git.ozlabs.org Git - ccan/blob - ccan/tdb2/free.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/ilog/ilog.h>
  21 #include <time.h>
  22 #include <assert.h>
  23 #include <limits.h>
  24
  25 static unsigned fls64(uint64_t val)
  26 {
  27         return ilog64(val);
  28 }
  29
  30 /* In which bucket would we find a particular record size? (ignoring header) */
  31 unsigned int size_to_bucket(unsigned int zone_bits, tdb_len_t data_len)
  32 {
  33         unsigned int bucket;
  34
  35         /* We can't have records smaller than this. */
  36         assert(data_len >= TDB_MIN_DATA_LEN);
  37
  38         /* Ignoring the header... */
  39         if (data_len - TDB_MIN_DATA_LEN <= 64) {
  40                 /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 8. */
  41                 bucket = (data_len - TDB_MIN_DATA_LEN) / 8;
  42         } else {
  43                 /* After that we go power of 2. */
  44                 bucket = fls64(data_len - TDB_MIN_DATA_LEN) + 2;
  45         }
  46
  47         if (unlikely(bucket > BUCKETS_FOR_ZONE(zone_bits)))
  48                 bucket = BUCKETS_FOR_ZONE(zone_bits);
  49         return bucket;
  50 }
  51
  52 /* Subtract 1-byte tailer and header.  Then round up to next power of 2. */
  53 static unsigned max_zone_bits(struct tdb_context *tdb)
  54 {
  55         return fls64(tdb->map_size-1-sizeof(struct tdb_header)-1) + 1;
  56 }
  57
  58 /* Start by using a random zone to spread the load: returns the offset. */
  59 static uint64_t random_zone(struct tdb_context *tdb)
  60 {
  61         struct free_zone_header zhdr;
  62         tdb_off_t off = sizeof(struct tdb_header);
  63         tdb_len_t half_bits;
  64         uint64_t randbits = 0;
  65         unsigned int i;
  66
  67         for (i = 0; i < 64; i += fls64(RAND_MAX))
  68                 randbits ^= ((uint64_t)random()) << i;
  69
  70         /* FIXME: Does this work?  Test! */
  71         half_bits = max_zone_bits(tdb) - 1;
  72         do {
  73                 /* Pick left or right side (not outside file) */
  74                 if ((randbits & 1)
  75                     && !tdb->methods->oob(tdb, off + (1ULL << half_bits)
  76                                           + sizeof(zhdr), true)) {
  77                         off += 1ULL << half_bits;
  78                 }
  79                 randbits >>= 1;
  80
  81                 if (tdb_read_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
  82                         return TDB_OFF_ERR;
  83
  84                 if (zhdr.zone_bits == half_bits)
  85                         return off;
  86
  87                 half_bits--;
  88         } while (half_bits >= INITIAL_ZONE_BITS);
  89
  90         tdb->ecode = TDB_ERR_CORRUPT;
  91         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
  92                  "random_zone: zone at %llu smaller than %u bits?",
  93                  (long long)off, INITIAL_ZONE_BITS);
  94         return TDB_OFF_ERR;
  95 }
  96
  97 int tdb_zone_init(struct tdb_context *tdb)
  98 {
  99         tdb->zone_off = random_zone(tdb);
 100         if (tdb->zone_off == TDB_OFF_ERR)
 101                 return -1;
 102         if (tdb_read_convert(tdb, tdb->zone_off,
 103                              &tdb->zhdr, sizeof(tdb->zhdr)) == -1)
 104                 return -1;
 105         return 0;
 106 }
 107
 108 /* Where's the header, given a zone size of 1 << zone_bits? */
 109 static tdb_off_t zone_off(tdb_off_t off, unsigned int zone_bits)
 110 {
 111         off -= sizeof(struct tdb_header);
 112         return (off & ~((1ULL << zone_bits) - 1)) + sizeof(struct tdb_header);
 113 }
 114
 115 /* Offset of a given bucket. */
 116 /* FIXME: bucket can be "unsigned" everywhere, or even uint8/16. */
 117 tdb_off_t bucket_off(tdb_off_t zone_off, tdb_off_t bucket)
 118 {
 119         return zone_off
 120                 + sizeof(struct free_zone_header)
 121                 + bucket * sizeof(tdb_off_t);
 122 }
 123
 124 /* Returns free_buckets + 1, or list number to search. */
 125 static tdb_off_t find_free_head(struct tdb_context *tdb, tdb_off_t bucket)
 126 {
 127         /* Speculatively search for a non-zero bucket. */
 128         return tdb_find_nonzero_off(tdb, bucket_off(tdb->zone_off, 0),
 129                                     bucket,
 130                                     BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits) + 1);
 131 }
 132
 133 /* Remove from free bucket. */
 134 static int remove_from_list(struct tdb_context *tdb,
 135                             tdb_off_t b_off, tdb_off_t r_off,
 136                             struct tdb_free_record *r)
 137 {
 138         tdb_off_t off;
 139
 140         /* Front of list? */
 141         if (r->prev == 0) {
 142                 off = b_off;
 143         } else {
 144                 off = r->prev + offsetof(struct tdb_free_record, next);
 145         }
 146
 147 #ifdef DEBUG
 148         if (tdb_read_off(tdb, off) != r_off) {
 149                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 150                          "remove_from_list: %llu bad prev in list %llu\n",
 151                          (long long)r_off, (long long)b_off);
 152                 return -1;
 153         }
 154 #endif
 155
 156         /* r->prev->next = r->next */
 157         if (tdb_write_off(tdb, off, r->next)) {
 158                 return -1;
 159         }
 160
 161         if (r->next != 0) {
 162                 off = r->next + offsetof(struct tdb_free_record, prev);
 163                 /* r->next->prev = r->prev */
 164
 165 #ifdef DEBUG
 166                 if (tdb_read_off(tdb, off) != r_off) {
 167                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 168                                  "remove_from_list: %llu bad list %llu\n",
 169                                  (long long)r_off, (long long)b_off);
 170                         return -1;
 171                 }
 172 #endif
 173
 174                 if (tdb_write_off(tdb, off, r->prev)) {
 175                         return -1;
 176                 }
 177         }
 178         return 0;
 179 }
 180
 181 /* Enqueue in this free bucket. */
 182 static int enqueue_in_free(struct tdb_context *tdb,
 183                            tdb_off_t b_off,
 184                            tdb_off_t off,
 185                            struct tdb_free_record *new)
 186 {
 187         new->prev = 0;
 188         /* new->next = head. */
 189         new->next = tdb_read_off(tdb, b_off);
 190         if (new->next == TDB_OFF_ERR)
 191                 return -1;
 192
 193         if (new->next) {
 194 #ifdef DEBUG
 195                 if (tdb_read_off(tdb,
 196                                  new->next
 197                                  + offsetof(struct tdb_free_record, prev))
 198                     != 0) {
 199                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 200                                  "enqueue_in_free: %llu bad head prev %llu\n",
 201                                  (long long)new->next, (long long)b_off);
 202                         return -1;
 203                 }
 204 #endif
 205                 /* next->prev = new. */
 206                 if (tdb_write_off(tdb, new->next
 207                                   + offsetof(struct tdb_free_record, prev),
 208                                   off) != 0)
 209                         return -1;
 210         }
 211         /* head = new */
 212         if (tdb_write_off(tdb, b_off, off) != 0)
 213                 return -1;
 214
 215         return tdb_write_convert(tdb, off, new, sizeof(*new));
 216 }
 217
 218 /* List need not be locked. */
 219 int add_free_record(struct tdb_context *tdb,
 220                     unsigned int zone_bits,
 221                     tdb_off_t off, tdb_len_t len_with_header)
 222 {
 223         struct tdb_free_record new;
 224         tdb_off_t b_off;
 225         int ret;
 226
 227         assert(len_with_header >= sizeof(new));
 228         assert(zone_bits < (1 << 6));
 229
 230         new.magic_and_meta = TDB_FREE_MAGIC | zone_bits;
 231         new.data_len = len_with_header - sizeof(struct tdb_used_record);
 232
 233         b_off = bucket_off(zone_off(off, zone_bits),
 234                            size_to_bucket(zone_bits, new.data_len));
 235         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
 236                 return -1;
 237
 238         ret = enqueue_in_free(tdb, b_off, off, &new);
 239         tdb_unlock_free_bucket(tdb, b_off);
 240         return ret;
 241 }
 242
 243 static size_t adjust_size(size_t keylen, size_t datalen, bool want_extra)
 244 {
 245         size_t size = keylen + datalen;
 246
 247         /* We want at least 50% growth for data. */
 248         if (want_extra)
 249                 size += datalen/2;
 250
 251         if (size < TDB_MIN_DATA_LEN)
 252                 size = TDB_MIN_DATA_LEN;
 253
 254         /* Round to next uint64_t boundary. */
 255         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
 256 }
 257
 258 /* If we have enough left over to be useful, split that off. */
 259 static int to_used_record(struct tdb_context *tdb,
 260                           unsigned int zone_bits,
 261                           tdb_off_t off,
 262                           tdb_len_t needed,
 263                           tdb_len_t total_len,
 264                           tdb_len_t *actual)
 265 {
 266         struct tdb_used_record used;
 267         tdb_len_t leftover;
 268
 269         leftover = total_len - needed;
 270         if (leftover < sizeof(struct tdb_free_record))
 271                 leftover = 0;
 272
 273         *actual = total_len - leftover;
 274
 275         if (leftover) {
 276                 if (add_free_record(tdb, zone_bits,
 277                                     off + sizeof(used) + *actual,
 278                                     total_len - needed))
 279                         return -1;
 280         }
 281         return 0;
 282 }
 283
 284 /* Note: we unlock the current bucket if we coalesce or fail. */
 285 static int coalesce(struct tdb_context *tdb,
 286                     tdb_off_t zone_off, unsigned zone_bits,
 287                     tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
 288 {
 289         struct tdb_free_record pad, *r;
 290         tdb_off_t end = off + sizeof(struct tdb_used_record) + data_len;
 291
 292         while (end < (zone_off + (1ULL << zone_bits))) {
 293                 tdb_off_t nb_off;
 294
 295                 /* FIXME: do tdb_get here and below really win? */
 296                 r = tdb_get(tdb, end, &pad, sizeof(pad));
 297                 if (!r)
 298                         goto err;
 299
 300                 if (frec_magic(r) != TDB_FREE_MAGIC)
 301                         break;
 302
 303                 nb_off = bucket_off(zone_off,
 304                                     size_to_bucket(zone_bits, r->data_len));
 305
 306                 /* We may be violating lock order here, so best effort. */
 307                 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1)
 308                         break;
 309
 310                 /* Now we have lock, re-check. */
 311                 r = tdb_get(tdb, end, &pad, sizeof(pad));
 312                 if (!r) {
 313                         tdb_unlock_free_bucket(tdb, nb_off);
 314                         goto err;
 315                 }
 316
 317                 if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
 318                         tdb_unlock_free_bucket(tdb, nb_off);
 319                         break;
 320                 }
 321
 322                 if (remove_from_list(tdb, nb_off, end, r) == -1) {
 323                         tdb_unlock_free_bucket(tdb, nb_off);
 324                         goto err;
 325                 }
 326
 327                 end += sizeof(struct tdb_used_record) + r->data_len;
 328                 tdb_unlock_free_bucket(tdb, nb_off);
 329         }
 330
 331         /* Didn't find any adjacent free? */
 332         if (end == off + sizeof(struct tdb_used_record) + data_len)
 333                 return 0;
 334
 335         /* OK, expand record */
 336         r = tdb_get(tdb, off, &pad, sizeof(pad));
 337         if (!r)
 338                 goto err;
 339
 340         if (r->data_len != data_len) {
 341                 tdb->ecode = TDB_ERR_CORRUPT;
 342                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 343                          "coalesce: expected data len %llu not %llu\n",
 344                          (long long)data_len, (long long)r->data_len);
 345                 goto err;
 346         }
 347
 348         if (remove_from_list(tdb, b_off, off, r) == -1)
 349                 goto err;
 350
 351         /* We have to drop this to avoid deadlocks. */
 352         tdb_unlock_free_bucket(tdb, b_off);
 353
 354         if (add_free_record(tdb, zone_bits, off, end - off) == -1)
 355                 return -1;
 356         return 1;
 357
 358 err:
 359         /* To unify error paths, we *always* unlock bucket on error. */
 360         tdb_unlock_free_bucket(tdb, b_off);
 361         return -1;
 362 }
 363
 364 /* We need size bytes to put our key and data in. */
 365 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 366                                 tdb_off_t zone_off,
 367                                 unsigned zone_bits,
 368                                 tdb_off_t bucket,
 369                                 size_t size,
 370                                 tdb_len_t *actual)
 371 {
 372         tdb_off_t off, b_off,best_off;
 373         struct tdb_free_record pad, best = { 0 }, *r;
 374         double multiplier;
 375
 376 again:
 377         b_off = bucket_off(zone_off, bucket);
 378
 379         /* FIXME: Try non-blocking wait first, to measure contention.
 380          * If we're contented, try switching zones, and don't enlarge zone
 381          * next time (we want more zones). */
 382         /* Lock this bucket. */
 383         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
 384                 return TDB_OFF_ERR;
 385         }
 386
 387         best.data_len = -1ULL;
 388         best_off = 0;
 389         /* FIXME: Start with larger multiplier if we're growing. */
 390         multiplier = 1.0;
 391
 392         /* Walk the list to see if any are large enough, getting less fussy
 393          * as we go. */
 394         off = tdb_read_off(tdb, b_off);
 395         if (unlikely(off == TDB_OFF_ERR))
 396                 goto unlock_err;
 397
 398         while (off) {
 399                 /* FIXME: Does tdb_get win anything here? */
 400                 r = tdb_get(tdb, off, &pad, sizeof(*r));
 401                 if (!r)
 402                         goto unlock_err;
 403
 404                 if (frec_magic(r) != TDB_FREE_MAGIC) {
 405                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 406                                  "lock_and_alloc: %llu non-free 0x%llx\n",
 407                                  (long long)off, (long long)r->magic_and_meta);
 408                         goto unlock_err;
 409                 }
 410
 411                 if (r->data_len >= size && r->data_len < best.data_len) {
 412                         best_off = off;
 413                         best = *r;
 414                 }
 415
 416                 if (best.data_len < size * multiplier && best_off)
 417                         goto use_best;
 418
 419                 multiplier *= 1.01;
 420
 421                 /* Since we're going slow anyway, try coalescing here. */
 422                 switch (coalesce(tdb, zone_off, zone_bits, off, b_off,
 423                                  r->data_len)) {
 424                 case -1:
 425                         /* This has already unlocked on error. */
 426                         return -1;
 427                 case 1:
 428                         /* This has unlocked list, restart. */
 429                         goto again;
 430                 }
 431                 off = r->next;
 432         }
 433
 434         /* If we found anything at all, use it. */
 435         if (best_off) {
 436         use_best:
 437                 /* We're happy with this size: take it. */
 438                 if (remove_from_list(tdb, b_off, best_off, &best) != 0)
 439                         goto unlock_err;
 440                 tdb_unlock_free_bucket(tdb, b_off);
 441
 442                 if (to_used_record(tdb, zone_bits, best_off, size,
 443                                    best.data_len, actual)) {
 444                         return -1;
 445                 }
 446                 return best_off;
 447         }
 448
 449         tdb_unlock_free_bucket(tdb, b_off);
 450         return 0;
 451
 452 unlock_err:
 453         tdb_unlock_free_bucket(tdb, b_off);
 454         return TDB_OFF_ERR;
 455 }
 456
 457 static bool next_zone(struct tdb_context *tdb)
 458 {
 459         tdb_off_t next = tdb->zone_off + (1ULL << tdb->zhdr.zone_bits);
 460
 461         /* We must have a header. */
 462         if (tdb->methods->oob(tdb, next + sizeof(tdb->zhdr), true))
 463                 return false;
 464
 465         tdb->zone_off = next;
 466         return tdb_read_convert(tdb, next, &tdb->zhdr, sizeof(tdb->zhdr)) == 0;
 467 }
 468
 469 /* Offset returned is within current zone (which it may alter). */
 470 static tdb_off_t get_free(struct tdb_context *tdb, size_t size,
 471                           tdb_len_t *actual)
 472 {
 473         tdb_off_t start_zone = tdb->zone_off, off;
 474         bool wrapped = false;
 475
 476         /* FIXME: If we don't get a hit in the first bucket we want,
 477          * try changing zones for next time.  That should help wear
 478          * zones evenly, so we don't need to search all of them before
 479          * expanding. */
 480         while (!wrapped || tdb->zone_off != start_zone) {
 481                 tdb_off_t b;
 482
 483                 /* Shortcut for really huge allocations... */
 484                 if ((size >> tdb->zhdr.zone_bits) != 0)
 485                         continue;
 486
 487                 /* Start at exact size bucket, and search up... */
 488                 b = size_to_bucket(tdb->zhdr.zone_bits, size);
 489                 for (b = find_free_head(tdb, b);
 490                      b <= BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits);
 491                      b += find_free_head(tdb, b + 1)) {
 492                         /* Try getting one from list. */
 493                         off = lock_and_alloc(tdb, tdb->zone_off,
 494                                              tdb->zhdr.zone_bits,
 495                                              b, size, actual);
 496                         if (off == TDB_OFF_ERR)
 497                                 return TDB_OFF_ERR;
 498                         if (off != 0)
 499                                 return off;
 500                         /* Didn't work.  Try next bucket. */
 501                 }
 502
 503                 /* Didn't work, try next zone, if it exists. */
 504                 if (!next_zone(tdb)) {
 505                         wrapped = true;
 506                         tdb->zone_off = sizeof(struct tdb_header);
 507                         if (tdb_read_convert(tdb, tdb->zone_off,
 508                                              &tdb->zhdr, sizeof(tdb->zhdr))) {
 509                                 return TDB_OFF_ERR;
 510                         }
 511                 }
 512         }
 513         return 0;
 514 }
 515
 516 int set_header(struct tdb_context *tdb,
 517                struct tdb_used_record *rec,
 518                uint64_t keylen, uint64_t datalen,
 519                uint64_t actuallen, unsigned hashlow,
 520                unsigned int zone_bits)
 521 {
 522         uint64_t keybits = (fls64(keylen) + 1) / 2;
 523
 524         /* Use bottom bits of hash, so it's independent of hash table size. */
 525         rec->magic_and_meta
 526                 = zone_bits
 527                 | ((hashlow & ((1 << 5)-1)) << 6)
 528                 | ((actuallen - (keylen + datalen)) << 11)
 529                 | (keybits << 43)
 530                 | (TDB_MAGIC << 48);
 531         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
 532
 533         /* Encoding can fail on big values. */
 534         if (rec_key_length(rec) != keylen
 535             || rec_data_length(rec) != datalen
 536             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
 537                 tdb->ecode = TDB_ERR_IO;
 538                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 539                          "Could not encode k=%llu,d=%llu,a=%llu\n",
 540                          (long long)keylen, (long long)datalen,
 541                          (long long)actuallen);
 542                 return -1;
 543         }
 544         return 0;
 545 }
 546
 547 static bool zones_happy(struct tdb_context *tdb)
 548 {
 549         /* FIXME: look at distribution of zones. */
 550         return true;
 551 }
 552
 553 /* Assume we want buckets up to the comfort factor. */
 554 static tdb_len_t overhead(unsigned int zone_bits)
 555 {
 556         return sizeof(struct free_zone_header)
 557                 + (BUCKETS_FOR_ZONE(zone_bits) + 1) * sizeof(tdb_off_t);
 558 }
 559
 560 /* Expand the database (by adding a zone). */
 561 static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 562 {
 563         uint64_t old_size;
 564         tdb_off_t off;
 565         uint8_t zone_bits;
 566         unsigned int num_buckets;
 567         tdb_len_t wanted;
 568         struct free_zone_header zhdr;
 569         bool enlarge_zone;
 570
 571         /* We need room for the record header too. */
 572         wanted = sizeof(struct tdb_used_record) + size;
 573
 574         /* Only one person can expand file at a time. */
 575         if (tdb_lock_expand(tdb, F_WRLCK) != 0)
 576                 return -1;
 577
 578         /* Someone else may have expanded the file, so retry. */
 579         old_size = tdb->map_size;
 580         tdb->methods->oob(tdb, tdb->map_size + 1, true);
 581         if (tdb->map_size != old_size)
 582                 goto success;
 583
 584         /* FIXME: Tailer is a bogus optimization, remove it. */
 585         /* zone bits tailer char is protected by EXPAND lock. */
 586         if (tdb->methods->read(tdb, old_size - 1, &zone_bits, 1) == -1)
 587                 goto fail;
 588
 589         /* If zones aren't working well, add larger zone if possible. */
 590         enlarge_zone = !zones_happy(tdb);
 591
 592         /* New zone can be between zone_bits or larger if we're on the right
 593          * boundary. */
 594         for (;;) {
 595                 /* Does this fit the allocation comfortably? */
 596                 if ((1ULL << zone_bits) >= overhead(zone_bits) + wanted) {
 597                         /* Only let enlarge_zone enlarge us once. */
 598                         if (!enlarge_zone)
 599                                 break;
 600                         enlarge_zone = false;
 601                 }
 602                 if ((old_size - 1 - sizeof(struct tdb_header))
 603                     & (1 << zone_bits))
 604                         break;
 605                 zone_bits++;
 606         }
 607
 608         zhdr.zone_bits = zone_bits;
 609         num_buckets = BUCKETS_FOR_ZONE(zone_bits);
 610
 611         /* FIXME: I don't think we need to expand to full zone, do we? */
 612         if (tdb->methods->expand_file(tdb, 1ULL << zone_bits) == -1)
 613                 goto fail;
 614
 615         /* Write new tailer. */
 616         if (tdb->methods->write(tdb, tdb->map_size - 1, &zone_bits, 1) == -1)
 617                 goto fail;
 618
 619         /* Write new zone header (just before old tailer). */
 620         off = old_size - 1;
 621         if (tdb_write_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
 622                 goto fail;
 623
 624         /* Now write empty buckets. */
 625         off += sizeof(zhdr);
 626         if (zero_out(tdb, off, (num_buckets+1) * sizeof(tdb_off_t)) == -1)
 627                 goto fail;
 628         off += (num_buckets+1) * sizeof(tdb_off_t);
 629
 630         /* Now add the rest as our free record. */
 631         if (add_free_record(tdb, zone_bits, off, tdb->map_size-1-off) == -1)
 632                 goto fail;
 633
 634         /* Try allocating from this zone now. */
 635         tdb->zone_off = old_size - 1;
 636         tdb->zhdr = zhdr;
 637
 638 success:
 639         tdb_unlock_expand(tdb, F_WRLCK);
 640         return 0;
 641
 642 fail:
 643         tdb_unlock_expand(tdb, F_WRLCK);
 644         return -1;
 645 }
 646
 647 /* This won't fail: it will expand the database if it has to. */
 648 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
 649                 uint64_t hash, bool growing)
 650 {
 651         tdb_off_t off;
 652         tdb_len_t size, actual;
 653         struct tdb_used_record rec;
 654
 655         /* We can't hold pointers during this: we could unmap! */
 656         assert(!tdb->direct_access);
 657
 658         size = adjust_size(keylen, datalen, growing);
 659
 660 again:
 661         off = get_free(tdb, size, &actual);
 662         if (unlikely(off == TDB_OFF_ERR))
 663                 return off;
 664
 665         if (unlikely(off == 0)) {
 666                 if (tdb_expand(tdb, size) == -1)
 667                         return TDB_OFF_ERR;
 668                 goto again;
 669         }
 670
 671         /* Some supergiant values can't be encoded. */
 672         /* FIXME: Check before, and limit actual in get_free. */
 673         if (set_header(tdb, &rec, keylen, datalen, actual, hash,
 674                        tdb->zhdr.zone_bits) != 0) {
 675                 add_free_record(tdb, tdb->zhdr.zone_bits, off,
 676                                 sizeof(rec) + actual);
 677                 return TDB_OFF_ERR;
 678         }
 679
 680         if (tdb_write_convert(tdb, off, &rec, sizeof(rec)) != 0)
 681                 return TDB_OFF_ERR;
 682
 683         return off;
 684 }