performance/io-cache: change the hash function used for rbtree based hash table.

- the earlier hash function does not distribute pages uniformly for offsets that
    fit into 32 bits. The reason is that the hash function just xors the contents of
    the key 4 bytes at a time with the current value of hash. Hence for keys that
    fit into 32 bits, the hash will be the key itself. Since we are using the
    rounded_offset (which is a multiple of 128KB) as the key, the key will
    be exactly divisible by the number of buckets configured (4096) resolving all
    the pages into the first bucket.

Signed-off-by: Anand V. Avati <avati@dev.gluster.com>

BUG: 335 (Io-cache optimization)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=335

2 files changed, 16 insertions, 0 deletions

diff --git a/libglusterfs/src/common-utils.c b/libglusterfs/src/common-utils.c
index de0cb774b74..3f52a90a010 100644
--- a/libglusterfs/src/common-utils.c
+++ b/libglusterfs/src/common-utils.c
@@ -56,6 +56,21 @@ struct dnscache6 {
 	struct addrinfo *next;
 };
 
+
+int
+gf_log2 (unsigned long x)
+{
+        int val = 0;
+
+        while (x != 1) {
+                x /= 2;
+                val++;
+        }
+
+        return val;
+}
+
+
 int32_t
 gf_resolve_ip6 (const char *hostname, 
 		uint16_t port, 
diff --git a/libglusterfs/src/common-utils.h b/libglusterfs/src/common-utils.h
index c5869f4a424..48788d29545 100644
--- a/libglusterfs/src/common-utils.h
+++ b/libglusterfs/src/common-utils.h
@@ -330,6 +330,7 @@ int gf_lockfd (int fd);
 int gf_unlockfd (int fd);
 
 int get_checksum_for_file (int fd, uint32_t *checksum);
+int gf_log2 (unsigned long x);
 
 #endif /* _COMMON_UTILS_H */