|
Hello,
On Sat, 15 Feb 2003, Roberto Nibali wrote:
> > Jacob provided me with tcpdump on the director. It shows
> > that director replies with ICMP but now the question is whether
> > these ICMP messages are dropped in client (it is known they are
> > received there), may be they have bad csum?
>
> The windows client? Could I see a windump of the windows client
> receiving those ICMP messages?
I hope Jacob will send you the same output.
> > ip_vs_tunnel_xmit(and friends) calls 'ip_send_check(old_iph);'
> > after the ICMP_FRAG_NEEDED is generated. The new datagram
> > will have correct iphdr sum but what about the embedded iphdr?
>
> According to ipv4/ipip.c and ipv4/ip_gre.c it should be correct, so
> there is no need to re-adjust csum. Then again I'm no expert with tunneling.
Yes, you are right. I don't see a good reason for wrong
checksum before reaching that point, I don't think someone already
fragmented/changed the datagram with DF bit. So, the interesting part is
what receives the client.
> I don't think fast_csum is needed, maybe a ip_rt_dump() could help
> what's wrong? Is ptmu enabled on his director?
This flag is used only for sockets, may be it does not play here
in LVS director.
> > icmp_send(skb, ICMP_DEST_UNREACH,ICMP_FRAG_NEEDED, htonl(mtu));
> > ip_rt_put(rt);
>
> Stupid question: I always wonder if ip_rt_put(rt) should not be before
> the icmp_send or is this only in the NAT case?
I don't remember for any rule concerning the
icmp_send ip_rt_put order in this case, so the order does not matter.
> > IP_VS_DBG_RL("ip_vs_tunnel_xmit(): frag needed\n");
>
> Does he see those messages when he turns on debug?
May be yes, at least the ICMP messages are shown from
tcpdump.
> Best regards,
> Roberto Nibali, ratz
Regards
--
Julian Anastasov <ja@xxxxxx>
|
