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>
9 struct tdb_data tdb_null = { .dptr = NULL, .dsize = 0 };
11 /* all contexts, to ensure no double-opens (fcntl locks don't nest!) */
12 static struct tdb_context *tdbs = NULL;
14 PRINTF_ATTRIBUTE(4, 5) static void
15 null_log_fn(struct tdb_context *tdb,
16 enum tdb_debug_level level, void *priv,
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
24 bool header_changed(struct tdb_context *tdb)
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");
33 /* We could get a partial update if we're not holding any locks. */
34 assert((tdb->flags & TDB_NOLOCK) || tdb_has_locks(tdb));
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));
46 int write_header(struct tdb_context *tdb)
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));
55 static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
58 return hash64_stable((const unsigned char *)key, length, seed);
61 uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
63 return tdb->khash(ptr, len, tdb->header.hash_seed, tdb->hash_priv);
66 static bool tdb_already_open(dev_t device, ino_t ino)
68 struct tdb_context *i;
70 for (i = tdbs; i; i = i->next) {
71 if (i->device == device && i->inode == ino) {
79 static uint64_t random_number(struct tdb_context *tdb)
85 fd = open("/dev/urandom", O_RDONLY);
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");
95 /* FIXME: Untested! Based on Wikipedia protocol description! */
96 fd = open("/dev/egd-pool", O_RDWR);
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));
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)) {
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");
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 */
135 /* initialise a new database */
136 static int tdb_new_database(struct tdb_context *tdb)
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);
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),
151 newdb.hdr.v.generation = 0;
153 /* The initial zone must cover the initial database size! */
154 BUILD_ASSERT((1ULL << INITIAL_ZONE_BITS) >= sizeof(newdb));
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));
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));
173 memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
174 strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
176 /* This creates an endian-converted database, as if read from disk */
178 (char *)&newdb.hdr + magic_off,
179 sizeof(newdb) - magic_off);
181 tdb->header = newdb.hdr;
183 if (tdb->flags & TDB_INTERNAL) {
184 tdb->map_size = sizeof(newdb);
185 tdb->map_ptr = malloc(tdb->map_size);
187 tdb->ecode = TDB_ERR_OOM;
190 memcpy(tdb->map_ptr, &newdb, tdb->map_size);
193 if (lseek(tdb->fd, 0, SEEK_SET) == -1)
196 if (ftruncate(tdb->fd, 0) == -1)
199 if (!tdb_pwrite_all(tdb->fd, &newdb, sizeof(newdb), 0)) {
200 tdb->ecode = TDB_ERR_IO;
206 struct tdb_context *tdb_open(const char *name, int tdb_flags,
207 int open_flags, mode_t mode,
208 union tdb_attribute *attr)
210 struct tdb_context *tdb;
216 tdb = malloc(sizeof(*tdb));
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. */
241 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
242 "tdb_open: attributes not yet supported\n");
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);
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;
259 tdb->read_only = false;
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!");
269 TEST_IT(tdb->flags & TDB_CONVERT);
270 tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
271 /* Zones don't matter for internal db. */
276 if ((tdb->fd = open(name, open_flags, mode)) == -1) {
277 tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
278 "tdb_open: could not open file %s: %s\n",
279 name, strerror(errno));
280 goto fail; /* errno set by open(2) */
283 /* on exec, don't inherit the fd */
284 v = fcntl(tdb->fd, F_GETFD, 0);
285 fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
287 /* ensure there is only one process initialising at once */
288 if (tdb_lock_open(tdb) == -1) {
289 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
290 "tdb_open: failed to get open lock on %s: %s\n",
291 name, strerror(errno));
292 goto fail; /* errno set by tdb_brlock */
295 if (!tdb_pread_all(tdb->fd, &tdb->header, sizeof(tdb->header), 0)
296 || strcmp(tdb->header.magic_food, TDB_MAGIC_FOOD) != 0) {
297 if (!(open_flags & O_CREAT) || tdb_new_database(tdb) == -1) {
299 errno = EIO; /* ie bad format or something */
303 } else if (tdb->header.version != TDB_VERSION) {
304 if (tdb->header.version == bswap_64(TDB_VERSION))
305 tdb->flags |= TDB_CONVERT;
308 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
309 "tdb_open: %s is unknown version 0x%llx\n",
310 name, (long long)tdb->header.version);
316 tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
317 hash_test = TDB_HASH_MAGIC;
318 hash_test = tdb->khash(&hash_test, sizeof(hash_test),
319 tdb->header.hash_seed, tdb->hash_priv);
320 if (tdb->header.hash_test != hash_test) {
321 /* wrong hash variant */
322 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
323 "tdb_open: %s uses a different hash function\n",
329 if (fstat(tdb->fd, &st) == -1)
332 /* Is it already in the open list? If so, fail. */
333 if (tdb_already_open(st.st_dev, st.st_ino)) {
335 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
336 "tdb_open: %s (%d,%d) is already open in this process\n",
337 name, (int)st.st_dev, (int)st.st_ino);
342 tdb->name = strdup(name);
348 tdb->map_size = st.st_size;
349 tdb->device = st.st_dev;
350 tdb->inode = st.st_ino;
352 tdb_unlock_open(tdb);
369 if (tdb->flags & TDB_INTERNAL) {
374 free((char *)tdb->name);
376 if (close(tdb->fd) != 0)
377 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
378 "tdb_open: failed to close tdb->fd"
385 static tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
387 return tdb->header.v.hash_off
388 + ((list & ((1ULL << tdb->header.v.hash_bits) - 1))
389 * sizeof(tdb_off_t));
392 /* Returns 0 if the entry is a zero (definitely not a match).
393 * Returns a valid entry offset if it's a match. Fills in rec.
394 * Otherwise returns TDB_OFF_ERR: keep searching. */
395 static tdb_off_t entry_matches(struct tdb_context *tdb,
398 const struct tdb_data *key,
399 struct tdb_used_record *rec)
403 const unsigned char *rkey;
405 off = tdb_read_off(tdb, tdb->header.v.hash_off
406 + list * sizeof(tdb_off_t));
407 if (off == 0 || off == TDB_OFF_ERR)
410 #if 0 /* FIXME: Check other bits. */
411 unsigned int bits, bitmask, hoffextra;
412 /* Bottom three bits show how many extra hash bits. */
413 bits = (off & ((1 << TDB_EXTRA_HASHBITS_NUM) - 1)) + 1;
414 bitmask = (1 << bits)-1;
415 hoffextra = ((off >> TDB_EXTRA_HASHBITS_NUM) & bitmask);
416 uint64_t hextra = hash >> tdb->header.v.hash_bits;
417 if ((hextra & bitmask) != hoffextra)
422 if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
425 /* FIXME: check extra bits in header! */
426 keylen = rec_key_length(rec);
427 if (keylen != key->dsize)
430 rkey = tdb_access_read(tdb, off + sizeof(*rec), keylen, false);
433 if (memcmp(rkey, key->dptr, keylen) != 0)
435 tdb_access_release(tdb, rkey);
439 /* FIXME: Optimize? */
440 static void unlock_lists(struct tdb_context *tdb,
441 tdb_off_t list, tdb_len_t num,
446 for (i = list; i < list + num; i++)
447 tdb_unlock_list(tdb, i, ltype);
450 /* FIXME: Optimize? */
451 static int lock_lists(struct tdb_context *tdb,
452 tdb_off_t list, tdb_len_t num,
457 for (i = list; i < list + num; i++) {
458 if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT) != 0) {
459 unlock_lists(tdb, list, i - list, ltype);
466 /* We lock hashes up to the next empty offset. We already hold the
467 * lock on the start bucket, but we may need to release and re-grab
468 * it. If we fail, we hold no locks at all! */
469 static tdb_len_t relock_hash_to_zero(struct tdb_context *tdb,
470 tdb_off_t start, int ltype)
472 tdb_len_t num, len, pre_locks;
475 num = 1ULL << tdb->header.v.hash_bits;
476 len = tdb_find_zero_off(tdb, hash_off(tdb, start), num - start);
477 if (unlikely(len == num - start)) {
478 /* We hit the end of the hash range. Drop lock: we have
479 to lock start of hash first. */
480 tdb_unlock_list(tdb, start, ltype);
481 /* Grab something, so header is stable. */
482 if (tdb_lock_list(tdb, 0, ltype, TDB_LOCK_WAIT))
484 len = tdb_find_zero_off(tdb, hash_off(tdb, 0), num);
485 if (lock_lists(tdb, 1, len, ltype) == -1) {
486 tdb_unlock_list(tdb, 0, ltype);
492 /* We already have lock on start. */
496 if (unlikely(lock_lists(tdb, start, len, ltype) == -1)) {
498 unlock_lists(tdb, 0, pre_locks, ltype);
500 tdb_unlock_list(tdb, start, ltype);
504 /* Now, did we lose the race, and it's not zero any more? */
505 if (unlikely(tdb_read_off(tdb, hash_off(tdb, pre_locks + len)) != 0)) {
506 unlock_lists(tdb, 0, pre_locks, ltype);
507 /* Leave the start locked, as expected. */
508 unlock_lists(tdb, start + 1, len - 1, ltype);
512 return pre_locks + len;
515 /* FIXME: modify, don't rewrite! */
516 static int update_rec_hdr(struct tdb_context *tdb,
520 struct tdb_used_record *rec,
523 uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
525 if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h))
528 return tdb_write_convert(tdb, off, rec, sizeof(*rec));
531 /* If we fail, others will try after us. */
532 static void enlarge_hash(struct tdb_context *tdb)
534 tdb_off_t newoff, oldoff, i;
536 uint64_t h, num = 1ULL << tdb->header.v.hash_bits;
537 struct tdb_used_record pad, *r;
539 /* FIXME: We should do this without holding locks throughout. */
540 if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
543 /* Someone else enlarged for us? Nothing to do. */
544 if ((1ULL << tdb->header.v.hash_bits) != num)
547 /* Allocate our new array. */
548 hlen = num * sizeof(tdb_off_t) * 2;
549 newoff = alloc(tdb, 0, hlen, 0, false);
550 if (unlikely(newoff == TDB_OFF_ERR))
552 if (unlikely(newoff == 0)) {
553 if (tdb_expand(tdb, 0, hlen, false) == -1)
555 newoff = alloc(tdb, 0, hlen, 0, false);
556 if (newoff == TDB_OFF_ERR || newoff == 0)
559 /* Step over record header! */
560 newoff += sizeof(struct tdb_used_record);
562 /* Starts all zero. */
563 if (zero_out(tdb, newoff, hlen) == -1)
566 /* FIXME: If the space before is empty, we know this is in its ideal
567 * location. Or steal a bit from the pointer to avoid rehash. */
568 for (i = tdb_find_nonzero_off(tdb, hash_off(tdb, 0), num);
570 i += tdb_find_nonzero_off(tdb, hash_off(tdb, i), num - i)) {
572 off = tdb_read_off(tdb, hash_off(tdb, i));
573 if (unlikely(off == TDB_OFF_ERR))
575 if (unlikely(!off)) {
576 tdb->ecode = TDB_ERR_CORRUPT;
577 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
578 "find_bucket_and_lock: zero hash bucket!\n");
581 h = hash_record(tdb, off);
582 /* FIXME: Encode extra hash bits! */
583 if (tdb_write_off(tdb, newoff
584 + (h & ((num * 2) - 1)) * sizeof(uint64_t),
589 /* Free up old hash. */
590 oldoff = tdb->header.v.hash_off - sizeof(*r);
591 r = tdb_get(tdb, oldoff, &pad, sizeof(*r));
594 add_free_record(tdb, oldoff,
595 sizeof(*r)+rec_data_length(r)+rec_extra_padding(r));
597 /* Now we write the modified header. */
598 tdb->header.v.hash_bits++;
599 tdb->header.v.hash_off = newoff;
602 tdb_allrecord_unlock(tdb, F_WRLCK);
605 int tdb_store(struct tdb_context *tdb,
606 struct tdb_data key, struct tdb_data dbuf, int flag)
608 tdb_off_t new_off, off, old_bucket, start, num_locks = 1;
609 struct tdb_used_record rec;
611 bool growing = false;
613 h = tdb_hash(tdb, key.dptr, key.dsize);
615 /* FIXME: can we avoid locks for some fast paths? */
616 start = tdb_lock_list(tdb, h, F_WRLCK, TDB_LOCK_WAIT);
617 if (start == TDB_OFF_ERR)
622 off = entry_matches(tdb, start, h, &key, &rec);
623 if (unlikely(off == TDB_OFF_ERR)) {
624 /* Slow path, need to grab more locks and search. */
627 /* Warning: this may drop the lock! Does that on error. */
628 num_locks = relock_hash_to_zero(tdb, start, F_WRLCK);
629 if (num_locks == TDB_OFF_ERR)
632 for (i = start; i < start + num_locks; i++) {
633 off = entry_matches(tdb, i, h, &key, &rec);
634 /* Empty entry or we found it? */
635 if (off == 0 || off != TDB_OFF_ERR) {
640 if (i == start + num_locks)
644 /* Now we have lock on this hash bucket. */
645 if (flag == TDB_INSERT) {
647 tdb->ecode = TDB_ERR_EXISTS;
652 if (rec_data_length(&rec) + rec_extra_padding(&rec)
655 if (update_rec_hdr(tdb, off,
656 key.dsize, dbuf.dsize,
661 /* FIXME: See if right record is free? */
662 /* Hint to allocator that we've realloced. */
665 if (flag == TDB_MODIFY) {
666 /* if the record doesn't exist and we
667 are in TDB_MODIFY mode then we should fail
669 tdb->ecode = TDB_ERR_NOEXIST;
675 /* Allocate a new record. */
676 new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
678 unlock_lists(tdb, start, num_locks, F_WRLCK);
679 /* Expand, then try again... */
680 if (tdb_expand(tdb, key.dsize, dbuf.dsize, growing) == -1)
682 return tdb_store(tdb, key, dbuf, flag);
685 /* We didn't like the existing one: remove it. */
687 add_free_record(tdb, off, sizeof(struct tdb_used_record)
688 + rec_key_length(&rec)
689 + rec_data_length(&rec)
690 + rec_extra_padding(&rec));
693 /* FIXME: Encode extra hash bits! */
694 if (tdb_write_off(tdb, hash_off(tdb, old_bucket), new_off) == -1)
698 off = new_off + sizeof(struct tdb_used_record);
699 if (tdb->methods->write(tdb, off, key.dptr, key.dsize) == -1)
702 if (tdb->methods->write(tdb, off, dbuf.dptr, dbuf.dsize) == -1)
705 /* FIXME: tdb_increment_seqnum(tdb); */
706 unlock_lists(tdb, start, num_locks, F_WRLCK);
708 /* FIXME: by simple simulation, this approximated 60% full.
709 * Check in real case! */
710 if (unlikely(num_locks > 4 * tdb->header.v.hash_bits - 31))
716 unlock_lists(tdb, start, num_locks, F_WRLCK);
720 struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
722 tdb_off_t off, start, num_locks = 1;
723 struct tdb_used_record rec;
727 h = tdb_hash(tdb, key.dptr, key.dsize);
729 /* FIXME: can we avoid locks for some fast paths? */
730 start = tdb_lock_list(tdb, h, F_RDLCK, TDB_LOCK_WAIT);
731 if (start == TDB_OFF_ERR)
735 off = entry_matches(tdb, start, h, &key, &rec);
736 if (unlikely(off == TDB_OFF_ERR)) {
737 /* Slow path, need to grab more locks and search. */
740 /* Warning: this may drop the lock! Does that on error. */
741 num_locks = relock_hash_to_zero(tdb, start, F_RDLCK);
742 if (num_locks == TDB_OFF_ERR)
745 for (i = start; i < start + num_locks; i++) {
746 off = entry_matches(tdb, i, h, &key, &rec);
747 /* Empty entry or we found it? */
748 if (off == 0 || off != TDB_OFF_ERR)
751 if (i == start + num_locks)
756 unlock_lists(tdb, start, num_locks, F_RDLCK);
757 tdb->ecode = TDB_ERR_NOEXIST;
761 ret.dsize = rec_data_length(&rec);
762 ret.dptr = tdb_alloc_read(tdb, off + sizeof(rec) + key.dsize,
764 unlock_lists(tdb, start, num_locks, F_RDLCK);
768 static int hash_add(struct tdb_context *tdb, uint64_t h, tdb_off_t off)
770 tdb_off_t i, hoff, len, num;
772 /* Look for next space. */
773 i = (h & ((1ULL << tdb->header.v.hash_bits) - 1));
774 len = (1ULL << tdb->header.v.hash_bits) - i;
775 num = tdb_find_zero_off(tdb, hash_off(tdb, i), len);
777 if (unlikely(num == len)) {
778 /* We wrapped. Look through start of hash table. */
779 hoff = hash_off(tdb, 0);
780 len = (1ULL << tdb->header.v.hash_bits);
781 num = tdb_find_zero_off(tdb, hoff, len);
783 tdb->ecode = TDB_ERR_CORRUPT;
784 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
785 "hash_add: full hash table!\n");
789 /* FIXME: Encode extra hash bits! */
790 return tdb_write_off(tdb, hash_off(tdb, i + num), off);
793 int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
795 tdb_off_t i, old_bucket, off, start, num_locks = 1;
796 struct tdb_used_record rec;
799 h = tdb_hash(tdb, key.dptr, key.dsize);
801 /* FIXME: can we avoid locks for some fast paths? */
802 start = tdb_lock_list(tdb, h, F_WRLCK, TDB_LOCK_WAIT);
803 if (start == TDB_OFF_ERR)
808 off = entry_matches(tdb, start, h, &key, &rec);
809 if (off && off != TDB_OFF_ERR) {
810 /* We can only really fastpath delete if next bucket
811 * is 0. Note that we haven't locked it, but our lock
812 * on this bucket stops anyone overflowing into it
814 if (tdb_read_off(tdb, hash_off(tdb, h+1)) == 0)
820 if (unlikely(off == TDB_OFF_ERR)) {
821 /* Slow path, need to grab more locks and search. */
824 /* Warning: this may drop the lock! Does that on error. */
825 num_locks = relock_hash_to_zero(tdb, start, F_WRLCK);
826 if (num_locks == TDB_OFF_ERR)
829 for (i = start; i < start + num_locks; i++) {
830 off = entry_matches(tdb, i, h, &key, &rec);
831 /* Empty entry or we found it? */
832 if (off == 0 || off != TDB_OFF_ERR) {
837 if (i == start + num_locks)
842 unlock_lists(tdb, start, num_locks, F_WRLCK);
843 tdb->ecode = TDB_ERR_NOEXIST;
848 /* This actually unlinks it. */
849 if (tdb_write_off(tdb, hash_off(tdb, old_bucket), 0) == -1)
852 /* Rehash anything following. */
853 for (i = hash_off(tdb, old_bucket+1);
854 i != hash_off(tdb, h + num_locks);
855 i += sizeof(tdb_off_t)) {
859 off2 = tdb_read_off(tdb, i);
860 if (unlikely(off2 == TDB_OFF_ERR))
863 /* Maybe use a bit to indicate it is in ideal place? */
864 h2 = hash_record(tdb, off2);
865 /* Is it happy where it is? */
866 if (hash_off(tdb, h2) == i)
870 if (tdb_write_off(tdb, i, 0) == -1)
874 if (hash_add(tdb, h2, off2) == -1)
878 /* Free the deleted entry. */
879 if (add_free_record(tdb, off,
880 sizeof(struct tdb_used_record)
881 + rec_key_length(&rec)
882 + rec_data_length(&rec)
883 + rec_extra_padding(&rec)) != 0)
886 unlock_lists(tdb, start, num_locks, F_WRLCK);
890 unlock_lists(tdb, start, num_locks, F_WRLCK);
894 int tdb_close(struct tdb_context *tdb)
896 struct tdb_context **i;
900 if (tdb->transaction) {
901 tdb_transaction_cancel(tdb);
904 tdb_trace(tdb, "tdb_close");
907 if (tdb->flags & TDB_INTERNAL)
912 free((char *)tdb->name);
914 ret = close(tdb->fd);
919 /* Remove from contexts list */
920 for (i = &tdbs; *i; i = &(*i)->next) {
935 enum TDB_ERROR tdb_error(struct tdb_context *tdb)