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 <time.h>
  21 #include <assert.h>
  22 #include <limits.h>
  23
  24 static unsigned fls64(uint64_t val)
  25 {
  26 #if HAVE_BUILTIN_CLZL
  27         if (val <= ULONG_MAX) {
  28                 /* This is significantly faster! */
  29                 return val ? sizeof(long) * CHAR_BIT - __builtin_clzl(val) : 0;
  30         } else {
  31 #endif
  32         uint64_t r = 64;
  33
  34         if (!val)
  35                 return 0;
  36         if (!(val & 0xffffffff00000000ull)) {
  37                 val <<= 32;
  38                 r -= 32;
  39         }
  40         if (!(val & 0xffff000000000000ull)) {
  41                 val <<= 16;
  42                 r -= 16;
  43         }
  44         if (!(val & 0xff00000000000000ull)) {
  45                 val <<= 8;
  46                 r -= 8;
  47         }
  48         if (!(val & 0xf000000000000000ull)) {
  49                 val <<= 4;
  50                 r -= 4;
  51         }
  52         if (!(val & 0xc000000000000000ull)) {
  53                 val <<= 2;
  54                 r -= 2;
  55         }
  56         if (!(val & 0x8000000000000000ull)) {
  57                 val <<= 1;
  58                 r -= 1;
  59         }
  60         return r;
  61 #if HAVE_BUILTIN_CLZL
  62         }
  63 #endif
  64 }
  65
  66 /* In which bucket would we find a particular record size? (ignoring header) */
  67 unsigned int size_to_bucket(unsigned int zone_bits, tdb_len_t data_len)
  68 {
  69         unsigned int bucket;
  70
  71         /* We can't have records smaller than this. */
  72         assert(data_len >= TDB_MIN_DATA_LEN);
  73
  74         /* Ignoring the header... */
  75         if (data_len - TDB_MIN_DATA_LEN <= 64) {
  76                 /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 8. */
  77                 bucket = (data_len - TDB_MIN_DATA_LEN) / 8;
  78         } else {
  79                 /* After that we go power of 2. */
  80                 bucket = fls64(data_len - TDB_MIN_DATA_LEN) + 2;
  81         }
  82
  83         if (unlikely(bucket > BUCKETS_FOR_ZONE(zone_bits)))
  84                 bucket = BUCKETS_FOR_ZONE(zone_bits);
  85         return bucket;
  86 }
  87
  88 /* Subtract 1-byte tailer and header.  Then round up to next power of 2. */
  89 static unsigned max_zone_bits(struct tdb_context *tdb)
  90 {
  91         return fls64(tdb->map_size-1-sizeof(struct tdb_header)-1) + 1;
  92 }
  93
  94 /* Start by using a random zone to spread the load: returns the offset. */
  95 static uint64_t random_zone(struct tdb_context *tdb)
  96 {
  97         struct free_zone_header zhdr;
  98         tdb_off_t off = sizeof(struct tdb_header);
  99         tdb_len_t half_bits;
 100         uint64_t randbits = 0;
 101         unsigned int i;
 102
 103         for (i = 0; i < 64; i += fls64(RAND_MAX))
 104                 randbits ^= ((uint64_t)random()) << i;
 105
 106         /* FIXME: Does this work?  Test! */
 107         half_bits = max_zone_bits(tdb) - 1;
 108         do {
 109                 /* Pick left or right side (not outside file) */
 110                 if ((randbits & 1)
 111                     && !tdb->methods->oob(tdb, off + (1ULL << half_bits)
 112                                           + sizeof(zhdr), true)) {
 113                         off += 1ULL << half_bits;
 114                 }
 115                 randbits >>= 1;
 116
 117                 if (tdb_read_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
 118                         return TDB_OFF_ERR;
 119
 120                 if (zhdr.zone_bits == half_bits)
 121                         return off;
 122
 123                 half_bits--;
 124         } while (half_bits >= INITIAL_ZONE_BITS);
 125
 126         tdb->ecode = TDB_ERR_CORRUPT;
 127         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 128                  "random_zone: zone at %llu smaller than %u bits?",
 129                  (long long)off, INITIAL_ZONE_BITS);
 130         return TDB_OFF_ERR;
 131 }
 132
 133 int tdb_zone_init(struct tdb_context *tdb)
 134 {
 135         tdb->zone_off = random_zone(tdb);
 136         if (tdb->zone_off == TDB_OFF_ERR)
 137                 return -1;
 138         if (tdb_read_convert(tdb, tdb->zone_off,
 139                              &tdb->zhdr, sizeof(tdb->zhdr)) == -1)
 140                 return -1;
 141         return 0;
 142 }
 143
 144 /* Where's the header, given a zone size of 1 << zone_bits? */
 145 static tdb_off_t zone_off(tdb_off_t off, unsigned int zone_bits)
 146 {
 147         off -= sizeof(struct tdb_header);
 148         return (off & ~((1ULL << zone_bits) - 1)) + sizeof(struct tdb_header);
 149 }
 150
 151 /* Offset of a given bucket. */
 152 /* FIXME: bucket can be "unsigned" everywhere, or even uint8/16. */
 153 tdb_off_t bucket_off(tdb_off_t zone_off, tdb_off_t bucket)
 154 {
 155         return zone_off
 156                 + sizeof(struct free_zone_header)
 157                 + bucket * sizeof(tdb_off_t);
 158 }
 159
 160 /* Returns free_buckets + 1, or list number to search. */
 161 static tdb_off_t find_free_head(struct tdb_context *tdb, tdb_off_t bucket)
 162 {
 163         /* Speculatively search for a non-zero bucket. */
 164         return tdb_find_nonzero_off(tdb, bucket_off(tdb->zone_off, 0),
 165                                     bucket,
 166                                     BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits) + 1);
 167 }
 168
 169 /* Remove from free bucket. */
 170 static int remove_from_list(struct tdb_context *tdb,
 171                             tdb_off_t b_off, struct tdb_free_record *r)
 172 {
 173         tdb_off_t off;
 174
 175         /* Front of list? */
 176         if (r->prev == 0) {
 177                 off = b_off;
 178         } else {
 179                 off = r->prev + offsetof(struct tdb_free_record, next);
 180         }
 181         /* r->prev->next = r->next */
 182         if (tdb_write_off(tdb, off, r->next)) {
 183                 return -1;
 184         }
 185
 186         if (r->next != 0) {
 187                 off = r->next + offsetof(struct tdb_free_record, prev);
 188                 /* r->next->prev = r->prev */
 189                 if (tdb_write_off(tdb, off, r->prev)) {
 190                         return -1;
 191                 }
 192         }
 193         return 0;
 194 }
 195
 196 /* Enqueue in this free bucket. */
 197 static int enqueue_in_free(struct tdb_context *tdb,
 198                            tdb_off_t b_off,
 199                            tdb_off_t off,
 200                            struct tdb_free_record *new)
 201 {
 202         new->prev = 0;
 203         /* new->next = head. */
 204         new->next = tdb_read_off(tdb, b_off);
 205         if (new->next == TDB_OFF_ERR)
 206                 return -1;
 207
 208         if (new->next) {
 209                 /* next->prev = new. */
 210                 if (tdb_write_off(tdb, new->next
 211                                   + offsetof(struct tdb_free_record, prev),
 212                                   off) != 0)
 213                         return -1;
 214         }
 215         /* head = new */
 216         if (tdb_write_off(tdb, b_off, off) != 0)
 217                 return -1;
 218
 219         return tdb_write_convert(tdb, off, new, sizeof(*new));
 220 }
 221
 222 /* List need not be locked. */
 223 int add_free_record(struct tdb_context *tdb,
 224                     unsigned int zone_bits,
 225                     tdb_off_t off, tdb_len_t len_with_header)
 226 {
 227         struct tdb_free_record new;
 228         tdb_off_t b_off;
 229         int ret;
 230
 231         assert(len_with_header >= sizeof(new));
 232         assert(zone_bits < (1 << 6));
 233
 234         new.magic_and_meta = TDB_FREE_MAGIC | zone_bits;
 235         new.data_len = len_with_header - sizeof(struct tdb_used_record);
 236
 237         b_off = bucket_off(zone_off(off, zone_bits),
 238                            size_to_bucket(zone_bits, new.data_len));
 239         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
 240                 return -1;
 241
 242         ret = enqueue_in_free(tdb, b_off, off, &new);
 243         tdb_unlock_free_bucket(tdb, b_off);
 244         return ret;
 245 }
 246
 247 /* If we have enough left over to be useful, split that off. */
 248 static int to_used_record(struct tdb_context *tdb,
 249                           unsigned int zone_bits,
 250                           tdb_off_t off,
 251                           tdb_len_t needed,
 252                           tdb_len_t total_len,
 253                           tdb_len_t *actual)
 254 {
 255         struct tdb_used_record used;
 256         tdb_len_t leftover;
 257
 258         leftover = total_len - needed;
 259         if (leftover < sizeof(struct tdb_free_record))
 260                 leftover = 0;
 261
 262         *actual = total_len - leftover;
 263
 264         if (leftover) {
 265                 if (add_free_record(tdb, zone_bits,
 266                                     off + sizeof(used) + *actual,
 267                                     total_len - needed))
 268                         return -1;
 269         }
 270         return 0;
 271 }
 272
 273 /* Note: we unlock the current bucket if we coalesce or fail. */
 274 static int coalesce(struct tdb_context *tdb,
 275                     tdb_off_t zone_off, unsigned zone_bits,
 276                     tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
 277 {
 278         struct tdb_free_record pad, *r;
 279         tdb_off_t end = off + sizeof(struct tdb_used_record) + data_len;
 280
 281         while (end < (zone_off + (1ULL << zone_bits))) {
 282                 tdb_off_t nb_off;
 283
 284                 /* FIXME: do tdb_get here and below really win? */
 285                 r = tdb_get(tdb, end, &pad, sizeof(pad));
 286                 if (!r)
 287                         goto err;
 288
 289                 if (frec_magic(r) != TDB_FREE_MAGIC)
 290                         break;
 291
 292                 nb_off = bucket_off(zone_off,
 293                                     size_to_bucket(zone_bits, r->data_len));
 294
 295                 /* We may be violating lock order here, so best effort. */
 296                 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1)
 297                         break;
 298
 299                 /* Now we have lock, re-check. */
 300                 r = tdb_get(tdb, end, &pad, sizeof(pad));
 301                 if (!r) {
 302                         tdb_unlock_free_bucket(tdb, nb_off);
 303                         goto err;
 304                 }
 305
 306                 if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
 307                         tdb_unlock_free_bucket(tdb, nb_off);
 308                         break;
 309                 }
 310
 311                 if (remove_from_list(tdb, nb_off, r) == -1) {
 312                         tdb_unlock_free_bucket(tdb, nb_off);
 313                         goto err;
 314                 }
 315
 316                 end += sizeof(struct tdb_used_record) + r->data_len;
 317                 tdb_unlock_free_bucket(tdb, nb_off);
 318         }
 319
 320         /* Didn't find any adjacent free? */
 321         if (end == off + sizeof(struct tdb_used_record) + data_len)
 322                 return 0;
 323
 324         /* OK, expand record */
 325         r = tdb_get(tdb, off, &pad, sizeof(pad));
 326         if (!r)
 327                 goto err;
 328
 329         if (r->data_len != data_len) {
 330                 tdb->ecode = TDB_ERR_CORRUPT;
 331                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 332                          "coalesce: expected data len %llu not %llu\n",
 333                          (long long)data_len, (long long)r->data_len);
 334                 goto err;
 335         }
 336
 337         if (remove_from_list(tdb, b_off, r) == -1)
 338                 goto err;
 339
 340         /* We have to drop this to avoid deadlocks. */
 341         tdb_unlock_free_bucket(tdb, b_off);
 342
 343         if (add_free_record(tdb, zone_bits, off, end - off) == -1)
 344                 return -1;
 345         return 1;
 346
 347 err:
 348         /* To unify error paths, we *always* unlock bucket on error. */
 349         tdb_unlock_free_bucket(tdb, b_off);
 350         return -1;
 351 }
 352
 353 /* We need size bytes to put our key and data in. */
 354 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 355                                 tdb_off_t zone_off,
 356                                 unsigned zone_bits,
 357                                 tdb_off_t bucket,
 358                                 size_t size,
 359                                 tdb_len_t *actual)
 360 {
 361         tdb_off_t off, b_off,best_off;
 362         struct tdb_free_record pad, best = { 0 }, *r;
 363         double multiplier;
 364
 365 again:
 366         b_off = bucket_off(zone_off, bucket);
 367
 368         /* FIXME: Try non-blocking wait first, to measure contention.
 369          * If we're contented, try switching zones, and don't enlarge zone
 370          * next time (we want more zones). */
 371         /* Lock this bucket. */
 372         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
 373                 return TDB_OFF_ERR;
 374         }
 375
 376         best.data_len = -1ULL;
 377         best_off = 0;
 378         /* FIXME: Start with larger multiplier if we're growing. */
 379         multiplier = 1.0;
 380
 381         /* Walk the list to see if any are large enough, getting less fussy
 382          * as we go. */
 383         off = tdb_read_off(tdb, b_off);
 384         if (unlikely(off == TDB_OFF_ERR))
 385                 goto unlock_err;
 386
 387         while (off) {
 388                 /* FIXME: Does tdb_get win anything here? */
 389                 r = tdb_get(tdb, off, &pad, sizeof(*r));
 390                 if (!r)
 391                         goto unlock_err;
 392
 393                 if (frec_magic(r) != TDB_FREE_MAGIC) {
 394                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 395                                  "lock_and_alloc: %llu non-free 0x%llx\n",
 396                                  (long long)off, (long long)r->magic_and_meta);
 397                         goto unlock_err;
 398                 }
 399
 400                 if (r->data_len >= size && r->data_len < best.data_len) {
 401                         best_off = off;
 402                         best = *r;
 403                 }
 404
 405                 if (best.data_len < size * multiplier && best_off)
 406                         goto use_best;
 407
 408                 multiplier *= 1.01;
 409
 410                 /* Since we're going slow anyway, try coalescing here. */
 411                 switch (coalesce(tdb, zone_off, zone_bits, off, b_off,
 412                                  r->data_len)) {
 413                 case -1:
 414                         /* This has already unlocked on error. */
 415                         return -1;
 416                 case 1:
 417                         /* This has unlocked list, restart. */
 418                         goto again;
 419                 }
 420                 off = r->next;
 421         }
 422
 423         /* If we found anything at all, use it. */
 424         if (best_off) {
 425         use_best:
 426                 /* We're happy with this size: take it. */
 427                 if (remove_from_list(tdb, b_off, &best) != 0)
 428                         goto unlock_err;
 429                 tdb_unlock_free_bucket(tdb, b_off);
 430
 431                 if (to_used_record(tdb, zone_bits, best_off, size,
 432                                    best.data_len, actual)) {
 433                         return -1;
 434                 }
 435                 return best_off;
 436         }
 437
 438         tdb_unlock_free_bucket(tdb, b_off);
 439         return 0;
 440
 441 unlock_err:
 442         tdb_unlock_free_bucket(tdb, b_off);
 443         return TDB_OFF_ERR;
 444 }
 445
 446 static bool next_zone(struct tdb_context *tdb)
 447 {
 448         tdb_off_t next = tdb->zone_off + (1ULL << tdb->zhdr.zone_bits);
 449
 450         /* We must have a header. */
 451         if (tdb->methods->oob(tdb, next + sizeof(tdb->zhdr), true))
 452                 return false;
 453
 454         tdb->zone_off = next;
 455         return tdb_read_convert(tdb, next, &tdb->zhdr, sizeof(tdb->zhdr)) == 0;
 456 }
 457
 458 /* Offset returned is within current zone (which it may alter). */
 459 static tdb_off_t get_free(struct tdb_context *tdb, size_t size,
 460                           tdb_len_t *actual)
 461 {
 462         tdb_off_t start_zone = tdb->zone_off, off;
 463         bool wrapped = false;
 464
 465         /* FIXME: If we don't get a hit in the first bucket we want,
 466          * try changing zones for next time.  That should help wear
 467          * zones evenly, so we don't need to search all of them before
 468          * expanding. */
 469         while (!wrapped || tdb->zone_off != start_zone) {
 470                 tdb_off_t b;
 471
 472                 /* Shortcut for really huge allocations... */
 473                 if ((size >> tdb->zhdr.zone_bits) != 0)
 474                         continue;
 475
 476                 /* Start at exact size bucket, and search up... */
 477                 b = size_to_bucket(tdb->zhdr.zone_bits, size);
 478                 for (b = find_free_head(tdb, b);
 479                      b <= BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits);
 480                      b += find_free_head(tdb, b + 1)) {
 481                         /* Try getting one from list. */
 482                         off = lock_and_alloc(tdb, tdb->zone_off,
 483                                              tdb->zhdr.zone_bits,
 484                                              b, size, actual);
 485                         if (off == TDB_OFF_ERR)
 486                                 return TDB_OFF_ERR;
 487                         if (off != 0)
 488                                 return off;
 489                         /* Didn't work.  Try next bucket. */
 490                 }
 491
 492                 /* Didn't work, try next zone, if it exists. */
 493                 if (!next_zone(tdb)) {
 494                         wrapped = true;
 495                         tdb->zone_off = sizeof(struct tdb_header);
 496                         if (tdb_read_convert(tdb, tdb->zone_off,
 497                                              &tdb->zhdr, sizeof(tdb->zhdr))) {
 498                                 return TDB_OFF_ERR;
 499                         }
 500                 }
 501         }
 502         return 0;
 503 }
 504
 505 int set_header(struct tdb_context *tdb,
 506                struct tdb_used_record *rec,
 507                uint64_t keylen, uint64_t datalen,
 508                uint64_t actuallen, uint64_t hash,
 509                unsigned int zone_bits)
 510 {
 511         uint64_t keybits = (fls64(keylen) + 1) / 2;
 512
 513         /* Use bottom bits of hash, so it's independent of hash table size. */
 514         rec->magic_and_meta
 515                 = zone_bits
 516                 | ((hash & ((1 << 5)-1)) << 6)
 517                 | ((actuallen - (keylen + datalen)) << 11)
 518                 | (keybits << 43)
 519                 | (TDB_MAGIC << 48);
 520         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
 521
 522         /* Encoding can fail on big values. */
 523         if (rec_key_length(rec) != keylen
 524             || rec_data_length(rec) != datalen
 525             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
 526                 tdb->ecode = TDB_ERR_IO;
 527                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 528                          "Could not encode k=%llu,d=%llu,a=%llu\n",
 529                          (long long)keylen, (long long)datalen,
 530                          (long long)actuallen);
 531                 return -1;
 532         }
 533         return 0;
 534 }
 535
 536 static bool zones_happy(struct tdb_context *tdb)
 537 {
 538         /* FIXME: look at distribution of zones. */
 539         return true;
 540 }
 541
 542 /* Assume we want buckets up to the comfort factor. */
 543 static tdb_len_t overhead(unsigned int zone_bits)
 544 {
 545         return sizeof(struct free_zone_header)
 546                 + (BUCKETS_FOR_ZONE(zone_bits) + 1) * sizeof(tdb_off_t);
 547 }
 548
 549 /* Expand the database (by adding a zone). */
 550 static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 551 {
 552         uint64_t old_size;
 553         tdb_off_t off;
 554         uint8_t zone_bits;
 555         unsigned int num_buckets;
 556         tdb_len_t wanted;
 557         struct free_zone_header zhdr;
 558         bool enlarge_zone;
 559
 560         /* We need room for the record header too. */
 561         wanted = sizeof(struct tdb_used_record) + size;
 562
 563         /* Only one person can expand file at a time. */
 564         if (tdb_lock_expand(tdb, F_WRLCK) != 0)
 565                 return -1;
 566
 567         /* Someone else may have expanded the file, so retry. */
 568         old_size = tdb->map_size;
 569         tdb->methods->oob(tdb, tdb->map_size + 1, true);
 570         if (tdb->map_size != old_size)
 571                 goto success;
 572
 573         /* FIXME: Tailer is a bogus optimization, remove it. */
 574         /* zone bits tailer char is protected by EXPAND lock. */
 575         if (tdb->methods->read(tdb, old_size - 1, &zone_bits, 1) == -1)
 576                 goto fail;
 577
 578         /* If zones aren't working well, add larger zone if possible. */
 579         enlarge_zone = !zones_happy(tdb);
 580
 581         /* New zone can be between zone_bits or larger if we're on the right
 582          * boundary. */
 583         for (;;) {
 584                 /* Does this fit the allocation comfortably? */
 585                 if ((1ULL << zone_bits) >= overhead(zone_bits) + wanted) {
 586                         /* Only let enlarge_zone enlarge us once. */
 587                         if (!enlarge_zone)
 588                                 break;
 589                         enlarge_zone = false;
 590                 }
 591                 if ((old_size - 1 - sizeof(struct tdb_header))
 592                     & (1 << zone_bits))
 593                         break;
 594                 zone_bits++;
 595         }
 596
 597         zhdr.zone_bits = zone_bits;
 598         num_buckets = BUCKETS_FOR_ZONE(zone_bits);
 599
 600         /* FIXME: I don't think we need to expand to full zone, do we? */
 601         if (tdb->methods->expand_file(tdb, 1ULL << zone_bits) == -1)
 602                 goto fail;
 603
 604         /* Write new tailer. */
 605         if (tdb->methods->write(tdb, tdb->map_size - 1, &zone_bits, 1) == -1)
 606                 goto fail;
 607
 608         /* Write new zone header (just before old tailer). */
 609         off = old_size - 1;
 610         if (tdb_write_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
 611                 goto fail;
 612
 613         /* Now write empty buckets. */
 614         off += sizeof(zhdr);
 615         if (zero_out(tdb, off, (num_buckets+1) * sizeof(tdb_off_t)) == -1)
 616                 goto fail;
 617         off += (num_buckets+1) * sizeof(tdb_off_t);
 618
 619         /* Now add the rest as our free record. */
 620         if (add_free_record(tdb, zone_bits, off, tdb->map_size-1-off) == -1)
 621                 goto fail;
 622
 623         /* Try allocating from this zone now. */
 624         tdb->zone_off = old_size - 1;
 625         tdb->zhdr = zhdr;
 626
 627 success:
 628         tdb_unlock_expand(tdb, F_WRLCK);
 629         return 0;
 630
 631 fail:
 632         tdb_unlock_expand(tdb, F_WRLCK);
 633         return -1;
 634 }
 635
 636 static tdb_len_t adjust_size(size_t keylen, size_t datalen, bool growing)
 637 {
 638         tdb_len_t size = keylen + datalen;
 639
 640         if (size < TDB_MIN_DATA_LEN)
 641                 size = TDB_MIN_DATA_LEN;
 642
 643         /* Overallocate if this is coming from an enlarging store. */
 644         if (growing)
 645                 size += datalen / 2;
 646
 647         /* Round to next uint64_t boundary. */
 648         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
 649 }
 650
 651 /* This won't fail: it will expand the database if it has to. */
 652 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
 653                 uint64_t hash, bool growing)
 654 {
 655         tdb_off_t off;
 656         tdb_len_t size, actual;
 657         struct tdb_used_record rec;
 658
 659         /* We can't hold pointers during this: we could unmap! */
 660         assert(!tdb->direct_access);
 661
 662         size = adjust_size(keylen, datalen, growing);
 663
 664 again:
 665         off = get_free(tdb, size, &actual);
 666         if (unlikely(off == TDB_OFF_ERR))
 667                 return off;
 668
 669         if (unlikely(off == 0)) {
 670                 if (tdb_expand(tdb, size) == -1)
 671                         return TDB_OFF_ERR;
 672                 goto again;
 673         }
 674
 675         /* Some supergiant values can't be encoded. */
 676         /* FIXME: Check before, and limit actual in get_free. */
 677         if (set_header(tdb, &rec, keylen, datalen, actual, hash,
 678                        tdb->zhdr.zone_bits) != 0) {
 679                 add_free_record(tdb, tdb->zhdr.zone_bits, off,
 680                                 sizeof(rec) + actual);
 681                 return TDB_OFF_ERR;
 682         }
 683
 684         if (tdb_write_convert(tdb, off, &rec, sizeof(rec)) != 0)
 685                 return TDB_OFF_ERR;
 686
 687         return off;
 688 }