DENOG18

IPv4 as a service on an IPv6-only network
2026-11-16 –, Saal A

Routing IPv4 packets using IPv6 next-hops has been around for some time now. It offers great opportunities to reduce IPv4 address usage in the distribution and access layer, and enables providing IPv4 as a pure add-on, keeping a clean IPv6 address structure underneath.

Despite its benefits, major challenges for deploying IPv4-with-IPv6 next-hops in practice persist. While, e.g., BGP supports such routes using RFC8950, many other IGPs still lack support. Other major issues are 'the last mile', i.e., auto-configuring clients (or, for this use-case, more likely servers), and the impact of IPv4-with-IPv6 next-hop usage on traceroutes.

In this talk, we will go over some current developments in these areas, which should make IPv4-with-IPv6 next-hops a viable option for practical deployments in the coming years. We will also expand further on possible deployment scenarios, where IPv4-with-IPv6 next-hops can remove further legacy Layer-2 infrastructure, e.g., LACP.

What actually changes in your network when you stop treating IPv4 as infrastructure and start treating it as a service?

draft-vanmook-intarea-ipv6-resolved-gateway (presented at IETF 126, WG adoption call pending) defines a single sentinel gateway address, provisionally 192.0.0.11 (requested from the IANA Special-Purpose Registry, not yet assigned), handed out as the ordinary DHCPv4 default gateway. Updated hosts resolve their first hop from the IPv6 neighbour cache; unmodified hosts ARP for the sentinel once and the router answers. That's the whole mechanism. Everything downstream of it is the interesting part.

This talk is about what that does to a running network: no IPv4 subnets anywhere, no ARP, no FHRP, IPv4 address pools that go flat — any /32 works anywhere in the footprint, nothing stranded by subnet sizing. The core forwards v4-via-v6 (RFC 8950) with no IPv4 configured on any router interface.

On the implementation side: host implementations for Linux (userspace daemon and native systemd-networkd), FreeBSD, macOS and Windows, with prebuilt packages; a namespace conformance lab with a zero-ARP assertion; verified unmodified-host behaviour across seven operating systems; and vendor configuration examples for IOS XR, JunOS, SR OS, EOS and RouterOS. On the router side, three working deployment models: a single-gateway reference, an active/active anycast VRRP pair, and a BGP-orchestrated multi-gateway fabric — fungible gateways, host /32s injected from inventory, and the entire DHCPv4 path carried as IPv4 values over IPv6 next-hops, with no IPv4 management network at all.

Plus the failure modes, the ND health prerequisites, and what breaks when you get it wrong.

No tunnels. No translation. No flag day. Keep the shape, change what it means.

See also: Slides (joined slides with Remco) (3.8 MB)

Tobias is a researcher at TU Wien, broadly working on everything related to the Internet; From pushing packets to the people getting the packets pushed. Besides his academic work, he enjoys running networks in his free time.

I thought I was done with system level engineering a long time ago, it turns out system level engineering wasn't done with me.

Technology executive with a passion for Internet and how the pieces come together - going down that rabbit hole for 25+ years. Prolific policy author, former chair of the Connect working group in the RIPE community and board member of RIPE NCC from 2010 to 2025.