diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 80175de..bdc9582 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -406,6 +406,11 @@
 
 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
 {
+	/*
+	 * Our caller must use css_get() first, because memcg_uncharge_kmem()
+	 * will call css_put() if it sees the memcg is dead.
+	 */
+	smp_wmb();
 	if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
 		set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
 }
@@ -3050,8 +3055,16 @@
 	if (res_counter_uncharge(&memcg->kmem, size))
 		return;
 
+	/*
+	 * Releases a reference taken in kmem_cgroup_css_offline in case
+	 * this last uncharge is racing with the offlining code or it is
+	 * outliving the memcg existence.
+	 *
+	 * The memory barrier imposed by test&clear is paired with the
+	 * explicit one in memcg_kmem_mark_dead().
+	 */
 	if (memcg_kmem_test_and_clear_dead(memcg))
-		mem_cgroup_put(memcg);
+		css_put(&memcg->css);
 }
 
 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
@@ -5183,14 +5196,6 @@
 		 * starts accounting before all call sites are patched
 		 */
 		memcg_kmem_set_active(memcg);
-
-		/*
-		 * kmem charges can outlive the cgroup. In the case of slab
-		 * pages, for instance, a page contain objects from various
-		 * processes, so it is unfeasible to migrate them away. We
-		 * need to reference count the memcg because of that.
-		 */
-		mem_cgroup_get(memcg);
 	} else
 		ret = res_counter_set_limit(&memcg->kmem, val);
 out:
@@ -5223,12 +5228,10 @@
 		goto out;
 
 	/*
-	 * destroy(), called if we fail, will issue static_key_slow_inc() and
-	 * mem_cgroup_put() if kmem is enabled. We have to either call them
-	 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
-	 * this more consistent, since it always leads to the same destroy path
+	 * __mem_cgroup_free() will issue static_key_slow_dec() because this
+	 * memcg is active already. If the later initialization fails then the
+	 * cgroup core triggers the cleanup so we do not have to do it here.
 	 */
-	mem_cgroup_get(memcg);
 	static_key_slow_inc(&memcg_kmem_enabled_key);
 
 	mutex_lock(&set_limit_mutex);
@@ -5913,23 +5916,43 @@
 	return mem_cgroup_sockets_init(memcg, ss);
 }
 
-static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
+static void memcg_destroy_kmem(struct mem_cgroup *memcg)
 {
 	mem_cgroup_sockets_destroy(memcg);
+}
+
+static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
+{
+	if (!memcg_kmem_is_active(memcg))
+		return;
+
+	/*
+	 * kmem charges can outlive the cgroup. In the case of slab
+	 * pages, for instance, a page contain objects from various
+	 * processes. As we prevent from taking a reference for every
+	 * such allocation we have to be careful when doing uncharge
+	 * (see memcg_uncharge_kmem) and here during offlining.
+	 *
+	 * The idea is that that only the _last_ uncharge which sees
+	 * the dead memcg will drop the last reference. An additional
+	 * reference is taken here before the group is marked dead
+	 * which is then paired with css_put during uncharge resp. here.
+	 *
+	 * Although this might sound strange as this path is called from
+	 * css_offline() when the referencemight have dropped down to 0
+	 * and shouldn't be incremented anymore (css_tryget would fail)
+	 * we do not have other options because of the kmem allocations
+	 * lifetime.
+	 */
+	css_get(&memcg->css);
 
 	memcg_kmem_mark_dead(memcg);
 
 	if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
 		return;
 
-	/*
-	 * Charges already down to 0, undo mem_cgroup_get() done in the charge
-	 * path here, being careful not to race with memcg_uncharge_kmem: it is
-	 * possible that the charges went down to 0 between mark_dead and the
-	 * res_counter read, so in that case, we don't need the put
-	 */
 	if (memcg_kmem_test_and_clear_dead(memcg))
-		mem_cgroup_put(memcg);
+		css_put(&memcg->css);
 }
 #else
 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
@@ -5937,7 +5960,11 @@
 	return 0;
 }
 
-static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
+static void memcg_destroy_kmem(struct mem_cgroup *memcg)
+{
+}
+
+static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
 {
 }
 #endif
@@ -6370,6 +6397,8 @@
 {
 	struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
 
+	kmem_cgroup_css_offline(memcg);
+
 	mem_cgroup_invalidate_reclaim_iterators(memcg);
 	mem_cgroup_reparent_charges(memcg);
 	mem_cgroup_destroy_all_caches(memcg);
@@ -6379,9 +6408,8 @@
 {
 	struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
 
-	kmem_cgroup_destroy(memcg);
-
-	mem_cgroup_put(memcg);
+	memcg_destroy_kmem(memcg);
+	__mem_cgroup_free(memcg);
 }
 
 #ifdef CONFIG_MMU