lib/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, ntdb_off_t *recovery,
  42                                    uint64_t *features, size_t *num_capabilities)
  43 {
  44         uint64_t hash_test;
  45         struct ntdb_header hdr;
  46         enum NTDB_ERROR ecode;
  47         ntdb_off_t off, next;
  48
  49         ecode = ntdb_read_convert(ntdb, 0, &hdr, sizeof(hdr));
  50         if (ecode != NTDB_SUCCESS) {
  51                 return ecode;
  52         }
  53         /* magic food should not be converted, so convert back. */
  54         ntdb_convert(ntdb, hdr.magic_food, sizeof(hdr.magic_food));
  55
  56         hash_test = NTDB_HASH_MAGIC;
  57         hash_test = ntdb_hash(ntdb, &hash_test, sizeof(hash_test));
  58         if (hdr.hash_test != hash_test) {
  59                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  60                                   "check: hash test %llu should be %llu",
  61                                   (long long)hdr.hash_test,
  62                                   (long long)hash_test);
  63         }
  64
  65         if (strcmp(hdr.magic_food, NTDB_MAGIC_FOOD) != 0) {
  66                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  67                                   "check: bad magic '%.*s'",
  68                                   (unsigned)sizeof(hdr.magic_food),
  69                                   hdr.magic_food);
  70         }
  71
  72         /* Features which are used must be a subset of features offered. */
  73         if (hdr.features_used & ~hdr.features_offered) {
  74                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  75                                   "check: features used (0x%llx) which"
  76                                   " are not offered (0x%llx)",
  77                                   (long long)hdr.features_used,
  78                                   (long long)hdr.features_offered);
  79         }
  80
  81         *features = hdr.features_offered;
  82         *recovery = hdr.recovery;
  83         if (*recovery) {
  84                 if (*recovery < sizeof(hdr)
  85                     || *recovery > ntdb->file->map_size) {
  86                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
  87                                           "ntdb_check:"
  88                                           " invalid recovery offset %zu",
  89                                           (size_t)*recovery);
  90                 }
  91         }
  92
  93         for (off = hdr.capabilities; off && ecode == NTDB_SUCCESS; off = next) {
  94                 const struct ntdb_capability *cap;
  95                 enum NTDB_ERROR e;
  96
  97                 cap = ntdb_access_read(ntdb, off, sizeof(*cap), true);
  98                 if (NTDB_PTR_IS_ERR(cap)) {
  99                         return NTDB_PTR_ERR(cap);
 100                 }
 101
 102                 /* All capabilities are unknown. */
 103                 e = unknown_capability(ntdb, "ntdb_check", cap->type);
 104                 next = cap->next;
 105                 ntdb_access_release(ntdb, cap);
 106                 if (e)
 107                         return e;
 108                 (*num_capabilities)++;
 109         }
 110
 111         /* Don't check reserved: they *can* be used later. */
 112         return NTDB_SUCCESS;
 113 }
 114
 115 static enum NTDB_ERROR check_hash_tree(struct ntdb_context *ntdb,
 116                                       ntdb_off_t off, unsigned int group_bits,
 117                                       uint64_t hprefix,
 118                                       unsigned hprefix_bits,
 119                                       ntdb_off_t used[],
 120                                       size_t num_used,
 121                                       size_t *num_found,
 122                                       enum NTDB_ERROR (*check)(NTDB_DATA,
 123                                                               NTDB_DATA, void *),
 124                                       void *data);
 125
 126 static enum NTDB_ERROR check_hash_chain(struct ntdb_context *ntdb,
 127                                        ntdb_off_t off,
 128                                        uint64_t hash,
 129                                        ntdb_off_t used[],
 130                                        size_t num_used,
 131                                        size_t *num_found,
 132                                        enum NTDB_ERROR (*check)(NTDB_DATA,
 133                                                                NTDB_DATA,
 134                                                                void *),
 135                                        void *data)
 136 {
 137         struct ntdb_used_record rec;
 138         enum NTDB_ERROR ecode;
 139
 140         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
 141         if (ecode != NTDB_SUCCESS) {
 142                 return ecode;
 143         }
 144
 145         if (rec_magic(&rec) != NTDB_CHAIN_MAGIC) {
 146                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 147                                   "ntdb_check: Bad hash chain magic %llu",
 148                                   (long long)rec_magic(&rec));
 149         }
 150
 151         if (rec_data_length(&rec) != sizeof(struct ntdb_chain)) {
 152                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 153                                   "ntdb_check:"
 154                                   " Bad hash chain length %llu vs %zu",
 155                                   (long long)rec_data_length(&rec),
 156                                   sizeof(struct ntdb_chain));
 157         }
 158         if (rec_key_length(&rec) != 0) {
 159                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 160                                   "ntdb_check: Bad hash chain key length %llu",
 161                                   (long long)rec_key_length(&rec));
 162         }
 163         if (rec_hash(&rec) != 0) {
 164                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 165                                   "ntdb_check: Bad hash chain hash value %llu",
 166                                   (long long)rec_hash(&rec));
 167         }
 168
 169         off += sizeof(rec);
 170         ecode = check_hash_tree(ntdb, off, 0, hash, 64,
 171                                 used, num_used, num_found, check, data);
 172         if (ecode != NTDB_SUCCESS) {
 173                 return ecode;
 174         }
 175
 176         off = ntdb_read_off(ntdb, off + offsetof(struct ntdb_chain, next));
 177         if (NTDB_OFF_IS_ERR(off)) {
 178                 return NTDB_OFF_TO_ERR(off);
 179         }
 180         if (off == 0)
 181                 return NTDB_SUCCESS;
 182         (*num_found)++;
 183         return check_hash_chain(ntdb, off, hash, used, num_used, num_found,
 184                                 check, data);
 185 }
 186
 187 static enum NTDB_ERROR check_hash_record(struct ntdb_context *ntdb,
 188                                         ntdb_off_t off,
 189                                         uint64_t hprefix,
 190                                         unsigned hprefix_bits,
 191                                         ntdb_off_t used[],
 192                                         size_t num_used,
 193                                         size_t *num_found,
 194                                         enum NTDB_ERROR (*check)(NTDB_DATA,
 195                                                                 NTDB_DATA,
 196                                                                 void *),
 197                                         void *data)
 198 {
 199         struct ntdb_used_record rec;
 200         enum NTDB_ERROR ecode;
 201
 202         if (hprefix_bits >= 64)
 203                 return check_hash_chain(ntdb, off, hprefix, used, num_used,
 204                                         num_found, check, data);
 205
 206         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
 207         if (ecode != NTDB_SUCCESS) {
 208                 return ecode;
 209         }
 210
 211         if (rec_magic(&rec) != NTDB_HTABLE_MAGIC) {
 212                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 213                                   "ntdb_check: Bad hash table magic %llu",
 214                                   (long long)rec_magic(&rec));
 215         }
 216         if (rec_data_length(&rec)
 217             != sizeof(ntdb_off_t) << NTDB_SUBLEVEL_HASH_BITS) {
 218                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 219                                   "ntdb_check:"
 220                                   " Bad hash table length %llu vs %llu",
 221                                   (long long)rec_data_length(&rec),
 222                                   (long long)sizeof(ntdb_off_t)
 223                                   << NTDB_SUBLEVEL_HASH_BITS);
 224         }
 225         if (rec_key_length(&rec) != 0) {
 226                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 227                                   "ntdb_check: Bad hash table key length %llu",
 228                                   (long long)rec_key_length(&rec));
 229         }
 230         if (rec_hash(&rec) != 0) {
 231                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 232                                   "ntdb_check: Bad hash table hash value %llu",
 233                                   (long long)rec_hash(&rec));
 234         }
 235
 236         off += sizeof(rec);
 237         return check_hash_tree(ntdb, off,
 238                                NTDB_SUBLEVEL_HASH_BITS-NTDB_HASH_GROUP_BITS,
 239                                hprefix, hprefix_bits,
 240                                used, num_used, num_found, check, data);
 241 }
 242
 243 static int off_cmp(const ntdb_off_t *a, const ntdb_off_t *b)
 244 {
 245         /* Can overflow an int. */
 246         return *a > *b ? 1
 247                 : *a < *b ? -1
 248                 : 0;
 249 }
 250
 251 static uint64_t get_bits(uint64_t h, unsigned num, unsigned *used)
 252 {
 253         *used += num;
 254
 255         return (h >> (64 - *used)) & ((1U << num) - 1);
 256 }
 257
 258 static enum NTDB_ERROR check_hash_tree(struct ntdb_context *ntdb,
 259                                       ntdb_off_t off, unsigned int group_bits,
 260                                       uint64_t hprefix,
 261                                       unsigned hprefix_bits,
 262                                       ntdb_off_t used[],
 263                                       size_t num_used,
 264                                       size_t *num_found,
 265                                       enum NTDB_ERROR (*check)(NTDB_DATA,
 266                                                               NTDB_DATA, void *),
 267                                       void *data)
 268 {
 269         unsigned int g, b;
 270         const ntdb_off_t *hash;
 271         struct ntdb_used_record rec;
 272         enum NTDB_ERROR ecode;
 273
 274         hash = ntdb_access_read(ntdb, off,
 275                                sizeof(ntdb_off_t)
 276                                << (group_bits + NTDB_HASH_GROUP_BITS),
 277                                true);
 278         if (NTDB_PTR_IS_ERR(hash)) {
 279                 return NTDB_PTR_ERR(hash);
 280         }
 281
 282         for (g = 0; g < (1 << group_bits); g++) {
 283                 const ntdb_off_t *group = hash + (g << NTDB_HASH_GROUP_BITS);
 284                 for (b = 0; b < (1 << NTDB_HASH_GROUP_BITS); b++) {
 285                         unsigned int bucket, i, used_bits;
 286                         uint64_t h;
 287                         ntdb_off_t *p;
 288                         if (group[b] == 0)
 289                                 continue;
 290
 291                         off = group[b] & NTDB_OFF_MASK;
 292                         p = asearch(&off, used, num_used, off_cmp);
 293                         if (!p) {
 294                                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 295                                                    NTDB_LOG_ERROR,
 296                                                    "ntdb_check: Invalid offset"
 297                                                    " %llu in hash",
 298                                                    (long long)off);
 299                                 goto fail;
 300                         }
 301                         /* Mark it invalid. */
 302                         *p ^= 1;
 303                         (*num_found)++;
 304
 305                         if (hprefix_bits == 64) {
 306                                 /* Chained entries are unordered. */
 307                                 if (is_subhash(group[b])) {
 308                                         ecode = NTDB_ERR_CORRUPT;
 309                                         ntdb_logerr(ntdb, ecode,
 310                                                    NTDB_LOG_ERROR,
 311                                                    "ntdb_check: Invalid chain"
 312                                                    " entry subhash");
 313                                         goto fail;
 314                                 }
 315                                 h = hash_record(ntdb, off);
 316                                 if (h != hprefix) {
 317                                         ecode = NTDB_ERR_CORRUPT;
 318                                         ntdb_logerr(ntdb, ecode,
 319                                                    NTDB_LOG_ERROR,
 320                                                    "check: bad hash chain"
 321                                                    " placement"
 322                                                    " 0x%llx vs 0x%llx",
 323                                                    (long long)h,
 324                                                    (long long)hprefix);
 325                                         goto fail;
 326                                 }
 327                                 ecode = ntdb_read_convert(ntdb, off, &rec,
 328                                                          sizeof(rec));
 329                                 if (ecode != NTDB_SUCCESS) {
 330                                         goto fail;
 331                                 }
 332                                 goto check;
 333                         }
 334
 335                         if (is_subhash(group[b])) {
 336                                 uint64_t subprefix;
 337                                 subprefix = (hprefix
 338                                      << (group_bits + NTDB_HASH_GROUP_BITS))
 339                                         + g * (1 << NTDB_HASH_GROUP_BITS) + b;
 340
 341                                 ecode = check_hash_record(ntdb,
 342                                                group[b] & NTDB_OFF_MASK,
 343                                                subprefix,
 344                                                hprefix_bits
 345                                                        + group_bits
 346                                                        + NTDB_HASH_GROUP_BITS,
 347                                                used, num_used, num_found,
 348                                                check, data);
 349                                 if (ecode != NTDB_SUCCESS) {
 350                                         goto fail;
 351                                 }
 352                                 continue;
 353                         }
 354                         /* A normal entry */
 355
 356                         /* Does it belong here at all? */
 357                         h = hash_record(ntdb, off);
 358                         used_bits = 0;
 359                         if (get_bits(h, hprefix_bits, &used_bits) != hprefix
 360                             && hprefix_bits) {
 361                                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 362                                                    NTDB_LOG_ERROR,
 363                                                    "check: bad hash placement"
 364                                                    " 0x%llx vs 0x%llx",
 365                                                    (long long)h,
 366                                                    (long long)hprefix);
 367                                 goto fail;
 368                         }
 369
 370                         /* Does it belong in this group? */
 371                         if (get_bits(h, group_bits, &used_bits) != g) {
 372                                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 373                                                    NTDB_LOG_ERROR,
 374                                                    "check: bad group %llu"
 375                                                    " vs %u",
 376                                                    (long long)h, g);
 377                                 goto fail;
 378                         }
 379
 380                         /* Are bucket bits correct? */
 381                         bucket = group[b] & NTDB_OFF_HASH_GROUP_MASK;
 382                         if (get_bits(h, NTDB_HASH_GROUP_BITS, &used_bits)
 383                             != bucket) {
 384                                 used_bits -= NTDB_HASH_GROUP_BITS;
 385                                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 386                                                    NTDB_LOG_ERROR,
 387                                                    "check: bad bucket %u vs %u",
 388                                                    (unsigned)get_bits(h,
 389                                                         NTDB_HASH_GROUP_BITS,
 390                                                         &used_bits),
 391                                                    bucket);
 392                                 goto fail;
 393                         }
 394
 395                         /* There must not be any zero entries between
 396                          * the bucket it belongs in and this one! */
 397                         for (i = bucket;
 398                              i != b;
 399                              i = (i + 1) % (1 << NTDB_HASH_GROUP_BITS)) {
 400                                 if (group[i] == 0) {
 401                                         ecode = NTDB_ERR_CORRUPT;
 402                                         ntdb_logerr(ntdb, ecode,
 403                                                    NTDB_LOG_ERROR,
 404                                                    "check: bad group placement"
 405                                                    " %u vs %u",
 406                                                    b, bucket);
 407                                         goto fail;
 408                                 }
 409                         }
 410
 411                         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
 412                         if (ecode != NTDB_SUCCESS) {
 413                                 goto fail;
 414                         }
 415
 416                         /* Bottom bits must match header. */
 417                         if ((h & ((1 << 11)-1)) != rec_hash(&rec)) {
 418                                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 419                                                    NTDB_LOG_ERROR,
 420                                                    "ntdb_check: Bad hash magic"
 421                                                    " at offset %llu"
 422                                                    " (0x%llx vs 0x%llx)",
 423                                                    (long long)off,
 424                                                    (long long)h,
 425                                                    (long long)rec_hash(&rec));
 426                                 goto fail;
 427                         }
 428
 429                 check:
 430                         if (check) {
 431                                 NTDB_DATA k, d;
 432                                 const unsigned char *kptr;
 433
 434                                 kptr = ntdb_access_read(ntdb,
 435                                                        off + sizeof(rec),
 436                                                        rec_key_length(&rec)
 437                                                        + rec_data_length(&rec),
 438                                                        false);
 439                                 if (NTDB_PTR_IS_ERR(kptr)) {
 440                                         ecode = NTDB_PTR_ERR(kptr);
 441                                         goto fail;
 442                                 }
 443
 444                                 k = ntdb_mkdata(kptr, rec_key_length(&rec));
 445                                 d = ntdb_mkdata(kptr + k.dsize,
 446                                                rec_data_length(&rec));
 447                                 ecode = check(k, d, data);
 448                                 ntdb_access_release(ntdb, kptr);
 449                                 if (ecode != NTDB_SUCCESS) {
 450                                         goto fail;
 451                                 }
 452                         }
 453                 }
 454         }
 455         ntdb_access_release(ntdb, hash);
 456         return NTDB_SUCCESS;
 457
 458 fail:
 459         ntdb_access_release(ntdb, hash);
 460         return ecode;
 461 }
 462
 463 static enum NTDB_ERROR check_hash(struct ntdb_context *ntdb,
 464                                  ntdb_off_t used[],
 465                                  size_t num_used, size_t num_other_used,
 466                                  enum NTDB_ERROR (*check)(NTDB_DATA, NTDB_DATA, void *),
 467                                  void *data)
 468 {
 469         /* Free tables and capabilities also show up as used. */
 470         size_t num_found = num_other_used;
 471         enum NTDB_ERROR ecode;
 472
 473         ecode = check_hash_tree(ntdb, offsetof(struct ntdb_header, hashtable),
 474                                 NTDB_TOPLEVEL_HASH_BITS-NTDB_HASH_GROUP_BITS,
 475                                 0, 0, used, num_used, &num_found,
 476                                 check, data);
 477         if (ecode == NTDB_SUCCESS) {
 478                 if (num_found != num_used) {
 479                         ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 480                                            "ntdb_check: Not all entries"
 481                                            " are in hash");
 482                 }
 483         }
 484         return ecode;
 485 }
 486
 487 static enum NTDB_ERROR check_free(struct ntdb_context *ntdb,
 488                                  ntdb_off_t off,
 489                                  const struct ntdb_free_record *frec,
 490                                  ntdb_off_t prev, unsigned int ftable,
 491                                  unsigned int bucket)
 492 {
 493         enum NTDB_ERROR ecode;
 494
 495         if (frec_magic(frec) != NTDB_FREE_MAGIC) {
 496                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 497                                   "ntdb_check: offset %llu bad magic 0x%llx",
 498                                   (long long)off,
 499                                   (long long)frec->magic_and_prev);
 500         }
 501         if (frec_ftable(frec) != ftable) {
 502                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 503                                   "ntdb_check: offset %llu bad freetable %u",
 504                                   (long long)off, frec_ftable(frec));
 505
 506         }
 507
 508         ecode = ntdb_oob(ntdb, off,
 509                          frec_len(frec) + sizeof(struct ntdb_used_record),
 510                          false);
 511         if (ecode != NTDB_SUCCESS) {
 512                 return ecode;
 513         }
 514         if (size_to_bucket(frec_len(frec)) != bucket) {
 515                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 516                                   "ntdb_check: offset %llu in wrong bucket"
 517                                   " (%u vs %u)",
 518                                   (long long)off,
 519                                   bucket, size_to_bucket(frec_len(frec)));
 520         }
 521         if (prev && prev != frec_prev(frec)) {
 522                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 523                                   "ntdb_check: offset %llu bad prev"
 524                                   " (%llu vs %llu)",
 525                                   (long long)off,
 526                                   (long long)prev, (long long)frec_len(frec));
 527         }
 528         return NTDB_SUCCESS;
 529 }
 530
 531 static enum NTDB_ERROR check_free_table(struct ntdb_context *ntdb,
 532                                        ntdb_off_t ftable_off,
 533                                        unsigned ftable_num,
 534                                        ntdb_off_t fr[],
 535                                        size_t num_free,
 536                                        size_t *num_found)
 537 {
 538         struct ntdb_freetable ft;
 539         ntdb_off_t h;
 540         unsigned int i;
 541         enum NTDB_ERROR ecode;
 542
 543         ecode = ntdb_read_convert(ntdb, ftable_off, &ft, sizeof(ft));
 544         if (ecode != NTDB_SUCCESS) {
 545                 return ecode;
 546         }
 547
 548         if (rec_magic(&ft.hdr) != NTDB_FTABLE_MAGIC
 549             || rec_key_length(&ft.hdr) != 0
 550             || rec_data_length(&ft.hdr) != sizeof(ft) - sizeof(ft.hdr)
 551             || rec_hash(&ft.hdr) != 0) {
 552                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 553                                   "ntdb_check: Invalid header on free table");
 554         }
 555
 556         for (i = 0; i < NTDB_FREE_BUCKETS; i++) {
 557                 ntdb_off_t off, prev = 0, *p, first = 0;
 558                 struct ntdb_free_record f;
 559
 560                 h = bucket_off(ftable_off, i);
 561                 for (off = ntdb_read_off(ntdb, h); off; off = f.next) {
 562                         if (NTDB_OFF_IS_ERR(off)) {
 563                                 return NTDB_OFF_TO_ERR(off);
 564                         }
 565                         if (!first) {
 566                                 off &= NTDB_OFF_MASK;
 567                                 first = off;
 568                         }
 569                         ecode = ntdb_read_convert(ntdb, off, &f, sizeof(f));
 570                         if (ecode != NTDB_SUCCESS) {
 571                                 return ecode;
 572                         }
 573                         ecode = check_free(ntdb, off, &f, prev, ftable_num, i);
 574                         if (ecode != NTDB_SUCCESS) {
 575                                 return ecode;
 576                         }
 577
 578                         /* FIXME: Check hash bits */
 579                         p = asearch(&off, fr, num_free, off_cmp);
 580                         if (!p) {
 581                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 582                                                   NTDB_LOG_ERROR,
 583                                                   "ntdb_check: Invalid offset"
 584                                                   " %llu in free table",
 585                                                   (long long)off);
 586                         }
 587                         /* Mark it invalid. */
 588                         *p ^= 1;
 589                         (*num_found)++;
 590                         prev = off;
 591                 }
 592
 593                 if (first) {
 594                         /* Now we can check first back pointer. */
 595                         ecode = ntdb_read_convert(ntdb, first, &f, sizeof(f));
 596                         if (ecode != NTDB_SUCCESS) {
 597                                 return ecode;
 598                         }
 599                         ecode = check_free(ntdb, first, &f, prev, ftable_num, i);
 600                         if (ecode != NTDB_SUCCESS) {
 601                                 return ecode;
 602                         }
 603                 }
 604         }
 605         return NTDB_SUCCESS;
 606 }
 607
 608 /* Slow, but should be very rare. */
 609 ntdb_off_t dead_space(struct ntdb_context *ntdb, ntdb_off_t off)
 610 {
 611         size_t len;
 612         enum NTDB_ERROR ecode;
 613
 614         for (len = 0; off + len < ntdb->file->map_size; len++) {
 615                 char c;
 616                 ecode = ntdb->io->tread(ntdb, off, &c, 1);
 617                 if (ecode != NTDB_SUCCESS) {
 618                         return NTDB_ERR_TO_OFF(ecode);
 619                 }
 620                 if (c != 0 && c != 0x43)
 621                         break;
 622         }
 623         return len;
 624 }
 625
 626 static enum NTDB_ERROR check_linear(struct ntdb_context *ntdb,
 627                                    ntdb_off_t **used, size_t *num_used,
 628                                    ntdb_off_t **fr, size_t *num_free,
 629                                    uint64_t features, ntdb_off_t recovery)
 630 {
 631         ntdb_off_t off;
 632         ntdb_len_t len;
 633         enum NTDB_ERROR ecode;
 634         bool found_recovery = false;
 635
 636         for (off = sizeof(struct ntdb_header);
 637              off < ntdb->file->map_size;
 638              off += len) {
 639                 union {
 640                         struct ntdb_used_record u;
 641                         struct ntdb_free_record f;
 642                         struct ntdb_recovery_record r;
 643                 } rec;
 644                 /* r is larger: only get that if we need to. */
 645                 ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec.f));
 646                 if (ecode != NTDB_SUCCESS) {
 647                         return ecode;
 648                 }
 649
 650                 /* If we crash after ftruncate, we can get zeroes or fill. */
 651                 if (rec.r.magic == NTDB_RECOVERY_INVALID_MAGIC
 652                     || rec.r.magic ==  0x4343434343434343ULL) {
 653                         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec.r));
 654                         if (ecode != NTDB_SUCCESS) {
 655                                 return ecode;
 656                         }
 657                         if (recovery == off) {
 658                                 found_recovery = true;
 659                                 len = sizeof(rec.r) + rec.r.max_len;
 660                         } else {
 661                                 len = dead_space(ntdb, off);
 662                                 if (NTDB_OFF_IS_ERR(len)) {
 663                                         return NTDB_OFF_TO_ERR(len);
 664                                 }
 665                                 if (len < sizeof(rec.r)) {
 666                                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 667                                                           NTDB_LOG_ERROR,
 668                                                           "ntdb_check: invalid"
 669                                                           " dead space at %zu",
 670                                                           (size_t)off);
 671                                 }
 672
 673                                 ntdb_logerr(ntdb, NTDB_SUCCESS, NTDB_LOG_WARNING,
 674                                            "Dead space at %zu-%zu (of %zu)",
 675                                            (size_t)off, (size_t)(off + len),
 676                                            (size_t)ntdb->file->map_size);
 677                         }
 678                 } else if (rec.r.magic == NTDB_RECOVERY_MAGIC) {
 679                         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec.r));
 680                         if (ecode != NTDB_SUCCESS) {
 681                                 return ecode;
 682                         }
 683                         if (recovery != off) {
 684                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 685                                                   NTDB_LOG_ERROR,
 686                                                   "ntdb_check: unexpected"
 687                                                   " recovery record at offset"
 688                                                   " %zu",
 689                                                   (size_t)off);
 690                         }
 691                         if (rec.r.len > rec.r.max_len) {
 692                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 693                                                   NTDB_LOG_ERROR,
 694                                                   "ntdb_check: invalid recovery"
 695                                                   " length %zu",
 696                                                   (size_t)rec.r.len);
 697                         }
 698                         if (rec.r.eof > ntdb->file->map_size) {
 699                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 700                                                   NTDB_LOG_ERROR,
 701                                                   "ntdb_check: invalid old EOF"
 702                                                   " %zu", (size_t)rec.r.eof);
 703                         }
 704                         found_recovery = true;
 705                         len = sizeof(rec.r) + rec.r.max_len;
 706                 } else if (frec_magic(&rec.f) == NTDB_FREE_MAGIC) {
 707                         len = sizeof(rec.u) + frec_len(&rec.f);
 708                         if (off + len > ntdb->file->map_size) {
 709                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 710                                                   NTDB_LOG_ERROR,
 711                                                   "ntdb_check: free overlength"
 712                                                   " %llu at offset %llu",
 713                                                   (long long)len,
 714                                                   (long long)off);
 715                         }
 716                         /* This record should be in free lists. */
 717                         if (frec_ftable(&rec.f) != NTDB_FTABLE_NONE
 718                             && !append(ntdb, fr, num_free, off)) {
 719                                 return ntdb_logerr(ntdb, NTDB_ERR_OOM,
 720                                                   NTDB_LOG_ERROR,
 721                                                   "ntdb_check: tracking %zu'th"
 722                                                   " free record.", *num_free);
 723                         }
 724                 } else if (rec_magic(&rec.u) == NTDB_USED_MAGIC
 725                            || rec_magic(&rec.u) == NTDB_CHAIN_MAGIC
 726                            || rec_magic(&rec.u) == NTDB_HTABLE_MAGIC
 727                            || rec_magic(&rec.u) == NTDB_FTABLE_MAGIC
 728                            || rec_magic(&rec.u) == NTDB_CAP_MAGIC) {
 729                         uint64_t klen, dlen, extra;
 730
 731                         /* This record is used! */
 732                         if (!append(ntdb, used, num_used, off)) {
 733                                 return ntdb_logerr(ntdb, NTDB_ERR_OOM,
 734                                                   NTDB_LOG_ERROR,
 735                                                   "ntdb_check: tracking %zu'th"
 736                                                   " used record.", *num_used);
 737                         }
 738
 739                         klen = rec_key_length(&rec.u);
 740                         dlen = rec_data_length(&rec.u);
 741                         extra = rec_extra_padding(&rec.u);
 742
 743                         len = sizeof(rec.u) + klen + dlen + extra;
 744                         if (off + len > ntdb->file->map_size) {
 745                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 746                                                   NTDB_LOG_ERROR,
 747                                                   "ntdb_check: used overlength"
 748                                                   " %llu at offset %llu",
 749                                                   (long long)len,
 750                                                   (long long)off);
 751                         }
 752
 753                         if (len < sizeof(rec.f)) {
 754                                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 755                                                   NTDB_LOG_ERROR,
 756                                                   "ntdb_check: too short record"
 757                                                   " %llu at %llu",
 758                                                   (long long)len,
 759                                                   (long long)off);
 760                         }
 761
 762                         /* Check that records have correct 0 at end (but may
 763                          * not in future). */
 764                         if (extra && !features
 765                             && rec_magic(&rec.u) != NTDB_CAP_MAGIC) {
 766                                 const char *p;
 767                                 char c;
 768                                 p = ntdb_access_read(ntdb, off + sizeof(rec.u)
 769                                                     + klen + dlen, 1, false);
 770                                 if (NTDB_PTR_IS_ERR(p))
 771                                         return NTDB_PTR_ERR(p);
 772                                 c = *p;
 773                                 ntdb_access_release(ntdb, p);
 774
 775                                 if (c != '\0') {
 776                                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 777                                                           NTDB_LOG_ERROR,
 778                                                           "ntdb_check:"
 779                                                           " non-zero extra"
 780                                                           " at %llu",
 781                                                           (long long)off);
 782                                 }
 783                         }
 784                 } else {
 785                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
 786                                           NTDB_LOG_ERROR,
 787                                           "ntdb_check: Bad magic 0x%llx"
 788                                           " at offset %zu",
 789                                           (long long)rec_magic(&rec.u),
 790                                           (size_t)off);
 791                 }
 792         }
 793
 794         /* We must have found recovery area if there was one. */
 795         if (recovery != 0 && !found_recovery) {
 796                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 797                                   "ntdb_check: expected a recovery area at %zu",
 798                                   (size_t)recovery);
 799         }
 800
 801         return NTDB_SUCCESS;
 802 }
 803
 804 _PUBLIC_ enum NTDB_ERROR ntdb_check_(struct ntdb_context *ntdb,
 805                           enum NTDB_ERROR (*check)(NTDB_DATA, NTDB_DATA, void *),
 806                           void *data)
 807 {
 808         ntdb_off_t *fr = NULL, *used = NULL, ft, recovery;
 809         size_t num_free = 0, num_used = 0, num_found = 0, num_ftables = 0,
 810                 num_capabilities = 0;
 811         uint64_t features;
 812         enum NTDB_ERROR ecode;
 813
 814         if (ntdb->flags & NTDB_CANT_CHECK) {
 815                 return ntdb_logerr(ntdb, NTDB_SUCCESS, NTDB_LOG_WARNING,
 816                                   "ntdb_check: database has unknown capability,"
 817                                   " cannot check.");
 818         }
 819
 820         ecode = ntdb_allrecord_lock(ntdb, F_RDLCK, NTDB_LOCK_WAIT, false);
 821         if (ecode != NTDB_SUCCESS) {
 822                 return ecode;
 823         }
 824
 825         ecode = ntdb_lock_expand(ntdb, F_RDLCK);
 826         if (ecode != NTDB_SUCCESS) {
 827                 ntdb_allrecord_unlock(ntdb, F_RDLCK);
 828                 return ecode;
 829         }
 830
 831         ecode = check_header(ntdb, &recovery, &features, &num_capabilities);
 832         if (ecode != NTDB_SUCCESS)
 833                 goto out;
 834
 835         /* First we do a linear scan, checking all records. */
 836         ecode = check_linear(ntdb, &used, &num_used, &fr, &num_free, features,
 837                              recovery);
 838         if (ecode != NTDB_SUCCESS)
 839                 goto out;
 840
 841         for (ft = first_ftable(ntdb); ft; ft = next_ftable(ntdb, ft)) {
 842                 if (NTDB_OFF_IS_ERR(ft)) {
 843                         ecode = NTDB_OFF_TO_ERR(ft);
 844                         goto out;
 845                 }
 846                 ecode = check_free_table(ntdb, ft, num_ftables, fr, num_free,
 847                                          &num_found);
 848                 if (ecode != NTDB_SUCCESS)
 849                         goto out;
 850                 num_ftables++;
 851         }
 852
 853         /* FIXME: Check key uniqueness? */
 854         ecode = check_hash(ntdb, used, num_used, num_ftables + num_capabilities,
 855                            check, data);
 856         if (ecode != NTDB_SUCCESS)
 857                 goto out;
 858
 859         if (num_found != num_free) {
 860                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 861                                    "ntdb_check: Not all entries are in"
 862                                    " free table");
 863         }
 864
 865 out:
 866         ntdb_allrecord_unlock(ntdb, F_RDLCK);
 867         ntdb_unlock_expand(ntdb, F_RDLCK);
 868         ntdb->free_fn(fr, ntdb->alloc_data);
 869         ntdb->free_fn(used, ntdb->alloc_data);
 870         return ecode;
 871 }