The ip addr output reveals duplicate IPv6 prefixes after a pppd restart.
*A custom x64 router running mainline Linux lost IPv6 connectivity whenever pppd renegotiated with the ISP. The failure traced to stale delegated prefixes and a missing RFC 9096 implementation in systemd-networkd.*
A hobbyist built a custom x64 router on mainline Linux, adding ad‑blocking and a full suite of daemons: networkd for routing, pppd for the ISP link, kea-dhcp4 for IPv4 leases, and blocky for DNS. The stack appeared rock solid until a routine pppd restart caused Google’s IPv6 address to resolve but never respond. Ping showed 100% loss while IPv4 traffic remained unaffected. The router’s ip addr output revealed two groups of global IPv6 addresses, the older one still holding a positive preferred lifetime. Restarting NetworkManager collapsed the duplicate group and restored connectivity, hinting at a stale address issue. The root cause lay in how the router handled delegated IPv6 prefixes during ISP‑initiated renumbering.
The router uses pppd to dial the ISP, networkd for routing, and kea-dhcp4 for IPv4 leases. When pppd restarts, the ISP hands a fresh IPv6 prefix to the router. Networkd does not drop the previously delegated prefix, so the client retains an old global address with a non‑zero preferred lifetime. DNS still resolves Google to its IPv6 address, but packets are sent from the stale source, leading to 100% packet loss. Restarting NetworkManager clears the extra address group, leaving only the new prefix and restoring connectivity. The symptom appears as a resolved IPv6 address that cannot be pinged, while the IPv4 path works unchanged.
RFC 9096, titled “Improving the Reaction of Customer Edge Routers to IPv6 Renumbering Events,” tightens the earlier RFC 7084. It mandates that when a router receives a new prefix and the upstream drops the old one, the router must advertise the old prefix with both preferred and valid lifetimes set to zero. This forces downstream devices to stop using the obsolete address as a source. The goal is to prevent traffic loss caused by lingering stale addresses after a prefix change.
Networkd’s automatic IPv6 range delegation lacks full RFC 7084 compliance. An issue tracking RFC 9096 support was opened in September 2026, but the patch remains unmerged. Similar requests on radvd and dnsmasq for OPNsense were closed as “not planned,” and VyOS’s dynamic‑prefix feature is still an open request. Consequently, networkd does not automatically advertise the old prefix with zero lifetimes, leaving clients with stale addresses after pppd renegotiation.
odhcpd, an OpenWrt package, implements RFC 9096 correctly. Unlike most OpenWrt services, odhcpd can operate without ubus by communicating directly with the kernel via netlink. After packaging odhcpd for NixOS and disabling networkd’s router advertisement, the author observed that a deliberate pppd restart left only the new, deprecated address on wlan0, and IPv6 connectivity to Google resumed. The solution avoids the need for a patched networkd and works within the existing NixOS router stack.
Restarting pppd forces the ISP to hand a new IPv6 prefix, but networkd does not drop the old address, leaving a stale source address that cannot reach the destination.
RFC 9096 requires routers to advertise the old prefix with preferred and valid lifetimes of zero so downstream devices stop using it as a source address.
odhcpd from OpenWrt implements RFC 9096 and can be used on NixOS to correctly handle IPv6 prefix changes.
The episode underscores a hidden fragility in many Linux‑based routers: without RFC 9096 support, automatic prefix delegation can silently break IPv6 traffic. Until systemd‑networkd integrates the required changes, operators must either patch the daemon or adopt a proven alternative like odhcpd. The stakes grow as IPv6 adoption expands, making proper renumbering compliance a non‑negotiable requirement for reliable connectivity.
Sources: [Hacker News – IPv6 delegation, an unusually subtle failure mode]