Need to shield the module from malformed packets further.

[ydahhrk]

Test case ID: N/A
Date: 2013/05/14
OS: N/A
Tester: AL
Error module: core, translate_packet
Description: Considering we'll eventually have to skip the kernel's packet defragmentation module, our validation routines are too weak. In order to avoid wild pointers, we'll have to ensure the next and length fields are consistent both in network and transport headers.
Observations: These validations need to be applied to ICMP inner packets too.

[magg]

Dude,

Why are you skipping the kernel's packet defragmentation function?

You could grab packets with the NF_INET_LOCAL_IN hook and receive only defragmented packets... or use conntrack but it seems it got removed... :(

[ydahhrk]

Because if the kernel reassembles the fragments for us, we lose the MTU computed during path MTU discovery. And that means we won't know the ideal fragment size when we need to refragment the packet. If the NAT64 picks an incorrect MTU, it could become a black hole.
IPv6 routers don't do fragmentation anyway (the NAT64 is a router), which means any reassembled packet from an incoming IPv4 interface would invariably be discarded as soon as it reaches IPv6 and needs to be fragmented.
I'm assuming these are the reasons why RFC 6146 never suggests the NAT64 should reassemble fragments (Except from zero-checksum UDP fragments). It says "translate all fragments at the same time", which is different.

[magg]

I always thought PMTUD was meant to avoid fragmentation...

source: http://packetlife.net/blog/2008/aug/18/path-mtu-discovery/

[ydahhrk]

Yup, my mistake; sorry.
But the IPv4-reassembled -> IPv6-refragment problem remains.
Since this issue is a little unrelated, I've opened a new one. Please continue the discussion in issue #59.