Hello, On Sun, 20 Oct 2013, Hannes Frederic Sowa wrote: > > Hm, maybe. I don't have too much insight into netfilter stack and > > what are the differences between OUTPUT and FORWARD path but plan to > > investigate. ;) > > It seems tables are processed with bh disabled, so no preemption while > recursing. So I guess the use of tee_active is safe for breaking the > tie here. May be, I'll check it again, for now I see only rcu_read_lock() in nf_hook_slow() which is preemptable. Looking at rcu_preempt_note_context_switch, many levels of RCU locks are preemptable too. > The reason I exhaust stack space is that we can actually send out packets > while looking up routes (rt6_probe). The nonreachability of the default > gateway and the to-teed-to box does the rest. In my test I used link route to local subnet, --gateway to IP that is not present. I'll try other variants. > We need to change the route lookup of the duplicated packet in xt_tee to not > cause ndisc probes to be generated. > > The more I review the patch the more I think it is ok. But we could actually > try to just always return rt6i_gateway, as we should always be handed a cloned > rt6_info where the gateway is already filled in, no? Yes, this patch is ok and after spending the whole saturday I'm preparing a new patch that will convert rt6_nexthop() to return just rt6i_gateway, without daddr. This can happen after filling rt6i_gateway in all places. For your concern for loopback, I don't see problem, local/anycast route will have rt6i_gateway=IP, they are simple DST_HOST routes. I'm preparing now the patches and will post them in following hours. Regards -- Julian Anastasov <ja@xxxxxx> -- To unsubscribe from this list: send the line "unsubscribe lvs-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
