Re: [RFC net-next 10/15] ipxlat: add 4to6 pre-fragmentation path

[Xavier HSINYUAN]

Hi Ralf,

>+int ipxlat_46_plan_prefrag(struct ipxlat_priv *ipxlat, struct sk_buff *skb)
>+{
>+	unsigned int pkt_len6, pmtu6, threshold6, frag_max_size, pkt_len4,
>+		old_l3_len, new_l3_len;
>+	struct ipxlat_cb *cb = ipxlat_skb_cb(skb);
>+	const struct iphdr *in4 = ip_hdr(skb);
>+	int l3_delta, frag_l3_delta;
>+
>+	if (unlikely(cb->frag_max_size)) {
>+		DEBUG_NET_WARN_ON_ONCE(1);
>+		cb->frag_max_size = 0;
>+	}
>+
>+	pkt_len4 = iph_totlen(skb, in4);
>+	old_l3_len = cb->l3_hdr_len;
>+	new_l3_len = sizeof(struct ipv6hdr) +
>+		     (ip_is_fragment(in4) ? sizeof(struct frag_hdr) : 0);
>+	l3_delta = (int)new_l3_len - (int)old_l3_len;
>+	pkt_len6 = pkt_len4 + l3_delta;
>+
>+	pmtu6 = ipxlat_46_lookup_pmtu6(ipxlat, skb, in4);
>+	threshold6 = min(pmtu6, READ_ONCE(ipxlat->lowest_ipv6_mtu));
>+
>+	if (likely(pkt_len6 <= threshold6))
>+		return 0;
>+
>+	/* df packets are never locally pre-fragmented */
>+	if (likely(be16_to_cpu(in4->frag_off) & IP_DF)) {
>+		/* Let the IPv6 forwarding path raise PTB when needed and rely
>+		 * on the reverse 6->4 ICMP translation path for feedback.
>+		 */
>+		return 0;
>+	}

How about putting the DF check before the PMTU lookup?
ipxlat_46_lookup_pmtu6() requires ip6_route_output() and dst_release(),
but the DF bit is much cheaper to test and is typically set on most
TCP/QUIC packets. It looks like a pure reorder but I have not verified
it carefully.

Best regards,
Xavier