Linux kernel network device driver and skb pointers - linux-kernel

Question

I am writing a network device driver.
Kernel 2.6.35.12
The device is supposed to be working when it is connected to a bridge port.
I am trying to intercept ICMPv6 RA and NS messages (Router/ Neighbor solicitation) forwarded to the interface from the bridge.
eth <–> br0 <–> mydevice
In the device start_xmit function I am doing to following:
Check that the protocol field after the Ethernet header is IPV6 (0x86dd)
Check that the ipv6 next header is ICMPv6 and check its type:
__u8 nexthdr = ipv6_hdr(skb)->nexthdr;
if (nexthdr == htons (IPPROTO_ICMPV6))
{
struct icmp6hdr *hdr = icmp6_hdr(skb);
u8 type = hdr->icmp6_type;
if(type == htons (NDISC_NEIGHBOUR_SOLICITATION) || type == htons (NDISC_ROUTER_SOLICITATION))
{
….Do something here…
}
}
When RS/NS are sent from within the device (e.g br0), I see that the code is working right.
The problem is when traffic is forwarded through the bridge from the other port.
I see that the icmp6_hdr(skb) returns an incorrect header.
Debugging some more, it seems that the
skb->network_header and the skb->transport_header are pointing to the same place.
icmp6_hdr is using the transport_header which explain why it is incorrect.
Dumping the skb data it looks that all the headers and payload are at the right offset (also compared it with tcpdump)
I suspect that it might be related to the bridge code, before going to dive into it,
I thought that maybe anyone had come up against anything similar or have any other ideas?

Answer 1

Part of the problem is that you are assuming that Netfilter did anything more than just figure out what was the next header. In my experience (albeit not very long) you want to do something like this:
struct icmp6hdr *icmp6;
// Obviously don't do this unless you check to make sure that it's the right protocol
struct ipv6_hdr *ip6hdr = (struct ipv6_hdr*)skb->network_header;
// You need to move the headers around
// Notice the memory address of skb->data and skb->network_header are the same
// that means that the IP header hasn't been "pulled"
skb->transport_header = skb_pull(skb, sizeof(struct ipv6_hdr));
if(ntohs(ip6hdr->nexthdr) == IPPROTO_ICMPV6) {
icmp6 = (struct icmp6hdr*)skb->transport_header;
// Doing this is more efficient, since you only are calling the
// Network to Host function once
__u8 type = ntohs(hdr->icmp6_type);
switch(type) {
case NDISC_NEIGHBOUR_SOLICITATION:
case NDISC_ROUTER_SOLICITATION:
// Do your stuff
break;
}
}
Hopefully this was helpful. I just started diving into writing Netfilter code, so I am not exactly certain 100%, but I found this out when I was trying to do something similar with IPv4 on the NF_IP_LOCAL_IN hook.