Re: [RFC net-next v2 11/18] net: ipxlat: add IPv4/IPv6 header and transport translation

[Ralf Lici <ralf@mandelbit.com>]

On Fri, 06 Nov 2026 21:37:52 +0800, Xavier HSINYUAN <XavierHsinyuan@outlook.com> wrote:
> Hi Ralf,
>
> > +static int ipxlat_46_translate_tcp_csum(struct sk_buff *skb,
> > +					const struct iphdr *in4)
> > +{
> > +	const struct ipxlat_cb *cb = ipxlat_skb_cb(skb);
> > +	const struct ipv6hdr *out6 = ipv6_hdr(skb);
> > +	struct tcphdr *tcp = tcp_hdr(skb);
> > +
> > +	/* CHECKSUM_PARTIAL contains a pseudo-header seed rather than a final
> > +	 * checksum. Replace it with the translated IPv6 pseudo-header and
> > +	 * leave completion to the later output path.
> > +	 */
> > +	if (skb->ip_summed == CHECKSUM_PARTIAL) {
> > +		tcp->check = ~tcp_v6_check(ipxlat_skb_l4_seg_len(skb, cb),
> > +					   &out6->saddr, &out6->daddr, 0);
> > +		return ipxlat_set_partial_csum(skb,
> > +					       offsetof(struct tcphdr, check));
> > +	}
> > +
> > +	tcp->check = ipxlat_46_update_l4_csum(tcp->check, in4, out6);
> > +	skb->ip_summed = CHECKSUM_NONE;
> > +	return 0;
> > +}
> > +
> > +static int ipxlat_46_translate_udp_csum(struct sk_buff *skb,
> > +					const struct iphdr *in4)
> > +{
> > +	const struct ipxlat_cb *cb = ipxlat_skb_cb(skb);
> > +	const struct ipv6hdr *out6 = ipv6_hdr(skb);
> > +	struct udphdr *udp = udp_hdr(skb);
> > +
> > +	if (skb->ip_summed == CHECKSUM_PARTIAL) {
> > +		udp->check = ~udp_v6_check(ipxlat_skb_l4_seg_len(skb, cb),
> > +					   &out6->saddr, &out6->daddr, 0);
> > +		return ipxlat_set_partial_csum(skb,
> > +					       offsetof(struct udphdr, check));
> > +	}
> > +
> > +	udp->check = ipxlat_46_update_l4_csum(udp->check, in4, out6);
> > +	if (unlikely(!udp->check))
> > +		udp->check = CSUM_MANGLED_0;
> > +	skb->ip_summed = CHECKSUM_NONE;
> > +	return 0;
> > +}
>
> Pseudo-header seed in CHECKSUM_PARTIAL uses per-segment length instead
> of the aggregate length, resulting in bad checksums for almost every
> segments.
>

You're right, just reproduced this locally. I confused the eventual
segment length with the length represented by the CHECKSUM_PARTIAL seed
on a GSO skb (the 6->4 path has the same bug). The current tests didn't
catch the bug because veth passes CHECKSUM_PARTIAL packets through, so
nothing ever computes or verifies the checksum.

> How about updating {tcp,udp}->check with a pseudo-header diff like
> inet_proto_csum_replace_by_diff() to preserve the existing seed?
>

Yes, the next revision will preserve the existing seed in both
directions by applying only the address delta, as the non-PARTIAL path
already does.

Thanks for catching this!

-- 
Ralf Lici
Mandelbit Srl