BugLink: http://bugs.launchpad.net/bugs/1566505

Currently a new mount of an existing hierarchy always reuses the
original super block, even when the new mount is in a cgroup
namespace. This sometimes conflicts with the user namespace mount
support, which requires a new mount of an existing super block to
be in the same user namespace as the original mount. When
mounting from non-init cgroup and user namespaces sget() will
fail.

To fix this we can pass a pointer to the cgroup ns to kernfs when
mounting, causing kernfs_test_super() to no longer match super
blocks from different cgroup namespaces.

However we do wish to continue sharing the cgroup_root between
mounts of the same heirarchy. The cgroup_root's lifetime is
governed by the reference count of its cgrp member, but this is
a percpu reference count and is not well suited to this new
situation. Instead a new reference count is added to the
cgroup_root structure to track the number of super blocks sharing
that root, and this refcnt is used to determine when to put the
cgroup reference.
Signed-off-by: Seth Forshee <seth.forshee@canonical.com>
Signed-off-by: Tim Gardner <tim.gardner@canonical.com>