git.ozlabs.org Git - ccan/blob - junkcode/rusty@rustcorp.com.au-ntdb/check.c

   1  /*
   2    Trivial Database 2: free list/block handling
   3    Copyright (C) Rusty Russell 2010
   4
   5    This library is free software; you can redistribute it and/or
   6    modify it under the terms of the GNU Lesser General Public
   7    License as published by the Free Software Foundation; either
   8    version 3 of the License, or (at your option) any later version.
   9
  10    This library is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13    Lesser General Public License for more details.
  14
  15    You should have received a copy of the GNU Lesser General Public
  16    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  17 */
  18 #include "private.h"
  19 #include <ccan/likely/likely.h>
  20 #include <ccan/asearch/asearch.h>
  21
  22 /* We keep an ordered array of offsets. */
  23 static bool append(struct ntdb_context *ntdb,
  24                    ntdb_off_t **arr, size_t *num, ntdb_off_t off)
  25 {
  26         ntdb_off_t *new;
  27
  28         if (*num == 0) {
  29                 new = ntdb->alloc_fn(ntdb, sizeof(ntdb_off_t), ntdb->alloc_data);
  30         } else {
  31                 new = ntdb->expand_fn(*arr, (*num + 1) * sizeof(ntdb_off_t),
  32                                   ntdb->alloc_data);
  33         }
  34         if (!new)
  35                 return false;
  36         new[(*num)++] = off;
  37         *arr = new;
  38         return true;
  39 }
  40
  41 static enum NTDB_ERROR check_header(struct ntdb_context *ntdb,
  42                                     ntdb_off_t *recovery,
  43                                     uint64_t *features,
  44                                     size_t *num_capabilities)
  45 {
  46         uint64_t hash_test;
  47         struct ntdb_header hdr;
  48         enum NTDB_ERROR ecode;
  49         ntdb_off_t off, next;
  50
  51         ecode = ntdb_read_convert(ntdb, 0, &hdr, sizeof(hdr));
  52         if (ecode != NTDB_SUCCESS) {
  53                 return ecode;
  54         }
  55         /* magic food should not be converted, so convert back. */
  56         ntdb_convert(ntdb, hdr.magic_food, sizeof(hdr.magic_food));
  57
  58         hash_test = NTDB_HASH_MAGIC;
  59         hash_test = ntdb_hash(ntdb, &hash_test, sizeof(hash_test));
  60         if (hdr.hash_test != hash_test) {
  61                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  62                                   "check: hash test %llu should be %llu",
  63                                   (long long)hdr.hash_test,
  64                                   (long long)hash_test);
  65         }
  66
  67         if (strcmp(hdr.magic_food, NTDB_MAGIC_FOOD) != 0) {
  68                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  69                                   "check: bad magic '%.*s'",
  70                                   (unsigned)sizeof(hdr.magic_food),
  71                                   hdr.magic_food);
  72         }
  73
  74         /* Features which are used must be a subset of features offered. */
  75         if (hdr.features_used & ~hdr.features_offered) {
  76                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  77                                   "check: features used (0x%llx) which"
  78                                   " are not offered (0x%llx)",
  79                                   (long long)hdr.features_used,
  80                                   (long long)hdr.features_offered);
  81         }
  82
  83         *features = hdr.features_offered;
  84         *recovery = hdr.recovery;
  85         if (*recovery) {
  86                 if (*recovery < sizeof(hdr)
  87                     || *recovery > ntdb->file->map_size) {
  88                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  89                                           "ntdb_check:"
  90                                           " invalid recovery offset %zu",
  91                                           (size_t)*recovery);
  92                 }
  93         }
  94
  95         for (off = hdr.capabilities; off && ecode == NTDB_SUCCESS; off = next) {
  96                 const struct ntdb_capability *cap;
  97                 enum NTDB_ERROR e;
  98
  99                 cap = ntdb_access_read(ntdb, off, sizeof(*cap), true);
 100                 if (NTDB_PTR_IS_ERR(cap)) {
 101                         return NTDB_PTR_ERR(cap);
 102                 }
 103
 104                 /* All capabilities are unknown. */
 105                 e = unknown_capability(ntdb, "ntdb_check", cap->type);
 106                 next = cap->next;
 107                 ntdb_access_release(ntdb, cap);
 108                 if (e)
 109                         return e;
 110                 (*num_capabilities)++;
 111         }
 112
 113         /* Don't check reserved: they *can* be used later. */
 114         return NTDB_SUCCESS;
 115 }
 116
 117 static int off_cmp(const ntdb_off_t *a, const ntdb_off_t *b, void *ctx)
 118 {
 119         /* Can overflow an int. */
 120         return *a > *b ? 1
 121                 : *a < *b ? -1
 122                 : 0;
 123 }
 124
 125 static enum NTDB_ERROR check_entry(struct ntdb_context *ntdb,
 126                                    ntdb_off_t off_and_hash,
 127                                    ntdb_len_t bucket,
 128                                    ntdb_off_t used[],
 129                                    size_t num_used,
 130                                    size_t *num_found,
 131                                    enum NTDB_ERROR (*check)(NTDB_DATA,
 132                                                             NTDB_DATA,
 133                                                             void *),
 134                                    void *data)
 135 {
 136         enum NTDB_ERROR ecode;
 137         const struct ntdb_used_record *r;
 138         const unsigned char *kptr;
 139         ntdb_len_t klen, dlen;
 140         uint32_t hash;
 141         ntdb_off_t off = off_and_hash & NTDB_OFF_MASK;
 142         ntdb_off_t *p;
 143
 144         /* Empty bucket is fine. */
 145         if (!off_and_hash) {
 146                 return NTDB_SUCCESS;
 147         }
 148
 149         /* This can't point to a chain, we handled those at toplevel. */
 150         if (off_and_hash & (1ULL << NTDB_OFF_CHAIN_BIT)) {
 151                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 152                                    "ntdb_check: Invalid chain bit in offset "
 153                                    " %llu", (long long)off_and_hash);
 154         }
 155
 156         p = asearch(&off, used, num_used, off_cmp, NULL);
 157         if (!p) {
 158                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 159                                    "ntdb_check: Invalid offset"
 160                                    " %llu in hash", (long long)off);
 161         }
 162         /* Mark it invalid. */
 163         *p ^= 1;
 164         (*num_found)++;
 165
 166         r = ntdb_access_read(ntdb, off, sizeof(*r), true);
 167         if (NTDB_PTR_IS_ERR(r)) {
 168                 return NTDB_PTR_ERR(r);
 169         }
 170         klen = rec_key_length(r);
 171         dlen = rec_data_length(r);
 172         ntdb_access_release(ntdb, r);
 173
 174         kptr = ntdb_access_read(ntdb, off + sizeof(*r), klen + dlen, false);
 175         if (NTDB_PTR_IS_ERR(kptr)) {
 176                 return NTDB_PTR_ERR(kptr);
 177         }
 178
 179         hash = ntdb_hash(ntdb, kptr, klen);
 180
 181         /* Are we in the right chain? */
 182         if (bits_from(hash, 0, ntdb->hash_bits) != bucket) {
 183                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 184                                     NTDB_LOG_ERROR,
 185                                     "ntdb_check: Bad bucket %u vs %llu",
 186                                     bits_from(hash, 0, ntdb->hash_bits),
 187                                     (long long)bucket);
 188         /* Next 8 bits should be the same as top bits of bucket. */
 189         } else if (bits_from(hash, ntdb->hash_bits, NTDB_OFF_UPPER_STEAL)
 190                    != bits_from(off_and_hash, 64-NTDB_OFF_UPPER_STEAL,
 191                                 NTDB_OFF_UPPER_STEAL)) {
 192                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 193                                     NTDB_LOG_ERROR,
 194                                     "ntdb_check: Bad hash bits %llu vs %llu",
 195                                     (long long)off_and_hash,
 196                                     (long long)hash);
 197         } else if (check) {
 198                 NTDB_DATA k, d;
 199
 200                 k = ntdb_mkdata(kptr, klen);
 201                 d = ntdb_mkdata(kptr + klen, dlen);
 202                 ecode = check(k, d, data);
 203         } else {
 204                 ecode = NTDB_SUCCESS;
 205         }
 206         ntdb_access_release(ntdb, kptr);
 207
 208         return ecode;
 209 }
 210
 211 static enum NTDB_ERROR check_hash_chain(struct ntdb_context *ntdb,
 212                                         ntdb_off_t off,
 213                                         ntdb_len_t bucket,
 214                                         ntdb_off_t used[],
 215                                         size_t num_used,
 216                                         size_t *num_found,
 217                                         enum NTDB_ERROR (*check)(NTDB_DATA,
 218                                                                  NTDB_DATA,
 219                                                                  void *),
 220                                         void *data)
 221 {
 222         struct ntdb_used_record rec;
 223         enum NTDB_ERROR ecode;
 224         const ntdb_off_t *entries;
 225         ntdb_len_t i, num;
 226
 227         /* This is a used entry. */
 228         (*num_found)++;
 229
 230         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
 231         if (ecode != NTDB_SUCCESS) {
 232                 return ecode;
 233         }
 234
 235         if (rec_magic(&rec) != NTDB_CHAIN_MAGIC) {
 236                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 237                                   "ntdb_check: Bad hash chain magic %llu",
 238                                   (long long)rec_magic(&rec));
 239         }
 240
 241         if (rec_data_length(&rec) % sizeof(ntdb_off_t)) {
 242                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 243                                   "ntdb_check: Bad hash chain data length %llu",
 244                                   (long long)rec_data_length(&rec));
 245         }
 246
 247         if (rec_key_length(&rec) != 0) {
 248                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 249                                   "ntdb_check: Bad hash chain key length %llu",
 250                                   (long long)rec_key_length(&rec));
 251         }
 252
 253         off += sizeof(rec);
 254         num = rec_data_length(&rec) / sizeof(ntdb_off_t);
 255         entries = ntdb_access_read(ntdb, off, rec_data_length(&rec), true);
 256         if (NTDB_PTR_IS_ERR(entries)) {
 257                 return NTDB_PTR_ERR(entries);
 258         }
 259
 260         /* Check each non-deleted entry in chain. */
 261         for (i = 0; i < num; i++) {
 262                 ecode = check_entry(ntdb, entries[i], bucket,
 263                                     used, num_used, num_found, check, data);
 264                 if (ecode) {
 265                         break;
 266                 }
 267         }
 268
 269         ntdb_access_release(ntdb, entries);
 270         return ecode;
 271 }
 272
 273 static enum NTDB_ERROR check_hash(struct ntdb_context *ntdb,
 274                                   ntdb_off_t used[],
 275                                   size_t num_used,
 276                                   size_t num_other_used,
 277                                   enum NTDB_ERROR (*check)(NTDB_DATA,
 278                                                            NTDB_DATA,
 279                                                            void *),
 280                                   void *data)
 281 {
 282         enum NTDB_ERROR ecode;
 283         struct ntdb_used_record rec;
 284         const ntdb_off_t *entries;
 285         ntdb_len_t i;
 286         /* Free tables and capabilities also show up as used, as do we. */
 287         size_t num_found = num_other_used + 1;
 288
 289         ecode = ntdb_read_convert(ntdb, NTDB_HASH_OFFSET, &rec, sizeof(rec));
 290         if (ecode != NTDB_SUCCESS) {
 291                 return ecode;
 292         }
 293
 294         if (rec_magic(&rec) != NTDB_HTABLE_MAGIC) {
 295                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 296                                   "ntdb_check: Bad hash table magic %llu",
 297                                   (long long)rec_magic(&rec));
 298         }
 299
 300         if (rec_data_length(&rec) != (sizeof(ntdb_off_t) << ntdb->hash_bits)) {
 301                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 302                                   "ntdb_check: Bad hash table data length %llu",
 303                                   (long long)rec_data_length(&rec));
 304         }
 305
 306         if (rec_key_length(&rec) != 0) {
 307                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 308                                   "ntdb_check: Bad hash table key length %llu",
 309                                   (long long)rec_key_length(&rec));
 310         }
 311
 312         entries = ntdb_access_read(ntdb, NTDB_HASH_OFFSET + sizeof(rec),
 313                                    rec_data_length(&rec), true);
 314         if (NTDB_PTR_IS_ERR(entries)) {
 315                 return NTDB_PTR_ERR(entries);
 316         }
 317
 318         for (i = 0; i < (1 << ntdb->hash_bits); i++) {
 319                 ntdb_off_t off = entries[i] & NTDB_OFF_MASK;
 320                 if (entries[i] & (1ULL << NTDB_OFF_CHAIN_BIT)) {
 321                         ecode = check_hash_chain(ntdb, off, i,
 322                                                  used, num_used, &num_found,
 323                                                  check, data);
 324                 } else {
 325                         ecode = check_entry(ntdb, entries[i], i,
 326                                             used, num_used, &num_found,
 327                                             check, data);
 328                 }
 329                 if (ecode) {
 330                         break;
 331                 }
 332         }
 333         ntdb_access_release(ntdb, entries);
 334
 335         if (ecode == NTDB_SUCCESS && num_found != num_used) {
 336                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 337                                     "ntdb_check: Not all entries are in hash");
 338         }
 339         return ecode;
 340 }
 341
 342 static enum NTDB_ERROR check_free(struct ntdb_context *ntdb,
 343                                  ntdb_off_t off,
 344                                  const struct ntdb_free_record *frec,
 345                                  ntdb_off_t prev, unsigned int ftable,
 346                                  unsigned int bucket)
 347 {
 348         enum NTDB_ERROR ecode;
 349
 350         if (frec_magic(frec) != NTDB_FREE_MAGIC) {
 351                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 352                                   "ntdb_check: offset %llu bad magic 0x%llx",
 353                                   (long long)off,
 354                                   (long long)frec->magic_and_prev);
 355         }
 356         if (frec_ftable(frec) != ftable) {
 357                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 358                                   "ntdb_check: offset %llu bad freetable %u",
 359                                   (long long)off, frec_ftable(frec));
 360
 361         }
 362
 363         ecode = ntdb_oob(ntdb, off,
 364                          frec_len(frec) + sizeof(struct ntdb_used_record),
 365                          false);
 366         if (ecode != NTDB_SUCCESS) {
 367                 return ecode;
 368         }
 369         if (size_to_bucket(frec_len(frec)) != bucket) {
 370                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 371                                   "ntdb_check: offset %llu in wrong bucket"
 372                                   " (%u vs %u)",
 373                                   (long long)off,
 374                                   bucket, size_to_bucket(frec_len(frec)));
 375         }
 376         if (prev && prev != frec_prev(frec)) {
 377                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 378                                   "ntdb_check: offset %llu bad prev"
 379                                   " (%llu vs %llu)",
 380                                   (long long)off,
 381                                   (long long)prev, (long long)frec_len(frec));
 382         }
 383         return NTDB_SUCCESS;
 384 }
 385
 386 static enum NTDB_ERROR check_free_table(struct ntdb_context *ntdb,
 387                                        ntdb_off_t ftable_off,
 388                                        unsigned ftable_num,
 389                                        ntdb_off_t fr[],
 390                                        size_t num_free,
 391                                        size_t *num_found)
 392 {
 393         struct ntdb_freetable ft;
 394         ntdb_off_t h;
 395         unsigned int i;
 396         enum NTDB_ERROR ecode;
 397
 398         ecode = ntdb_read_convert(ntdb, ftable_off, &ft, sizeof(ft));
 399         if (ecode != NTDB_SUCCESS) {
 400                 return ecode;
 401         }
 402
 403         if (rec_magic(&ft.hdr) != NTDB_FTABLE_MAGIC
 404             || rec_key_length(&ft.hdr) != 0
 405             || rec_data_length(&ft.hdr) != sizeof(ft) - sizeof(ft.hdr)) {
 406                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 407                                   "ntdb_check: Invalid header on free table");
 408         }
 409
 410         for (i = 0; i < NTDB_FREE_BUCKETS; i++) {
 411                 ntdb_off_t off, prev = 0, *p, first = 0;
 412                 struct ntdb_free_record f;
 413
 414                 h = bucket_off(ftable_off, i);
 415                 for (off = ntdb_read_off(ntdb, h); off; off = f.next) {
 416                         if (NTDB_OFF_IS_ERR(off)) {
 417                                 return NTDB_OFF_TO_ERR(off);
 418                         }
 419                         if (!first) {
 420                                 off &= NTDB_OFF_MASK;
 421                                 first = off;
 422                         }
 423                         ecode = ntdb_read_convert(ntdb, off, &f, sizeof(f));
 424                         if (ecode != NTDB_SUCCESS) {
 425                                 return ecode;
 426                         }
 427                         ecode = check_free(ntdb, off, &f, prev, ftable_num, i);
 428                         if (ecode != NTDB_SUCCESS) {
 429                                 return ecode;
 430                         }
 431
 432                         /* FIXME: Check hash bits */
 433                         p = asearch(&off, fr, num_free, off_cmp, NULL);
 434                         if (!p) {
 435                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 436                                                   NTDB_LOG_ERROR,
 437                                                   "ntdb_check: Invalid offset"
 438                                                   " %llu in free table",
 439                                                   (long long)off);
 440                         }
 441                         /* Mark it invalid. */
 442                         *p ^= 1;
 443                         (*num_found)++;
 444                         prev = off;
 445                 }
 446
 447                 if (first) {
 448                         /* Now we can check first back pointer. */
 449                         ecode = ntdb_read_convert(ntdb, first, &f, sizeof(f));
 450                         if (ecode != NTDB_SUCCESS) {
 451                                 return ecode;
 452                         }
 453                         ecode = check_free(ntdb, first, &f, prev, ftable_num, i);
 454                         if (ecode != NTDB_SUCCESS) {
 455                                 return ecode;
 456                         }
 457                 }
 458         }
 459         return NTDB_SUCCESS;
 460 }
 461
 462 /* Slow, but should be very rare. */
 463 ntdb_off_t dead_space(struct ntdb_context *ntdb, ntdb_off_t off)
 464 {
 465         size_t len;
 466         enum NTDB_ERROR ecode;
 467
 468         for (len = 0; off + len < ntdb->file->map_size; len++) {
 469                 char c;
 470                 ecode = ntdb->io->tread(ntdb, off, &c, 1);
 471                 if (ecode != NTDB_SUCCESS) {
 472                         return NTDB_ERR_TO_OFF(ecode);
 473                 }
 474                 if (c != 0 && c != 0x43)
 475                         break;
 476         }
 477         return len;
 478 }
 479
 480 static enum NTDB_ERROR check_linear(struct ntdb_context *ntdb,
 481                                    ntdb_off_t **used, size_t *num_used,
 482                                    ntdb_off_t **fr, size_t *num_free,
 483                                    uint64_t features, ntdb_off_t recovery)
 484 {
 485         ntdb_off_t off;
 486         ntdb_len_t len;
 487         enum NTDB_ERROR ecode;
 488         bool found_recovery = false;
 489
 490         for (off = sizeof(struct ntdb_header);
 491              off < ntdb->file->map_size;
 492              off += len) {
 493                 union {
 494                         struct ntdb_used_record u;
 495                         struct ntdb_free_record f;
 496                         struct ntdb_recovery_record r;
 497                 } rec;
 498                 /* r is larger: only get that if we need to. */
 499                 ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec.f));
 500                 if (ecode != NTDB_SUCCESS) {
 501                         return ecode;
 502                 }
 503
 504                 /* If we crash after ftruncate, we can get zeroes or fill. */
 505                 if (rec.r.magic == NTDB_RECOVERY_INVALID_MAGIC
 506                     || rec.r.magic ==  0x4343434343434343ULL) {
 507                         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec.r));
 508                         if (ecode != NTDB_SUCCESS) {
 509                                 return ecode;
 510                         }
 511                         if (recovery == off) {
 512                                 found_recovery = true;
 513                                 len = sizeof(rec.r) + rec.r.max_len;
 514                         } else {
 515                                 len = dead_space(ntdb, off);
 516                                 if (NTDB_OFF_IS_ERR(len)) {
 517                                         return NTDB_OFF_TO_ERR(len);
 518                                 }
 519                                 if (len < sizeof(rec.r)) {
 520                                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 521                                                           NTDB_LOG_ERROR,
 522                                                           "ntdb_check: invalid"
 523                                                           " dead space at %zu",
 524                                                           (size_t)off);
 525                                 }
 526
 527                                 ntdb_logerr(ntdb, NTDB_SUCCESS, NTDB_LOG_WARNING,
 528                                            "Dead space at %zu-%zu (of %zu)",
 529                                            (size_t)off, (size_t)(off + len),
 530                                            (size_t)ntdb->file->map_size);
 531                         }
 532                 } else if (rec.r.magic == NTDB_RECOVERY_MAGIC) {
 533                         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec.r));
 534                         if (ecode != NTDB_SUCCESS) {
 535                                 return ecode;
 536                         }
 537                         if (recovery != off) {
 538                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 539                                                   NTDB_LOG_ERROR,
 540                                                   "ntdb_check: unexpected"
 541                                                   " recovery record at offset"
 542                                                   " %zu",
 543                                                   (size_t)off);
 544                         }
 545                         if (rec.r.len > rec.r.max_len) {
 546                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 547                                                   NTDB_LOG_ERROR,
 548                                                   "ntdb_check: invalid recovery"
 549                                                   " length %zu",
 550                                                   (size_t)rec.r.len);
 551                         }
 552                         if (rec.r.eof > ntdb->file->map_size) {
 553                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 554                                                   NTDB_LOG_ERROR,
 555                                                   "ntdb_check: invalid old EOF"
 556                                                   " %zu", (size_t)rec.r.eof);
 557                         }
 558                         found_recovery = true;
 559                         len = sizeof(rec.r) + rec.r.max_len;
 560                 } else if (frec_magic(&rec.f) == NTDB_FREE_MAGIC) {
 561                         len = sizeof(rec.u) + frec_len(&rec.f);
 562                         if (off + len > ntdb->file->map_size) {
 563                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 564                                                   NTDB_LOG_ERROR,
 565                                                   "ntdb_check: free overlength"
 566                                                   " %llu at offset %llu",
 567                                                   (long long)len,
 568                                                   (long long)off);
 569                         }
 570                         /* This record should be in free lists. */
 571                         if (frec_ftable(&rec.f) != NTDB_FTABLE_NONE
 572                             && !append(ntdb, fr, num_free, off)) {
 573                                 return ntdb_logerr(ntdb, NTDB_ERR_OOM,
 574                                                   NTDB_LOG_ERROR,
 575                                                   "ntdb_check: tracking %zu'th"
 576                                                   " free record.", *num_free);
 577                         }
 578                 } else if (rec_magic(&rec.u) == NTDB_USED_MAGIC
 579                            || rec_magic(&rec.u) == NTDB_CHAIN_MAGIC
 580                            || rec_magic(&rec.u) == NTDB_HTABLE_MAGIC
 581                            || rec_magic(&rec.u) == NTDB_FTABLE_MAGIC
 582                            || rec_magic(&rec.u) == NTDB_CAP_MAGIC) {
 583                         uint64_t klen, dlen, extra;
 584
 585                         /* This record is used! */
 586                         if (!append(ntdb, used, num_used, off)) {
 587                                 return ntdb_logerr(ntdb, NTDB_ERR_OOM,
 588                                                   NTDB_LOG_ERROR,
 589                                                   "ntdb_check: tracking %zu'th"
 590                                                   " used record.", *num_used);
 591                         }
 592
 593                         klen = rec_key_length(&rec.u);
 594                         dlen = rec_data_length(&rec.u);
 595                         extra = rec_extra_padding(&rec.u);
 596
 597                         len = sizeof(rec.u) + klen + dlen + extra;
 598                         if (off + len > ntdb->file->map_size) {
 599                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 600                                                   NTDB_LOG_ERROR,
 601                                                   "ntdb_check: used overlength"
 602                                                   " %llu at offset %llu",
 603                                                   (long long)len,
 604                                                   (long long)off);
 605                         }
 606
 607                         if (len < sizeof(rec.f)) {
 608                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 609                                                   NTDB_LOG_ERROR,
 610                                                   "ntdb_check: too short record"
 611                                                   " %llu at %llu",
 612                                                   (long long)len,
 613                                                   (long long)off);
 614                         }
 615
 616                         /* Check that records have correct 0 at end (but may
 617                          * not in future). */
 618                         if (extra && !features
 619                             && rec_magic(&rec.u) != NTDB_CAP_MAGIC) {
 620                                 const char *p;
 621                                 char c;
 622                                 p = ntdb_access_read(ntdb, off + sizeof(rec.u)
 623                                                     + klen + dlen, 1, false);
 624                                 if (NTDB_PTR_IS_ERR(p))
 625                                         return NTDB_PTR_ERR(p);
 626                                 c = *p;
 627                                 ntdb_access_release(ntdb, p);
 628
 629                                 if (c != '\0') {
 630                                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 631                                                           NTDB_LOG_ERROR,
 632                                                           "ntdb_check:"
 633                                                           " non-zero extra"
 634                                                           " at %llu",
 635                                                           (long long)off);
 636                                 }
 637                         }
 638                 } else {
 639                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 640                                           NTDB_LOG_ERROR,
 641                                           "ntdb_check: Bad magic 0x%llx"
 642                                           " at offset %zu",
 643                                           (long long)rec_magic(&rec.u),
 644                                           (size_t)off);
 645                 }
 646         }
 647
 648         /* We must have found recovery area if there was one. */
 649         if (recovery != 0 && !found_recovery) {
 650                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 651                                   "ntdb_check: expected a recovery area at %zu",
 652                                   (size_t)recovery);
 653         }
 654
 655         return NTDB_SUCCESS;
 656 }
 657
 658 _PUBLIC_ enum NTDB_ERROR ntdb_check_(struct ntdb_context *ntdb,
 659                           enum NTDB_ERROR (*check)(NTDB_DATA, NTDB_DATA, void *),
 660                           void *data)
 661 {
 662         ntdb_off_t *fr = NULL, *used = NULL;
 663         ntdb_off_t ft = 0, recovery = 0;
 664         size_t num_free = 0, num_used = 0, num_found = 0, num_ftables = 0,
 665                 num_capabilities = 0;
 666         uint64_t features = 0;
 667         enum NTDB_ERROR ecode;
 668
 669         if (ntdb->flags & NTDB_CANT_CHECK) {
 670                 return ntdb_logerr(ntdb, NTDB_SUCCESS, NTDB_LOG_WARNING,
 671                                   "ntdb_check: database has unknown capability,"
 672                                   " cannot check.");
 673         }
 674
 675         ecode = ntdb_allrecord_lock(ntdb, F_RDLCK, NTDB_LOCK_WAIT, false);
 676         if (ecode != NTDB_SUCCESS) {
 677                 return ecode;
 678         }
 679
 680         ecode = ntdb_lock_expand(ntdb, F_RDLCK);
 681         if (ecode != NTDB_SUCCESS) {
 682                 ntdb_allrecord_unlock(ntdb, F_RDLCK);
 683                 return ecode;
 684         }
 685
 686         ecode = check_header(ntdb, &recovery, &features, &num_capabilities);
 687         if (ecode != NTDB_SUCCESS)
 688                 goto out;
 689
 690         /* First we do a linear scan, checking all records. */
 691         ecode = check_linear(ntdb, &used, &num_used, &fr, &num_free, features,
 692                              recovery);
 693         if (ecode != NTDB_SUCCESS)
 694                 goto out;
 695
 696         for (ft = first_ftable(ntdb); ft; ft = next_ftable(ntdb, ft)) {
 697                 if (NTDB_OFF_IS_ERR(ft)) {
 698                         ecode = NTDB_OFF_TO_ERR(ft);
 699                         goto out;
 700                 }
 701                 ecode = check_free_table(ntdb, ft, num_ftables, fr, num_free,
 702                                          &num_found);
 703                 if (ecode != NTDB_SUCCESS)
 704                         goto out;
 705                 num_ftables++;
 706         }
 707
 708         /* FIXME: Check key uniqueness? */
 709         ecode = check_hash(ntdb, used, num_used, num_ftables + num_capabilities,
 710                            check, data);
 711         if (ecode != NTDB_SUCCESS)
 712                 goto out;
 713
 714         if (num_found != num_free) {
 715                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 716                                    "ntdb_check: Not all entries are in"
 717                                    " free table");
 718         }
 719
 720 out:
 721         ntdb_allrecord_unlock(ntdb, F_RDLCK);
 722         ntdb_unlock_expand(ntdb, F_RDLCK);
 723         ntdb->free_fn(fr, ntdb->alloc_data);
 724         ntdb->free_fn(used, ntdb->alloc_data);
 725         return ecode;
 726 }