The __qdisc_destroy rcu-callback doesn't do any locking when calling
ops->reset and ops->destroy. qdisc_destroy is often called from both
of these functions and it changes dev->qdisc_list. This patch adds proper
locking to __qdisc_destroy. Unfortunately when using qdisc_tree_lock in
process context we now also need to disable local bh's to avoid beeing
interrupted by the rcu-callback. I'm not sure if RCU callback can be
scheduled while the kernel is running in process context, so this may
be unneccessary.