Index: sys/sysctl.h
===================================================================
--- sys/sysctl.h	(revision 194811)
+++ sys/sysctl.h	(working copy)
@@ -179,6 +179,7 @@ int sysctl_handle_string(SYSCTL_HANDLER_ARGS);
 int sysctl_handle_opaque(SYSCTL_HANDLER_ARGS);
 
 int sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS);
+int sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS);
 int sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS);
 
 #ifdef VIMAGE
Index: sys/sched.h
===================================================================
--- sys/sched.h	(revision 194811)
+++ sys/sched.h	(working copy)
@@ -162,14 +162,27 @@ sched_unpin(void)
 
 /* Scheduler stats. */
 #ifdef SCHED_STATS
-extern long sched_switch_stats[SWT_COUNT];
+DPCPU_DECLARE(long, sched_switch_stats[SWT_COUNT]);
 
 #define	SCHED_STAT_DEFINE_VAR(name, ptr, descr)				\
-    SYSCTL_LONG(_kern_sched_stats, OID_AUTO, name, CTLFLAG_RD, ptr, 0, descr)
+static void name ## _add_proc(void *dummy __unused)			\
+{									\
+									\
+	SYSCTL_ADD_PROC(NULL,						\
+	    SYSCTL_STATIC_CHILDREN(_kern_sched_stats), OID_AUTO,	\
+	    #name, CTLTYPE_LONG|CTLFLAG_RD|CTLFLAG_MPSAFE,		\
+	    ptr, 0, sysctl_dpcpu_long, "LU", descr);			\
+}									\
+SYSINIT(name, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, name ## _add_proc, NULL);
+
 #define	SCHED_STAT_DEFINE(name, descr)					\
-    unsigned long name;							\
-    SCHED_STAT_DEFINE_VAR(name, &name, descr)
-#define SCHED_STAT_INC(var)     atomic_add_long(&(var), 1)
+    DPCPU_DEFINE(unsigned long, name);					\
+    SCHED_STAT_DEFINE_VAR(name, &DPCPU_NAME(name), descr)
+/*
+ * Sched stats are always incremented in critical sections so no atomic
+ * is necesssary to increment them.
+ */
+#define SCHED_STAT_INC(var)     DPCPU_GET(var)++;
 #else
 #define	SCHED_STAT_DEFINE_VAR(name, descr, ptr)
 #define	SCHED_STAT_DEFINE(name, descr)
Index: kern/kern_switch.c
===================================================================
--- kern/kern_switch.c	(revision 194811)
+++ kern/kern_switch.c	(working copy)
@@ -79,31 +79,45 @@ SYSCTL_INT(_kern_sched, OID_AUTO, preemption, CTLF
  * with SCHED_STAT_DEFINE().
  */
 #ifdef SCHED_STATS
-long sched_switch_stats[SWT_COUNT];	/* Switch reasons from mi_switch(). */
-
 SYSCTL_NODE(_kern_sched, OID_AUTO, stats, CTLFLAG_RW, 0, "switch stats");
-SCHED_STAT_DEFINE_VAR(uncategorized, &sched_switch_stats[SWT_NONE], "");
-SCHED_STAT_DEFINE_VAR(preempt, &sched_switch_stats[SWT_PREEMPT], "");
-SCHED_STAT_DEFINE_VAR(owepreempt, &sched_switch_stats[SWT_OWEPREEMPT], "");
-SCHED_STAT_DEFINE_VAR(turnstile, &sched_switch_stats[SWT_TURNSTILE], "");
-SCHED_STAT_DEFINE_VAR(sleepq, &sched_switch_stats[SWT_SLEEPQ], "");
-SCHED_STAT_DEFINE_VAR(sleepqtimo, &sched_switch_stats[SWT_SLEEPQTIMO], "");
-SCHED_STAT_DEFINE_VAR(relinquish, &sched_switch_stats[SWT_RELINQUISH], "");
-SCHED_STAT_DEFINE_VAR(needresched, &sched_switch_stats[SWT_NEEDRESCHED], "");
-SCHED_STAT_DEFINE_VAR(idle, &sched_switch_stats[SWT_IDLE], "");
-SCHED_STAT_DEFINE_VAR(iwait, &sched_switch_stats[SWT_IWAIT], "");
-SCHED_STAT_DEFINE_VAR(suspend, &sched_switch_stats[SWT_SUSPEND], "");
-SCHED_STAT_DEFINE_VAR(remotepreempt, &sched_switch_stats[SWT_REMOTEPREEMPT],
-    "");
-SCHED_STAT_DEFINE_VAR(remotewakeidle, &sched_switch_stats[SWT_REMOTEWAKEIDLE],
-    "");
 
+/* Switch reasons from mi_switch(). */
+DPCPU_DEFINE(long, sched_switch_stats[SWT_COUNT]);
+SCHED_STAT_DEFINE_VAR(uncategorized,
+    &DPCPU_NAME(sched_switch_stats[SWT_NONE]), "");
+SCHED_STAT_DEFINE_VAR(preempt,
+    &DPCPU_NAME(sched_switch_stats[SWT_PREEMPT]), "");
+SCHED_STAT_DEFINE_VAR(owepreempt,
+    &DPCPU_NAME(sched_switch_stats[SWT_OWEPREEMPT]), "");
+SCHED_STAT_DEFINE_VAR(turnstile,
+    &DPCPU_NAME(sched_switch_stats[SWT_TURNSTILE]), "");
+SCHED_STAT_DEFINE_VAR(sleepq,
+    &DPCPU_NAME(sched_switch_stats[SWT_SLEEPQ]), "");
+SCHED_STAT_DEFINE_VAR(sleepqtimo,
+    &DPCPU_NAME(sched_switch_stats[SWT_SLEEPQTIMO]), "");
+SCHED_STAT_DEFINE_VAR(relinquish, 
+    &DPCPU_NAME(sched_switch_stats[SWT_RELINQUISH]), "");
+SCHED_STAT_DEFINE_VAR(needresched,
+    &DPCPU_NAME(sched_switch_stats[SWT_NEEDRESCHED]), "");
+SCHED_STAT_DEFINE_VAR(idle, 
+    &DPCPU_NAME(sched_switch_stats[SWT_IDLE]), "");
+SCHED_STAT_DEFINE_VAR(iwait,
+    &DPCPU_NAME(sched_switch_stats[SWT_IWAIT]), "");
+SCHED_STAT_DEFINE_VAR(suspend,
+    &DPCPU_NAME(sched_switch_stats[SWT_SUSPEND]), "");
+SCHED_STAT_DEFINE_VAR(remotepreempt,
+    &DPCPU_NAME(sched_switch_stats[SWT_REMOTEPREEMPT]), "");
+SCHED_STAT_DEFINE_VAR(remotewakeidle,
+    &DPCPU_NAME(sched_switch_stats[SWT_REMOTEWAKEIDLE]), "");
+
 static int
 sysctl_stats_reset(SYSCTL_HANDLER_ARGS)
 {
 	struct sysctl_oid *p;
+	uintptr_t counter;
         int error;
 	int val;
+	int i;
 
         val = 0;
         error = sysctl_handle_int(oidp, &val, 0, req);
@@ -118,7 +132,12 @@ sysctl_stats_reset(SYSCTL_HANDLER_ARGS)
 	SLIST_FOREACH(p, oidp->oid_parent, oid_link) {
 		if (p == oidp || p->oid_arg1 == NULL)
 			continue;
-		*(long *)p->oid_arg1 = 0;
+		counter = (uintptr_t)p->oid_arg1;
+		for (i = 0; i <= mp_maxid; i++) {
+			if (CPU_ABSENT(i))
+				continue;
+			*(long *)(dpcpu_off[i] + counter) = 0;
+		}
 	}
 	return (0);
 }
Index: kern/subr_pcpu.c
===================================================================
--- kern/subr_pcpu.c	(revision 194811)
+++ kern/subr_pcpu.c	(working copy)
@@ -264,9 +264,8 @@ pcpu_find(u_int cpuid)
 int
 sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
 {
+	uintptr_t dpcpu;
 	int64_t count;
-#ifdef SMP
-	uintptr_t dpcpu;
 	int i;
 
 	count = 0;
@@ -276,18 +275,31 @@ sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS)
 			continue;
 		count += *(int64_t *)(dpcpu + (uintptr_t)arg1);
 	}
-#else
-	count = *(int64_t *)(dpcpu_off[0] + (uintptr_t)arg1);
-#endif
 	return (SYSCTL_OUT(req, &count, sizeof(count)));
 }
 
 int
+sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS)
+{
+	uintptr_t dpcpu;
+	long count;
+	int i;
+
+	count = 0;
+	for (i = 0; i < mp_ncpus; ++i) {
+		dpcpu = dpcpu_off[i];
+		if (dpcpu == 0)
+			continue;
+		count += *(long *)(dpcpu + (uintptr_t)arg1);
+	}
+	return (SYSCTL_OUT(req, &count, sizeof(count)));
+}
+
+int
 sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
 {
+	uintptr_t dpcpu;
 	int count;
-#ifdef SMP
-	uintptr_t dpcpu;
 	int i;
 
 	count = 0;
@@ -297,9 +309,6 @@ sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS)
 			continue;
 		count += *(int *)(dpcpu + (uintptr_t)arg1);
 	}
-#else
-	count = *(int *)(dpcpu_off[0] + (uintptr_t)arg1);
-#endif
 	return (SYSCTL_OUT(req, &count, sizeof(count)));
 }