[patch] Source entries removing is awfully slow.

[Kajetan Staszkiewicz]

Hi,

I think I have the answer.

1. Some traffic creates a nat src node and some states.
2. Those states are properly linked to src_node->state_list, each has a proper
   pointer to nat_src_node.
3. At some point insertion of state (I do not for what reason) fails in this
   code:

3970         if (pf_state_insert(BOUND_IFACE(r, kif), skw, sks, s)) {
3971                 if (pd->proto == IPPROTO_TCP)
3972                         pf_normalize_tcp_cleanup(s);
3973                 REASON_SET(&reason, PFRES_STATEINS);
3974                 pf_src_tree_remove_state(s);
3975                 STATE_DEC_COUNTERS(s);
3976 #ifdef __FreeBSD__
3977                 pool_put(&V_pf_state_pl, s);

This state already has nat_src_node properly pointing to the src node.

pf_src_tree_remove_state() is called:
- s->nat_src_node is not NULL
- TAILQ_EMPTY is false, as the src_node has a state_list containing some
  previously and properly created states
- TAILQ_REMOVE fails because state s is not in the list, s->srcnode_link is
  {NULL,NULL}, src_node->state_list's head gets broken, giving the result as in
  my previous post and kernel panic.

With calling TAILQ_INSERT_HEAD before any pf_src_tree_remove_state is 
potentally called, I have a kernel running stable since the last week.

-- 
| pozdrawiam / greetings | powered by Debian, CentOS and FreeBSD |
|  Kajetan Staszkiewicz  | jabber,email: vegeta()tuxpowered net  |
|        Vegeta          | www: http://vegeta.tuxpowered.net     |
`------------------------^---------------------------------------'