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

   1  /*
   2    Trivial Database 2: hash 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 <assert.h>
  20
  21 uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
  22 {
  23         return tdb->hash_fn(ptr, len, tdb->hash_seed, tdb->hash_data);
  24 }
  25
  26 uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off)
  27 {
  28         const struct tdb_used_record *r;
  29         const void *key;
  30         uint64_t klen, hash;
  31
  32         r = tdb_access_read(tdb, off, sizeof(*r), true);
  33         if (TDB_PTR_IS_ERR(r)) {
  34                 /* FIXME */
  35                 return 0;
  36         }
  37
  38         klen = rec_key_length(r);
  39         tdb_access_release(tdb, r);
  40
  41         key = tdb_access_read(tdb, off + sizeof(*r), klen, false);
  42         if (TDB_PTR_IS_ERR(key)) {
  43                 return 0;
  44         }
  45
  46         hash = tdb_hash(tdb, key, klen);
  47         tdb_access_release(tdb, key);
  48         return hash;
  49 }
  50
  51 /* Get bits from a value. */
  52 static uint32_t bits_from(uint64_t val, unsigned start, unsigned num)
  53 {
  54         assert(num <= 32);
  55         return (val >> start) & ((1U << num) - 1);
  56 }
  57
  58 /* We take bits from the top: that way we can lock whole sections of the hash
  59  * by using lock ranges. */
  60 static uint32_t use_bits(struct hash_info *h, unsigned num)
  61 {
  62         h->hash_used += num;
  63         return bits_from(h->h, 64 - h->hash_used, num);
  64 }
  65
  66 static tdb_bool_err key_matches(struct tdb_context *tdb,
  67                                 const struct tdb_used_record *rec,
  68                                 tdb_off_t off,
  69                                 const struct tdb_data *key)
  70 {
  71         tdb_bool_err ret = false;
  72         const char *rkey;
  73
  74         if (rec_key_length(rec) != key->dsize) {
  75                 tdb->stats.compare_wrong_keylen++;
  76                 return ret;
  77         }
  78
  79         rkey = tdb_access_read(tdb, off + sizeof(*rec), key->dsize, false);
  80         if (TDB_PTR_IS_ERR(rkey)) {
  81                 return TDB_PTR_ERR(rkey);
  82         }
  83         if (memcmp(rkey, key->dptr, key->dsize) == 0)
  84                 ret = true;
  85         else
  86                 tdb->stats.compare_wrong_keycmp++;
  87         tdb_access_release(tdb, rkey);
  88         return ret;
  89 }
  90
  91 /* Does entry match? */
  92 static tdb_bool_err match(struct tdb_context *tdb,
  93                           struct hash_info *h,
  94                           const struct tdb_data *key,
  95                           tdb_off_t val,
  96                           struct tdb_used_record *rec)
  97 {
  98         tdb_off_t off;
  99         enum TDB_ERROR ecode;
 100
 101         tdb->stats.compares++;
 102         /* Desired bucket must match. */
 103         if (h->home_bucket != (val & TDB_OFF_HASH_GROUP_MASK)) {
 104                 tdb->stats.compare_wrong_bucket++;
 105                 return false;
 106         }
 107
 108         /* Top bits of offset == next bits of hash. */
 109         if (bits_from(val, TDB_OFF_HASH_EXTRA_BIT, TDB_OFF_UPPER_STEAL_EXTRA)
 110             != bits_from(h->h, 64 - h->hash_used - TDB_OFF_UPPER_STEAL_EXTRA,
 111                     TDB_OFF_UPPER_STEAL_EXTRA)) {
 112                 tdb->stats.compare_wrong_offsetbits++;
 113                 return false;
 114         }
 115
 116         off = val & TDB_OFF_MASK;
 117         ecode = tdb_read_convert(tdb, off, rec, sizeof(*rec));
 118         if (ecode != TDB_SUCCESS) {
 119                 return ecode;
 120         }
 121
 122         if ((h->h & ((1 << 11)-1)) != rec_hash(rec)) {
 123                 tdb->stats.compare_wrong_rechash++;
 124                 return false;
 125         }
 126
 127         return key_matches(tdb, rec, off, key);
 128 }
 129
 130 static tdb_off_t hbucket_off(tdb_off_t group_start, unsigned bucket)
 131 {
 132         return group_start
 133                 + (bucket % (1 << TDB_HASH_GROUP_BITS)) * sizeof(tdb_off_t);
 134 }
 135
 136 bool is_subhash(tdb_off_t val)
 137 {
 138         return (val >> TDB_OFF_UPPER_STEAL_SUBHASH_BIT) & 1;
 139 }
 140
 141 /* FIXME: Guess the depth, don't over-lock! */
 142 static tdb_off_t hlock_range(tdb_off_t group, tdb_off_t *size)
 143 {
 144         *size = 1ULL << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
 145         return group << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
 146 }
 147
 148 static tdb_off_t COLD find_in_chain(struct tdb_context *tdb,
 149                                     struct tdb_data key,
 150                                     tdb_off_t chain,
 151                                     struct hash_info *h,
 152                                     struct tdb_used_record *rec,
 153                                     struct traverse_info *tinfo)
 154 {
 155         tdb_off_t off, next;
 156         enum TDB_ERROR ecode;
 157
 158         /* In case nothing is free, we set these to zero. */
 159         h->home_bucket = h->found_bucket = 0;
 160
 161         for (off = chain; off; off = next) {
 162                 unsigned int i;
 163
 164                 h->group_start = off;
 165                 ecode = tdb_read_convert(tdb, off, h->group, sizeof(h->group));
 166                 if (ecode != TDB_SUCCESS) {
 167                         return ecode;
 168                 }
 169
 170                 for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 171                         tdb_off_t recoff;
 172                         if (!h->group[i]) {
 173                                 /* Remember this empty bucket. */
 174                                 h->home_bucket = h->found_bucket = i;
 175                                 continue;
 176                         }
 177
 178                         /* We can insert extra bits via add_to_hash
 179                          * empty bucket logic. */
 180                         recoff = h->group[i] & TDB_OFF_MASK;
 181                         ecode = tdb_read_convert(tdb, recoff, rec,
 182                                                  sizeof(*rec));
 183                         if (ecode != TDB_SUCCESS) {
 184                                 return ecode;
 185                         }
 186
 187                         ecode = key_matches(tdb, rec, recoff, &key);
 188                         if (ecode < 0) {
 189                                 return ecode;
 190                         }
 191                         if (ecode == 1) {
 192                                 h->home_bucket = h->found_bucket = i;
 193
 194                                 if (tinfo) {
 195                                         tinfo->levels[tinfo->num_levels]
 196                                                 .hashtable = off;
 197                                         tinfo->levels[tinfo->num_levels]
 198                                                 .total_buckets
 199                                                 = 1 << TDB_HASH_GROUP_BITS;
 200                                         tinfo->levels[tinfo->num_levels].entry
 201                                                 = i;
 202                                         tinfo->num_levels++;
 203                                 }
 204                                 return recoff;
 205                         }
 206                 }
 207                 next = tdb_read_off(tdb, off
 208                                     + offsetof(struct tdb_chain, next));
 209                 if (TDB_OFF_IS_ERR(next)) {
 210                         return next;
 211                 }
 212                 if (next)
 213                         next += sizeof(struct tdb_used_record);
 214         }
 215         return 0;
 216 }
 217
 218 /* This is the core routine which searches the hashtable for an entry.
 219  * On error, no locks are held and -ve is returned.
 220  * Otherwise, hinfo is filled in (and the optional tinfo).
 221  * If not found, the return value is 0.
 222  * If found, the return value is the offset, and *rec is the record. */
 223 tdb_off_t find_and_lock(struct tdb_context *tdb,
 224                         struct tdb_data key,
 225                         int ltype,
 226                         struct hash_info *h,
 227                         struct tdb_used_record *rec,
 228                         struct traverse_info *tinfo)
 229 {
 230         uint32_t i, group;
 231         tdb_off_t hashtable;
 232         enum TDB_ERROR ecode;
 233
 234         h->h = tdb_hash(tdb, key.dptr, key.dsize);
 235         h->hash_used = 0;
 236         group = use_bits(h, TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS);
 237         h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
 238
 239         h->hlock_start = hlock_range(group, &h->hlock_range);
 240         ecode = tdb_lock_hashes(tdb, h->hlock_start, h->hlock_range, ltype,
 241                                 TDB_LOCK_WAIT);
 242         if (ecode != TDB_SUCCESS) {
 243                 return ecode;
 244         }
 245
 246         hashtable = offsetof(struct tdb_header, hashtable);
 247         if (tinfo) {
 248                 tinfo->toplevel_group = group;
 249                 tinfo->num_levels = 1;
 250                 tinfo->levels[0].entry = 0;
 251                 tinfo->levels[0].hashtable = hashtable
 252                         + (group << TDB_HASH_GROUP_BITS) * sizeof(tdb_off_t);
 253                 tinfo->levels[0].total_buckets = 1 << TDB_HASH_GROUP_BITS;
 254         }
 255
 256         while (h->hash_used <= 64) {
 257                 /* Read in the hash group. */
 258                 h->group_start = hashtable
 259                         + group * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 260
 261                 ecode = tdb_read_convert(tdb, h->group_start, &h->group,
 262                                          sizeof(h->group));
 263                 if (ecode != TDB_SUCCESS) {
 264                         goto fail;
 265                 }
 266
 267                 /* Pointer to another hash table?  Go down... */
 268                 if (is_subhash(h->group[h->home_bucket])) {
 269                         hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK)
 270                                 + sizeof(struct tdb_used_record);
 271                         if (tinfo) {
 272                                 /* When we come back, use *next* bucket */
 273                                 tinfo->levels[tinfo->num_levels-1].entry
 274                                         += h->home_bucket + 1;
 275                         }
 276                         group = use_bits(h, TDB_SUBLEVEL_HASH_BITS
 277                                          - TDB_HASH_GROUP_BITS);
 278                         h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
 279                         if (tinfo) {
 280                                 tinfo->levels[tinfo->num_levels].hashtable
 281                                         = hashtable;
 282                                 tinfo->levels[tinfo->num_levels].total_buckets
 283                                         = 1 << TDB_SUBLEVEL_HASH_BITS;
 284                                 tinfo->levels[tinfo->num_levels].entry
 285                                         = group << TDB_HASH_GROUP_BITS;
 286                                 tinfo->num_levels++;
 287                         }
 288                         continue;
 289                 }
 290
 291                 /* It's in this group: search (until 0 or all searched) */
 292                 for (i = 0, h->found_bucket = h->home_bucket;
 293                      i < (1 << TDB_HASH_GROUP_BITS);
 294                      i++, h->found_bucket = ((h->found_bucket+1)
 295                                              % (1 << TDB_HASH_GROUP_BITS))) {
 296                         tdb_bool_err berr;
 297                         if (is_subhash(h->group[h->found_bucket]))
 298                                 continue;
 299
 300                         if (!h->group[h->found_bucket])
 301                                 break;
 302
 303                         berr = match(tdb, h, &key, h->group[h->found_bucket],
 304                                      rec);
 305                         if (berr < 0) {
 306                                 ecode = berr;
 307                                 goto fail;
 308                         }
 309                         if (berr) {
 310                                 if (tinfo) {
 311                                         tinfo->levels[tinfo->num_levels-1].entry
 312                                                 += h->found_bucket;
 313                                 }
 314                                 return h->group[h->found_bucket] & TDB_OFF_MASK;
 315                         }
 316                 }
 317                 /* Didn't find it: h indicates where it would go. */
 318                 return 0;
 319         }
 320
 321         return find_in_chain(tdb, key, hashtable, h, rec, tinfo);
 322
 323 fail:
 324         tdb_unlock_hashes(tdb, h->hlock_start, h->hlock_range, ltype);
 325         return ecode;
 326 }
 327
 328 /* I wrote a simple test, expanding a hash to 2GB, for the following
 329  * cases:
 330  * 1) Expanding all the buckets at once,
 331  * 2) Expanding the bucket we wanted to place the new entry into.
 332  * 3) Expanding the most-populated bucket,
 333  *
 334  * I measured the worst/average/best density during this process.
 335  * 1) 3%/16%/30%
 336  * 2) 4%/20%/38%
 337  * 3) 6%/22%/41%
 338  *
 339  * So we figure out the busiest bucket for the moment.
 340  */
 341 static unsigned fullest_bucket(struct tdb_context *tdb,
 342                                const tdb_off_t *group,
 343                                unsigned new_bucket)
 344 {
 345         unsigned counts[1 << TDB_HASH_GROUP_BITS] = { 0 };
 346         unsigned int i, best_bucket;
 347
 348         /* Count the new entry. */
 349         counts[new_bucket]++;
 350         best_bucket = new_bucket;
 351
 352         for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 353                 unsigned this_bucket;
 354
 355                 if (is_subhash(group[i]))
 356                         continue;
 357                 this_bucket = group[i] & TDB_OFF_HASH_GROUP_MASK;
 358                 if (++counts[this_bucket] > counts[best_bucket])
 359                         best_bucket = this_bucket;
 360         }
 361
 362         return best_bucket;
 363 }
 364
 365 static bool put_into_group(tdb_off_t *group,
 366                            unsigned bucket, tdb_off_t encoded)
 367 {
 368         unsigned int i;
 369
 370         for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 371                 unsigned b = (bucket + i) % (1 << TDB_HASH_GROUP_BITS);
 372
 373                 if (group[b] == 0) {
 374                         group[b] = encoded;
 375                         return true;
 376                 }
 377         }
 378         return false;
 379 }
 380
 381 static void force_into_group(tdb_off_t *group,
 382                              unsigned bucket, tdb_off_t encoded)
 383 {
 384         if (!put_into_group(group, bucket, encoded))
 385                 abort();
 386 }
 387
 388 static tdb_off_t encode_offset(tdb_off_t new_off, struct hash_info *h)
 389 {
 390         return h->home_bucket
 391                 | new_off
 392                 | ((uint64_t)bits_from(h->h,
 393                                   64 - h->hash_used - TDB_OFF_UPPER_STEAL_EXTRA,
 394                                   TDB_OFF_UPPER_STEAL_EXTRA)
 395                    << TDB_OFF_HASH_EXTRA_BIT);
 396 }
 397
 398 /* Simply overwrite the hash entry we found before. */
 399 enum TDB_ERROR replace_in_hash(struct tdb_context *tdb,
 400                                struct hash_info *h,
 401                                tdb_off_t new_off)
 402 {
 403         return tdb_write_off(tdb, hbucket_off(h->group_start, h->found_bucket),
 404                              encode_offset(new_off, h));
 405 }
 406
 407 /* We slot in anywhere that's empty in the chain. */
 408 static enum TDB_ERROR COLD add_to_chain(struct tdb_context *tdb,
 409                                         tdb_off_t subhash,
 410                                         tdb_off_t new_off)
 411 {
 412         tdb_off_t entry;
 413         enum TDB_ERROR ecode;
 414
 415         entry = tdb_find_zero_off(tdb, subhash, 1<<TDB_HASH_GROUP_BITS);
 416         if (TDB_OFF_IS_ERR(entry)) {
 417                 return entry;
 418         }
 419
 420         if (entry == 1 << TDB_HASH_GROUP_BITS) {
 421                 tdb_off_t next;
 422
 423                 next = tdb_read_off(tdb, subhash
 424                                     + offsetof(struct tdb_chain, next));
 425                 if (TDB_OFF_IS_ERR(next)) {
 426                         return next;
 427                 }
 428
 429                 if (!next) {
 430                         next = alloc(tdb, 0, sizeof(struct tdb_chain), 0,
 431                                      TDB_CHAIN_MAGIC, false);
 432                         if (TDB_OFF_IS_ERR(next))
 433                                 return next;
 434                         ecode = zero_out(tdb,
 435                                          next+sizeof(struct tdb_used_record),
 436                                          sizeof(struct tdb_chain));
 437                         if (ecode != TDB_SUCCESS) {
 438                                 return ecode;
 439                         }
 440                         ecode = tdb_write_off(tdb, subhash
 441                                               + offsetof(struct tdb_chain,
 442                                                          next),
 443                                               next);
 444                         if (ecode != TDB_SUCCESS) {
 445                                 return ecode;
 446                         }
 447                 }
 448                 return add_to_chain(tdb, next, new_off);
 449         }
 450
 451         return tdb_write_off(tdb, subhash + entry * sizeof(tdb_off_t),
 452                              new_off);
 453 }
 454
 455 /* Add into a newly created subhash. */
 456 static enum TDB_ERROR add_to_subhash(struct tdb_context *tdb, tdb_off_t subhash,
 457                                      unsigned hash_used, tdb_off_t val)
 458 {
 459         tdb_off_t off = (val & TDB_OFF_MASK), *group;
 460         struct hash_info h;
 461         unsigned int gnum;
 462
 463         h.hash_used = hash_used;
 464
 465         if (hash_used + TDB_SUBLEVEL_HASH_BITS > 64)
 466                 return add_to_chain(tdb, subhash, off);
 467
 468         h.h = hash_record(tdb, off);
 469         gnum = use_bits(&h, TDB_SUBLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS);
 470         h.group_start = subhash
 471                 + gnum * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 472         h.home_bucket = use_bits(&h, TDB_HASH_GROUP_BITS);
 473
 474         group = tdb_access_write(tdb, h.group_start,
 475                                  sizeof(*group) << TDB_HASH_GROUP_BITS, true);
 476         if (TDB_PTR_IS_ERR(group)) {
 477                 return TDB_PTR_ERR(group);
 478         }
 479         force_into_group(group, h.home_bucket, encode_offset(off, &h));
 480         return tdb_access_commit(tdb, group);
 481 }
 482
 483 static enum TDB_ERROR expand_group(struct tdb_context *tdb, struct hash_info *h)
 484 {
 485         unsigned bucket, num_vals, i, magic;
 486         size_t subsize;
 487         tdb_off_t subhash;
 488         tdb_off_t vals[1 << TDB_HASH_GROUP_BITS];
 489         enum TDB_ERROR ecode;
 490
 491         /* Attach new empty subhash under fullest bucket. */
 492         bucket = fullest_bucket(tdb, h->group, h->home_bucket);
 493
 494         if (h->hash_used == 64) {
 495                 tdb->stats.alloc_chain++;
 496                 subsize = sizeof(struct tdb_chain);
 497                 magic = TDB_CHAIN_MAGIC;
 498         } else {
 499                 tdb->stats.alloc_subhash++;
 500                 subsize = (sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS);
 501                 magic = TDB_HTABLE_MAGIC;
 502         }
 503
 504         subhash = alloc(tdb, 0, subsize, 0, magic, false);
 505         if (TDB_OFF_IS_ERR(subhash)) {
 506                 return subhash;
 507         }
 508
 509         ecode = zero_out(tdb, subhash + sizeof(struct tdb_used_record),
 510                          subsize);
 511         if (ecode != TDB_SUCCESS) {
 512                 return ecode;
 513         }
 514
 515         /* Remove any which are destined for bucket or are in wrong place. */
 516         num_vals = 0;
 517         for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 518                 unsigned home_bucket = h->group[i] & TDB_OFF_HASH_GROUP_MASK;
 519                 if (!h->group[i] || is_subhash(h->group[i]))
 520                         continue;
 521                 if (home_bucket == bucket || home_bucket != i) {
 522                         vals[num_vals++] = h->group[i];
 523                         h->group[i] = 0;
 524                 }
 525         }
 526         /* FIXME: This assert is valid, but we do this during unit test :( */
 527         /* assert(num_vals); */
 528
 529         /* Overwrite expanded bucket with subhash pointer. */
 530         h->group[bucket] = subhash | (1ULL << TDB_OFF_UPPER_STEAL_SUBHASH_BIT);
 531
 532         /* Point to actual contents of record. */
 533         subhash += sizeof(struct tdb_used_record);
 534
 535         /* Put values back. */
 536         for (i = 0; i < num_vals; i++) {
 537                 unsigned this_bucket = vals[i] & TDB_OFF_HASH_GROUP_MASK;
 538
 539                 if (this_bucket == bucket) {
 540                         ecode = add_to_subhash(tdb, subhash, h->hash_used,
 541                                                vals[i]);
 542                         if (ecode != TDB_SUCCESS)
 543                                 return ecode;
 544                 } else {
 545                         /* There should be room to put this back. */
 546                         force_into_group(h->group, this_bucket, vals[i]);
 547                 }
 548         }
 549         return TDB_SUCCESS;
 550 }
 551
 552 enum TDB_ERROR delete_from_hash(struct tdb_context *tdb, struct hash_info *h)
 553 {
 554         unsigned int i, num_movers = 0;
 555         tdb_off_t movers[1 << TDB_HASH_GROUP_BITS];
 556
 557         h->group[h->found_bucket] = 0;
 558         for (i = 1; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 559                 unsigned this_bucket;
 560
 561                 this_bucket = (h->found_bucket+i) % (1 << TDB_HASH_GROUP_BITS);
 562                 /* Empty bucket?  We're done. */
 563                 if (!h->group[this_bucket])
 564                         break;
 565
 566                 /* Ignore subhashes. */
 567                 if (is_subhash(h->group[this_bucket]))
 568                         continue;
 569
 570                 /* If this one is not happy where it is, we'll move it. */
 571                 if ((h->group[this_bucket] & TDB_OFF_HASH_GROUP_MASK)
 572                     != this_bucket) {
 573                         movers[num_movers++] = h->group[this_bucket];
 574                         h->group[this_bucket] = 0;
 575                 }
 576         }
 577
 578         /* Put back the ones we erased. */
 579         for (i = 0; i < num_movers; i++) {
 580                 force_into_group(h->group, movers[i] & TDB_OFF_HASH_GROUP_MASK,
 581                                  movers[i]);
 582         }
 583
 584         /* Now we write back the hash group */
 585         return tdb_write_convert(tdb, h->group_start,
 586                                  h->group, sizeof(h->group));
 587 }
 588
 589 enum TDB_ERROR add_to_hash(struct tdb_context *tdb, struct hash_info *h,
 590                            tdb_off_t new_off)
 591 {
 592         enum TDB_ERROR ecode;
 593
 594         /* We hit an empty bucket during search?  That's where it goes. */
 595         if (!h->group[h->found_bucket]) {
 596                 h->group[h->found_bucket] = encode_offset(new_off, h);
 597                 /* Write back the modified group. */
 598                 return tdb_write_convert(tdb, h->group_start,
 599                                          h->group, sizeof(h->group));
 600         }
 601
 602         if (h->hash_used > 64)
 603                 return add_to_chain(tdb, h->group_start, new_off);
 604
 605         /* We're full.  Expand. */
 606         ecode = expand_group(tdb, h);
 607         if (ecode != TDB_SUCCESS) {
 608                 return ecode;
 609         }
 610
 611         if (is_subhash(h->group[h->home_bucket])) {
 612                 /* We were expanded! */
 613                 tdb_off_t hashtable;
 614                 unsigned int gnum;
 615
 616                 /* Write back the modified group. */
 617                 ecode = tdb_write_convert(tdb, h->group_start, h->group,
 618                                           sizeof(h->group));
 619                 if (ecode != TDB_SUCCESS) {
 620                         return ecode;
 621                 }
 622
 623                 /* Move hashinfo down a level. */
 624                 hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK)
 625                         + sizeof(struct tdb_used_record);
 626                 gnum = use_bits(h,TDB_SUBLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS);
 627                 h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
 628                 h->group_start = hashtable
 629                         + gnum * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 630                 ecode = tdb_read_convert(tdb, h->group_start, &h->group,
 631                                          sizeof(h->group));
 632                 if (ecode != TDB_SUCCESS) {
 633                         return ecode;
 634                 }
 635         }
 636
 637         /* Expanding the group must have made room if it didn't choose this
 638          * bucket. */
 639         if (put_into_group(h->group, h->home_bucket, encode_offset(new_off,h))){
 640                 return tdb_write_convert(tdb, h->group_start,
 641                                          h->group, sizeof(h->group));
 642         }
 643
 644         /* This can happen if all hashes in group (and us) dropped into same
 645          * group in subhash. */
 646         return add_to_hash(tdb, h, new_off);
 647 }
 648
 649 /* Traverse support: returns offset of record, or 0 or -ve error. */
 650 static tdb_off_t iterate_hash(struct tdb_context *tdb,
 651                               struct traverse_info *tinfo)
 652 {
 653         tdb_off_t off, val, i;
 654         struct traverse_level *tlevel;
 655
 656         tlevel = &tinfo->levels[tinfo->num_levels-1];
 657
 658 again:
 659         for (i = tdb_find_nonzero_off(tdb, tlevel->hashtable,
 660                                       tlevel->entry, tlevel->total_buckets);
 661              i != tlevel->total_buckets;
 662              i = tdb_find_nonzero_off(tdb, tlevel->hashtable,
 663                                       i+1, tlevel->total_buckets)) {
 664                 if (TDB_OFF_IS_ERR(i)) {
 665                         return i;
 666                 }
 667
 668                 val = tdb_read_off(tdb, tlevel->hashtable+sizeof(tdb_off_t)*i);
 669                 if (TDB_OFF_IS_ERR(val)) {
 670                         return val;
 671                 }
 672
 673                 off = val & TDB_OFF_MASK;
 674
 675                 /* This makes the delete-all-in-traverse case work
 676                  * (and simplifies our logic a little). */
 677                 if (off == tinfo->prev)
 678                         continue;
 679
 680                 tlevel->entry = i;
 681
 682                 if (!is_subhash(val)) {
 683                         /* Found one. */
 684                         tinfo->prev = off;
 685                         return off;
 686                 }
 687
 688                 /* When we come back, we want the next one */
 689                 tlevel->entry++;
 690                 tinfo->num_levels++;
 691                 tlevel++;
 692                 tlevel->hashtable = off + sizeof(struct tdb_used_record);
 693                 tlevel->entry = 0;
 694                 /* Next level is a chain? */
 695                 if (unlikely(tinfo->num_levels == TDB_MAX_LEVELS + 1))
 696                         tlevel->total_buckets = (1 << TDB_HASH_GROUP_BITS);
 697                 else
 698                         tlevel->total_buckets = (1 << TDB_SUBLEVEL_HASH_BITS);
 699                 goto again;
 700         }
 701
 702         /* Nothing there? */
 703         if (tinfo->num_levels == 1)
 704                 return 0;
 705
 706         /* Handle chained entries. */
 707         if (unlikely(tinfo->num_levels == TDB_MAX_LEVELS + 1)) {
 708                 tlevel->hashtable = tdb_read_off(tdb, tlevel->hashtable
 709                                                  + offsetof(struct tdb_chain,
 710                                                             next));
 711                 if (TDB_OFF_IS_ERR(tlevel->hashtable)) {
 712                         return tlevel->hashtable;
 713                 }
 714                 if (tlevel->hashtable) {
 715                         tlevel->hashtable += sizeof(struct tdb_used_record);
 716                         tlevel->entry = 0;
 717                         goto again;
 718                 }
 719         }
 720
 721         /* Go back up and keep searching. */
 722         tinfo->num_levels--;
 723         tlevel--;
 724         goto again;
 725 }
 726
 727 /* Return success if we find something, TDB_ERR_NOEXIST if none. */
 728 enum TDB_ERROR next_in_hash(struct tdb_context *tdb,
 729                             struct traverse_info *tinfo,
 730                             TDB_DATA *kbuf, size_t *dlen)
 731 {
 732         const unsigned group_bits = TDB_TOPLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS;
 733         tdb_off_t hl_start, hl_range, off;
 734         enum TDB_ERROR ecode;
 735
 736         while (tinfo->toplevel_group < (1 << group_bits)) {
 737                 hl_start = (tdb_off_t)tinfo->toplevel_group
 738                         << (64 - group_bits);
 739                 hl_range = 1ULL << group_bits;
 740                 ecode = tdb_lock_hashes(tdb, hl_start, hl_range, F_RDLCK,
 741                                         TDB_LOCK_WAIT);
 742                 if (ecode != TDB_SUCCESS) {
 743                         return ecode;
 744                 }
 745
 746                 off = iterate_hash(tdb, tinfo);
 747                 if (off) {
 748                         struct tdb_used_record rec;
 749
 750                         if (TDB_OFF_IS_ERR(off)) {
 751                                 ecode = off;
 752                                 goto fail;
 753                         }
 754
 755                         ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
 756                         if (ecode != TDB_SUCCESS) {
 757                                 goto fail;
 758                         }
 759                         if (rec_magic(&rec) != TDB_USED_MAGIC) {
 760                                 ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT,
 761                                                    TDB_LOG_ERROR,
 762                                                    "next_in_hash:"
 763                                                    " corrupt record at %llu",
 764                                                    (long long)off);
 765                                 goto fail;
 766                         }
 767
 768                         kbuf->dsize = rec_key_length(&rec);
 769
 770                         /* They want data as well? */
 771                         if (dlen) {
 772                                 *dlen = rec_data_length(&rec);
 773                                 kbuf->dptr = tdb_alloc_read(tdb,
 774                                                             off + sizeof(rec),
 775                                                             kbuf->dsize
 776                                                             + *dlen);
 777                         } else {
 778                                 kbuf->dptr = tdb_alloc_read(tdb,
 779                                                             off + sizeof(rec),
 780                                                             kbuf->dsize);
 781                         }
 782                         tdb_unlock_hashes(tdb, hl_start, hl_range, F_RDLCK);
 783                         if (TDB_PTR_IS_ERR(kbuf->dptr)) {
 784                                 return TDB_PTR_ERR(kbuf->dptr);
 785                         }
 786                         return TDB_SUCCESS;
 787                 }
 788
 789                 tdb_unlock_hashes(tdb, hl_start, hl_range, F_RDLCK);
 790
 791                 tinfo->toplevel_group++;
 792                 tinfo->levels[0].hashtable
 793                         += (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 794                 tinfo->levels[0].entry = 0;
 795         }
 796         return TDB_ERR_NOEXIST;
 797
 798 fail:
 799         tdb_unlock_hashes(tdb, hl_start, hl_range, F_RDLCK);
 800         return ecode;
 801
 802 }
 803
 804 enum TDB_ERROR first_in_hash(struct tdb_context *tdb,
 805                              struct traverse_info *tinfo,
 806                              TDB_DATA *kbuf, size_t *dlen)
 807 {
 808         tinfo->prev = 0;
 809         tinfo->toplevel_group = 0;
 810         tinfo->num_levels = 1;
 811         tinfo->levels[0].hashtable = offsetof(struct tdb_header, hashtable);
 812         tinfo->levels[0].entry = 0;
 813         tinfo->levels[0].total_buckets = (1 << TDB_HASH_GROUP_BITS);
 814
 815         return next_in_hash(tdb, tinfo, kbuf, dlen);
 816 }
 817
 818 /* Even if the entry isn't in this hash bucket, you'd have to lock this
 819  * bucket to find it. */
 820 static enum TDB_ERROR chainlock(struct tdb_context *tdb, const TDB_DATA *key,
 821                                 int ltype, enum tdb_lock_flags waitflag,
 822                                 const char *func)
 823 {
 824         enum TDB_ERROR ecode;
 825         uint64_t h = tdb_hash(tdb, key->dptr, key->dsize);
 826         tdb_off_t lockstart, locksize;
 827         unsigned int group, gbits;
 828
 829         gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS;
 830         group = bits_from(h, 64 - gbits, gbits);
 831
 832         lockstart = hlock_range(group, &locksize);
 833
 834         ecode = tdb_lock_hashes(tdb, lockstart, locksize, ltype, waitflag);
 835         tdb_trace_1rec(tdb, func, *key);
 836         return ecode;
 837 }
 838
 839 /* lock/unlock one hash chain. This is meant to be used to reduce
 840    contention - it cannot guarantee how many records will be locked */
 841 enum TDB_ERROR tdb_chainlock(struct tdb_context *tdb, TDB_DATA key)
 842 {
 843         return tdb->last_error = chainlock(tdb, &key, F_WRLCK, TDB_LOCK_WAIT,
 844                                            "tdb_chainlock");
 845 }
 846
 847 void tdb_chainunlock(struct tdb_context *tdb, TDB_DATA key)
 848 {
 849         uint64_t h = tdb_hash(tdb, key.dptr, key.dsize);
 850         tdb_off_t lockstart, locksize;
 851         unsigned int group, gbits;
 852
 853         gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS;
 854         group = bits_from(h, 64 - gbits, gbits);
 855
 856         lockstart = hlock_range(group, &locksize);
 857
 858         tdb_trace_1rec(tdb, "tdb_chainunlock", key);
 859         tdb_unlock_hashes(tdb, lockstart, locksize, F_WRLCK);
 860 }
 861
 862 enum TDB_ERROR tdb_chainlock_read(struct tdb_context *tdb, TDB_DATA key)
 863 {
 864         return tdb->last_error = chainlock(tdb, &key, F_RDLCK, TDB_LOCK_WAIT,
 865                                            "tdb_chainlock_read");
 866 }
 867
 868 void tdb_chainunlock_read(struct tdb_context *tdb, TDB_DATA key)
 869 {
 870         uint64_t h = tdb_hash(tdb, key.dptr, key.dsize);
 871         tdb_off_t lockstart, locksize;
 872         unsigned int group, gbits;
 873
 874         gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS;
 875         group = bits_from(h, 64 - gbits, gbits);
 876
 877         lockstart = hlock_range(group, &locksize);
 878
 879         tdb_trace_1rec(tdb, "tdb_chainunlock_read", key);
 880         tdb_unlock_hashes(tdb, lockstart, locksize, F_RDLCK);
 881 }