netfilter: don't track fragmented packets

Andrey reports syzkaller splat caused by

NF_CT_ASSERT(!ip_is_fragment(ip_hdr(skb)));

in ipv4 nat.  But this assertion (and the comment) are wrong, this function
does see fragments when IP_NODEFRAG setsockopt is used.

As conntrack doesn't track packets without complete l4 header, only the
first fragment is tracked.

Because applying nat to first packet but not the rest makes no sense this
also turns off tracking of all fragments.

Reported-by: Andrey Konovalov <andreyknvl@google.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>

net/ipv4/netfilter/nf_conntrack_l3proto_ipv4.c

@@ -165,6 +165,10 @@ static unsigned int ipv4_conntrack_local(void *priv,
 	if (skb->len < sizeof(struct iphdr) ||
 	    ip_hdrlen(skb) < sizeof(struct iphdr))
 		return NF_ACCEPT;
+
+	if (ip_is_fragment(ip_hdr(skb))) /* IP_NODEFRAG setsockopt set */
+		return NF_ACCEPT;
+
 	return nf_conntrack_in(state->net, PF_INET, state->hook, skb);
 }
 

net/ipv4/netfilter/nf_nat_l3proto_ipv4.c

@@ -255,11 +255,6 @@ nf_nat_ipv4_fn(void *priv, struct sk_buff *skb,
 	/* maniptype == SRC for postrouting. */
 	enum nf_nat_manip_type maniptype = HOOK2MANIP(state->hook);
 
-	/* We never see fragments: conntrack defrags on pre-routing
-	 * and local-out, and nf_nat_out protects post-routing.
-	 */
-	NF_CT_ASSERT(!ip_is_fragment(ip_hdr(skb)));
-
 	ct = nf_ct_get(skb, &ctinfo);
 	/* Can't track?  It's not due to stress, or conntrack would
 	 * have dropped it.  Hence it's the user's responsibilty to