diff --git a/src/kern/npf_nat.c b/src/kern/npf_nat.c
index 1143c1f..b1caf43 100644
--- a/src/kern/npf_nat.c
+++ b/src/kern/npf_nat.c
@@ -257,9 +257,11 @@ npf_natpolicy_release(npf_natpolicy_t *np)
 {
 	KASSERT(atomic_load_relaxed(&np->n_refcnt) > 0);
 
+	membar_release();
 	if (atomic_dec_uint_nv(&np->n_refcnt) != 0) {
 		return;
 	}
+	membar_acquire();
 	KASSERT(LIST_EMPTY(&np->n_nat_list));
 	mutex_destroy(&np->n_lock);
 	kmem_free(np, sizeof(npf_natpolicy_t));
diff --git a/src/kern/npf_rproc.c b/src/kern/npf_rproc.c
index 82133cd..50a454d 100644
--- a/src/kern/npf_rproc.c
+++ b/src/kern/npf_rproc.c
@@ -308,9 +308,11 @@ npf_rproc_release(npf_rproc_t *rp)
 {
 	KASSERT(atomic_load_relaxed(&rp->rp_refcnt) > 0);
 
+	membar_release();
 	if (atomic_dec_uint_nv(&rp->rp_refcnt) != 0) {
 		return;
 	}
+	membar_acquire();
 	/* XXXintr */
 	for (unsigned i = 0; i < rp->rp_ext_count; i++) {
 		npf_ext_t *ext = rp->rp_ext[i];
diff --git a/src/kern/npf_tableset.c b/src/kern/npf_tableset.c
index 8492b17..23ddca1 100644
--- a/src/kern/npf_tableset.c
+++ b/src/kern/npf_tableset.c
@@ -136,9 +136,13 @@ npf_tableset_destroy(npf_tableset_t *ts)
 	for (u_int tid = 0; tid < ts->ts_nitems; tid++) {
 		npf_table_t *t = ts->ts_map[tid];
 
-		if (t && atomic_dec_uint_nv(&t->t_refcnt) == 0) {
-			npf_table_destroy(t);
-		}
+		if (t == NULL)
+			continue;
+		membar_release();
+		if (atomic_dec_uint_nv(&t->t_refcnt) > 0)
+			continue;
+		membar_acquire();
+		npf_table_destroy(t);
 	}
 	kmem_free(ts, NPF_TABLESET_SIZE(ts->ts_nitems));
 }
diff --git a/src/kern/stand/npf_stand.h b/src/kern/stand/npf_stand.h
index 9538530..5c068d6 100644
--- a/src/kern/stand/npf_stand.h
+++ b/src/kern/stand/npf_stand.h
@@ -136,6 +136,8 @@ npfkern_atomic_swap_ptr(volatile void *ptr, void *newval)
 #define	membar_sync()		__sync_synchronize()
 #define	membar_consumer()	__sync_synchronize()
 #define	membar_producer()	__sync_synchronize()
+#define	membar_release()	__sync_synchronize()
+#define	membar_acquire()	__sync_synchronize()
 #define	atomic_inc_uint(x)	__sync_fetch_and_add((x), 1)
 #define	atomic_inc_uint_nv(x)	__sync_add_and_fetch((x), 1)
 #define	atomic_inc_ulong_nv(x)	__sync_add_and_fetch((x), 1)