uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off)
{
- struct tdb_used_record pad, *r;
+ const struct tdb_used_record *r;
const void *key;
uint64_t klen, hash;
- r = tdb_get(tdb, off, &pad, sizeof(pad));
+ r = tdb_access_read(tdb, off, sizeof(*r), true);
if (!r)
/* FIXME */
return 0;
klen = rec_key_length(r);
- key = tdb_access_read(tdb, off + sizeof(pad), klen, false);
+ tdb_access_release(tdb, r);
+
+ key = tdb_access_read(tdb, off + sizeof(*r), klen, false);
if (!key)
return 0;
return val >> (64-TDB_OFF_UPPER_STEAL) == (1<<TDB_OFF_UPPER_STEAL) - 1;
}
+/* FIXME: Guess the depth, don't over-lock! */
+static tdb_off_t hlock_range(tdb_off_t group, tdb_off_t *size)
+{
+ *size = 1ULL << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
+ return group << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
+}
+
/* 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, hinfo is filled in (and the optional tinfo).
group = use_bits(h, TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS);
h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
- /* FIXME: Guess the depth, don't over-lock! */
- h->hlock_start = (tdb_off_t)group
- << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
- h->hlock_range = 1ULL << (64 - (TDB_TOPLEVEL_HASH_BITS
- - TDB_HASH_GROUP_BITS));
+ h->hlock_start = hlock_range(group, &h->hlock_range);
if (tdb_lock_hashes(tdb, h->hlock_start, h->hlock_range, ltype,
TDB_LOCK_WAIT))
return TDB_OFF_ERR;
if (subhash == TDB_OFF_ERR)
return -1;
+ add_stat(tdb, alloc_subhash, 1);
if (zero_out(tdb, subhash + sizeof(struct tdb_used_record),
sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS) == -1)
return -1;
/* Return 1 if we find something, 0 if not, -1 on error. */
int next_in_hash(struct tdb_context *tdb, int ltype,
struct traverse_info *tinfo,
- TDB_DATA *kbuf, unsigned int *dlen)
+ TDB_DATA *kbuf, size_t *dlen)
{
const unsigned group_bits = TDB_TOPLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS;
tdb_off_t hlock_start, hlock_range, off;
ltype);
return -1;
}
+ if (rec_magic(&rec) != TDB_MAGIC) {
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "next_in_hash:"
+ " corrupt record at %llu\n",
+ (long long)off);
+ return -1;
+ }
+
kbuf->dsize = rec_key_length(&rec);
/* They want data as well? */
/* Return 1 if we find something, 0 if not, -1 on error. */
int first_in_hash(struct tdb_context *tdb, int ltype,
struct traverse_info *tinfo,
- TDB_DATA *kbuf, unsigned int *dlen)
+ TDB_DATA *kbuf, size_t *dlen)
{
tinfo->prev = 0;
tinfo->toplevel_group = 0;
return next_in_hash(tdb, ltype, tinfo, kbuf, dlen);
}
-TDB_DATA tdb_firstkey(struct tdb_context *tdb)
+/* Even if the entry isn't in this hash bucket, you'd have to lock this
+ * bucket to find it. */
+static int chainlock(struct tdb_context *tdb, const TDB_DATA *key,
+ int ltype, enum tdb_lock_flags waitflag,
+ const char *func)
{
- struct traverse_info tinfo;
- struct tdb_data k;
- switch (first_in_hash(tdb, F_RDLCK, &tinfo, &k, NULL)) {
- case 1:
- return k;
- case 0:
- tdb->ecode = TDB_SUCCESS;
- /* Fall thru... */
- default:
- return tdb_null;
- }
-}
+ int ret;
+ uint64_t h = tdb_hash(tdb, key->dptr, key->dsize);
+ tdb_off_t lockstart, locksize;
+ unsigned int group, gbits;
+
+ gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS;
+ group = bits(h, 64 - gbits, gbits);
+
+ lockstart = hlock_range(group, &locksize);
+
+ ret = tdb_lock_hashes(tdb, lockstart, locksize, ltype, waitflag);
+ tdb_trace_1rec(tdb, func, *key);
+ return ret;
+}
-/* We lock twice, not very efficient. We could keep last key & tinfo cached. */
-TDB_DATA tdb_nextkey(struct tdb_context *tdb, TDB_DATA key)
+/* lock/unlock one hash chain. This is meant to be used to reduce
+ contention - it cannot guarantee how many records will be locked */
+int tdb_chainlock(struct tdb_context *tdb, TDB_DATA key)
{
- struct traverse_info tinfo;
- struct hash_info h;
- struct tdb_used_record rec;
-
- tinfo.prev = find_and_lock(tdb, key, F_RDLCK, &h, &rec, &tinfo);
- if (unlikely(tinfo.prev == TDB_OFF_ERR))
- return tdb_null;
- tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_RDLCK);
-
- switch (next_in_hash(tdb, F_RDLCK, &tinfo, &key, NULL)) {
- case 1:
- return key;
- case 0:
- tdb->ecode = TDB_SUCCESS;
- /* Fall thru... */
- default:
- return tdb_null;
- }
-}
+ return chainlock(tdb, &key, F_WRLCK, TDB_LOCK_WAIT, "tdb_chainlock");
+}
+
+int tdb_chainunlock(struct tdb_context *tdb, TDB_DATA key)
+{
+ uint64_t h = tdb_hash(tdb, key.dptr, key.dsize);
+ tdb_off_t lockstart, locksize;
+ unsigned int group, gbits;
+
+ gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS;
+ group = bits(h, 64 - gbits, gbits);
+
+ lockstart = hlock_range(group, &locksize);
+
+ tdb_trace_1rec(tdb, "tdb_chainunlock", key);
+ return tdb_unlock_hashes(tdb, lockstart, locksize, F_WRLCK);
+}
projects / ccan / blobdiff