summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--source4/lib/tdb/common/tdb.c37
1 files changed, 37 insertions, 0 deletions
diff --git a/source4/lib/tdb/common/tdb.c b/source4/lib/tdb/common/tdb.c
index 4c2d9a1add..8e8e3ce3b3 100644
--- a/source4/lib/tdb/common/tdb.c
+++ b/source4/lib/tdb/common/tdb.c
@@ -1250,6 +1250,43 @@ static int tdb_next_lock(TDB_CONTEXT *tdb, struct tdb_traverse_lock *tlock,
/* Lock each chain from the start one. */
for (; tlock->hash < tdb->header.hash_size; tlock->hash++) {
+
+ /* this is an optimisation for the common case where
+ the hash chain is empty, which is particularly
+ common for the use of tdb with ldb, where large
+ hashes are used. In that case we spend most of our
+ time in tdb_brlock(), locking empty hash chains.
+
+ To avoid this, we do an unlocked pre-check to see
+ if the hash chain is empty before starting to look
+ inside it. If it is empty then we can avoid that
+ hash chain. If it isn't empty then we can't believe
+ the value we get back, as we read it without a
+ lock, so instead we get the lock and re-fetch the
+ value below.
+
+ Notice that not doing this optimisation on the
+ first hash chain is critical. We must guarantee
+ that we have done at least one fcntl lock at the
+ start of a search to guarantee that memory is
+ coherent on SMP systems. If records are added by
+ others during the search then thats OK, and we
+ could possibly miss those with this trick, but we
+ could miss them anyway without this trick, so the
+ semantics don't change.
+
+ With a non-indexed ldb search this trick gains us a
+ factor of more than 10 in speed on a linux 2.6.x
+ system.
+ */
+ if (!tlock->off && tlock->hash != 0) {
+ u32 off;
+ if (ofs_read(tdb, TDB_HASH_TOP(tlock->hash), &off) == 0 &&
+ off == 0) {
+ continue;
+ }
+ }
+
if (tdb_lock(tdb, tlock->hash, F_WRLCK) == -1)
return -1;