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(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 >= TDB_FREE_BUCKETS))
  48                 bucket = TDB_FREE_BUCKETS - 1;
  49         return bucket;
  50 }
  51
  52 tdb_off_t first_flist(struct tdb_context *tdb)
  53 {
  54         return tdb_read_off(tdb, offsetof(struct tdb_header, free_list));
  55 }
  56
  57 tdb_off_t next_flist(struct tdb_context *tdb, tdb_off_t flist)
  58 {
  59         return tdb_read_off(tdb, flist + offsetof(struct tdb_freelist, next));
  60 }
  61
  62 int tdb_flist_init(struct tdb_context *tdb)
  63 {
  64         /* Use reservoir sampling algorithm to select a free list at random. */
  65         unsigned int rnd, max = 0, count = 0;
  66         tdb_off_t off;
  67
  68         tdb->flist_off = off = first_flist(tdb);
  69         tdb->flist = 0;
  70
  71         while (off) {
  72                 if (off == TDB_OFF_ERR)
  73                         return -1;
  74
  75                 rnd = random();
  76                 if (rnd >= max) {
  77                         tdb->flist_off = off;
  78                         tdb->flist = count;
  79                         max = rnd;
  80                 }
  81
  82                 off = next_flist(tdb, off);
  83                 count++;
  84         }
  85         return 0;
  86 }
  87
  88 /* Offset of a given bucket. */
  89 tdb_off_t bucket_off(tdb_off_t flist_off, unsigned bucket)
  90 {
  91         return flist_off + offsetof(struct tdb_freelist, buckets)
  92                 + bucket * sizeof(tdb_off_t);
  93 }
  94
  95 /* Returns free_buckets + 1, or list number to search. */
  96 static tdb_off_t find_free_head(struct tdb_context *tdb,
  97                                 tdb_off_t flist_off,
  98                                 tdb_off_t bucket)
  99 {
 100         /* Speculatively search for a non-zero bucket. */
 101         return tdb_find_nonzero_off(tdb, bucket_off(flist_off, 0),
 102                                     bucket, TDB_FREE_BUCKETS);
 103 }
 104
 105 /* Remove from free bucket. */
 106 static int remove_from_list(struct tdb_context *tdb,
 107                             tdb_off_t b_off, tdb_off_t r_off,
 108                             struct tdb_free_record *r)
 109 {
 110         tdb_off_t off;
 111
 112         /* Front of list? */
 113         if (frec_prev(r) == 0) {
 114                 off = b_off;
 115         } else {
 116                 off = frec_prev(r) + offsetof(struct tdb_free_record, next);
 117         }
 118
 119 #ifdef DEBUG
 120         if (tdb_read_off(tdb, off) != r_off) {
 121                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 122                          "remove_from_list: %llu bad prev in list %llu\n",
 123                          (long long)r_off, (long long)b_off);
 124                 return -1;
 125         }
 126 #endif
 127
 128         /* r->prev->next = r->next */
 129         if (tdb_write_off(tdb, off, r->next)) {
 130                 return -1;
 131         }
 132
 133         if (r->next != 0) {
 134                 off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
 135                 /* r->next->prev = r->prev */
 136
 137 #ifdef DEBUG
 138                 if (tdb_read_off(tdb, off) & TDB_OFF_MASK != r_off) {
 139                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 140                                  "remove_from_list: %llu bad list %llu\n",
 141                                  (long long)r_off, (long long)b_off);
 142                         return -1;
 143                 }
 144 #endif
 145
 146                 if (tdb_write_off(tdb, off, r->magic_and_prev)) {
 147                         return -1;
 148                 }
 149         }
 150         return 0;
 151 }
 152
 153 /* Enqueue in this free bucket. */
 154 static int enqueue_in_free(struct tdb_context *tdb,
 155                            tdb_off_t b_off,
 156                            tdb_off_t off,
 157                            tdb_len_t len)
 158 {
 159         struct tdb_free_record new;
 160         uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
 161
 162         /* We only need to set flist_and_len; rest is set in enqueue_in_free */
 163         new.flist_and_len = ((uint64_t)tdb->flist << (64 - TDB_OFF_UPPER_STEAL))
 164                 | len;
 165         /* prev = 0. */
 166         new.magic_and_prev = magic;
 167
 168         /* new->next = head. */
 169         new.next = tdb_read_off(tdb, b_off);
 170         if (new.next == TDB_OFF_ERR)
 171                 return -1;
 172
 173         if (new.next) {
 174 #ifdef DEBUG
 175                 if (tdb_read_off(tdb,
 176                                  new.next + offsetof(struct tdb_free_record,
 177                                                      magic_and_prev))
 178                     != magic) {
 179                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 180                                  "enqueue_in_free: %llu bad head prev %llu\n",
 181                                  (long long)new.next, (long long)b_off);
 182                         return -1;
 183                 }
 184 #endif
 185                 /* next->prev = new. */
 186                 if (tdb_write_off(tdb, new.next
 187                                   + offsetof(struct tdb_free_record,
 188                                              magic_and_prev),
 189                                   off | magic) != 0)
 190                         return -1;
 191         }
 192         /* head = new */
 193         if (tdb_write_off(tdb, b_off, off) != 0)
 194                 return -1;
 195
 196         return tdb_write_convert(tdb, off, &new, sizeof(new));
 197 }
 198
 199 /* List need not be locked. */
 200 int add_free_record(struct tdb_context *tdb,
 201                     tdb_off_t off, tdb_len_t len_with_header)
 202 {
 203         tdb_off_t b_off;
 204         tdb_len_t len;
 205         int ret;
 206
 207         assert(len_with_header >= sizeof(struct tdb_free_record));
 208
 209         len = len_with_header - sizeof(struct tdb_used_record);
 210
 211         b_off = bucket_off(tdb->flist_off, size_to_bucket(len));
 212         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
 213                 return -1;
 214
 215         ret = enqueue_in_free(tdb, b_off, off, len);
 216         tdb_unlock_free_bucket(tdb, b_off);
 217         return ret;
 218 }
 219
 220 static size_t adjust_size(size_t keylen, size_t datalen)
 221 {
 222         size_t size = keylen + datalen;
 223
 224         if (size < TDB_MIN_DATA_LEN)
 225                 size = TDB_MIN_DATA_LEN;
 226
 227         /* Round to next uint64_t boundary. */
 228         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
 229 }
 230
 231 /* If we have enough left over to be useful, split that off. */
 232 static size_t record_leftover(size_t keylen, size_t datalen,
 233                               bool want_extra, size_t total_len)
 234 {
 235         ssize_t leftover;
 236
 237         if (want_extra)
 238                 datalen += datalen / 2;
 239         leftover = total_len - adjust_size(keylen, datalen);
 240
 241         if (leftover < (ssize_t)sizeof(struct tdb_free_record))
 242                 return 0;
 243
 244         return leftover;
 245 }
 246
 247 static tdb_off_t flist_offset(struct tdb_context *tdb, unsigned int flist)
 248 {
 249         tdb_off_t off;
 250         unsigned int i;
 251
 252         if (likely(tdb->flist == flist))
 253                 return tdb->flist_off;
 254
 255         off = first_flist(tdb);
 256         for (i = 0; i < flist; i++)
 257                 off = next_flist(tdb, off);
 258         return off;
 259 }
 260
 261 /* Note: we unlock the current bucket if we coalesce or fail. */
 262 static int coalesce(struct tdb_context *tdb,
 263                     tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
 264 {
 265         struct tdb_free_record pad, *r;
 266         tdb_off_t end;
 267
 268         add_stat(tdb, alloc_coalesce_tried, 1);
 269         end = off + sizeof(struct tdb_used_record) + data_len;
 270
 271         while (end < tdb->map_size) {
 272                 tdb_off_t nb_off;
 273                 unsigned flist, bucket;
 274
 275                 /* FIXME: do tdb_get here and below really win? */
 276                 r = tdb_get(tdb, end, &pad, sizeof(pad));
 277                 if (!r)
 278                         goto err;
 279
 280                 if (frec_magic(r) != TDB_FREE_MAGIC)
 281                         break;
 282
 283                 flist = frec_flist(r);
 284                 bucket = size_to_bucket(frec_len(r));
 285                 nb_off = bucket_off(flist_offset(tdb, flist), bucket);
 286
 287                 /* We may be violating lock order here, so best effort. */
 288                 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1) {
 289                         add_stat(tdb, alloc_coalesce_lockfail, 1);
 290                         break;
 291                 }
 292
 293                 /* Now we have lock, re-check. */
 294                 r = tdb_get(tdb, end, &pad, sizeof(pad));
 295                 if (!r) {
 296                         tdb_unlock_free_bucket(tdb, nb_off);
 297                         goto err;
 298                 }
 299
 300                 if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
 301                         add_stat(tdb, alloc_coalesce_race, 1);
 302                         tdb_unlock_free_bucket(tdb, nb_off);
 303                         break;
 304                 }
 305
 306                 if (unlikely(frec_flist(r) != flist)
 307                     || unlikely(size_to_bucket(frec_len(r)) != bucket)) {
 308                         add_stat(tdb, alloc_coalesce_race, 1);
 309                         tdb_unlock_free_bucket(tdb, nb_off);
 310                         break;
 311                 }
 312
 313                 if (remove_from_list(tdb, nb_off, end, r) == -1) {
 314                         tdb_unlock_free_bucket(tdb, nb_off);
 315                         goto err;
 316                 }
 317
 318                 end += sizeof(struct tdb_used_record) + frec_len(r);
 319                 tdb_unlock_free_bucket(tdb, nb_off);
 320                 add_stat(tdb, alloc_coalesce_num_merged, 1);
 321         }
 322
 323         /* Didn't find any adjacent free? */
 324         if (end == off + sizeof(struct tdb_used_record) + data_len)
 325                 return 0;
 326
 327         /* OK, expand record */
 328         r = tdb_get(tdb, off, &pad, sizeof(pad));
 329         if (!r)
 330                 goto err;
 331
 332         if (frec_len(r) != data_len) {
 333                 tdb->ecode = TDB_ERR_CORRUPT;
 334                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 335                          "coalesce: expected data len %llu not %llu\n",
 336                          (long long)data_len, (long long)frec_len(r));
 337                 goto err;
 338         }
 339
 340         if (remove_from_list(tdb, b_off, off, r) == -1)
 341                 goto err;
 342
 343         r = tdb_access_write(tdb, off, sizeof(*r), true);
 344         if (!r)
 345                 goto err;
 346
 347         /* We have to drop this to avoid deadlocks, so make sure record
 348          * doesn't get coalesced by someone else! */
 349         r->magic_and_prev = TDB_COALESCING_MAGIC << (64 - TDB_OFF_UPPER_STEAL);
 350         /* FIXME: Use 255 as invalid free list? */
 351         r->flist_and_len = end - off - sizeof(struct tdb_used_record);
 352         if (tdb_access_commit(tdb, r) != 0)
 353                 goto err;
 354
 355         add_stat(tdb, alloc_coalesce_succeeded, 1);
 356         tdb_unlock_free_bucket(tdb, b_off);
 357
 358         if (add_free_record(tdb, off, end - off) == -1)
 359                 return -1;
 360         return 1;
 361
 362 err:
 363         /* To unify error paths, we *always* unlock bucket on error. */
 364         tdb_unlock_free_bucket(tdb, b_off);
 365         return -1;
 366 }
 367
 368 /* We need size bytes to put our key and data in. */
 369 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 370                                 tdb_off_t flist_off,
 371                                 tdb_off_t bucket,
 372                                 size_t keylen, size_t datalen,
 373                                 bool want_extra,
 374                                 unsigned hashlow)
 375 {
 376         tdb_off_t off, b_off,best_off;
 377         struct tdb_free_record pad, best = { 0 }, *r;
 378         double multiplier;
 379         size_t size = adjust_size(keylen, datalen);
 380
 381         add_stat(tdb, allocs, 1);
 382 again:
 383         b_off = bucket_off(flist_off, bucket);
 384
 385         /* FIXME: Try non-blocking wait first, to measure contention. */
 386         /* Lock this bucket. */
 387         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
 388                 return TDB_OFF_ERR;
 389         }
 390
 391         best.flist_and_len = -1ULL;
 392         best_off = 0;
 393
 394         /* Get slack if we're after extra. */
 395         if (want_extra)
 396                 multiplier = 1.5;
 397         else
 398                 multiplier = 1.0;
 399
 400         /* Walk the list to see if any are large enough, getting less fussy
 401          * as we go. */
 402         off = tdb_read_off(tdb, b_off);
 403         if (unlikely(off == TDB_OFF_ERR))
 404                 goto unlock_err;
 405
 406         while (off) {
 407                 /* FIXME: Does tdb_get win anything here? */
 408                 r = tdb_get(tdb, off, &pad, sizeof(*r));
 409                 if (!r)
 410                         goto unlock_err;
 411
 412                 if (frec_magic(r) != TDB_FREE_MAGIC) {
 413                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 414                                  "lock_and_alloc: %llu non-free 0x%llx\n",
 415                                  (long long)off, (long long)r->magic_and_prev);
 416                         goto unlock_err;
 417                 }
 418
 419                 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
 420                         best_off = off;
 421                         best = *r;
 422                 }
 423
 424                 if (frec_len(&best) < size * multiplier && best_off)
 425                         break;
 426
 427                 multiplier *= 1.01;
 428
 429                 /* Since we're going slow anyway, try coalescing here. */
 430                 switch (coalesce(tdb, off, b_off, frec_len(r))) {
 431                 case -1:
 432                         /* This has already unlocked on error. */
 433                         return -1;
 434                 case 1:
 435                         /* This has unlocked list, restart. */
 436                         goto again;
 437                 }
 438                 off = r->next;
 439         }
 440
 441         /* If we found anything at all, use it. */
 442         if (best_off) {
 443                 struct tdb_used_record rec;
 444                 size_t leftover;
 445
 446                 /* We're happy with this size: take it. */
 447                 if (remove_from_list(tdb, b_off, best_off, &best) != 0)
 448                         goto unlock_err;
 449
 450                 leftover = record_leftover(keylen, datalen, want_extra,
 451                                            frec_len(&best));
 452
 453                 assert(keylen + datalen + leftover <= frec_len(&best));
 454                 /* We need to mark non-free before we drop lock, otherwise
 455                  * coalesce() could try to merge it! */
 456                 if (set_used_header(tdb, &rec, keylen, datalen,
 457                                     frec_len(&best) - leftover,
 458                                     hashlow) != 0)
 459                         goto unlock_err;
 460
 461                 if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
 462                         goto unlock_err;
 463
 464                 /* Bucket of leftover will be <= current bucket, so nested
 465                  * locking is allowed. */
 466                 if (leftover) {
 467                         add_stat(tdb, alloc_leftover, 1);
 468                         if (add_free_record(tdb,
 469                                             best_off + sizeof(rec)
 470                                             + frec_len(&best) - leftover,
 471                                             leftover))
 472                                 best_off = TDB_OFF_ERR;
 473                 }
 474                 tdb_unlock_free_bucket(tdb, b_off);
 475
 476                 return best_off;
 477         }
 478
 479         tdb_unlock_free_bucket(tdb, b_off);
 480         return 0;
 481
 482 unlock_err:
 483         tdb_unlock_free_bucket(tdb, b_off);
 484         return TDB_OFF_ERR;
 485 }
 486
 487 /* Get a free block from current free list, or 0 if none. */
 488 static tdb_off_t get_free(struct tdb_context *tdb,
 489                           size_t keylen, size_t datalen, bool want_extra,
 490                           unsigned hashlow)
 491 {
 492         tdb_off_t off, flist_off;
 493         unsigned start_b, b, flist;
 494         bool wrapped = false;
 495
 496         /* If they are growing, add 50% to get to higher bucket. */
 497         if (want_extra)
 498                 start_b = size_to_bucket(adjust_size(keylen,
 499                                                      datalen + datalen / 2));
 500         else
 501                 start_b = size_to_bucket(adjust_size(keylen, datalen));
 502
 503         flist_off = tdb->flist_off;
 504         flist = tdb->flist;
 505         while (!wrapped || flist_off != tdb->flist_off) {
 506                 /* Start at exact size bucket, and search up... */
 507                 for (b = find_free_head(tdb, flist_off, start_b);
 508                      b < TDB_FREE_BUCKETS;
 509                      b = find_free_head(tdb, flist_off, b + 1)) {
 510                         /* Try getting one from list. */
 511                         off = lock_and_alloc(tdb, flist_off,
 512                                              b, keylen, datalen, want_extra,
 513                                              hashlow);
 514                         if (off == TDB_OFF_ERR)
 515                                 return TDB_OFF_ERR;
 516                         if (off != 0) {
 517                                 if (b == start_b)
 518                                         add_stat(tdb, alloc_bucket_exact, 1);
 519                                 if (b == TDB_FREE_BUCKETS - 1)
 520                                         add_stat(tdb, alloc_bucket_max, 1);
 521                                 /* Worked?  Stay using this list. */
 522                                 tdb->flist_off = flist_off;
 523                                 tdb->flist = flist;
 524                                 return off;
 525                         }
 526                         /* Didn't work.  Try next bucket. */
 527                 }
 528
 529                 /* Hmm, try next list. */
 530                 flist_off = next_flist(tdb, flist_off);
 531                 flist++;
 532
 533                 if (flist_off == 0) {
 534                         wrapped = true;
 535                         flist_off = first_flist(tdb);
 536                         flist = 0;
 537                 }
 538         }
 539
 540         return 0;
 541 }
 542
 543 int set_used_header(struct tdb_context *tdb,
 544                     struct tdb_used_record *rec,
 545                     uint64_t keylen, uint64_t datalen,
 546                     uint64_t actuallen, unsigned hashlow)
 547 {
 548         uint64_t keybits = (fls64(keylen) + 1) / 2;
 549
 550         /* Use bottom bits of hash, so it's independent of hash table size. */
 551         rec->magic_and_meta = (hashlow & ((1 << 11)-1))
 552                 | ((actuallen - (keylen + datalen)) << 11)
 553                 | (keybits << 43)
 554                 | (TDB_MAGIC << 48);
 555         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
 556
 557         /* Encoding can fail on big values. */
 558         if (rec_key_length(rec) != keylen
 559             || rec_data_length(rec) != datalen
 560             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
 561                 tdb->ecode = TDB_ERR_IO;
 562                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 563                          "Could not encode k=%llu,d=%llu,a=%llu\n",
 564                          (long long)keylen, (long long)datalen,
 565                          (long long)actuallen);
 566                 return -1;
 567         }
 568         return 0;
 569 }
 570
 571 /* Expand the database. */
 572 static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 573 {
 574         uint64_t old_size;
 575         tdb_len_t wanted;
 576
 577         /* We need room for the record header too. */
 578         wanted = sizeof(struct tdb_used_record) + size;
 579
 580         /* Need to hold a hash lock to expand DB: transactions rely on it. */
 581         if (!(tdb->flags & TDB_NOLOCK)
 582             && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
 583                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 584                          "tdb_expand: must hold lock during expand\n");
 585                 return -1;
 586         }
 587
 588         /* always make room for at least 100 more records, and at
 589            least 25% more space. */
 590         if (size * TDB_EXTENSION_FACTOR > tdb->map_size / 4)
 591                 wanted = size * TDB_EXTENSION_FACTOR;
 592         else
 593                 wanted = tdb->map_size / 4;
 594         wanted = adjust_size(0, wanted);
 595
 596         /* Only one person can expand file at a time. */
 597         if (tdb_lock_expand(tdb, F_WRLCK) != 0)
 598                 return -1;
 599
 600         /* Someone else may have expanded the file, so retry. */
 601         old_size = tdb->map_size;
 602         tdb->methods->oob(tdb, tdb->map_size + 1, true);
 603         if (tdb->map_size != old_size) {
 604                 tdb_unlock_expand(tdb, F_WRLCK);
 605                 return 0;
 606         }
 607
 608         if (tdb->methods->expand_file(tdb, wanted) == -1) {
 609                 tdb_unlock_expand(tdb, F_WRLCK);
 610                 return -1;
 611         }
 612
 613         /* We need to drop this lock before adding free record. */
 614         tdb_unlock_expand(tdb, F_WRLCK);
 615
 616         add_stat(tdb, expands, 1);
 617         return add_free_record(tdb, old_size, wanted);
 618 }
 619
 620 /* This won't fail: it will expand the database if it has to. */
 621 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
 622                 uint64_t hash, bool growing)
 623 {
 624         tdb_off_t off;
 625
 626         /* We can't hold pointers during this: we could unmap! */
 627         assert(!tdb->direct_access);
 628
 629         for (;;) {
 630                 off = get_free(tdb, keylen, datalen, growing, hash);
 631                 if (likely(off != 0))
 632                         break;
 633
 634                 if (tdb_expand(tdb, adjust_size(keylen, datalen)))
 635                         return TDB_OFF_ERR;
 636         }
 637
 638         return off;
 639 }