lockdep: Always try to set ->class_cache in register_lock_class() lockdep_init_map()

Commit ["62016250 lockdep: Add improved subclass caching"] tries to
improve performance (expecially to reduce the cost of rq->lock)
when using lockdep, but it fails due to lockdep_init_map() in
which ->class_cache is cleared. The typical caller is
lock_set_subclass(), after that class will not be cached anymore.

This patch tries to achive the goal of commit 62016250 by always
setting ->class_cache in register_lock_class().

=== Score comparison of benchmarks ===

for i in `seq 1 10`; do ./perf bench -f simple sched messaging; done

before:  min: 0.604, max: 0.660, avg: 0.622
after:   min: 0.414, max: 0.473, avg: 0.427

for i in `seq 1 10`; do ./perf bench -f simple sched messaging -g 40; done

before:  min: 2.347, max: 2.421, avg: 2.391
after:   min: 1.652, max: 1.699, avg: 1.671

Signed-off-by: Yong Zhang <yong.zhang0@gmail.com>
Cc: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20111109080714.GC8124@zhy
Signed-off-by: Ingo Molnar <mingo@elte.hu>

kernel/lockdep.c

@@ -722,7 +722,7 @@ register_lock_class(struct lockdep_map *lock, unsigned int subclass, int force)
 
 	class = look_up_lock_class(lock, subclass);
 	if (likely(class))
-		return class;
+		goto out_set_class_cache;
 
 	/*
 	 * Debug-check: all keys must be persistent!
@@ -807,6 +807,7 @@ out_unlock_set:
 	graph_unlock();
 	raw_local_irq_restore(flags);
 
+out_set_class_cache:
 	if (!subclass || force)
 		lock->class_cache[0] = class;
 	else if (subclass < NR_LOCKDEP_CACHING_CLASSES)