The locking of release_lock was stupid; t should have been be called with
fence_lock_irq if it was legitimately used. Unfortunately it never protected
anything except the fence implementation correctly.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>