return ret;
}
-struct new_database {
+struct new_db_head {
struct tdb_header hdr;
+ struct free_zone_header zhdr;
+ tdb_off_t free[BUCKETS_FOR_ZONE(INITIAL_ZONE_BITS) + 1];
struct tdb_used_record hrec;
tdb_off_t hash[1ULL << INITIAL_HASH_BITS];
- struct tdb_used_record frec;
- tdb_off_t free[INITIAL_FREE_BUCKETS + 1]; /* One overflow bucket */
+ struct tdb_free_record frec;
+};
+
+struct new_database {
+ struct new_db_head h;
+ /* Rest up to 1 << INITIAL_ZONE_BITS is empty. */
+ char space[(1 << INITIAL_ZONE_BITS)
+ - (sizeof(struct new_db_head) - sizeof(struct tdb_header))];
+ uint8_t tailer;
+ /* Don't count final padding! */
};
/* initialise a new database */
{
/* We make it up in memory, then write it out if not internal */
struct new_database newdb;
- unsigned int magic_off = offsetof(struct tdb_header, magic_food);
+ unsigned int bucket, magic_off, dbsize;
- /* Fill in the header */
- newdb.hdr.version = TDB_VERSION;
- newdb.hdr.hash_seed = random_number(tdb);
- newdb.hdr.hash_test = TDB_HASH_MAGIC;
- newdb.hdr.hash_test = tdb->khash(&newdb.hdr.hash_test,
- sizeof(newdb.hdr.hash_test),
- newdb.hdr.hash_seed,
- tdb->hash_priv);
-
- newdb.hdr.v.generation = 0;
-
- /* The initial zone must cover the initial database size! */
- BUILD_ASSERT((1ULL << INITIAL_ZONE_BITS) >= sizeof(newdb));
-
- /* Free array has 1 zone, 10 buckets. All buckets empty. */
- newdb.hdr.v.num_zones = 1;
- newdb.hdr.v.zone_bits = INITIAL_ZONE_BITS;
- newdb.hdr.v.free_buckets = INITIAL_FREE_BUCKETS;
- newdb.hdr.v.free_off = offsetof(struct new_database, free);
- set_header(tdb, &newdb.frec, 0,
- sizeof(newdb.free), sizeof(newdb.free), 0);
- memset(newdb.free, 0, sizeof(newdb.free));
+ /* Don't want any extra padding! */
+ dbsize = offsetof(struct new_database, tailer) + sizeof(newdb.tailer);
+ /* Fill in the header */
+ newdb.h.hdr.version = TDB_VERSION;
+ newdb.h.hdr.hash_seed = random_number(tdb);
+ newdb.h.hdr.hash_test = TDB_HASH_MAGIC;
+ newdb.h.hdr.hash_test = tdb->khash(&newdb.h.hdr.hash_test,
+ sizeof(newdb.h.hdr.hash_test),
+ newdb.h.hdr.hash_seed,
+ tdb->hash_priv);
+ memset(newdb.h.hdr.reserved, 0, sizeof(newdb.h.hdr.reserved));
+ newdb.h.hdr.v.generation = 0;
/* Initial hashes are empty. */
- newdb.hdr.v.hash_bits = INITIAL_HASH_BITS;
- newdb.hdr.v.hash_off = offsetof(struct new_database, hash);
- set_header(tdb, &newdb.hrec, 0,
- sizeof(newdb.hash), sizeof(newdb.hash), 0);
- memset(newdb.hash, 0, sizeof(newdb.hash));
+ newdb.h.hdr.v.hash_bits = INITIAL_HASH_BITS;
+ newdb.h.hdr.v.hash_off = offsetof(struct new_database, h.hash);
+ set_header(tdb, &newdb.h.hrec, 0,
+ sizeof(newdb.h.hash), sizeof(newdb.h.hash), 0,
+ INITIAL_ZONE_BITS);
+ memset(newdb.h.hash, 0, sizeof(newdb.h.hash));
+
+ /* Create the single free entry. */
+ newdb.h.frec.magic_and_meta = TDB_FREE_MAGIC | INITIAL_ZONE_BITS;
+ newdb.h.frec.data_len = (sizeof(newdb.h.frec)
+ - sizeof(struct tdb_used_record)
+ + sizeof(newdb.space));
+
+ /* Free is mostly empty... */
+ newdb.h.zhdr.zone_bits = INITIAL_ZONE_BITS;
+ memset(newdb.h.free, 0, sizeof(newdb.h.free));
+
+ /* ... except for this one bucket. */
+ bucket = size_to_bucket(INITIAL_ZONE_BITS, newdb.h.frec.data_len);
+ newdb.h.free[bucket] = offsetof(struct new_database, h.frec);
+ newdb.h.frec.next = newdb.h.frec.prev = 0;
+
+ /* Tailer contains maximum number of free_zone bits. */
+ newdb.tailer = INITIAL_ZONE_BITS;
/* Magic food */
- memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
- strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
+ memset(newdb.h.hdr.magic_food, 0, sizeof(newdb.h.hdr.magic_food));
+ strcpy(newdb.h.hdr.magic_food, TDB_MAGIC_FOOD);
/* This creates an endian-converted database, as if read from disk */
+ magic_off = offsetof(struct tdb_header, magic_food);
tdb_convert(tdb,
- (char *)&newdb.hdr + magic_off,
- sizeof(newdb) - magic_off);
+ (char *)&newdb.h.hdr + magic_off,
+ dbsize - 1 - magic_off);
- tdb->header = newdb.hdr;
+ tdb->header = newdb.h.hdr;
if (tdb->flags & TDB_INTERNAL) {
- tdb->map_size = sizeof(newdb);
+ tdb->map_size = dbsize;
tdb->map_ptr = malloc(tdb->map_size);
if (!tdb->map_ptr) {
tdb->ecode = TDB_ERR_OOM;
if (ftruncate(tdb->fd, 0) == -1)
return -1;
- if (!tdb_pwrite_all(tdb->fd, &newdb, sizeof(newdb), 0)) {
+ if (!tdb_pwrite_all(tdb->fd, &newdb, dbsize, 0)) {
tdb->ecode = TDB_ERR_IO;
return -1;
}
tdb->name = NULL;
tdb->map_ptr = NULL;
tdb->fd = -1;
- /* map_size will be set below. */
+ tdb->map_size = sizeof(struct tdb_header);
tdb->ecode = TDB_SUCCESS;
/* header will be read in below. */
tdb->header_uptodate = false;
}
TEST_IT(tdb->flags & TDB_CONVERT);
tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
- /* Zones don't matter for internal db. */
- tdb->last_zone = 0;
+ tdb_zone_init(tdb);
return tdb;
}
goto fail;
}
- tdb->map_size = st.st_size;
tdb->device = st.st_dev;
tdb->inode = st.st_ino;
- tdb_mmap(tdb);
tdb_unlock_open(tdb);
- tdb_zone_init(tdb);
+
+ /* This make sure we have current map_size and mmap. */
+ tdb->methods->oob(tdb, tdb->map_size + 1, true);
+
+ /* Now we can pick a random free zone to start from. */
+ if (tdb_zone_init(tdb) == -1)
+ goto fail;
tdb->next = tdbs;
tdbs = tdb;
return NULL;
}
-static tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
+tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
{
return tdb->header.v.hash_off
+ ((list & ((1ULL << tdb->header.v.hash_bits) - 1))
{
uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
- if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h))
+ if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h,
+ rec_zone_bits(rec)))
return -1;
return tdb_write_convert(tdb, off, rec, sizeof(*rec));
if (unlikely(num == len)) {
/* We wrapped. Look through start of hash table. */
+ i = 0;
hoff = hash_off(tdb, 0);
len = (1ULL << tdb->header.v.hash_bits);
num = tdb_find_zero_off(tdb, hoff, len);
- if (i == len) {
+ if (num == len) {
tdb->ecode = TDB_ERR_CORRUPT;
tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
"hash_add: full hash table!\n");
return -1;
}
}
+ if (tdb_read_off(tdb, hash_off(tdb, i + num)) != 0) {
+ tdb->ecode = TDB_ERR_CORRUPT;
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "hash_add: overwriting hash table?\n");
+ return -1;
+ }
+
/* FIXME: Encode extra hash bits! */
return tdb_write_off(tdb, hash_off(tdb, i + num), off);
}
tdb_len_t hlen;
uint64_t num = 1ULL << tdb->header.v.hash_bits;
struct tdb_used_record pad, *r;
+ unsigned int records = 0;
/* FIXME: We should do this without holding locks throughout. */
if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
goto oldheader;
if (off && hash_add(tdb, hash_record(tdb, off), off) == -1)
goto oldheader;
+ if (off)
+ records++;
}
+ tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
+ "enlarge_hash: moved %u records from %llu buckets.\n",
+ records, (long long)num);
+
/* Free up old hash. */
r = tdb_get(tdb, oldoff - sizeof(*r), &pad, sizeof(*r));
if (!r)
goto oldheader;
- add_free_record(tdb, oldoff - sizeof(*r),
+ add_free_record(tdb, rec_zone_bits(r), oldoff - sizeof(*r),
sizeof(*r)+rec_data_length(r)+rec_extra_padding(r));
/* Now we write the modified header. */
goto unlock;
}
+
+/* This is the slow version of the routine which searches the
+ * hashtable for an entry.
+ * We lock every hash bucket up to and including the next zero one.
+ */
+static tdb_off_t find_and_lock_slow(struct tdb_context *tdb,
+ struct tdb_data key,
+ uint64_t h,
+ int ltype,
+ tdb_off_t *start_lock,
+ tdb_len_t *num_locks,
+ tdb_off_t *bucket,
+ struct tdb_used_record *rec)
+{
+ /* Warning: this may drop the lock on *bucket! */
+ *num_locks = relock_hash_to_zero(tdb, *start_lock, ltype);
+ if (*num_locks == TDB_OFF_ERR)
+ return TDB_OFF_ERR;
+
+ for (*bucket = *start_lock;
+ *bucket < *start_lock + *num_locks;
+ (*bucket)++) {
+ tdb_off_t off = entry_matches(tdb, *bucket, h, &key, rec);
+ /* Empty entry or we found it? */
+ if (off == 0 || off != TDB_OFF_ERR)
+ return off;
+ }
+
+ /* We didn't find a zero entry? Something went badly wrong... */
+ unlock_lists(tdb, *start_lock, *start_lock + *num_locks, ltype);
+ tdb->ecode = TDB_ERR_CORRUPT;
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "find_and_lock: expected to find an empty hash bucket!\n");
+ return TDB_OFF_ERR;
+}
+
/* This is the core routine which searches the hashtable for an entry.
* On error, no locks are held and TDB_OFF_ERR is returned.
* Otherwise, *num_locks locks of type ltype from *start_lock are held.
}
/* Slow path, need to grab more locks and search. */
-
- /* Warning: this may drop the lock on *bucket! */
- *num_locks = relock_hash_to_zero(tdb, *start_lock, ltype);
- if (*num_locks == TDB_OFF_ERR)
- return TDB_OFF_ERR;
-
- for (*bucket = *start_lock;
- *bucket < *start_lock + *num_locks;
- (*bucket)++) {
- off = entry_matches(tdb, *bucket, h, &key, rec);
- /* Empty entry or we found it? */
- if (off == 0 || off != TDB_OFF_ERR)
- return off;
- }
-
- /* We didn't find a zero entry? Something went badly wrong... */
- unlock_lists(tdb, *start_lock, *start_lock + *num_locks, ltype);
- tdb->ecode = TDB_ERR_CORRUPT;
- tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
- "find_and_lock: expected to find an empty hash bucket!\n");
- return TDB_OFF_ERR;
+ return find_and_lock_slow(tdb, key, h, ltype, start_lock, num_locks,
+ bucket, rec);
}
/* Returns -1 on error, 0 on OK, 1 on "expand and retry." */
uint64_t h, struct tdb_data key, struct tdb_data dbuf,
tdb_off_t bucket,
tdb_off_t old_off, tdb_len_t old_room,
+ unsigned old_zone,
bool growing)
{
tdb_off_t new_off;
/* Allocate a new record. */
new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
- if (new_off == 0)
+ if (unlikely(new_off == TDB_OFF_ERR))
+ return -1;
+
+ if (unlikely(new_off == 0))
return 1;
/* We didn't like the existing one: remove it. */
if (old_off)
- add_free_record(tdb, old_off,
+ add_free_record(tdb, old_zone, old_off,
sizeof(struct tdb_used_record)
+ key.dsize + old_room);
}
/* If we didn't use the old record, this implies we're growing. */
- ret = replace_data(tdb, h, key, dbuf, bucket, off, old_room, off != 0);
+ ret = replace_data(tdb, h, key, dbuf, bucket, off, old_room,
+ rec_zone_bits(&rec), off != 0);
unlock_lists(tdb, start, num, F_WRLCK);
if (unlikely(ret == 1)) {
}
/* If they're using tdb_append(), it implies they're growing record. */
- ret = replace_data(tdb, h, key, new_dbuf, bucket, off, old_room, true);
+ ret = replace_data(tdb, h, key, new_dbuf, bucket, off, old_room,
+ rec_zone_bits(&rec), true);
unlock_lists(tdb, start, num, F_WRLCK);
free(newdata);
if (unlikely(num > 4 * tdb->header.v.hash_bits - 30))
enlarge_hash(tdb);
- return 0;
+ return ret;
fail:
unlock_lists(tdb, start, num, F_WRLCK);
uint64_t h;
h = tdb_hash(tdb, key.dptr, key.dsize);
- off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
+ start = tdb_lock_list(tdb, h, F_WRLCK, TDB_LOCK_WAIT);
+ if (unlikely(start == TDB_OFF_ERR))
+ return -1;
+
+ /* FIXME: Fastpath: if next is zero, we can delete without lock,
+ * since this lock protects us. */
+ off = find_and_lock_slow(tdb, key, h, F_WRLCK,
+ &start, &num, &bucket, &rec);
if (unlikely(off == TDB_OFF_ERR))
return -1;
if (!off) {
+ /* FIXME: We could optimize not found case if it mattered, by
+ * reading offset after first lock: if it's zero, goto here. */
unlock_lists(tdb, start, num, F_WRLCK);
tdb->ecode = TDB_ERR_NOEXIST;
return -1;
}
+ /* Since we found the entry, we must have locked it and a zero. */
+ assert(num >= 2);
/* This actually unlinks it. */
if (tdb_write_off(tdb, hash_off(tdb, bucket), 0) == -1)
goto unlock_err;
/* Rehash anything following. */
- for (i = hash_off(tdb, bucket+1);
- i != hash_off(tdb, h + num - 1);
- i += sizeof(tdb_off_t)) {
- tdb_off_t off2;
+ for (i = bucket+1; i != bucket + num - 1; i++) {
+ tdb_off_t hoff, off2;
uint64_t h2;
- off2 = tdb_read_off(tdb, i);
+ hoff = hash_off(tdb, i);
+ off2 = tdb_read_off(tdb, hoff);
if (unlikely(off2 == TDB_OFF_ERR))
goto unlock_err;
/* Maybe use a bit to indicate it is in ideal place? */
h2 = hash_record(tdb, off2);
/* Is it happy where it is? */
- if (hash_off(tdb, h2) == i)
+ if (hash_off(tdb, h2) == hoff)
continue;
/* Remove it. */
- if (tdb_write_off(tdb, i, 0) == -1)
+ if (tdb_write_off(tdb, hoff, 0) == -1)
goto unlock_err;
/* Rehash it. */
}
/* Free the deleted entry. */
- if (add_free_record(tdb, off,
+ if (add_free_record(tdb, rec_zone_bits(&rec), off,
sizeof(struct tdb_used_record)
+ rec_key_length(&rec)
+ rec_data_length(&rec)