /* write the recovery header offset and sync - we can sync without a race here
 715            as the magic ptr in the recovery record has not been set */
 716         tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
 717         ecode = methods->twrite(tdb, offsetof(struct tdb_header, recovery),
 718                                 &recovery_head, sizeof(tdb_off_t));
 719         if (ecode != TDB_SUCCESS) {
 720                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 721                                   "tdb_recovery_allocate:"
 722                                   " failed to write recovery head");
 723         }
 724         transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
 725                                    &recovery_head,
 726                                    sizeof(tdb_off_t));
 727         return TDB_SUCCESS;
 728 }
 729
 730 /* Set up header for the recovery record. */
 731 static void set_recovery_header(struct tdb_recovery_record *rec,
 732                                 uint64_t magic,
 733                                 uint64_t datalen, uint64_t actuallen,
 734                                 uint64_t oldsize)
 735 {
 736         rec->magic = magic;
 737         rec->max_len = actuallen;
 738         rec->len = datalen;
 739         rec->eof = oldsize;
 740 }
 741
 742 /*
 743   setup the recovery data that will be used on a crash during commit
 744 */
 745 static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
 746                                                  tdb_off_t *magic_offset)
 747 {
 748         /* Initialized for GCC's 4.4.5 overzealous uninitialized warnings. */
 749         tdb_len_t recovery_size = 0;
 750         tdb_off_t recovery_offset = 0, recovery_max_size = 0;
 751         unsigned char *data, *p;
 752         const struct tdb_methods *methods = tdb->transaction->io_methods;
 753         struct tdb_recovery_record *rec;
 754         tdb_off_t old_map_size = tdb->transaction->old_map_size;
 755         uint64_t magic, tailer;
 756         int i;
 757         enum TDB_ERROR ecode;
 758
 759         /*
 760           check that the recovery area has enough space
 761         */
 762         ecode = tdb_recovery_allocate(tdb, &recovery_size,
 763                                       &recovery_offset, &recovery_max_size);
 764         if (ecode != TDB_SUCCESS) {
 765                 return ecode;
 766         }
 767
 768         data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
 769         if (data == NULL) {
 770                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 771                                   "transaction_setup_recovery:"
 772                                   " cannot allocate");
 773         }
 774
 775         rec = (struct tdb_recovery_record *)data;
 776         set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
 777                             recovery_size, recovery_max_size, old_map_size);
 778         tdb_convert(tdb, rec, sizeof(*rec));
 779
 780         /* build the recovery data into a single blob to allow us to do a single
 781            large write, which should be more efficient */
 782         p = data + sizeof(*rec);
 783         for (i=0;i<tdb->transaction->num_blocks;i++) {
 784                 tdb_off_t offset;
 785                 tdb_len_t length;
 786
 787                 if (tdb->transaction->blocks[i] == NULL) {
 788                         continue;
 789                 }
 790
 791                 offset = i * getpagesize();
 792                 length = getpagesize();
 793                 if (i == tdb->transaction->num_blocks-1) {
 794                         length = tdb->transaction->last_block_size;
 795                 }
 796
 797                 if (offset >= old_map_size) {
 798                         continue;
 799                 }
 800                 if (offset + length > tdb->file->map_size) {
 801                         free(data);
 802                         return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
 803                                           "tdb_transaction_setup_recovery:"
 804                                           " transaction data over new region"
 805                                           " boundary");
 806                 }
 807                 memcpy(p, &offset, sizeof(offset));
 808                 memcpy(p + sizeof(offset), &length, sizeof(length));
 809                 tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
 810
 811                 /* the recovery area contains the old data, not the
 812                    new data, so we have to call the original tdb_read
 813                    method to get it */
 814                 ecode = methods->tread(tdb, offset,
 815                                        p + sizeof(offset) + sizeof(length),
 816                                        length);
 817                 if (ecode != TDB_SUCCESS) {
 818                         free(data);
 819                         return ecode;
 820                 }
 821                 p += sizeof(offset) + sizeof(length) + length;
 822         }
 823
 824         /* and the tailer */
 825         tailer = sizeof(*rec) + recovery_max_size;
 826         memcpy(p, &tailer, sizeof(tailer));
 827         tdb_convert(tdb, p, sizeof(tailer));
 828
 829         /* write the recovery data to the recovery area */
 830         ecode = methods->twrite(tdb, recovery_offset, data,
 831                                 sizeof(*rec) + recovery_size);
 832         if (ecode != TDB_SUCCESS) {
 833                 free(data);
 834                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 835                                   "tdb_transaction_setup_recovery:"
 836                                   " failed to write recovery data");
 837         }
 838         transaction_write_existing(tdb, recovery_offset, data,
 839                                    sizeof(*rec) + recovery_size);
 840
 841         /* as we don't have ordered writes, we have to sync the recovery
 842            data before we update the magic to indicate that the recovery
 843            data is present */
 844         ecode = transaction_sync(tdb, recovery_offset,
 845                                  sizeof(*rec) + recovery_size);
 846         if (ecode != TDB_SUCCESS) {
 847                 free(data);
 848                 return ecode;
 849         }
 850
 851         free(data);
 852
 853         magic = TDB_RECOVERY_MAGIC;
 854         tdb_convert(tdb, &magic, sizeof(magic));
 855
 856         *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
 857                                                    magic);
 858
 859         ecode = methods->twrite(tdb, *magic_offset, &magic, sizeof(magic));
 860         if (ecode != TDB_SUCCESS) {
 861                 return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 862                                   "tdb_transaction_setup_recovery:"
 863                                   " failed to write recovery magic");
 864         }
 865         transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
 866
 867         /* ensure the recovery magic marker is on disk */
 868         return transaction_sync(tdb, *magic_offset, sizeof(magic));
 869 }
 870
 871 static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 872 {
 873         const struct tdb_methods *methods;
 874         enum TDB_ERROR ecode;
 875
 876         if (tdb->transaction == NULL) {
 877                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 878                                   "tdb_transaction_prepare_commit:"
 879                                   " no transaction");
 880         }
 881
 882         if (tdb->transaction->prepared) {
 883                 _tdb_transaction_cancel(tdb);
 884                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 885                                   "tdb_transaction_prepare_commit:"
 886                                   " transaction already prepared");
 887         }
 888
 889         if (tdb->transaction->transaction_error) {
 890                 _tdb_transaction_cancel(tdb);
 891                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_ERROR,
 892                                   "tdb_transaction_prepare_commit:"
 893                                   " transaction error pending");
 894         }
 895
 896
 897         if (tdb->transaction->nesting != 0) {
 898                 tdb->transaction->nesting--;
 899                 return TDB_SUCCESS;
 900         }
 901
 902         /* check for a null transaction */
 903         if (tdb->transaction->blocks == NULL) {
 904                 return TDB_SUCCESS;
 905         }
 906
 907         methods = tdb->transaction->io_methods;
 908
 909         /* upgrade the main transaction lock region to a write lock */
 910         ecode = tdb_allrecord_upgrade(tdb);
 911         if (ecode != TDB_SUCCESS) {
 912                 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 913                          "tdb_transaction_prepare_commit:"
 914                          " failed to upgrade hash locks");
 915                 _tdb_transaction_cancel(tdb);
 916                 return ecode;
 917         }
 918
 919         /* get the open lock - this prevents new users attaching to the database
 920            during the commit */
 921         ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 922         if (ecode != TDB_SUCCESS) {
 923                 tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 924                            "tdb_transaction_prepare_commit:"
 925                            " failed to get open lock");
 926                 _tdb_transaction_cancel(tdb);
 927                 return ecode;
 928         }
 929
 930         /* Since we have whole db locked, we don't need the expansion lock. */
 931         if (!(tdb->flags & TDB_NOSYNC)) {
 932                 /* write the recovery data to the end of the file */
 933                 ecode = transaction_setup_recovery(tdb,
 934                                                    &tdb->transaction
 935                                                    ->magic_offset);
 936                 if (ecode != TDB_SUCCESS) {
 937                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 938                                  "tdb_transaction_prepare_commit:"
 939                                  " failed to setup recovery data");
 940                         _tdb_transaction_cancel(tdb);
 941                         return ecode;
 942                 }
 943         }
 944
 945         tdb->transaction->prepared = true;
 946
 947         /* expand the file to the new size if needed */
 948         if (tdb->file->map_size != tdb->transaction->old_map_size) {
 949                 tdb_len_t add;
 950
 951                 add = tdb->file->map_size - tdb->transaction->old_map_size;
 952                 /* Restore original map size for tdb_expand_file */
 953                 tdb->file->map_size = tdb->transaction->old_map_size;
 954                 ecode = methods->expand_file(tdb, add);
 955                 if (ecode != TDB_SUCCESS) {
 956                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 957                                  "tdb_transaction_prepare_commit:"
 958                                  " expansion failed");
 959                         _tdb_transaction_cancel(tdb);
 960                         return ecode;
 961                 }
 962         }
 963
 964         /* Keep the open lock until the actual commit */
 965         return TDB_SUCCESS;
 966 }
 967
 968 /*
 969    prepare to commit the current transaction
 970 */
 971 enum TDB_ERROR tdb_transaction_prepare_commit(struct tdb_context *tdb)
 972 {
 973         return _tdb_transaction_prepare_commit(tdb);
 974 }
 975
 976 /*
 977   commit the current transaction
 978 */
 979 enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
 980 {
 981         const struct tdb_methods *methods;
 982         int i;
 983         enum TDB_ERROR ecode;
 984
 985         if (tdb->transaction == NULL) {
 986                 return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 987                                   "tdb_transaction_commit: no transaction");
 988         }
 989
 990         tdb_trace(tdb, "tdb_transaction_commit");
 991
 992         if (tdb->transaction->nesting != 0) {
 993                 tdb->transaction->nesting--;
 994                 return TDB_SUCCESS;
 995         }
 996
 997         /* check for a null transaction */
 998         if (tdb->transaction->blocks == NULL) {
 999                 _tdb_transaction_cancel(tdb);
1000                 return TDB_SUCCESS;
1001         }
1002
1003         if (!tdb->transaction->prepared) {
1004                 ecode = _tdb_transaction_prepare_commit(tdb);
1005                 if (ecode != TDB_SUCCESS)
1006                         return ecode;
1007         }
1008
1009         methods = tdb->transaction->io_methods;
1010
1011         /* perform all the writes */
1012         for (i=0;i<tdb->transaction->num_blocks;i++) {
1013                 tdb_off_t offset;
1014                 tdb_len_t length;
1015
1016                 if (tdb->transaction->blocks[i] == NULL) {
1017                         continue;
1018                 }
1019
1020                 offset = i * getpagesize();
1021                 length = getpagesize();
1022                 if (i == tdb->transaction->num_blocks-1) {
1023                         length = tdb->transaction->last_block_size;
1024                 }
1025
1026                 ecode = methods->twrite(tdb, offset,
1027                                         tdb->transaction->blocks[i], length);
1028                 if (ecode != TDB_SUCCESS) {
1029                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
1030                                    "tdb_transaction_commit:"
1031                                    " write failed during commit");
1032
1033                         /* we've overwritten part of the data and
1034                            possibly expanded the file, so we need to
1035                            run the crash recovery code */
1036                         tdb->methods = methods;
1037                         tdb_transaction_recover(tdb);
1038
1039                         _tdb_transaction_cancel(tdb);
1040
1041                         return ecode;
1042                 }
1043                 SAFE_FREE(tdb->transaction->blocks[i]);
1044         }
1045
1046         SAFE_FREE(tdb->transaction->blocks);
1047         tdb->transaction->num_blocks = 0;
1048
1049         /* ensure the new data is on disk */
1050         ecode = transaction_sync(tdb, 0, tdb->file->map_size);
1051         if (ecode != TDB_SUCCESS) {
1052                 return ecode;
1053         }
1054
1055         /*
1056           TODO: maybe write to some dummy hdr field, or write to magic
1057           offset without mmap, before the last sync, instead of the
1058           utime() call
1059         */
1060
1061         /* on some systems (like Linux 2.6.x) changes via mmap/msync
1062            don't change the mtime of the file, this means the file may
1063            not be backed up (as tdb rounding to block sizes means that
1064            file size changes are quite rare too). The following forces
1065            mtime changes when a transaction completes */
1066 #if HAVE_UTIME
1067         utime(tdb->name, NULL);
1068 #endif
1069
1070         /* use a transaction cancel to free memory and remove the
1071            transaction locks */
1072         _tdb_transaction_cancel(tdb);
1073
1074         return 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 }