diff --git a/lib/lockref.c b/lib/lockref.c
index 2afe4c5d8919..56f4419f593d 100644
--- a/lib/lockref.c
+++ b/lib/lockref.c
@@ -26,6 +26,17 @@
 	}									\
 } while (0)
 
+/*
+ * The compiler isn't smart enough to the the count
+ * increment in the high 32 bits of the 64-bit value,
+ * so do this optimization by hand.
+ */
+#if defined(__LITTLE_ENDIAN) && BITS_PER_LONG == 64
+ #define LOCKREF_ADD(n,x) ((n).lock_count += (unsigned long)(x)<<32)
+#else
+ #define LOCKREF_ADD(n,x) ((n).count += (unsigned long)(x)<<32)
+#endif
+
 #else
 
 #define CMPXCHG_LOOP(CODE, SUCCESS) do { } while (0)
@@ -42,7 +53,7 @@
 void lockref_get(struct lockref *lockref)
 {
 	CMPXCHG_LOOP(
-		new.count++;
+		LOCKREF_ADD(new,1);
 	,
 		return;
 	);
@@ -63,7 +74,7 @@ int lockref_get_not_zero(struct lockref *lockref)
 	int retval;
 
 	CMPXCHG_LOOP(
-		new.count++;
+		LOCKREF_ADD(new,1);
 		if (old.count <= 0)
 			return 0;
 	,
@@ -91,7 +102,7 @@ int lockref_put_not_zero(struct lockref *lockref)
 	int retval;
 
 	CMPXCHG_LOOP(
-		new.count--;
+		LOCKREF_ADD(new,-1);
 		if (old.count <= 1)
 			return 0;
 	,
@@ -119,7 +130,7 @@ EXPORT_SYMBOL(lockref_put_not_zero);
 int lockref_put_return(struct lockref *lockref)
 {
 	CMPXCHG_LOOP(
-		new.count--;
+		LOCKREF_ADD(new,-1);
 		if (old.count <= 0)
 			return -1;
 	,
@@ -137,7 +148,7 @@ EXPORT_SYMBOL(lockref_put_return);
 int lockref_put_or_lock(struct lockref *lockref)
 {
 	CMPXCHG_LOOP(
-		new.count--;
+		LOCKREF_ADD(new,-1);
 		if (old.count <= 1)
 			break;
 	,
@@ -174,7 +185,7 @@ int lockref_get_not_dead(struct lockref *lockref)
 	int retval;
 
 	CMPXCHG_LOOP(
-		new.count++;
+		LOCKREF_ADD(new,1);
 		if (old.count < 0)
 			return 0;
 	,