memcg: split part of memcg creation to css_online
authorGlauber Costa <glommer@parallels.com>
Sat, 23 Feb 2013 00:34:52 +0000 (16:34 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sun, 24 Feb 2013 01:50:18 +0000 (17:50 -0800)
This patch is a preparatory work for later locking rework to get rid of
big cgroup lock from memory controller code.

The memory controller uses some tunables to adjust its operation.  Those
tunables are inherited from parent to children upon children
intialization.  For most of them, the value cannot be changed after the
parent has a new children.

cgroup core splits initialization in two phases: css_alloc and css_online.
After css_alloc, the memory allocation and basic initialization are done.
But the new group is not yet visible anywhere, not even for cgroup core
code.  It is only somewhere between css_alloc and css_online that it is
inserted into the internal children lists.  Copying tunable values in
css_alloc will lead to inconsistent values: the children will copy the old
parent values, that can change between the copy and the moment in which
the groups is linked to any data structure that can indicate the presence
of children.

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

index 8e0988fee888cd369d5b00991972157e24c70bad..a9318963d5d5f1a13b30328973049a071b404513 100644 (file)
@@ -6093,7 +6093,7 @@ err_cleanup:
 static struct cgroup_subsys_state * __ref
 mem_cgroup_css_alloc(struct cgroup *cont)
 {
-       struct mem_cgroup *memcg, *parent;
+       struct mem_cgroup *memcg;
        long error = -ENOMEM;
        int node;
 
@@ -6108,7 +6108,7 @@ mem_cgroup_css_alloc(struct cgroup *cont)
        /* root ? */
        if (cont->parent == NULL) {
                int cpu;
-               parent = NULL;
+
                if (mem_cgroup_soft_limit_tree_init())
                        goto free_out;
                root_mem_cgroup = memcg;
@@ -6117,13 +6117,43 @@ mem_cgroup_css_alloc(struct cgroup *cont)
                                                &per_cpu(memcg_stock, cpu);
                        INIT_WORK(&stock->work, drain_local_stock);
                }
-       } else {
-               parent = mem_cgroup_from_cont(cont->parent);
-               memcg->use_hierarchy = parent->use_hierarchy;
-               memcg->oom_kill_disable = parent->oom_kill_disable;
+
+               res_counter_init(&memcg->res, NULL);
+               res_counter_init(&memcg->memsw, NULL);
+               res_counter_init(&memcg->kmem, NULL);
        }
 
-       if (parent && parent->use_hierarchy) {
+       memcg->last_scanned_node = MAX_NUMNODES;
+       INIT_LIST_HEAD(&memcg->oom_notify);
+       atomic_set(&memcg->refcnt, 1);
+       memcg->move_charge_at_immigrate = 0;
+       mutex_init(&memcg->thresholds_lock);
+       spin_lock_init(&memcg->move_lock);
+
+       return &memcg->css;
+
+free_out:
+       __mem_cgroup_free(memcg);
+       return ERR_PTR(error);
+}
+
+static int
+mem_cgroup_css_online(struct cgroup *cont)
+{
+       struct mem_cgroup *memcg, *parent;
+       int error = 0;
+
+       if (!cont->parent)
+               return 0;
+
+       memcg = mem_cgroup_from_cont(cont);
+       parent = mem_cgroup_from_cont(cont->parent);
+
+       memcg->use_hierarchy = parent->use_hierarchy;
+       memcg->oom_kill_disable = parent->oom_kill_disable;
+       memcg->swappiness = mem_cgroup_swappiness(parent);
+
+       if (parent->use_hierarchy) {
                res_counter_init(&memcg->res, &parent->res);
                res_counter_init(&memcg->memsw, &parent->memsw);
                res_counter_init(&memcg->kmem, &parent->kmem);
@@ -6144,18 +6174,9 @@ mem_cgroup_css_alloc(struct cgroup *cont)
                 * much sense so let cgroup subsystem know about this
                 * unfortunate state in our controller.
                 */
-               if (parent && parent != root_mem_cgroup)
+               if (parent != root_mem_cgroup)
                        mem_cgroup_subsys.broken_hierarchy = true;
        }
-       memcg->last_scanned_node = MAX_NUMNODES;
-       INIT_LIST_HEAD(&memcg->oom_notify);
-
-       if (parent)
-               memcg->swappiness = mem_cgroup_swappiness(parent);
-       atomic_set(&memcg->refcnt, 1);
-       memcg->move_charge_at_immigrate = 0;
-       mutex_init(&memcg->thresholds_lock);
-       spin_lock_init(&memcg->move_lock);
 
        error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
        if (error) {
@@ -6165,12 +6186,8 @@ mem_cgroup_css_alloc(struct cgroup *cont)
                 * call __mem_cgroup_free, so return directly
                 */
                mem_cgroup_put(memcg);
-               return ERR_PTR(error);
        }
-       return &memcg->css;
-free_out:
-       __mem_cgroup_free(memcg);
-       return ERR_PTR(error);
+       return error;
 }
 
 static void mem_cgroup_css_offline(struct cgroup *cont)
@@ -6780,6 +6797,7 @@ struct cgroup_subsys mem_cgroup_subsys = {
        .name = "memory",
        .subsys_id = mem_cgroup_subsys_id,
        .css_alloc = mem_cgroup_css_alloc,
+       .css_online = mem_cgroup_css_online,
        .css_offline = mem_cgroup_css_offline,
        .css_free = mem_cgroup_css_free,
        .can_attach = mem_cgroup_can_attach,