lib/ntdb/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(ntdb_len_t data_len)
  32 {
  33         unsigned int bucket;
  34
  35         /* We can't have records smaller than this. */
  36         assert(data_len >= NTDB_MIN_DATA_LEN);
  37
  38         /* Ignoring the header... */
  39         if (data_len - NTDB_MIN_DATA_LEN <= 64) {
  40                 /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 8. */
  41                 bucket = (data_len - NTDB_MIN_DATA_LEN) / 8;
  42         } else {
  43                 /* After that we go power of 2. */
  44                 bucket = fls64(data_len - NTDB_MIN_DATA_LEN) + 2;
  45         }
  46
  47         if (unlikely(bucket >= NTDB_FREE_BUCKETS))
  48                 bucket = NTDB_FREE_BUCKETS - 1;
  49         return bucket;
  50 }
  51
  52 ntdb_off_t first_ftable(struct ntdb_context *ntdb)
  53 {
  54         return ntdb_read_off(ntdb, offsetof(struct ntdb_header, free_table));
  55 }
  56
  57 ntdb_off_t next_ftable(struct ntdb_context *ntdb, ntdb_off_t ftable)
  58 {
  59         return ntdb_read_off(ntdb, ftable + offsetof(struct ntdb_freetable,next));
  60 }
  61
  62 enum NTDB_ERROR ntdb_ftable_init(struct ntdb_context *ntdb)
  63 {
  64         /* Use reservoir sampling algorithm to select a free list at random. */
  65         unsigned int rnd, max = 0, count = 0;
  66         ntdb_off_t off;
  67
  68         ntdb->ftable_off = off = first_ftable(ntdb);
  69         ntdb->ftable = 0;
  70
  71         while (off) {
  72                 if (NTDB_OFF_IS_ERR(off)) {
  73                         return NTDB_OFF_TO_ERR(off);
  74                 }
  75
  76                 rnd = random();
  77                 if (rnd >= max) {
  78                         ntdb->ftable_off = off;
  79                         ntdb->ftable = count;
  80                         max = rnd;
  81                 }
  82
  83                 off = next_ftable(ntdb, off);
  84                 count++;
  85         }
  86         return NTDB_SUCCESS;
  87 }
  88
  89 /* Offset of a given bucket. */
  90 ntdb_off_t bucket_off(ntdb_off_t ftable_off, unsigned bucket)
  91 {
  92         return ftable_off + offsetof(struct ntdb_freetable, buckets)
  93                 + bucket * sizeof(ntdb_off_t);
  94 }
  95
  96 /* Returns free_buckets + 1, or list number to search, or -ve error. */
  97 static ntdb_off_t find_free_head(struct ntdb_context *ntdb,
  98                                 ntdb_off_t ftable_off,
  99                                 ntdb_off_t bucket)
 100 {
 101         /* Speculatively search for a non-zero bucket. */
 102         return ntdb_find_nonzero_off(ntdb, bucket_off(ftable_off, 0),
 103                                     bucket, NTDB_FREE_BUCKETS);
 104 }
 105
 106 static void check_list(struct ntdb_context *ntdb, ntdb_off_t b_off)
 107 {
 108 #ifdef CCAN_NTDB_DEBUG
 109         ntdb_off_t off, prev = 0, first;
 110         struct ntdb_free_record r;
 111
 112         first = off = (ntdb_read_off(ntdb, b_off) & NTDB_OFF_MASK);
 113         while (off != 0) {
 114                 ntdb_read_convert(ntdb, off, &r, sizeof(r));
 115                 if (frec_magic(&r) != NTDB_FREE_MAGIC)
 116                         abort();
 117                 if (prev && frec_prev(&r) != prev)
 118                         abort();
 119                 prev = off;
 120                 off = r.next;
 121         }
 122
 123         if (first) {
 124                 ntdb_read_convert(ntdb, first, &r, sizeof(r));
 125                 if (frec_prev(&r) != prev)
 126                         abort();
 127         }
 128 #endif
 129 }
 130
 131 /* Remove from free bucket. */
 132 static enum NTDB_ERROR remove_from_list(struct ntdb_context *ntdb,
 133                                        ntdb_off_t b_off, ntdb_off_t r_off,
 134                                        const struct ntdb_free_record *r)
 135 {
 136         ntdb_off_t off, prev_next, head;
 137         enum NTDB_ERROR ecode;
 138
 139         /* Is this only element in list?  Zero out bucket, and we're done. */
 140         if (frec_prev(r) == r_off)
 141                 return ntdb_write_off(ntdb, b_off, 0);
 142
 143         /* off = &r->prev->next */
 144         off = frec_prev(r) + offsetof(struct ntdb_free_record, next);
 145
 146         /* Get prev->next */
 147         prev_next = ntdb_read_off(ntdb, off);
 148         if (NTDB_OFF_IS_ERR(prev_next))
 149                 return NTDB_OFF_TO_ERR(prev_next);
 150
 151         /* If prev->next == 0, we were head: update bucket to point to next. */
 152         if (prev_next == 0) {
 153                 /* We must preserve upper bits. */
 154                 head = ntdb_read_off(ntdb, b_off);
 155                 if (NTDB_OFF_IS_ERR(head))
 156                         return NTDB_OFF_TO_ERR(head);
 157
 158                 if ((head & NTDB_OFF_MASK) != r_off) {
 159                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 160                                           "remove_from_list:"
 161                                           " %llu head %llu on list %llu",
 162                                           (long long)r_off,
 163                                           (long long)head,
 164                                           (long long)b_off);
 165                 }
 166                 head = ((head & ~NTDB_OFF_MASK) | r->next);
 167                 ecode = ntdb_write_off(ntdb, b_off, head);
 168                 if (ecode != NTDB_SUCCESS)
 169                         return ecode;
 170         } else {
 171                 /* r->prev->next = r->next */
 172                 ecode = ntdb_write_off(ntdb, off, r->next);
 173                 if (ecode != NTDB_SUCCESS)
 174                         return ecode;
 175         }
 176
 177         /* If we were the tail, off = &head->prev. */
 178         if (r->next == 0) {
 179                 head = ntdb_read_off(ntdb, b_off);
 180                 if (NTDB_OFF_IS_ERR(head))
 181                         return NTDB_OFF_TO_ERR(head);
 182                 head &= NTDB_OFF_MASK;
 183                 off = head + offsetof(struct ntdb_free_record, magic_and_prev);
 184         } else {
 185                 /* off = &r->next->prev */
 186                 off = r->next + offsetof(struct ntdb_free_record,
 187                                          magic_and_prev);
 188         }
 189
 190 #ifdef CCAN_NTDB_DEBUG
 191         /* *off == r */
 192         if ((ntdb_read_off(ntdb, off) & NTDB_OFF_MASK) != r_off) {
 193                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 194                                   "remove_from_list:"
 195                                   " %llu bad prev in list %llu",
 196                                   (long long)r_off, (long long)b_off);
 197         }
 198 #endif
 199         /* r->next->prev = r->prev */
 200         return ntdb_write_off(ntdb, off, r->magic_and_prev);
 201 }
 202
 203 /* Enqueue in this free bucket: sets coalesce if we've added 128
 204  * entries to it. */
 205 static enum NTDB_ERROR enqueue_in_free(struct ntdb_context *ntdb,
 206                                       ntdb_off_t b_off,
 207                                       ntdb_off_t off,
 208                                       ntdb_len_t len,
 209                                       bool *coalesce)
 210 {
 211         struct ntdb_free_record new;
 212         enum NTDB_ERROR ecode;
 213         ntdb_off_t prev, head;
 214         uint64_t magic = (NTDB_FREE_MAGIC << (64 - NTDB_OFF_UPPER_STEAL));
 215
 216         head = ntdb_read_off(ntdb, b_off);
 217         if (NTDB_OFF_IS_ERR(head))
 218                 return NTDB_OFF_TO_ERR(head);
 219
 220         /* We only need to set ftable_and_len; rest is set in enqueue_in_free */
 221         new.ftable_and_len = ((uint64_t)ntdb->ftable
 222                               << (64 - NTDB_OFF_UPPER_STEAL))
 223                 | len;
 224
 225         /* new->next = head. */
 226         new.next = (head & NTDB_OFF_MASK);
 227
 228         /* First element?  Prev points to ourselves. */
 229         if (!new.next) {
 230                 new.magic_and_prev = (magic | off);
 231         } else {
 232                 /* new->prev = next->prev */
 233                 prev = ntdb_read_off(ntdb,
 234                                     new.next + offsetof(struct ntdb_free_record,
 235                                                         magic_and_prev));
 236                 new.magic_and_prev = prev;
 237                 if (frec_magic(&new) != NTDB_FREE_MAGIC) {
 238                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 239                                           "enqueue_in_free: %llu bad head"
 240                                           " prev %llu",
 241                                           (long long)new.next,
 242                                           (long long)prev);
 243                 }
 244                 /* next->prev = new. */
 245                 ecode = ntdb_write_off(ntdb, new.next
 246                                       + offsetof(struct ntdb_free_record,
 247                                                  magic_and_prev),
 248                                       off | magic);
 249                 if (ecode != NTDB_SUCCESS) {
 250                         return ecode;
 251                 }
 252
 253 #ifdef CCAN_NTDB_DEBUG
 254                 prev = ntdb_read_off(ntdb, frec_prev(&new)
 255                                     + offsetof(struct ntdb_free_record, next));
 256                 if (prev != 0) {
 257                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 258                                           "enqueue_in_free:"
 259                                           " %llu bad tail next ptr %llu",
 260                                           (long long)frec_prev(&new)
 261                                           + offsetof(struct ntdb_free_record,
 262                                                      next),
 263                                           (long long)prev);
 264                 }
 265 #endif
 266         }
 267
 268         /* Update enqueue count, but don't set high bit: see NTDB_OFF_IS_ERR */
 269         if (*coalesce)
 270                 head += (1ULL << (64 - NTDB_OFF_UPPER_STEAL));
 271         head &= ~(NTDB_OFF_MASK | (1ULL << 63));
 272         head |= off;
 273
 274         ecode = ntdb_write_off(ntdb, b_off, head);
 275         if (ecode != NTDB_SUCCESS) {
 276                 return ecode;
 277         }
 278
 279         /* It's time to coalesce if counter wrapped. */
 280         if (*coalesce)
 281                 *coalesce = ((head & ~NTDB_OFF_MASK) == 0);
 282
 283         return ntdb_write_convert(ntdb, off, &new, sizeof(new));
 284 }
 285
 286 static ntdb_off_t ftable_offset(struct ntdb_context *ntdb, unsigned int ftable)
 287 {
 288         ntdb_off_t off;
 289         unsigned int i;
 290
 291         if (likely(ntdb->ftable == ftable))
 292                 return ntdb->ftable_off;
 293
 294         off = first_ftable(ntdb);
 295         for (i = 0; i < ftable; i++) {
 296                 if (NTDB_OFF_IS_ERR(off)) {
 297                         break;
 298                 }
 299                 off = next_ftable(ntdb, off);
 300         }
 301         return off;
 302 }
 303
 304 /* Note: we unlock the current bucket if fail (-ve), or coalesce (+ve) and
 305  * need to blatt the *protect record (which is set to an error). */
 306 static ntdb_len_t coalesce(struct ntdb_context *ntdb,
 307                           ntdb_off_t off, ntdb_off_t b_off,
 308                           ntdb_len_t data_len,
 309                           ntdb_off_t *protect)
 310 {
 311         ntdb_off_t end;
 312         struct ntdb_free_record rec;
 313         enum NTDB_ERROR ecode;
 314
 315         ntdb->stats.alloc_coalesce_tried++;
 316         end = off + sizeof(struct ntdb_used_record) + data_len;
 317
 318         while (end < ntdb->file->map_size) {
 319                 const struct ntdb_free_record *r;
 320                 ntdb_off_t nb_off;
 321                 unsigned ftable, bucket;
 322
 323                 r = ntdb_access_read(ntdb, end, sizeof(*r), true);
 324                 if (NTDB_PTR_IS_ERR(r)) {
 325                         ecode = NTDB_PTR_ERR(r);
 326                         goto err;
 327                 }
 328
 329                 if (frec_magic(r) != NTDB_FREE_MAGIC
 330                     || frec_ftable(r) == NTDB_FTABLE_NONE) {
 331                         ntdb_access_release(ntdb, r);
 332                         break;
 333                 }
 334
 335                 ftable = frec_ftable(r);
 336                 bucket = size_to_bucket(frec_len(r));
 337                 nb_off = ftable_offset(ntdb, ftable);
 338                 if (NTDB_OFF_IS_ERR(nb_off)) {
 339                         ntdb_access_release(ntdb, r);
 340                         ecode = NTDB_OFF_TO_ERR(nb_off);
 341                         goto err;
 342                 }
 343                 nb_off = bucket_off(nb_off, bucket);
 344                 ntdb_access_release(ntdb, r);
 345
 346                 /* We may be violating lock order here, so best effort. */
 347                 if (ntdb_lock_free_bucket(ntdb, nb_off, NTDB_LOCK_NOWAIT)
 348                     != NTDB_SUCCESS) {
 349                         ntdb->stats.alloc_coalesce_lockfail++;
 350                         break;
 351                 }
 352
 353                 /* Now we have lock, re-check. */
 354                 ecode = ntdb_read_convert(ntdb, end, &rec, sizeof(rec));
 355                 if (ecode != NTDB_SUCCESS) {
 356                         ntdb_unlock_free_bucket(ntdb, nb_off);
 357                         goto err;
 358                 }
 359
 360                 if (unlikely(frec_magic(&rec) != NTDB_FREE_MAGIC)) {
 361                         ntdb->stats.alloc_coalesce_race++;
 362                         ntdb_unlock_free_bucket(ntdb, nb_off);
 363                         break;
 364                 }
 365
 366                 if (unlikely(frec_ftable(&rec) != ftable)
 367                     || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
 368                         ntdb->stats.alloc_coalesce_race++;
 369                         ntdb_unlock_free_bucket(ntdb, nb_off);
 370                         break;
 371                 }
 372
 373                 /* Did we just mess up a record you were hoping to use? */
 374                 if (end == *protect) {
 375                         ntdb->stats.alloc_coalesce_iterate_clash++;
 376                         *protect = NTDB_ERR_TO_OFF(NTDB_ERR_NOEXIST);
 377                 }
 378
 379                 ecode = remove_from_list(ntdb, nb_off, end, &rec);
 380                 check_list(ntdb, nb_off);
 381                 if (ecode != NTDB_SUCCESS) {
 382                         ntdb_unlock_free_bucket(ntdb, nb_off);
 383                         goto err;
 384                 }
 385
 386                 end += sizeof(struct ntdb_used_record) + frec_len(&rec);
 387                 ntdb_unlock_free_bucket(ntdb, nb_off);
 388                 ntdb->stats.alloc_coalesce_num_merged++;
 389         }
 390
 391         /* Didn't find any adjacent free? */
 392         if (end == off + sizeof(struct ntdb_used_record) + data_len)
 393                 return 0;
 394
 395         /* Before we expand, check this isn't one you wanted protected? */
 396         if (off == *protect) {
 397                 *protect = NTDB_ERR_TO_OFF(NTDB_ERR_EXISTS);
 398                 ntdb->stats.alloc_coalesce_iterate_clash++;
 399         }
 400
 401         /* OK, expand initial record */
 402         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
 403         if (ecode != NTDB_SUCCESS) {
 404                 goto err;
 405         }
 406
 407         if (frec_len(&rec) != data_len) {
 408                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 409                                    "coalesce: expected data len %zu not %zu",
 410                                    (size_t)data_len, (size_t)frec_len(&rec));
 411                 goto err;
 412         }
 413
 414         ecode = remove_from_list(ntdb, b_off, off, &rec);
 415         check_list(ntdb, b_off);
 416         if (ecode != NTDB_SUCCESS) {
 417                 goto err;
 418         }
 419
 420         /* Try locking violation first.  We don't allow coalesce recursion! */
 421         ecode = add_free_record(ntdb, off, end - off, NTDB_LOCK_NOWAIT, false);
 422         if (ecode != NTDB_SUCCESS) {
 423                 /* Need to drop lock.  Can't rely on anything stable. */
 424                 ntdb->stats.alloc_coalesce_lockfail++;
 425                 *protect = NTDB_ERR_TO_OFF(NTDB_ERR_CORRUPT);
 426
 427                 /* We have to drop this to avoid deadlocks, so make sure record
 428                  * doesn't get coalesced by someone else! */
 429                 rec.ftable_and_len = (NTDB_FTABLE_NONE
 430                                       << (64 - NTDB_OFF_UPPER_STEAL))
 431                         | (end - off - sizeof(struct ntdb_used_record));
 432                 ecode = ntdb_write_off(ntdb,
 433                                       off + offsetof(struct ntdb_free_record,
 434                                                      ftable_and_len),
 435                                       rec.ftable_and_len);
 436                 if (ecode != NTDB_SUCCESS) {
 437                         goto err;
 438                 }
 439
 440                 ntdb_unlock_free_bucket(ntdb, b_off);
 441
 442                 ecode = add_free_record(ntdb, off, end - off, NTDB_LOCK_WAIT,
 443                                         false);
 444                 if (ecode != NTDB_SUCCESS) {
 445                         return NTDB_ERR_TO_OFF(ecode);
 446                 }
 447         } else if (NTDB_OFF_IS_ERR(*protect)) {
 448                 /* For simplicity, we always drop lock if they can't continue */
 449                 ntdb_unlock_free_bucket(ntdb, b_off);
 450         }
 451         ntdb->stats.alloc_coalesce_succeeded++;
 452
 453         /* Return usable length. */
 454         return end - off - sizeof(struct ntdb_used_record);
 455
 456 err:
 457         /* To unify error paths, we *always* unlock bucket on error. */
 458         ntdb_unlock_free_bucket(ntdb, b_off);
 459         return NTDB_ERR_TO_OFF(ecode);
 460 }
 461
 462 /* List is locked: we unlock it. */
 463 static enum NTDB_ERROR coalesce_list(struct ntdb_context *ntdb,
 464                                     ntdb_off_t ftable_off,
 465                                     ntdb_off_t b_off,
 466                                     unsigned int limit)
 467 {
 468         enum NTDB_ERROR ecode;
 469         ntdb_off_t off;
 470
 471         off = ntdb_read_off(ntdb, b_off);
 472         if (NTDB_OFF_IS_ERR(off)) {
 473                 ecode = NTDB_OFF_TO_ERR(off);
 474                 goto unlock_err;
 475         }
 476         /* A little bit of paranoia: counter should be 0. */
 477         off &= NTDB_OFF_MASK;
 478
 479         while (off && limit--) {
 480                 struct ntdb_free_record rec;
 481                 ntdb_len_t coal;
 482                 ntdb_off_t next;
 483
 484                 ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
 485                 if (ecode != NTDB_SUCCESS)
 486                         goto unlock_err;
 487
 488                 next = rec.next;
 489                 coal = coalesce(ntdb, off, b_off, frec_len(&rec), &next);
 490                 if (NTDB_OFF_IS_ERR(coal)) {
 491                         /* This has already unlocked on error. */
 492                         return NTDB_OFF_TO_ERR(coal);
 493                 }
 494                 if (NTDB_OFF_IS_ERR(next)) {
 495                         /* Coalescing had to unlock, so stop. */
 496                         return NTDB_SUCCESS;
 497                 }
 498                 /* Keep going if we're doing well... */
 499                 limit += size_to_bucket(coal / 16 + NTDB_MIN_DATA_LEN);
 500                 off = next;
 501         }
 502
 503         /* Now, move those elements to the tail of the list so we get something
 504          * else next time. */
 505         if (off) {
 506                 struct ntdb_free_record oldhrec, newhrec, oldtrec, newtrec;
 507                 ntdb_off_t oldhoff, oldtoff, newtoff;
 508
 509                 /* The record we were up to is the new head. */
 510                 ecode = ntdb_read_convert(ntdb, off, &newhrec, sizeof(newhrec));
 511                 if (ecode != NTDB_SUCCESS)
 512                         goto unlock_err;
 513
 514                 /* Get the new tail. */
 515                 newtoff = frec_prev(&newhrec);
 516                 ecode = ntdb_read_convert(ntdb, newtoff, &newtrec,
 517                                          sizeof(newtrec));
 518                 if (ecode != NTDB_SUCCESS)
 519                         goto unlock_err;
 520
 521                 /* Get the old head. */
 522                 oldhoff = ntdb_read_off(ntdb, b_off);
 523                 if (NTDB_OFF_IS_ERR(oldhoff)) {
 524                         ecode = NTDB_OFF_TO_ERR(oldhoff);
 525                         goto unlock_err;
 526                 }
 527
 528                 /* This could happen if they all coalesced away. */
 529                 if (oldhoff == off)
 530                         goto out;
 531
 532                 ecode = ntdb_read_convert(ntdb, oldhoff, &oldhrec,
 533                                          sizeof(oldhrec));
 534                 if (ecode != NTDB_SUCCESS)
 535                         goto unlock_err;
 536
 537                 /* Get the old tail. */
 538                 oldtoff = frec_prev(&oldhrec);
 539                 ecode = ntdb_read_convert(ntdb, oldtoff, &oldtrec,
 540                                          sizeof(oldtrec));
 541                 if (ecode != NTDB_SUCCESS)
 542                         goto unlock_err;
 543
 544                 /* Old tail's next points to old head. */
 545                 oldtrec.next = oldhoff;
 546
 547                 /* Old head's prev points to old tail. */
 548                 oldhrec.magic_and_prev
 549                         = (NTDB_FREE_MAGIC << (64 - NTDB_OFF_UPPER_STEAL))
 550                         | oldtoff;
 551
 552                 /* New tail's next is 0. */
 553                 newtrec.next = 0;
 554
 555                 /* Write out the modified versions. */
 556                 ecode = ntdb_write_convert(ntdb, oldtoff, &oldtrec,
 557                                           sizeof(oldtrec));
 558                 if (ecode != NTDB_SUCCESS)
 559                         goto unlock_err;
 560
 561                 ecode = ntdb_write_convert(ntdb, oldhoff, &oldhrec,
 562                                           sizeof(oldhrec));
 563                 if (ecode != NTDB_SUCCESS)
 564                         goto unlock_err;
 565
 566                 ecode = ntdb_write_convert(ntdb, newtoff, &newtrec,
 567                                           sizeof(newtrec));
 568                 if (ecode != NTDB_SUCCESS)
 569                         goto unlock_err;
 570
 571                 /* And finally link in new head. */
 572                 ecode = ntdb_write_off(ntdb, b_off, off);
 573                 if (ecode != NTDB_SUCCESS)
 574                         goto unlock_err;
 575         }
 576 out:
 577         ntdb_unlock_free_bucket(ntdb, b_off);
 578         return NTDB_SUCCESS;
 579
 580 unlock_err:
 581         ntdb_unlock_free_bucket(ntdb, b_off);
 582         return ecode;
 583 }
 584
 585 /* List must not be locked if coalesce_ok is set. */
 586 enum NTDB_ERROR add_free_record(struct ntdb_context *ntdb,
 587                                ntdb_off_t off, ntdb_len_t len_with_header,
 588                                enum ntdb_lock_flags waitflag,
 589                                bool coalesce_ok)
 590 {
 591         ntdb_off_t b_off;
 592         ntdb_len_t len;
 593         enum NTDB_ERROR ecode;
 594
 595         assert(len_with_header >= sizeof(struct ntdb_free_record));
 596
 597         len = len_with_header - sizeof(struct ntdb_used_record);
 598
 599         b_off = bucket_off(ntdb->ftable_off, size_to_bucket(len));
 600         ecode = ntdb_lock_free_bucket(ntdb, b_off, waitflag);
 601         if (ecode != NTDB_SUCCESS) {
 602                 return ecode;
 603         }
 604
 605         ecode = enqueue_in_free(ntdb, b_off, off, len, &coalesce_ok);
 606         check_list(ntdb, b_off);
 607
 608         /* Coalescing unlocks free list. */
 609         if (!ecode && coalesce_ok)
 610                 ecode = coalesce_list(ntdb, ntdb->ftable_off, b_off, 2);
 611         else
 612                 ntdb_unlock_free_bucket(ntdb, b_off);
 613         return ecode;
 614 }
 615
 616 static size_t adjust_size(size_t keylen, size_t datalen)
 617 {
 618         size_t size = keylen + datalen;
 619
 620         if (size < NTDB_MIN_DATA_LEN)
 621                 size = NTDB_MIN_DATA_LEN;
 622
 623         /* Round to next uint64_t boundary. */
 624         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
 625 }
 626
 627 /* If we have enough left over to be useful, split that off. */
 628 static size_t record_leftover(size_t keylen, size_t datalen,
 629                               bool want_extra, size_t total_len)
 630 {
 631         ssize_t leftover;
 632
 633         if (want_extra)
 634                 datalen += datalen / 2;
 635         leftover = total_len - adjust_size(keylen, datalen);
 636
 637         if (leftover < (ssize_t)sizeof(struct ntdb_free_record))
 638                 return 0;
 639
 640         return leftover;
 641 }
 642
 643 /* We need size bytes to put our key and data in. */
 644 static ntdb_off_t lock_and_alloc(struct ntdb_context *ntdb,
 645                                 ntdb_off_t ftable_off,
 646                                 ntdb_off_t bucket,
 647                                 size_t keylen, size_t datalen,
 648                                 bool want_extra,
 649                                 unsigned magic,
 650                                 unsigned hashlow)
 651 {
 652         ntdb_off_t off, b_off,best_off;
 653         struct ntdb_free_record best = { 0 };
 654         double multiplier;
 655         size_t size = adjust_size(keylen, datalen);
 656         enum NTDB_ERROR ecode;
 657
 658         ntdb->stats.allocs++;
 659         b_off = bucket_off(ftable_off, bucket);
 660
 661         /* FIXME: Try non-blocking wait first, to measure contention. */
 662         /* Lock this bucket. */
 663         ecode = ntdb_lock_free_bucket(ntdb, b_off, NTDB_LOCK_WAIT);
 664         if (ecode != NTDB_SUCCESS) {
 665                 return NTDB_ERR_TO_OFF(ecode);
 666         }
 667
 668         best.ftable_and_len = -1ULL;
 669         best_off = 0;
 670
 671         /* Get slack if we're after extra. */
 672         if (want_extra)
 673                 multiplier = 1.5;
 674         else
 675                 multiplier = 1.0;
 676
 677         /* Walk the list to see if any are large enough, getting less fussy
 678          * as we go. */
 679         off = ntdb_read_off(ntdb, b_off);
 680         if (NTDB_OFF_IS_ERR(off)) {
 681                 ecode = NTDB_OFF_TO_ERR(off);
 682                 goto unlock_err;
 683         }
 684         off &= NTDB_OFF_MASK;
 685
 686         while (off) {
 687                 const struct ntdb_free_record *r;
 688                 ntdb_len_t len;
 689                 ntdb_off_t next;
 690
 691                 r = ntdb_access_read(ntdb, off, sizeof(*r), true);
 692                 if (NTDB_PTR_IS_ERR(r)) {
 693                         ecode = NTDB_PTR_ERR(r);
 694                         goto unlock_err;
 695                 }
 696
 697                 if (frec_magic(r) != NTDB_FREE_MAGIC) {
 698                         ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
 699                                            "lock_and_alloc:"
 700                                            " %llu non-free 0x%llx",
 701                                            (long long)off,
 702                                            (long long)r->magic_and_prev);
 703                         ntdb_access_release(ntdb, r);
 704                         goto unlock_err;
 705                 }
 706
 707                 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
 708                         best_off = off;
 709                         best = *r;
 710                 }
 711
 712                 if (frec_len(&best) <= size * multiplier && best_off) {
 713                         ntdb_access_release(ntdb, r);
 714                         break;
 715                 }
 716
 717                 multiplier *= 1.01;
 718
 719                 next = r->next;
 720                 len = frec_len(r);
 721                 ntdb_access_release(ntdb, r);
 722                 off = next;
 723         }
 724
 725         /* If we found anything at all, use it. */
 726         if (best_off) {
 727                 struct ntdb_used_record rec;
 728                 size_t leftover;
 729
 730                 /* We're happy with this size: take it. */
 731                 ecode = remove_from_list(ntdb, b_off, best_off, &best);
 732                 check_list(ntdb, b_off);
 733                 if (ecode != NTDB_SUCCESS) {
 734                         goto unlock_err;
 735                 }
 736
 737                 leftover = record_leftover(keylen, datalen, want_extra,
 738                                            frec_len(&best));
 739
 740                 assert(keylen + datalen + leftover <= frec_len(&best));
 741                 /* We need to mark non-free before we drop lock, otherwise
 742                  * coalesce() could try to merge it! */
 743                 ecode = set_header(ntdb, &rec, magic, keylen, datalen,
 744                                    frec_len(&best) - leftover, hashlow);
 745                 if (ecode != NTDB_SUCCESS) {
 746                         goto unlock_err;
 747                 }
 748
 749                 ecode = ntdb_write_convert(ntdb, best_off, &rec, sizeof(rec));
 750                 if (ecode != NTDB_SUCCESS) {
 751                         goto unlock_err;
 752                 }
 753
 754                 /* For futureproofing, we put a 0 in any unused space. */
 755                 if (rec_extra_padding(&rec)) {
 756                         ecode = ntdb->io->twrite(ntdb, best_off + sizeof(rec)
 757                                                 + keylen + datalen, "", 1);
 758                         if (ecode != NTDB_SUCCESS) {
 759                                 goto unlock_err;
 760                         }
 761                 }
 762
 763                 /* Bucket of leftover will be <= current bucket, so nested
 764                  * locking is allowed. */
 765                 if (leftover) {
 766                         ntdb->stats.alloc_leftover++;
 767                         ecode = add_free_record(ntdb,
 768                                                 best_off + sizeof(rec)
 769                                                 + frec_len(&best) - leftover,
 770                                                 leftover, NTDB_LOCK_WAIT, false);
 771                         if (ecode != NTDB_SUCCESS) {
 772                                 best_off = NTDB_ERR_TO_OFF(ecode);
 773                         }
 774                 }
 775                 ntdb_unlock_free_bucket(ntdb, b_off);
 776
 777                 return best_off;
 778         }
 779
 780         ntdb_unlock_free_bucket(ntdb, b_off);
 781         return 0;
 782
 783 unlock_err:
 784         ntdb_unlock_free_bucket(ntdb, b_off);
 785         return NTDB_ERR_TO_OFF(ecode);
 786 }
 787
 788 /* Get a free block from current free list, or 0 if none, -ve on error. */
 789 static ntdb_off_t get_free(struct ntdb_context *ntdb,
 790                           size_t keylen, size_t datalen, bool want_extra,
 791                           unsigned magic, unsigned hashlow)
 792 {
 793         ntdb_off_t off, ftable_off;
 794         ntdb_off_t start_b, b, ftable;
 795         bool wrapped = false;
 796
 797         /* If they are growing, add 50% to get to higher bucket. */
 798         if (want_extra)
 799                 start_b = size_to_bucket(adjust_size(keylen,
 800                                                      datalen + datalen / 2));
 801         else
 802                 start_b = size_to_bucket(adjust_size(keylen, datalen));
 803
 804         ftable_off = ntdb->ftable_off;
 805         ftable = ntdb->ftable;
 806         while (!wrapped || ftable_off != ntdb->ftable_off) {
 807                 /* Start at exact size bucket, and search up... */
 808                 for (b = find_free_head(ntdb, ftable_off, start_b);
 809                      b < NTDB_FREE_BUCKETS;
 810                      b = find_free_head(ntdb, ftable_off, b + 1)) {
 811                         /* Try getting one from list. */
 812                         off = lock_and_alloc(ntdb, ftable_off,
 813                                              b, keylen, datalen, want_extra,
 814                                              magic, hashlow);
 815                         if (NTDB_OFF_IS_ERR(off))
 816                                 return off;
 817                         if (off != 0) {
 818                                 if (b == start_b)
 819                                         ntdb->stats.alloc_bucket_exact++;
 820                                 if (b == NTDB_FREE_BUCKETS - 1)
 821                                         ntdb->stats.alloc_bucket_max++;
 822                                 /* Worked?  Stay using this list. */
 823                                 ntdb->ftable_off = ftable_off;
 824                                 ntdb->ftable = ftable;
 825                                 return off;
 826                         }
 827                         /* Didn't work.  Try next bucket. */
 828                 }
 829
 830                 if (NTDB_OFF_IS_ERR(b)) {
 831                         return b;
 832                 }
 833
 834                 /* Hmm, try next table. */
 835                 ftable_off = next_ftable(ntdb, ftable_off);
 836                 if (NTDB_OFF_IS_ERR(ftable_off)) {
 837                         return ftable_off;
 838                 }
 839                 ftable++;
 840
 841                 if (ftable_off == 0) {
 842                         wrapped = true;
 843                         ftable_off = first_ftable(ntdb);
 844                         if (NTDB_OFF_IS_ERR(ftable_off)) {
 845                                 return ftable_off;
 846                         }
 847                         ftable = 0;
 848                 }
 849         }
 850
 851         return 0;
 852 }
 853
 854 enum NTDB_ERROR set_header(struct ntdb_context *ntdb,
 855                           struct ntdb_used_record *rec,
 856                           unsigned magic, uint64_t keylen, uint64_t datalen,
 857                           uint64_t actuallen, unsigned hashlow)
 858 {
 859         uint64_t keybits = (fls64(keylen) + 1) / 2;
 860
 861         /* Use bottom bits of hash, so it's independent of hash table size. */
 862         rec->magic_and_meta = (hashlow & ((1 << 11)-1))
 863                 | ((actuallen - (keylen + datalen)) << 11)
 864                 | (keybits << 43)
 865                 | ((uint64_t)magic << 48);
 866         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
 867
 868         /* Encoding can fail on big values. */
 869         if (rec_key_length(rec) != keylen
 870             || rec_data_length(rec) != datalen
 871             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
 872                 return ntdb_logerr(ntdb, NTDB_ERR_IO, NTDB_LOG_ERROR,
 873                                   "Could not encode k=%llu,d=%llu,a=%llu",
 874                                   (long long)keylen, (long long)datalen,
 875                                   (long long)actuallen);
 876         }
 877         return NTDB_SUCCESS;
 878 }
 879
 880 /* You need 'size', this tells you how much you should expand by. */
 881 ntdb_off_t ntdb_expand_adjust(ntdb_off_t map_size, ntdb_off_t size)
 882 {
 883         ntdb_off_t new_size, top_size;
 884
 885         /* limit size in order to avoid using up huge amounts of memory for
 886          * in memory tdbs if an oddball huge record creeps in */
 887         if (size > 100 * 1024) {
 888                 top_size = map_size + size * 2;
 889         } else {
 890                 top_size = map_size + size * 100;
 891         }
 892
 893         /* always make room for at least top_size more records, and at
 894            least 25% more space. if the DB is smaller than 100MiB,
 895            otherwise grow it by 10% only. */
 896         if (map_size > 100 * 1024 * 1024) {
 897                 new_size = map_size * 1.10;
 898         } else {
 899                 new_size = map_size * 1.25;
 900         }
 901
 902         if (new_size < top_size)
 903                 new_size = top_size;
 904
 905         /* We always make the file a multiple of transaction page
 906          * size.  This guarantees that the transaction recovery area
 907          * is always aligned, otherwise the transaction code can overwrite
 908          * itself. */
 909         new_size = (new_size + NTDB_PGSIZE-1) & ~(NTDB_PGSIZE-1);
 910         return new_size - map_size;
 911 }
 912
 913 /* Expand the database. */
 914 static enum NTDB_ERROR ntdb_expand(struct ntdb_context *ntdb, ntdb_len_t size)
 915 {
 916         uint64_t old_size;
 917         ntdb_len_t wanted;
 918         enum NTDB_ERROR ecode;
 919
 920         /* Need to hold a hash lock to expand DB: transactions rely on it. */
 921         if (!(ntdb->flags & NTDB_NOLOCK)
 922             && !ntdb->file->allrecord_lock.count && !ntdb_has_hash_locks(ntdb)) {
 923                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 924                                   "ntdb_expand: must hold lock during expand");
 925         }
 926
 927         /* Only one person can expand file at a time. */
 928         ecode = ntdb_lock_expand(ntdb, F_WRLCK);
 929         if (ecode != NTDB_SUCCESS) {
 930                 return ecode;
 931         }
 932
 933         /* Someone else may have expanded the file, so retry. */
 934         old_size = ntdb->file->map_size;
 935         ntdb_oob(ntdb, ntdb->file->map_size, 1, true);
 936         if (ntdb->file->map_size != old_size) {
 937                 ntdb_unlock_expand(ntdb, F_WRLCK);
 938                 return NTDB_SUCCESS;
 939         }
 940
 941         /* We need room for the record header too. */
 942         size = adjust_size(0, sizeof(struct ntdb_used_record) + size);
 943         /* Overallocate. */
 944         wanted = ntdb_expand_adjust(old_size, size);
 945
 946         ecode = ntdb->io->expand_file(ntdb, wanted);
 947         if (ecode != NTDB_SUCCESS) {
 948                 ntdb_unlock_expand(ntdb, F_WRLCK);
 949                 return ecode;
 950         }
 951
 952         /* We need to drop this lock before adding free record. */
 953         ntdb_unlock_expand(ntdb, F_WRLCK);
 954
 955         ntdb->stats.expands++;
 956         return add_free_record(ntdb, old_size, wanted, NTDB_LOCK_WAIT, true);
 957 }
 958
 959 /* This won't fail: it will expand the database if it has to. */
 960 ntdb_off_t alloc(struct ntdb_context *ntdb, size_t keylen, size_t datalen,
 961                 uint64_t hash, unsigned magic, bool growing)
 962 {
 963         ntdb_off_t off;
 964
 965         /* We can't hold pointers during this: we could unmap! */
 966         assert(!ntdb->direct_access);
 967
 968         for (;;) {
 969                 enum NTDB_ERROR ecode;
 970                 off = get_free(ntdb, keylen, datalen, growing, magic, hash);
 971                 if (likely(off != 0))
 972                         break;
 973
 974                 ecode = ntdb_expand(ntdb, adjust_size(keylen, datalen));
 975                 if (ecode != NTDB_SUCCESS) {
 976                         return NTDB_ERR_TO_OFF(ecode);
 977                 }
 978         }
 979
 980         return off;
 981 }