On Wed, 2 Oct 2019 09:38:55 -0700 Eric Dumazet wrote:
> This began with a syzbot report. syzkaller was injecting
> IPv6 TCP SYN packets having a v4mapped source address.
>
> After an unsuccessful 4-tuple lookup, TCP creates a request
> socket (SYN_RECV) and calls reqsk_queue_hash_req()
>
> reqsk_queue_hash_req() calls sk_ehashfn(sk)
>
> At this point we have AF_INET6 sockets, and the heuristic
> used by sk_ehashfn() to either hash the IPv4 or IPv6 addresses
> is to use ipv6_addr_v4mapped(&sk->sk_v6_daddr)
>
> For the particular spoofed packet, we end up hashing V4 addresses
> which were not initialized by the TCP IPv6 stack, so KMSAN fired
> a warning.
>
> I first fixed sk_ehashfn() to test both source and destination addresses,
> but then faced various problems, including user-space programs
> like packetdrill that had similar assumptions.
>
> Instead of trying to fix the whole ecosystem, it is better
> to admit that we have a dual stack behavior, and that we
> can not build linux kernels without V4 stack anyway.
>
> The dual stack API automatically forces the traffic to be IPv4
> if v4mapped addresses are used at bind() or connect(), so it makes
> no sense to allow IPv6 traffic to use the same v4mapped class.
>
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Signed-off-by: Eric Dumazet <[email protected]>
> Cc: Florian Westphal <[email protected]>
> Cc: Hannes Frederic Sowa <[email protected]>
> Reported-by: syzbot <[email protected]>
FTR this appears to break an UDP/sFlow application which used to work
fine with mapped addresses. Given the IETF memo perhaps a sysctl would
be appropriate, but even with that we're back to problems in TCP if the
sysctl is flipped :S
> diff --git a/net/ipv6/ip6_input.c b/net/ipv6/ip6_input.c
> index
> d432d0011c160f41aec09640e95179dd7b364cfc..2bb0b66181a741c7fb73cacbdf34c5160f52d186
> 100644
> --- a/net/ipv6/ip6_input.c
> +++ b/net/ipv6/ip6_input.c
> @@ -223,6 +223,16 @@ static struct sk_buff *ip6_rcv_core(struct sk_buff *skb,
> struct net_device *dev,
> if (ipv6_addr_is_multicast(&hdr->saddr))
> goto err;
>
> + /* While RFC4291 is not explicit about v4mapped addresses
> + * in IPv6 headers, it seems clear linux dual-stack
> + * model can not deal properly with these.
> + * Security models could be fooled by ::ffff:127.0.0.1 for example.
> + *
> + * https://tools.ietf.org/html/draft-itojun-v6ops-v4mapped-harmful-02
> + */
> + if (ipv6_addr_v4mapped(&hdr->saddr))
> + goto err;
> +
> skb->transport_header = skb->network_header + sizeof(*hdr);
> IP6CB(skb)->nhoff = offsetof(struct ipv6hdr, nexthdr);
>
