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;
tdb_off_t val,
struct tdb_used_record *rec)
{
- bool ret;
+ bool ret = false;
const unsigned char *rkey;
tdb_off_t off;
- /* FIXME: Handle hash value truncated. */
- if (bits(val, TDB_OFF_HASH_TRUNCATED_BIT, 1))
- abort();
-
+ add_stat(tdb, compares, 1);
/* Desired bucket must match. */
- if (h->home_bucket != (val & TDB_OFF_HASH_GROUP_MASK))
- return false;
+ if (h->home_bucket != (val & TDB_OFF_HASH_GROUP_MASK)) {
+ add_stat(tdb, compare_wrong_bucket, 1);
+ return ret;
+ }
/* Top bits of offset == next bits of hash. */
if (bits(val, TDB_OFF_HASH_EXTRA_BIT, TDB_OFF_UPPER_STEAL_EXTRA)
!= bits(h->h, 64 - h->hash_used - TDB_OFF_UPPER_STEAL_EXTRA,
- TDB_OFF_UPPER_STEAL_EXTRA))
- return false;
+ TDB_OFF_UPPER_STEAL_EXTRA)) {
+ add_stat(tdb, compare_wrong_offsetbits, 1);
+ return ret;
+ }
off = val & TDB_OFF_MASK;
if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
- return false;
+ return ret;
- /* FIXME: check extra bits in header? */
- if (rec_key_length(rec) != key->dsize)
+ if (rec_key_length(rec) != key->dsize) {
+ add_stat(tdb, compare_wrong_keylen, 1);
+ return ret;
+ }
+
+ if ((h->h & ((1 << 11)-1)) != rec_hash(rec)) {
+ add_stat(tdb, compare_wrong_rechash, 1);
return false;
+ }
rkey = tdb_access_read(tdb, off + sizeof(*rec), key->dsize, false);
if (!rkey)
- return false;
- ret = (memcmp(rkey, key->dptr, key->dsize) == 0);
+ return ret;
+ if (memcmp(rkey, key->dptr, key->dsize) == 0)
+ ret = true;
+ else
+ add_stat(tdb, compare_wrong_keycmp, 1);
tdb_access_release(tdb, rkey);
return ret;
}
+ (bucket % (1 << TDB_HASH_GROUP_BITS)) * sizeof(tdb_off_t);
}
-/* Truncated hashes can't be all 1: that's how we spot a sub-hash */
bool is_subhash(tdb_off_t val)
{
- return val >> (64-TDB_OFF_UPPER_STEAL) == (1<<TDB_OFF_UPPER_STEAL) - 1;
+ return (val >> TDB_OFF_UPPER_STEAL_SUBHASH_BIT) & 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.
+ * Otherwise, hinfo is filled in (and the optional tinfo).
* If not found, the return value is 0.
* If found, the return value is the offset, and *rec is the record. */
tdb_off_t find_and_lock(struct tdb_context *tdb,
struct tdb_data key,
int ltype,
struct hash_info *h,
- struct tdb_used_record *rec)
+ struct tdb_used_record *rec,
+ struct traverse_info *tinfo)
{
uint32_t i, group;
tdb_off_t hashtable;
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;
hashtable = offsetof(struct tdb_header, hashtable);
+ if (tinfo) {
+ tinfo->toplevel_group = group;
+ tinfo->num_levels = 1;
+ tinfo->levels[0].entry = 0;
+ tinfo->levels[0].hashtable = hashtable
+ + (group << TDB_HASH_GROUP_BITS) * sizeof(tdb_off_t);
+ tinfo->levels[0].total_buckets = 1 << TDB_HASH_GROUP_BITS;
+ }
while (likely(h->hash_used < 64)) {
/* Read in the hash group. */
if (is_subhash(h->group[h->home_bucket])) {
hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK)
+ sizeof(struct tdb_used_record);
+ if (tinfo) {
+ /* When we come back, use *next* bucket */
+ tinfo->levels[tinfo->num_levels-1].entry
+ += h->home_bucket + 1;
+ }
group = use_bits(h, TDB_SUBLEVEL_HASH_BITS
- TDB_HASH_GROUP_BITS);
h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
+ if (tinfo) {
+ tinfo->levels[tinfo->num_levels].hashtable
+ = hashtable;
+ tinfo->levels[tinfo->num_levels].total_buckets
+ = 1 << TDB_SUBLEVEL_HASH_BITS;
+ tinfo->levels[tinfo->num_levels].entry
+ = group << TDB_HASH_GROUP_BITS;
+ tinfo->num_levels++;
+ }
continue;
}
if (!h->group[h->found_bucket])
break;
- if (match(tdb, h, &key, h->group[h->found_bucket], rec))
+ if (match(tdb, h, &key, h->group[h->found_bucket],
+ rec)) {
+ if (tinfo) {
+ tinfo->levels[tinfo->num_levels-1].entry
+ += h->found_bucket;
+ }
return h->group[h->found_bucket] & TDB_OFF_MASK;
+ }
}
/* Didn't find it: h indicates where it would go. */
return 0;
/* I wrote a simple test, expanding a hash to 2GB, for the following
* cases:
* 1) Expanding all the buckets at once,
- * 2) Expanding the most-populated bucket,
- * 3) Expanding the bucket we wanted to place the new entry ito.
+ * 2) Expanding the bucket we wanted to place the new entry into.
+ * 3) Expanding the most-populated bucket,
*
* I measured the worst/average/best density during this process.
* 1) 3%/16%/30%
if (hash_used + TDB_SUBLEVEL_HASH_BITS > 64)
abort();
- /* FIXME: Do truncated hash bits if we can! */
h.h = hash_record(tdb, off);
gnum = use_bits(&h, TDB_SUBLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS);
h.group_start = subhash + sizeof(struct tdb_used_record)
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;
/* assert(num_vals); */
/* Overwrite expanded bucket with subhash pointer. */
- h->group[bucket] = subhash | ~((1ULL << (64 - TDB_OFF_UPPER_STEAL))-1);
+ h->group[bucket] = subhash | (1ULL << TDB_OFF_UPPER_STEAL_SUBHASH_BIT);
/* Put values back. */
for (i = 0; i < num_vals; i++) {
if (unlikely(val == TDB_OFF_ERR))
return TDB_OFF_ERR;
+ off = val & TDB_OFF_MASK;
+
/* This makes the delete-all-in-traverse case work
* (and simplifies our logic a little). */
- if (val == tinfo->prev)
+ if (off == tinfo->prev)
continue;
tlevel->entry = i;
- off = val & TDB_OFF_MASK;
if (!is_subhash(val)) {
/* Found one. */
- tinfo->prev = val;
+ tinfo->prev = off;
return off;
}
/* 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_logerr(tdb, TDB_ERR_CORRUPT,
+ TDB_DEBUG_FATAL,
+ "next_in_hash:"
+ " corrupt record at %llu",
+ (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);
}
+
+/* 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)
+{
+ 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;
+}
+
+/* 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)
+{
+ 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);
+}