git.ozlabs.org Git - ccan/blob - ccan/tdb2/tdb.c

   1 #include "private.h"
   2 #include <ccan/tdb2/tdb2.h>
   3 #include <ccan/hash/hash.h>
   4 #include <ccan/build_assert/build_assert.h>
   5 #include <ccan/likely/likely.h>
   6 #include <assert.h>
   7
   8 /* The null return. */
   9 struct tdb_data tdb_null = { .dptr = NULL, .dsize = 0 };
  10
  11 /* all contexts, to ensure no double-opens (fcntl locks don't nest!) */
  12 static struct tdb_context *tdbs = NULL;
  13
  14 PRINTF_ATTRIBUTE(4, 5) static void
  15 null_log_fn(struct tdb_context *tdb,
  16             enum tdb_debug_level level, void *priv,
  17             const char *fmt, ...)
  18 {
  19 }
  20
  21 /* We do a lot of work assuming our copy of the header volatile area
  22  * is uptodate, and usually it is.  However, once we grab a lock, we have to
  23  * re-check it. */
  24 bool header_changed(struct tdb_context *tdb)
  25 {
  26         uint64_t gen;
  27
  28         if (!(tdb->flags & TDB_NOLOCK) && tdb->header_uptodate) {
  29                 tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
  30                          "warning: header uptodate already\n");
  31         }
  32
  33         /* We could get a partial update if we're not holding any locks. */
  34         assert((tdb->flags & TDB_NOLOCK) || tdb_has_locks(tdb));
  35
  36         tdb->header_uptodate = true;
  37         gen = tdb_read_off(tdb, offsetof(struct tdb_header, v.generation));
  38         if (unlikely(gen != tdb->header.v.generation)) {
  39                 tdb_read_convert(tdb, offsetof(struct tdb_header, v),
  40                                  &tdb->header.v, sizeof(tdb->header.v));
  41                 return true;
  42         }
  43         return false;
  44 }
  45
  46 int write_header(struct tdb_context *tdb)
  47 {
  48         assert(tdb_read_off(tdb, offsetof(struct tdb_header, v.generation))
  49                == tdb->header.v.generation);
  50         tdb->header.v.generation++;
  51         return tdb_write_convert(tdb, offsetof(struct tdb_header, v),
  52                                  &tdb->header.v, sizeof(tdb->header.v));
  53 }
  54
  55 static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
  56                              void *arg)
  57 {
  58         return hash64_stable((const unsigned char *)key, length, seed);
  59 }
  60
  61 uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
  62 {
  63         return tdb->khash(ptr, len, tdb->header.hash_seed, tdb->hash_priv);
  64 }
  65
  66 static bool tdb_already_open(dev_t device, ino_t ino)
  67 {
  68         struct tdb_context *i;
  69
  70         for (i = tdbs; i; i = i->next) {
  71                 if (i->device == device && i->inode == ino) {
  72                         return true;
  73                 }
  74         }
  75
  76         return false;
  77 }
  78
  79 static uint64_t random_number(struct tdb_context *tdb)
  80 {
  81         int fd;
  82         uint64_t ret = 0;
  83         struct timeval now;
  84
  85         fd = open("/dev/urandom", O_RDONLY);
  86         if (fd >= 0) {
  87                 if (tdb_read_all(fd, &ret, sizeof(ret))) {
  88                         tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
  89                                  "tdb_open: random from /dev/urandom\n");
  90                         close(fd);
  91                         return ret;
  92                 }
  93                 close(fd);
  94         }
  95         /* FIXME: Untested!  Based on Wikipedia protocol description! */
  96         fd = open("/dev/egd-pool", O_RDWR);
  97         if (fd >= 0) {
  98                 /* Command is 1, next byte is size we want to read. */
  99                 char cmd[2] = { 1, sizeof(uint64_t) };
 100                 if (write(fd, cmd, sizeof(cmd)) == sizeof(cmd)) {
 101                         char reply[1 + sizeof(uint64_t)];
 102                         int r = read(fd, reply, sizeof(reply));
 103                         if (r > 1) {
 104                                 tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
 105                                          "tdb_open: %u random bytes from"
 106                                          " /dev/egd-pool\n", r-1);
 107                                 /* Copy at least some bytes. */
 108                                 memcpy(&ret, reply+1, r - 1);
 109                                 if (reply[0] == sizeof(uint64_t)
 110                                     && r == sizeof(reply)) {
 111                                         close(fd);
 112                                         return ret;
 113                                 }
 114                         }
 115                 }
 116                 close(fd);
 117         }
 118
 119         /* Fallback: pid and time. */
 120         gettimeofday(&now, NULL);
 121         ret = getpid() * 100132289ULL + now.tv_sec * 1000000ULL + now.tv_usec;
 122         tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
 123                  "tdb_open: random from getpid and time\n");
 124         return ret;
 125 }
 126
 127 struct new_database {
 128         struct tdb_header hdr;
 129         struct tdb_used_record hrec;
 130         tdb_off_t hash[1ULL << INITIAL_HASH_BITS];
 131         struct tdb_used_record frec;
 132         tdb_off_t free[INITIAL_FREE_BUCKETS + 1]; /* One overflow bucket */
 133 };
 134
 135 /* initialise a new database */
 136 static int tdb_new_database(struct tdb_context *tdb)
 137 {
 138         /* We make it up in memory, then write it out if not internal */
 139         struct new_database newdb;
 140         unsigned int magic_off = offsetof(struct tdb_header, magic_food);
 141
 142         /* Fill in the header */
 143         newdb.hdr.version = TDB_VERSION;
 144         newdb.hdr.hash_seed = random_number(tdb);
 145         newdb.hdr.hash_test = TDB_HASH_MAGIC;
 146         newdb.hdr.hash_test = tdb->khash(&newdb.hdr.hash_test,
 147                                          sizeof(newdb.hdr.hash_test),
 148                                          newdb.hdr.hash_seed,
 149                                          tdb->hash_priv);
 150
 151         newdb.hdr.v.generation = 0;
 152
 153         /* The initial zone must cover the initial database size! */
 154         BUILD_ASSERT((1ULL << INITIAL_ZONE_BITS) >= sizeof(newdb));
 155
 156         /* Free array has 1 zone, 10 buckets.  All buckets empty. */
 157         newdb.hdr.v.num_zones = 1;
 158         newdb.hdr.v.zone_bits = INITIAL_ZONE_BITS;
 159         newdb.hdr.v.free_buckets = INITIAL_FREE_BUCKETS;
 160         newdb.hdr.v.free_off = offsetof(struct new_database, free);
 161         set_header(tdb, &newdb.frec, 0,
 162                    sizeof(newdb.free), sizeof(newdb.free), 0);
 163         memset(newdb.free, 0, sizeof(newdb.free));
 164
 165         /* Initial hashes are empty. */
 166         newdb.hdr.v.hash_bits = INITIAL_HASH_BITS;
 167         newdb.hdr.v.hash_off = offsetof(struct new_database, hash);
 168         set_header(tdb, &newdb.hrec, 0,
 169                    sizeof(newdb.hash), sizeof(newdb.hash), 0);
 170         memset(newdb.hash, 0, sizeof(newdb.hash));
 171
 172         /* Magic food */
 173         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 174         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 175
 176         /* This creates an endian-converted database, as if read from disk */
 177         tdb_convert(tdb,
 178                     (char *)&newdb.hdr + magic_off,
 179                     sizeof(newdb) - magic_off);
 180
 181         tdb->header = newdb.hdr;
 182
 183         if (tdb->flags & TDB_INTERNAL) {
 184                 tdb->map_size = sizeof(newdb);
 185                 tdb->map_ptr = malloc(tdb->map_size);
 186                 if (!tdb->map_ptr) {
 187                         tdb->ecode = TDB_ERR_OOM;
 188                         return -1;
 189                 }
 190                 memcpy(tdb->map_ptr, &newdb, tdb->map_size);
 191                 return 0;
 192         }
 193         if (lseek(tdb->fd, 0, SEEK_SET) == -1)
 194                 return -1;
 195
 196         if (ftruncate(tdb->fd, 0) == -1)
 197                 return -1;
 198
 199         if (!tdb_pwrite_all(tdb->fd, &newdb, sizeof(newdb), 0)) {
 200                 tdb->ecode = TDB_ERR_IO;
 201                 return -1;
 202         }
 203         return 0;
 204 }
 205
 206 struct tdb_context *tdb_open(const char *name, int tdb_flags,
 207                              int open_flags, mode_t mode,
 208                              union tdb_attribute *attr)
 209 {
 210         struct tdb_context *tdb;
 211         struct stat st;
 212         int save_errno;
 213         uint64_t hash_test;
 214         unsigned v;
 215
 216         tdb = malloc(sizeof(*tdb));
 217         if (!tdb) {
 218                 /* Can't log this */
 219                 errno = ENOMEM;
 220                 goto fail;
 221         }
 222         tdb->name = NULL;
 223         tdb->map_ptr = NULL;
 224         tdb->fd = -1;
 225         /* map_size will be set below. */
 226         tdb->ecode = TDB_SUCCESS;
 227         /* header will be read in below. */
 228         tdb->header_uptodate = false;
 229         tdb->flags = tdb_flags;
 230         tdb->log = null_log_fn;
 231         tdb->log_priv = NULL;
 232         tdb->khash = jenkins_hash;
 233         tdb->hash_priv = NULL;
 234         tdb->transaction = NULL;
 235         /* last_zone will be set below. */
 236         tdb_io_init(tdb);
 237         tdb_lock_init(tdb);
 238
 239         /* FIXME */
 240         if (attr) {
 241                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 242                          "tdb_open: attributes not yet supported\n");
 243                 errno = EINVAL;
 244                 goto fail;
 245         }
 246
 247         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 248                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 249                          "tdb_open: can't open tdb %s write-only\n", name);
 250                 errno = EINVAL;
 251                 goto fail;
 252         }
 253
 254         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 255                 tdb->read_only = true;
 256                 /* read only databases don't do locking */
 257                 tdb->flags |= TDB_NOLOCK;
 258         } else
 259                 tdb->read_only = false;
 260
 261         /* internal databases don't need any of the rest. */
 262         if (tdb->flags & TDB_INTERNAL) {
 263                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 264                 if (tdb_new_database(tdb) != 0) {
 265                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 266                                  "tdb_open: tdb_new_database failed!");
 267                         goto fail;
 268                 }
 269                 TEST_IT(tdb->flags & TDB_CONVERT);
 270                 tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
 271                 return tdb;
 272         }
 273
 274         if ((tdb->fd = open(name, open_flags, mode)) == -1) {
 275                 tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
 276                          "tdb_open: could not open file %s: %s\n",
 277                          name, strerror(errno));
 278                 goto fail;      /* errno set by open(2) */
 279         }
 280
 281         /* on exec, don't inherit the fd */
 282         v = fcntl(tdb->fd, F_GETFD, 0);
 283         fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
 284
 285         /* ensure there is only one process initialising at once */
 286         if (tdb_lock_open(tdb) == -1) {
 287                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 288                          "tdb_open: failed to get open lock on %s: %s\n",
 289                          name, strerror(errno));
 290                 goto fail;      /* errno set by tdb_brlock */
 291         }
 292
 293         if (!tdb_pread_all(tdb->fd, &tdb->header, sizeof(tdb->header), 0)
 294             || strcmp(tdb->header.magic_food, TDB_MAGIC_FOOD) != 0) {
 295                 if (!(open_flags & O_CREAT) || tdb_new_database(tdb) == -1) {
 296                         if (errno == 0) {
 297                                 errno = EIO; /* ie bad format or something */
 298                         }
 299                         goto fail;
 300                 }
 301         } else if (tdb->header.version != TDB_VERSION) {
 302                 if (tdb->header.version == bswap_64(TDB_VERSION))
 303                         tdb->flags |= TDB_CONVERT;
 304                 else {
 305                         /* wrong version */
 306                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 307                                  "tdb_open: %s is unknown version 0x%llx\n",
 308                                  name, (long long)tdb->header.version);
 309                         errno = EIO;
 310                         goto fail;
 311                 }
 312         }
 313
 314         tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
 315         hash_test = TDB_HASH_MAGIC;
 316         hash_test = tdb->khash(&hash_test, sizeof(hash_test),
 317                                tdb->header.hash_seed, tdb->hash_priv);
 318         if (tdb->header.hash_test != hash_test) {
 319                 /* wrong hash variant */
 320                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 321                          "tdb_open: %s uses a different hash function\n",
 322                          name);
 323                 errno = EIO;
 324                 goto fail;
 325         }
 326
 327         if (fstat(tdb->fd, &st) == -1)
 328                 goto fail;
 329
 330         /* Is it already in the open list?  If so, fail. */
 331         if (tdb_already_open(st.st_dev, st.st_ino)) {
 332                 /* FIXME */
 333                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 334                          "tdb_open: %s (%d,%d) is already open in this process\n",
 335                          name, (int)st.st_dev, (int)st.st_ino);
 336                 errno = EBUSY;
 337                 goto fail;
 338         }
 339
 340         tdb->name = strdup(name);
 341         if (!tdb->name) {
 342                 errno = ENOMEM;
 343                 goto fail;
 344         }
 345
 346         tdb->map_size = st.st_size;
 347         tdb->device = st.st_dev;
 348         tdb->inode = st.st_ino;
 349         tdb_mmap(tdb);
 350         tdb_unlock_open(tdb);
 351         tdb_zone_init(tdb);
 352
 353         tdb->next = tdbs;
 354         tdbs = tdb;
 355         return tdb;
 356
 357  fail:
 358         save_errno = errno;
 359
 360         if (!tdb)
 361                 return NULL;
 362
 363 #ifdef TDB_TRACE
 364         close(tdb->tracefd);
 365 #endif
 366         if (tdb->map_ptr) {
 367                 if (tdb->flags & TDB_INTERNAL) {
 368                         free(tdb->map_ptr);
 369                 } else
 370                         tdb_munmap(tdb);
 371         }
 372         free((char *)tdb->name);
 373         if (tdb->fd != -1)
 374                 if (close(tdb->fd) != 0)
 375                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 376                                  "tdb_open: failed to close tdb->fd"
 377                                  " on error!\n");
 378         free(tdb);
 379         errno = save_errno;
 380         return NULL;
 381 }
 382
 383 static int tdb_key_compare(TDB_DATA key, TDB_DATA data, void *private_data)
 384 {
 385         return memcmp(data.dptr, key.dptr, data.dsize) == 0;
 386 }
 387
 388 static void unlock_lists(struct tdb_context *tdb,
 389                          uint64_t start, uint64_t end, int ltype)
 390 {
 391         for (;;) {
 392                 tdb_unlock_list(tdb, start, ltype);
 393                 if (start == end)
 394                         return;
 395                 start = (start + 1) & ((1ULL << tdb->header.v.hash_bits) - 1);
 396         }
 397 }
 398
 399 /* FIXME: Return header copy? */
 400 /* Returns -1 or offset of entry (0 if not found).
 401  * Locks hash entried from *start to *end (where the entry was found). */
 402 static tdb_off_t find_bucket_and_lock(struct tdb_context *tdb,
 403                                       const struct tdb_data *key,
 404                                       uint64_t hash,
 405                                       uint64_t *start,
 406                                       uint64_t *end,
 407                                       uint64_t *room,
 408                                       int ltype)
 409 {
 410         uint64_t hextra;
 411         tdb_off_t off;
 412
 413         /* FIXME: can we avoid locks for some fast paths? */
 414         *start = tdb_lock_list(tdb, hash, ltype, TDB_LOCK_WAIT);
 415         if (*start == TDB_OFF_ERR)
 416                 return TDB_OFF_ERR;
 417
 418         *end = *start;
 419         hextra = hash >> tdb->header.v.hash_bits;
 420
 421         while ((off = tdb_read_off(tdb, tdb->header.v.hash_off
 422                                    + *end * sizeof(tdb_off_t)))
 423                != TDB_OFF_ERR) {
 424                 struct tdb_used_record pad, *r;
 425                 uint64_t keylen;
 426
 427                 /* Didn't find it? */
 428                 if (!off)
 429                         return 0;
 430
 431 #if 0 /* FIXME: Check other bits. */
 432                 unsigned int bits, bitmask, hoffextra;
 433                 /* Bottom three bits show how many extra hash bits. */
 434                 bits = (off & ((1 << TDB_EXTRA_HASHBITS_NUM) - 1)) + 1;
 435                 bitmask = (1 << bits)-1;
 436                 hoffextra = ((off >> TDB_EXTRA_HASHBITS_NUM) & bitmask);
 437                 if ((hextra & bitmask) != hoffextra)
 438                         goto lock_next;
 439 #endif
 440
 441                 r = tdb_get(tdb, off, &pad, sizeof(*r));
 442                 if (!r)
 443                         goto unlock_err;
 444
 445                 if (rec_magic(r) != TDB_MAGIC) {
 446                         tdb->ecode = TDB_ERR_CORRUPT;
 447                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 448                                  "find_bucket_and_lock: bad magic 0x%llx"
 449                                  " at offset %llu!\n",
 450                                  (long long)rec_magic(r), (long long)off);
 451                         goto unlock_err;
 452                 }
 453
 454                 /* FIXME: check extra bits in header! */
 455                 keylen = rec_key_length(r);
 456                 if (keylen != key->dsize)
 457                         goto lock_next;
 458
 459                 switch (tdb_parse_data(tdb, *key, off + sizeof(*r), key->dsize,
 460                                        tdb_key_compare, NULL)) {
 461                 case 1:
 462                         /* Match! */
 463                         *room = rec_data_length(r) + rec_extra_padding(r);
 464                         return off >> TDB_EXTRA_HASHBITS_NUM;
 465                 case 0:
 466                         break;
 467                 default:
 468                         goto unlock_err;
 469                 }
 470
 471         lock_next:
 472                 /* Lock next bucket. */
 473                 /* FIXME: We can deadlock if this wraps! */
 474                 off = tdb_lock_list(tdb, ++hash, ltype, TDB_LOCK_WAIT);
 475                 if (off == TDB_OFF_ERR)
 476                         goto unlock_err;
 477                 *end = off;
 478         }
 479
 480 unlock_err:
 481         TEST_IT(*end < *start);
 482         unlock_lists(tdb, *start, *end, ltype);
 483         return TDB_OFF_ERR;
 484 }
 485
 486 static int update_rec_hdr(struct tdb_context *tdb,
 487                           tdb_off_t off,
 488                           tdb_len_t keylen,
 489                           tdb_len_t datalen,
 490                           tdb_len_t room,
 491                           uint64_t h)
 492 {
 493         struct tdb_used_record rec;
 494
 495         if (set_header(tdb, &rec, keylen, datalen, room - datalen, h))
 496                 return -1;
 497
 498         return tdb_write_convert(tdb, off, &rec, sizeof(rec));
 499 }
 500
 501 /* If we fail, others will try after us. */
 502 static void enlarge_hash(struct tdb_context *tdb)
 503 {
 504         tdb_off_t newoff, i;
 505         uint64_t h, num = 1ULL << tdb->header.v.hash_bits;
 506         struct tdb_used_record pad, *r;
 507
 508         /* FIXME: We should do this without holding locks throughout. */
 509         if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
 510                 return;
 511
 512         /* Someone else enlarged for us?  Nothing to do. */
 513         if ((1ULL << tdb->header.v.hash_bits) != num)
 514                 goto unlock;
 515
 516         newoff = alloc(tdb, 0, num * 2, 0, false);
 517         if (unlikely(newoff == TDB_OFF_ERR))
 518                 goto unlock;
 519         if (unlikely(newoff == 0)) {
 520                 if (tdb_expand(tdb, 0, num * 2, false) == -1)
 521                         goto unlock;
 522                 newoff = alloc(tdb, 0, num * 2, 0, false);
 523                 if (newoff == TDB_OFF_ERR || newoff == 0)
 524                         goto unlock;
 525         }
 526
 527         /* FIXME: If the space before is empty, we know this is in its ideal
 528          * location.  We can steal a bit from the pointer to avoid rehash. */
 529         for (i = tdb_find_nonzero_off(tdb, tdb->header.v.hash_off, num);
 530              i < num;
 531              i += tdb_find_nonzero_off(tdb, tdb->header.v.hash_off
 532                                        + i*sizeof(tdb_off_t), num - i)) {
 533                 tdb_off_t off;
 534                 off = tdb_read_off(tdb, tdb->header.v.hash_off
 535                                    + i*sizeof(tdb_off_t));
 536                 if (unlikely(off == TDB_OFF_ERR))
 537                         goto unlock;
 538                 if (unlikely(!off)) {
 539                         tdb->ecode = TDB_ERR_CORRUPT;
 540                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 541                                  "find_bucket_and_lock: zero hash bucket!\n");
 542                         goto unlock;
 543                 }
 544                 h = hash_record(tdb, off);
 545                 /* FIXME: Encode extra hash bits! */
 546                 if (tdb_write_off(tdb, newoff
 547                                   + (h & ((num * 2) - 1)) * sizeof(uint64_t),
 548                                   off) == -1)
 549                         goto unlock;
 550         }
 551
 552         /* Free up old hash. */
 553         r = tdb_get(tdb, tdb->header.v.hash_off, &pad, sizeof(*r));
 554         if (!r)
 555                 goto unlock;
 556         add_free_record(tdb, tdb->header.v.hash_off,
 557                         rec_data_length(r) + rec_extra_padding(r));
 558
 559         /* Now we write the modified header. */
 560         tdb->header.v.generation++;
 561         tdb->header.v.hash_bits++;
 562         tdb->header.v.hash_off = newoff;
 563         tdb_write_convert(tdb, offsetof(struct tdb_header, v),
 564                           &tdb->header.v, sizeof(tdb->header.v));
 565 unlock:
 566         tdb_allrecord_unlock(tdb, F_WRLCK);
 567 }
 568
 569 int tdb_store(struct tdb_context *tdb,
 570               struct tdb_data key, struct tdb_data dbuf, int flag)
 571 {
 572         tdb_off_t new_off, off, start, end, room;
 573         uint64_t h;
 574         bool growing = false;
 575
 576         h = tdb_hash(tdb, key.dptr, key.dsize);
 577         off = find_bucket_and_lock(tdb, &key, h, &start, &end, &room, F_WRLCK);
 578         if (off == TDB_OFF_ERR)
 579                 return -1;
 580
 581         /* Now we have lock on this hash bucket. */
 582         if (flag == TDB_INSERT) {
 583                 if (off) {
 584                         tdb->ecode = TDB_ERR_EXISTS;
 585                         goto fail;
 586                 }
 587         } else {
 588                 if (off) {
 589                         if (room >= key.dsize + dbuf.dsize) {
 590                                 new_off = off;
 591                                 if (update_rec_hdr(tdb, off,
 592                                                    key.dsize, dbuf.dsize,
 593                                                    room, h))
 594                                         goto fail;
 595                                 goto write;
 596                         }
 597                         /* FIXME: See if right record is free? */
 598                         /* Hint to allocator that we've realloced. */
 599                         growing = true;
 600                 } else {
 601                         if (flag == TDB_MODIFY) {
 602                                 /* if the record doesn't exist and we
 603                                    are in TDB_MODIFY mode then we should fail
 604                                    the store */
 605                                 tdb->ecode = TDB_ERR_NOEXIST;
 606                                 goto fail;
 607                         }
 608                 }
 609         }
 610
 611         /* Allocate a new record. */
 612         new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
 613         if (new_off == 0) {
 614                 unlock_lists(tdb, start, end, F_WRLCK);
 615                 /* Expand, then try again... */
 616                 if (tdb_expand(tdb, key.dsize, dbuf.dsize, growing) == -1)
 617                         return -1;
 618                 return tdb_store(tdb, key, dbuf, flag);
 619         }
 620
 621         /* We didn't like the existing one: remove it. */
 622         if (off) {
 623                 add_free_record(tdb, off, sizeof(struct tdb_used_record)
 624                                 + key.dsize + room);
 625         }
 626
 627 write:
 628         off = tdb->header.v.hash_off + end * sizeof(tdb_off_t);
 629         /* FIXME: Encode extra hash bits! */
 630         if (tdb_write_off(tdb, off, new_off) == -1)
 631                 goto fail;
 632
 633         off = new_off + sizeof(struct tdb_used_record);
 634         if (tdb->methods->write(tdb, off, key.dptr, key.dsize) == -1)
 635                 goto fail;
 636         off += key.dsize;
 637         if (tdb->methods->write(tdb, off, dbuf.dptr, dbuf.dsize) == -1)
 638                 goto fail;
 639
 640         /* FIXME: tdb_increment_seqnum(tdb); */
 641         unlock_lists(tdb, start, end, F_WRLCK);
 642
 643         /* By simple trial and error, this roughly approximates a 60%
 644          * full measure. */
 645         if (unlikely(end - start > 4 * tdb->header.v.hash_bits - 32))
 646                 enlarge_hash(tdb);
 647
 648         return 0;
 649
 650 fail:
 651         unlock_lists(tdb, start, end, F_WRLCK);
 652         return -1;
 653 }
 654
 655 struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
 656 {
 657         tdb_off_t off, start, end, room;
 658         uint64_t h;
 659         struct tdb_used_record pad, *r;
 660         struct tdb_data ret;
 661
 662         h = tdb_hash(tdb, key.dptr, key.dsize);
 663         off = find_bucket_and_lock(tdb, &key, h, &start, &end, &room, F_RDLCK);
 664         if (off == TDB_OFF_ERR)
 665                 return tdb_null;
 666
 667         if (!off) {
 668                 unlock_lists(tdb, start, end, F_RDLCK);
 669                 tdb->ecode = TDB_SUCCESS;
 670                 return tdb_null;
 671         }
 672
 673         r = tdb_get(tdb, off, &pad, sizeof(*r));
 674         if (!r) {
 675                 unlock_lists(tdb, start, end, F_RDLCK);
 676                 return tdb_null;
 677         }
 678
 679         ret.dsize = rec_data_length(r);
 680         ret.dptr = tdb_alloc_read(tdb, off + sizeof(*r) + key.dsize,
 681                                   ret.dsize);
 682         unlock_lists(tdb, start, end, F_RDLCK);
 683         return ret;
 684 }
 685
 686 static int hash_add(struct tdb_context *tdb, uint64_t h, tdb_off_t off)
 687 {
 688         tdb_off_t i, hoff, len, num;
 689
 690         i = (h & ((1ULL << tdb->header.v.hash_bits) - 1));
 691         hoff = tdb->header.v.hash_off + i * sizeof(tdb_off_t);
 692         len = (1ULL << tdb->header.v.hash_bits) - i;
 693
 694         /* Look for next space. */
 695         num = tdb_find_zero_off(tdb, hoff, len);
 696         if (unlikely(num == len)) {
 697                 hoff = tdb->header.v.hash_off;
 698                 len = (1ULL << tdb->header.v.hash_bits);
 699                 num = tdb_find_zero_off(tdb, hoff, len);
 700                 if (i == len)
 701                         return -1;
 702         }
 703         /* FIXME: Encode extra hash bits! */
 704         return tdb_write_off(tdb, hoff + num * sizeof(tdb_off_t), off);
 705 }
 706
 707 static int unlink_used_record(struct tdb_context *tdb, tdb_off_t chain,
 708                               uint64_t *extra_locks)
 709 {
 710         tdb_off_t num, len, i, hoff;
 711
 712         /* FIXME: Maybe lock more in search?  Maybe don't lock if scan
 713          * finds none? */
 714 again:
 715         len = (1ULL << tdb->header.v.hash_bits) - (chain + 1);
 716         hoff = tdb->header.v.hash_off + (chain + 1) * sizeof(tdb_off_t);
 717         num = tdb_find_zero_off(tdb, hoff, len);
 718
 719         /* We want to lock the zero entry, too.  In the wrap case,
 720          * this locks one extra.  That's harmless. */
 721         num++;
 722
 723         for (i = chain + 1; i < chain + 1 + num; i++) {
 724                 if (tdb_lock_list(tdb, i, F_WRLCK, TDB_LOCK_WAIT)
 725                     == TDB_OFF_ERR) {
 726                         if (i != chain + 1)
 727                                 unlock_lists(tdb, chain + 1, i-1, F_WRLCK);
 728                         return -1;
 729                 }
 730         }
 731
 732         /* The wrap case: we need those locks out of order! */
 733         if (unlikely(num == len + 1)) {
 734                 *extra_locks = tdb_find_zero_off(tdb, tdb->header.v.hash_off,
 735                                                  1ULL << tdb->header.v.hash_bits);
 736                 (*extra_locks)++;
 737                 for (i = 0; i < *extra_locks; i++) {
 738                         if (tdb_lock_list(tdb, i, F_WRLCK, TDB_LOCK_NOWAIT)
 739                             == TDB_OFF_ERR) {
 740                                 /* Failed.  Caller must lock in order. */
 741                                 if (i)
 742                                         unlock_lists(tdb, 0, i-1, F_WRLCK);
 743                                 unlock_lists(tdb, chain + 1, chain + num,
 744                                              F_WRLCK);
 745                                 return 1;
 746                         }
 747                 }
 748                 num += *extra_locks;
 749         }
 750
 751         /* Now we have the locks, be certain that offset is still 0! */
 752         hoff = tdb->header.v.hash_off
 753                 + (((chain + num) * sizeof(tdb_off_t))
 754                    & ((1ULL << tdb->header.v.hash_bits) - 1));
 755
 756         if (unlikely(tdb_read_off(tdb, hoff) != 0)) {
 757                 unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
 758                 goto again;
 759         }
 760
 761         /* OK, all locked.  Unlink first one. */
 762         hoff = tdb->header.v.hash_off + chain * sizeof(tdb_off_t);
 763         if (tdb_write_off(tdb, hoff, 0) == -1)
 764                 goto unlock_err;
 765
 766         /* Rehash the rest. */
 767         for (i = 1; i < num; i++) {
 768                 tdb_off_t off;
 769                 uint64_t h;
 770
 771                 hoff = tdb->header.v.hash_off
 772                         + (((chain + i) * sizeof(tdb_off_t))
 773                            & ((1ULL << tdb->header.v.hash_bits) - 1));
 774                 off = tdb_read_off(tdb, hoff);
 775                 if (unlikely(off == TDB_OFF_ERR))
 776                         goto unlock_err;
 777
 778                 /* Maybe use a bit to indicate it is in ideal place? */
 779                 h = hash_record(tdb, off);
 780                 /* Is it happy where it is? */
 781                 if ((h & ((1ULL << tdb->header.v.hash_bits)-1)) == (chain + i))
 782                         continue;
 783
 784                 /* Remove it. */
 785                 if (tdb_write_off(tdb, hoff, 0) == -1)
 786                         goto unlock_err;
 787
 788                 /* Rehash it. */
 789                 if (hash_add(tdb, h, off) == -1)
 790                         goto unlock_err;
 791         }
 792         unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
 793         return 0;
 794
 795 unlock_err:
 796         unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
 797         return -1;
 798 }
 799
 800 int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
 801 {
 802         tdb_off_t off, start, end, room, extra_locks = 0;
 803         uint64_t h;
 804         int ret;
 805
 806         h = tdb_hash(tdb, key.dptr, key.dsize);
 807         off = find_bucket_and_lock(tdb, &key, h, &start, &end, &room, F_WRLCK);
 808         if (off == TDB_OFF_ERR)
 809                 return -1;
 810
 811         if (off == 0) {
 812                 unlock_lists(tdb, start, end, F_WRLCK);
 813                 tdb->ecode = TDB_ERR_NOEXIST;
 814                 return -1;
 815         }
 816
 817         ret = unlink_used_record(tdb, end, &extra_locks);
 818         if (unlikely(ret == 1)) {
 819                 unsigned int i;
 820
 821                 unlock_lists(tdb, start, end, F_WRLCK);
 822
 823                 /* We need extra locks at the start. */
 824                 for (i = 0; i < extra_locks; i++) {
 825                         if (tdb_lock_list(tdb, i, F_WRLCK, TDB_LOCK_WAIT)
 826                             == TDB_OFF_ERR) {
 827                                 if (i)
 828                                         unlock_lists(tdb, 0, i-1, F_WRLCK);
 829                                 return -1;
 830                         }
 831                 }
 832                 /* Try again now we're holding more locks. */
 833                 ret = tdb_delete(tdb, key);
 834                 unlock_lists(tdb, 0, i, F_WRLCK);
 835                 return ret;
 836         }
 837         unlock_lists(tdb, start, end, F_WRLCK);
 838         return ret;
 839 }
 840
 841 int tdb_close(struct tdb_context *tdb)
 842 {
 843         struct tdb_context **i;
 844         int ret = 0;
 845
 846         /* FIXME:
 847         if (tdb->transaction) {
 848                 tdb_transaction_cancel(tdb);
 849         }
 850         */
 851         tdb_trace(tdb, "tdb_close");
 852
 853         if (tdb->map_ptr) {
 854                 if (tdb->flags & TDB_INTERNAL)
 855                         free(tdb->map_ptr);
 856                 else
 857                         tdb_munmap(tdb);
 858         }
 859         free((char *)tdb->name);
 860         if (tdb->fd != -1) {
 861                 ret = close(tdb->fd);
 862                 tdb->fd = -1;
 863         }
 864         free(tdb->lockrecs);
 865
 866         /* Remove from contexts list */
 867         for (i = &tdbs; *i; i = &(*i)->next) {
 868                 if (*i == tdb) {
 869                         *i = tdb->next;
 870                         break;
 871                 }
 872         }
 873
 874 #ifdef TDB_TRACE
 875         close(tdb->tracefd);
 876 #endif
 877         free(tdb);
 878
 879         return ret;
 880 }
 881
 882 enum TDB_ERROR tdb_error(struct tdb_context *tdb)
 883 {
 884         return tdb->ecode;
 885 }