cgroup: don't online subsystems before cgroup_name/path() are operati…

…onal

While refactoring cgroup creation, a5bca21 ("cgroup: factor out
cgroup_create() out of cgroup_mkdir()") incorrectly onlined subsystems
before the new cgroup is associated with it kernfs_node.  This is fine
for cgroup proper but cgroup_name/path() depend on the associated
kernfs_node and if a subsystem makes the new cgroup_subsys_state
visible, which they're allowed to after onlining, it can lead to NULL
dereference.

The current code performs cgroup creation and subsystem onlining in
cgroup_create() and cgroup_mkdir() makes the cgroup and subsystems
visible afterwards.  There's no reason to online the subsystems early
and we can simply drop cgroup_apply_control_enable() call from
cgroup_create() so that the subsystems are onlined and made visible at
the same time.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Fixes: a5bca21 ("cgroup: factor out cgroup_create() out of cgroup_mkdir()") 
Cc: stable@vger.kernel.org # v4.6+

kernel/cgroup.c

@@ -5221,6 +5221,11 @@ static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
 	return ERR_PTR(err);
 }
 
+/*
+ * The returned cgroup is fully initialized including its control mask, but
+ * it isn't associated with its kernfs_node and doesn't have the control
+ * mask applied.
+ */
 static struct cgroup *cgroup_create(struct cgroup *parent)
 {
 	struct cgroup_root *root = parent->root;
@@ -5288,21 +5293,13 @@ static struct cgroup *cgroup_create(struct cgroup *parent)
 
 	cgroup_propagate_control(cgrp);
 
-	/* @cgrp doesn't have dir yet so the following will only create csses */
-	ret = cgroup_apply_control_enable(cgrp);
-	if (ret)
-		goto out_destroy;
-
 	return cgrp;
 
 out_cancel_ref:
 	percpu_ref_exit(&cgrp->self.refcnt);
 out_free_cgrp:
 	kfree(cgrp);
 	return ERR_PTR(ret);
-out_destroy:
-	cgroup_destroy_locked(cgrp);
-	return ERR_PTR(ret);
 }
 
 static int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name,