cgroup: fix rmdir EBUSY regression in 3.11

On 3.11-rc we are seeing cgroup directories left behind when they should
have been removed.  Here's a trivial reproducer:

cd /sys/fs/cgroup/memory
mkdir parent parent/child; rmdir parent/child parent
rmdir: failed to remove `parent': Device or resource busy

It's because cgroup_destroy_locked() (step 1 of destruction) leaves
cgroup on parent's children list, letting cgroup_offline_fn() (step 2 of
destruction) remove it; but step 2 is run by work queue, which may not
yet have removed the children when parent destruction checks the list.

Fix that by checking through a non-empty list of children: if every one
of them has already been marked CGRP_DEAD, then it's safe to proceed:
those children are invisible to userspace, and should not obstruct rmdir.

(I didn't see any reason to keep the cgrp->children checks under the
unrelated css_set_lock, so moved them out.)

tj: Flattened nested ifs a bit and updated comment so that it's
    correct on both for-3.11-fixes and for-3.12.

Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Tejun Heo <tj@kernel.org>

1 files changed, 18 insertions, 1 deletions

diff --git a/kernel/cgroup.c b/kernel/cgroup.c
index 781845a013ab..e91963302c0d 100644
--- a/kernel/cgroup.c
+++ b/kernel/cgroup.c
@@ -4480,6 +4480,7 @@ static int cgroup_destroy_locked(struct cgroup *cgrp)
         struct dentry *d = cgrp->dentry;
         struct cgroup_event *event, *tmp;
         struct cgroup_subsys *ss;
+        struct cgroup *child;
         bool empty;
 
         lockdep_assert_held(&d->d_inode->i_mutex);
@@ -4490,12 +4491,28 @@ static int cgroup_destroy_locked(struct cgroup *cgrp)
          * @cgrp from being removed while __put_css_set() is in progress.
          */
         read_lock(&css_set_lock);
-        empty = list_empty(&cgrp->cset_links) && list_empty(&cgrp->children);
+        empty = list_empty(&cgrp->cset_links);
         read_unlock(&css_set_lock);
         if (!empty)
                 return -EBUSY;
 
         /*
+         * Make sure there's no live children.  We can't test ->children
+         * emptiness as dead children linger on it while being destroyed;
+         * otherwise, "rmdir parent/child parent" may fail with -EBUSY.
+         */
+        empty = true;
+        rcu_read_lock();
+        list_for_each_entry_rcu(child, &cgrp->children, sibling) {
+                empty = cgroup_is_dead(child);
+                if (!empty)
+                        break;
+        }
+        rcu_read_unlock();
+        if (!empty)
+                return -EBUSY;
+
+        /*
          * Block new css_tryget() by killing css refcnts.  cgroup core
          * guarantees that, by the time ->css_offline() is invoked, no new
          * css reference will be given out via css_tryget().  We can't