diff --git a/include/linux/pcounter.h b/include/linux/pcounter.h
index 9c4760a328f..a82d9f2628c 100644
--- a/include/linux/pcounter.h
+++ b/include/linux/pcounter.h
@@ -1,41 +1,39 @@
 #ifndef __LINUX_PCOUNTER_H
 #define __LINUX_PCOUNTER_H
-
+/*
+ * Using a dynamic percpu 'int' variable has a cost :
+ * 1) Extra dereference
+ * Current per_cpu_ptr() implementation uses an array per 'percpu variable'.
+ * 2) memory cost of NR_CPUS*(32+sizeof(void *)) instead of num_possible_cpus()*4
+ *
+ * This pcounter implementation is an abstraction to be able to use
+ * either a static or a dynamic per cpu variable.
+ * One dynamic per cpu variable gets a fast & cheap implementation, we can
+ * change pcounter implementation too.
+ */
 struct pcounter {
 #ifdef CONFIG_SMP
 	void		(*add)(struct pcounter *self, int inc);
-	int		(*getval)(const struct pcounter *self);
+	int		(*getval)(const struct pcounter *self, int cpu);
 	int		*per_cpu_values;
 #else
 	int		val;
 #endif
 };
 
-/*
- * Special macros to let pcounters use a fast version of {getvalue|add}
- * using a static percpu variable per pcounter instead of an allocated one,
- * saving one dereference.
- * This might be changed if/when dynamic percpu vars become fast.
- */
 #ifdef CONFIG_SMP
-#include <linux/cpumask.h>
 #include <linux/percpu.h>
 
-#define DEFINE_PCOUNTER(NAME)					\
-static DEFINE_PER_CPU(int, NAME##_pcounter_values);		\
-static void NAME##_pcounter_add(struct pcounter *self, int inc)	\
-{								\
-       __get_cpu_var(NAME##_pcounter_values) += inc;		\
-}								\
-								\
-static int NAME##_pcounter_getval(const struct pcounter *self)	\
-{								\
-       int res = 0, cpu;					\
-								\
-       for_each_possible_cpu(cpu)				\
-               res += per_cpu(NAME##_pcounter_values, cpu);	\
-       return res;						\
-}
+#define DEFINE_PCOUNTER(NAME)						\
+static DEFINE_PER_CPU(int, NAME##_pcounter_values);			\
+static void NAME##_pcounter_add(struct pcounter *self, int val)		\
+{									\
+       __get_cpu_var(NAME##_pcounter_values) += val;			\
+}									\
+static int NAME##_pcounter_getval(const struct pcounter *self, int cpu)	\
+{									\
+	return per_cpu(NAME##_pcounter_values, cpu);			\
+}									\
 
 #define PCOUNTER_MEMBER_INITIALIZER(NAME, MEMBER)		\
 	MEMBER = {						\
@@ -43,42 +41,16 @@ static int NAME##_pcounter_getval(const struct pcounter *self)	\
 		.getval = NAME##_pcounter_getval,		\
 	}
 
-extern void pcounter_def_add(struct pcounter *self, int inc);
-extern int pcounter_def_getval(const struct pcounter *self);
-
-static inline int pcounter_alloc(struct pcounter *self)
-{
-	int rc = 0;
-	if (self->add == NULL) {
-		self->per_cpu_values = alloc_percpu(int);
-		if (self->per_cpu_values != NULL) {
-			self->add    = pcounter_def_add;
-			self->getval = pcounter_def_getval;
-		} else
-			rc = 1;
-	}
-	return rc;
-}
-
-static inline void pcounter_free(struct pcounter *self)
-{
-	if (self->per_cpu_values != NULL) {
-		free_percpu(self->per_cpu_values);
-		self->per_cpu_values = NULL;
-		self->getval = NULL;
-		self->add = NULL;
-	}
-}
 
 static inline void pcounter_add(struct pcounter *self, int inc)
 {
 	self->add(self, inc);
 }
 
-static inline int pcounter_getval(const struct pcounter *self)
-{
-	return self->getval(self);
-}
+extern int pcounter_getval(const struct pcounter *self);
+extern int pcounter_alloc(struct pcounter *self);
+extern void pcounter_free(struct pcounter *self);
+
 
 #else /* CONFIG_SMP */
 
diff --git a/lib/pcounter.c b/lib/pcounter.c
index 93feea59825..9b56807da93 100644
--- a/lib/pcounter.c
+++ b/lib/pcounter.c
@@ -7,20 +7,52 @@
 #include <linux/module.h>
 #include <linux/pcounter.h>
 #include <linux/smp.h>
+#include <linux/cpumask.h>
 
-void pcounter_def_add(struct pcounter *self, int inc)
+static void pcounter_dyn_add(struct pcounter *self, int inc)
 {
 	per_cpu_ptr(self->per_cpu_values, smp_processor_id())[0] += inc;
 }
 
-EXPORT_SYMBOL_GPL(pcounter_def_add);
-
-int pcounter_def_getval(const struct pcounter *self)
+static int pcounter_dyn_getval(const struct pcounter *self, int cpu)
 {
-	int res = 0, cpu;
-	for_each_possible_cpu(cpu)
-		res += per_cpu_ptr(self->per_cpu_values, cpu)[0];
-	return res;
+	return per_cpu_ptr(self->per_cpu_values, cpu)[0];
 }
 
-EXPORT_SYMBOL_GPL(pcounter_def_getval);
+int pcounter_getval(const struct pcounter *self)
+{
+	int res = 0, cpu;
+
+	for_each_possible_cpu(cpu)
+		res += self->getval(self, cpu);
+
+	return res;
+}
+EXPORT_SYMBOL_GPL(pcounter_getval);
+
+int pcounter_alloc(struct pcounter *self)
+{
+	int rc = 0;
+	if (self->add == NULL) {
+		self->per_cpu_values = alloc_percpu(int);
+		if (self->per_cpu_values != NULL) {
+			self->add    = pcounter_dyn_add;
+			self->getval = pcounter_dyn_getval;
+		} else
+			rc = 1;
+	}
+	return rc;
+}
+EXPORT_SYMBOL_GPL(pcounter_alloc);
+
+void pcounter_free(struct pcounter *self)
+{
+	if (self->per_cpu_values != NULL) {
+		free_percpu(self->per_cpu_values);
+		self->per_cpu_values = NULL;
+		self->getval = NULL;
+		self->add = NULL;
+	}
+}
+EXPORT_SYMBOL_GPL(pcounter_free);
+