Re: 2.6.17-rc4: netfilter LOG messages truncated via NETCONSOLE

[Patrick McHardy <kaber@trash.net>]

[-- Attachment #1: Type: text/plain, Size: 3581 bytes --]

Frank van Maarseveen wrote:
> I have two machines named "porvoo" and "espoo". The first one
> has netconsole configured to send kernel messages to UDP port 514
> (a.k.a. syslog) on the other machine.
> 
> Somewhere between 2.6.13.2 and 2.6.17-rc4 there is a regression causing
> the netconsole messages which originate from netfilter to be truncated
> right after the MAC addresses. For example, /var/log/messages on the
> sending machine says:
> 
> May 31 09:28:11 porvoo kernel: IN=eth0 OUT= MAC=00:12:3f:85:9f:92:00:04:9a:a0:1d:d1:08:00 SRC=192.168.100.30 DST=172.17.1.113 LEN=60 TOS=0x00 PREC=0x00 TTL=54 ID=51496 DF PROTO=TCP SPT=50868 DPT=22 WINDOW=5840 RES=0x00 SYN URGP=0 
> 
> but netconsole messages captured in /var/log/messages on the receiving
> machine:
> 
> May 31 09:28:11 porvoo IN=eth0 OUT= 
> May 31 09:28:11 porvoo MAC=
> May 31 09:28:11 porvoo 00:
> May 31 09:28:11 porvoo 12:
> May 31 09:28:11 porvoo 3f:
> May 31 09:28:11 porvoo 85:
> May 31 09:28:11 porvoo 9f:
> May 31 09:28:11 porvoo 92:
> May 31 09:28:11 porvoo 00:
> May 31 09:28:11 porvoo 04:
> May 31 09:28:11 porvoo 9a:
> May 31 09:28:11 porvoo a0:
> May 31 09:28:11 porvoo 1d:
> May 31 09:28:11 porvoo d1:
> May 31 09:28:11 porvoo 08:
> May 31 09:28:11 porvoo 00 
> May 31 09:49:06 espoo -- MARK --
> 
> I ran a tcpdump on the sending machine to verify(?) what goes out but in
> that case the 2.6.17-rc4 kernel starts to report "protocol 0000 is buggy":
> 
> May 31 11:00:49 porvoo kernel: device eth0 entered promiscuous mode
> May 31 11:03:31 porvoo kernel: IN=eth0 OUT= MAC=00:12:3f:85:9f:92:00:12:3f:85:17:52:08:00 SRC=172.17.1.64 DST=172.17.1.113 LEN=60 TOS=0x00 PREC=0x00 TTL=64 ID=60618 DF PROTO=TCP SPT=34984 DPT=21212 WINDOW=5840 RES=0x00 SYN URGP=0
> May 31 11:03:31 porvoo kernel: protocol 0000 is buggy, dev eth0
> May 31 11:03:31 porvoo last message repeated 9 times
> 
> the netconsole output captured on the receiving machine:
> 
> May 31 11:00:49 porvoo device eth0 entered promiscuous mode 
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo IN=eth0 OUT= 
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo MAC=
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 00:
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 12:
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 3f:
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 85:
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 9f:
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 92:
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 00:
> May 31 11:03:31 porvoo protocol 0000 is buggy, dev eth0 
> May 31 11:03:31 porvoo 12:
> May 31 11:03:31 porvoo 3f:
> May 31 11:03:31 porvoo 85:
> May 31 11:03:31 porvoo 17:
> May 31 11:03:31 porvoo 52:
> May 31 11:03:31 porvoo 08:
> May 31 11:03:31 porvoo 00 
> 
> again 9 packets are missing and there are 9 "protocol 0000 is buggy"
> messages. Netfilter has almost everything configured. IPv6 is left out
> everywhere. Above has been produced using this rule on the netconsole
> sending machine:
> 
> iptables -I INPUT -p tcp -s espoo --dport 21212 -j LOG


The message means that there was recursion and netpoll fell back
to dev_queue_xmit This patch should fix the "protocol is buggy"
messages, netpoll didn't set skb->nh.raw. Please try if it also
makes the other problem go away.

[-- Attachment #2: x --]
[-- Type: text/plain, Size: 452 bytes --]

diff --git a/net/core/netpoll.c b/net/core/netpoll.c
index e8e05ce..7a4787c 100644
--- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -329,7 +329,7 @@ void netpoll_send_udp(struct netpoll *np
 	udph->len = htons(udp_len);
 	udph->check = 0;
 
-	iph = (struct iphdr *)skb_push(skb, sizeof(*iph));
+	skb->nh.iph = iph = (struct iphdr *)skb_push(skb, sizeof(*iph));
 
 	/* iph->version = 4; iph->ihl = 5; */
 	put_unaligned(0x45, (unsigned char *)iph);