Curtis Taylor and Jon Maxwell reported and debugged a crash on 3.10
based kernel.

Crash occurs in ctnetlink_conntrack_events because net->nfnl socket is
NULL.  The nfnl socket was set to NULL by netns destruction running on
another cpu.

The exiting network namespace calls the relevant destructors in the
following order:

1. ctnetlink_net_exit_batch

This nulls out the event callback pointer in struct netns.

2. nfnetlink_net_exit_batch

This nulls net->nfnl socket and frees it.

3. nf_conntrack_cleanup_net_list

This removes all remaining conntrack entries.

This is order is correct. The only explanation for the crash so ar is:

cpu1: conntrack is dying, eviction occurs:
 -> nf_ct_delete()
   -> nf_conntrack_event_report \
     -> nf_conntrack_eventmask_report
       -> notify->fcn() (== ctnetlink_conntrack_events).

cpu1: a. fetches rcu protected pointer to obtain ctnetlink event callback.
      b. gets interrupted.
 cpu2: runs netns exit handlers:
     a runs ctnetlink destructor, event cb pointer set to NULL.
     b runs nfnetlink destructor, nfnl socket is closed and set to NULL.
cpu1: c. resumes and trips over NULL net->nfnl.

Problem appears to be that ctnetlink_net_exit_batch only prevents future
callers of nf_conntrack_eventmask_report() from obtaining the callback.
It doesn't wait of other cpus that might have already obtained the
callbacks address.

I don't see anything in upstream kernels that would prevent similar
crash: We need to wait for all cpus to have exited the event callback.

Fixes: 9592a5c0 ("netfilter: ctnetlink: netns support")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>