After converting to use rcu for conntrack hash, one CPU may update
the ct->status via ctnetlink, while another CPU may process the
packets and update the ct->status.

So the non-atomic operation "ct->status |= status;" via ctnetlink
becomes unsafe, and this may clear the IPS_DYING_BIT bit set by
another CPU unexpectedly. For example:
         CPU0                            CPU1
  ctnetlink_change_status        __nf_conntrack_find_get
      old = ct->status              nf_ct_gc_expired
          -                         nf_ct_kill
          -                      test_and_set_bit(IPS_DYING_BIT
      new = old | status;                 -
  ct->status = new; <-- oops, _DYING_ is cleared!

Now using a series of atomic bit operation to solve the above issue.

Also note, user shouldn't set IPS_TEMPLATE, IPS_SEQ_ADJUST directly,
so make these two bits be unchangable too.

If we set the IPS_TEMPLATE_BIT, ct will be freed by nf_ct_tmpl_free,
but actually it is alloced by nf_conntrack_alloc.
If we set the IPS_SEQ_ADJUST_BIT, this may cause the NULL pointer
deference, as the nfct_seqadj(ct) maybe NULL.

Last, add some comments to describe the logic change due to the
commit a963d710 ("netfilter: ctnetlink: Fix regression in CTA_STATUS
processing"), which makes me feel a little confusing.

Fixes: 76507f69 ("[NETFILTER]: nf_conntrack: use RCU for conntrack hash")
Signed-off-by: Liping Zhang <zlpnobody@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>