[PATCH v4 23/25] memcg: propagate kmem limiting information to children

[Glauber Costa <glommer@parallels.com>]

The current memcg slab cache management fails to present satisfatory hierarchical
behavior in the following scenario:

-> /cgroups/memory/A/B/C

* kmem limit set at A
* A and B empty taskwise
* bash in C does find /

Because kmem_accounted is a boolean that was not set for C, no accounting
would be done. This is, however, not what we expect.

The basic idea, is that when a cgroup is limited, we walk the tree
upwards (something Kame and I already thought about doing for other purposes),
and make sure that we store the information about the parent being limited in
kmem_accounted (that is turned into a bitmap: two booleans would not be space
efficient). The code for that is taken from sched/core.c. My reasons for not
putting it into a common place is to dodge the type issues that would arise
from a common implementation between memcg and the scheduler - but I think
that it should ultimately happen, so if you want me to do it now, let me
know.

We do the reverse operation when a formerly limited cgroup becomes unlimited.

Signed-off-by: Glauber Costa <glommer@parallels.com>
CC: Christoph Lameter <cl@linux.com>
CC: Pekka Enberg <penberg@cs.helsinki.fi>
CC: Michal Hocko <mhocko@suse.cz>
CC: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
CC: Johannes Weiner <hannes@cmpxchg.org>
CC: Suleiman Souhlal <suleiman@google.com>
---
 mm/memcontrol.c |  147 +++++++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 131 insertions(+), 16 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index e32b53e..972e83f 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -259,6 +259,9 @@ struct mem_cgroup {
 	 * the counter to account for kernel memory usage.
 	 */
 	struct res_counter kmem;
+
+	struct list_head children;
+	struct list_head siblings;
 	/*
 	 * Per cgroup active and inactive list, similar to the
 	 * per zone LRU lists.
@@ -274,7 +277,11 @@ struct mem_cgroup {
 	 * Should the accounting and control be hierarchical, per subtree?
 	 */
 	bool use_hierarchy;
-	bool kmem_accounted;
+	/*
+	 * bit0: accounted by this cgroup
+	 * bit1: accounted by a parent.
+	 */
+	volatile unsigned long kmem_accounted;
 
 	bool		oom_lock;
 	atomic_t	under_oom;
@@ -332,6 +339,9 @@ struct mem_cgroup {
 #endif
 };
 
+#define KMEM_ACCOUNTED_THIS	0
+#define KMEM_ACCOUNTED_PARENT	1
+
 int memcg_css_id(struct mem_cgroup *memcg)
 {
 	return css_id(&memcg->css);
@@ -474,7 +484,7 @@ void sock_release_memcg(struct sock *sk)
 
 static void disarm_static_keys(struct mem_cgroup *memcg)
 {
-	if (memcg->kmem_accounted)
+	if (test_bit(KMEM_ACCOUNTED_THIS, &memcg->kmem_accounted))
 		static_key_slow_dec(&mem_cgroup_kmem_enabled_key);
 	/*
 	 * This check can't live in kmem destruction function,
@@ -4418,6 +4428,110 @@ static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
 	len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
 	return simple_read_from_buffer(buf, nbytes, ppos, str, len);
 }
+
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
+typedef int (*memcg_visitor)(struct mem_cgroup*, void *);
+
+/*
+ * This is mostly "inspired" by the code in sched/core.c. I decided to copy it,
+ * instead of factoring it, because of all the typing issues we'd run into.
+ * In particular, grabbing the parent is very different for memcg, because we
+ * may or may not have hierarchy, while cpu cgroups always do. That would lead
+ * to either indirect calls - this is not a fast path for us, but can be for
+ * the scheduler - or a big and ugly macro.
+ *
+ * If we ever get rid of hierarchy, we could iterate over struct cgroup, and
+ * then it would cease to be a problem.
+ */
+int walk_tree_from(struct mem_cgroup *from,
+		   memcg_visitor down, memcg_visitor up, void *data)
+{
+	struct mem_cgroup *parent, *child;
+	int ret;
+
+
+	parent = from;
+down:
+	ret = (*down)(parent, data);
+	if (ret)
+		goto out;
+
+	list_for_each_entry_rcu(child, &parent->children, siblings) {
+		parent = child;
+		goto down;
+
+up:
+		continue;
+	}
+	ret = (*up)(parent, data);
+	if (ret || parent == from)
+		goto out;
+
+	child = parent;
+	parent = parent_mem_cgroup(parent);
+	if (parent)
+		goto up;
+out:
+	return ret;
+}
+
+static int memcg_nop(struct mem_cgroup *memcg, void *data)
+{
+	return 0;
+}
+
+static int memcg_parent_account(struct mem_cgroup *memcg, void *data)
+{
+	if (memcg == data)
+		return 0;
+
+	set_bit(KMEM_ACCOUNTED_PARENT, &memcg->kmem_accounted);
+	return 0;
+}
+
+static int memcg_parent_no_account(struct mem_cgroup *memcg, void *data)
+{
+	if (memcg == data)
+		return 0;
+
+	clear_bit(KMEM_ACCOUNTED_PARENT, &memcg->kmem_accounted);
+	/*
+	 * Stop propagation if we are accounted: our children should
+	 * be parent-accounted
+	 */
+	return test_bit(KMEM_ACCOUNTED_THIS, &memcg->kmem_accounted);
+}
+
+static void mem_cgroup_update_kmem_limit(struct mem_cgroup *memcg, u64 val)
+{
+	mutex_lock(&set_limit_mutex);
+	if (!test_and_set_bit(KMEM_ACCOUNTED_THIS, &memcg->kmem_accounted) &&
+		val != RESOURCE_MAX) {
+
+		/*
+		 * Once enabled, can't be disabled. We could in theory
+		 * disable it if we haven't yet created any caches, or
+		 * if we can shrink them all to death.
+		 *
+		 * But it is not worth the trouble
+		 */
+		static_key_slow_inc(&mem_cgroup_kmem_enabled_key);
+
+		rcu_read_lock();
+		walk_tree_from(memcg, memcg_parent_account, memcg_nop, memcg);
+		rcu_read_unlock();
+	} else if (test_and_clear_bit(KMEM_ACCOUNTED_THIS, &memcg->kmem_accounted)
+		&& val == RESOURCE_MAX) {
+
+		rcu_read_lock();
+		walk_tree_from(memcg, memcg_parent_no_account,
+			       memcg_nop, memcg);
+		rcu_read_unlock();
+	}
+
+	mutex_unlock(&set_limit_mutex);
+}
+#endif
 /*
  * The user of this function is...
  * RES_LIMIT.
@@ -4455,20 +4569,8 @@ static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
 			ret = res_counter_set_limit(&memcg->kmem, val);
 			if (ret)
 				break;
-			/*
-			 * Once enabled, can't be disabled. We could in theory
-			 * disable it if we haven't yet created any caches, or
-			 * if we can shrink them all to death.
-			 *
-			 * But it is not worth the trouble
-			 */
-			mutex_lock(&set_limit_mutex);
-			if (!memcg->kmem_accounted && val != RESOURCE_MAX
-			    && !memcg->kmem_accounted) {
-				static_key_slow_inc(&mem_cgroup_kmem_enabled_key);
-				memcg->kmem_accounted = true;
-			}
-			mutex_unlock(&set_limit_mutex);
+			mem_cgroup_update_kmem_limit(memcg, val);
+			break;
 		}
 #endif
 		else
@@ -5527,6 +5629,8 @@ err_cleanup:
 
 }
 
+static DEFINE_MUTEX(memcg_list_mutex);
+
 static struct cgroup_subsys_state * __ref
 mem_cgroup_create(struct cgroup *cont)
 {
@@ -5542,6 +5646,7 @@ mem_cgroup_create(struct cgroup *cont)
 		if (alloc_mem_cgroup_per_zone_info(memcg, node))
 			goto free_out;
 
+	INIT_LIST_HEAD(&memcg->children);
 	/* root ? */
 	if (cont->parent == NULL) {
 		int cpu;
@@ -5580,6 +5685,10 @@ mem_cgroup_create(struct cgroup *cont)
 		 * mem_cgroup(see mem_cgroup_put).
 		 */
 		mem_cgroup_get(parent);
+
+		mutex_lock(&memcg_list_mutex);
+		list_add_rcu(&memcg->siblings, &parent->children);
+		mutex_unlock(&memcg_list_mutex);
 	} else {
 		res_counter_init(&memcg->res, NULL);
 		res_counter_init(&memcg->memsw, NULL);
@@ -5622,9 +5731,15 @@ static int mem_cgroup_pre_destroy(struct cgroup *cont)
 static void mem_cgroup_destroy(struct cgroup *cont)
 {
 	struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
+	struct mem_cgroup *parent = parent_mem_cgroup(memcg);
 
 	kmem_cgroup_destroy(memcg);
 
+	mutex_lock(&memcg_list_mutex);
+	if (parent)
+		list_del_rcu(&memcg->siblings);
+	mutex_unlock(&memcg_list_mutex);
+
 	mem_cgroup_put(memcg);
 }
 
-- 
1.7.10.2