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