mm: memcontrol: factor out reclaim iterator loading and updating
authorJohannes Weiner <hannes@cmpxchg.org>
Wed, 3 Jul 2013 22:04:51 +0000 (15:04 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 3 Jul 2013 23:07:40 +0000 (16:07 -0700)
mem_cgroup_iter() is too hard to follow.  Factor out the lockless reclaim
iterator loading and updating so it's easier to follow the big picture.

Also document the iterator invalidation mechanism a bit more extensively.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Glauber Costa <glommer@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memcontrol.c

index 4748966b1511c94956658c42d539d6b2b0ca3c31..2e851f453814529d0a3e032e9e5e1dd009a4b34a 100644 (file)
@@ -1148,6 +1148,58 @@ skip_node:
        return NULL;
 }
 
+static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
+{
+       /*
+        * When a group in the hierarchy below root is destroyed, the
+        * hierarchy iterator can no longer be trusted since it might
+        * have pointed to the destroyed group.  Invalidate it.
+        */
+       atomic_inc(&root->dead_count);
+}
+
+static struct mem_cgroup *
+mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
+                    struct mem_cgroup *root,
+                    int *sequence)
+{
+       struct mem_cgroup *position = NULL;
+       /*
+        * A cgroup destruction happens in two stages: offlining and
+        * release.  They are separated by a RCU grace period.
+        *
+        * If the iterator is valid, we may still race with an
+        * offlining.  The RCU lock ensures the object won't be
+        * released, tryget will fail if we lost the race.
+        */
+       *sequence = atomic_read(&root->dead_count);
+       if (iter->last_dead_count == *sequence) {
+               smp_rmb();
+               position = iter->last_visited;
+               if (position && !css_tryget(&position->css))
+                       position = NULL;
+       }
+       return position;
+}
+
+static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
+                                  struct mem_cgroup *last_visited,
+                                  struct mem_cgroup *new_position,
+                                  int sequence)
+{
+       if (last_visited)
+               css_put(&last_visited->css);
+       /*
+        * We store the sequence count from the time @last_visited was
+        * loaded successfully instead of rereading it here so that we
+        * don't lose destruction events in between.  We could have
+        * raced with the destruction of @new_position after all.
+        */
+       iter->last_visited = new_position;
+       smp_wmb();
+       iter->last_dead_count = sequence;
+}
+
 /**
  * mem_cgroup_iter - iterate over memory cgroup hierarchy
  * @root: hierarchy root
@@ -1171,7 +1223,6 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
 {
        struct mem_cgroup *memcg = NULL;
        struct mem_cgroup *last_visited = NULL;
-       unsigned long uninitialized_var(dead_count);
 
        if (mem_cgroup_disabled())
                return NULL;
@@ -1191,6 +1242,7 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
        rcu_read_lock();
        while (!memcg) {
                struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
+               int uninitialized_var(seq);
 
                if (reclaim) {
                        int nid = zone_to_nid(reclaim->zone);
@@ -1204,37 +1256,13 @@ struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
                                goto out_unlock;
                        }
 
-                       /*
-                        * If the dead_count mismatches, a destruction
-                        * has happened or is happening concurrently.
-                        * If the dead_count matches, a destruction
-                        * might still happen concurrently, but since
-                        * we checked under RCU, that destruction
-                        * won't free the object until we release the
-                        * RCU reader lock.  Thus, the dead_count
-                        * check verifies the pointer is still valid,
-                        * css_tryget() verifies the cgroup pointed to
-                        * is alive.
-                        */
-                       dead_count = atomic_read(&root->dead_count);
-                       if (dead_count == iter->last_dead_count) {
-                               smp_rmb();
-                               last_visited = iter->last_visited;
-                               if (last_visited &&
-                                   !css_tryget(&last_visited->css))
-                                       last_visited = NULL;
-                       }
+                       last_visited = mem_cgroup_iter_load(iter, root, &seq);
                }
 
                memcg = __mem_cgroup_iter_next(root, last_visited);
 
                if (reclaim) {
-                       if (last_visited)
-                               css_put(&last_visited->css);
-
-                       iter->last_visited = memcg;
-                       smp_wmb();
-                       iter->last_dead_count = dead_count;
+                       mem_cgroup_iter_update(iter, last_visited, memcg, seq);
 
                        if (!memcg)
                                iter->generation++;
@@ -6318,14 +6346,14 @@ static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
        struct mem_cgroup *parent = memcg;
 
        while ((parent = parent_mem_cgroup(parent)))
-               atomic_inc(&parent->dead_count);
+               mem_cgroup_iter_invalidate(parent);
 
        /*
         * if the root memcg is not hierarchical we have to check it
         * explicitely.
         */
        if (!root_mem_cgroup->use_hierarchy)
-               atomic_inc(&root_mem_cgroup->dead_count);
+               mem_cgroup_iter_invalidate(root_mem_cgroup);
 }
 
 static void mem_cgroup_css_offline(struct cgroup *cont)