⏱️ Reading time: 8 min
When Savyasachee Jha restarted his home router’s PPP link, Google stopped loading. The name still resolved to a valid IPv6 address, but every packet was lost: four ping attempts, four failures, zero received. The problem wasn’t DNS or the ISP: it was how his router handles IPv6 prefix delegation, the mechanism that assigns it a new network block every time the connection with the ISP is renegotiated (source).
📑 En este artículo
- What IPv6 Prefix Delegation Is
- The Real Case: DNS Resolves, but the Packet Never Arrives
- Why It Happens: The Announcement RFC 9096 Requires and Almost No Router Sends
- What This Means for You if You Run Your Own Router
- odhcpd vs. networkd, radvd and dnsmasq
- What the Community Is Debating: Renumbering Less Matters More Than Announcing Well
- Limitations and When This Doesn’t Affect You
- References
What IPv6 Prefix Delegation Is
In IPv4, your router shares a single public address across the whole house through NAT. IPv6 works differently: the ISP assigns your router an entire block of addresses (a prefix, typically a /56 or /60) to distribute among its internal networks. That distribution runs on two protocols working together. DHCPv6-PD (Prefix Delegation) is the negotiation between the ISP and your router: it hands over a block for you to use for the duration of that session. Router Advertisement (RA) is what your router then tells the laptops, phones and servers on the local network: this is the subnet you get, and this address is valid for this many seconds (source).
The weak point sits right at the handoff between those two protocols. If the ISP takes back the old prefix and hands your router a new one, something has to tell every laptop and phone that the address they had no longer works. If nobody does, the device keeps using it as if nothing changed.
The Real Case: DNS Resolves, but the Packet Never Arrives
Jha runs his own router on NixOS: one daemon handles RA, another dials the ISP, a third hands out IPv4 addresses, and a fourth resolves DNS on both the router and his laptop. Each piece, on its own, is proven, stable software: nothing in that stack looked suspicious at first (source).
The diagnosis started with the obvious: a ping to google.com resolved to the correct IPv6 address but lost 100% of its packets. Checking the laptop’s addresses revealed the key detail: the interface held two sets of global IPv6 addresses at once, one newer and one older, and the older set was still marked valid even though its preferred lifetime was already running out. Restarting the laptop’s network manager fixed it immediately: only the new address set remained, and Google responded again (source).
What ruled out a one-off laptop glitch was seeing the same behavior on an iPhone connected to the same network. Two devices with completely different network stacks failing identically points to something they share, and the only thing in common is the router, specifically its role in advertising addresses via RA.
Why It Happens: The Announcement RFC 9096 Requires and Almost No Router Sends
The cause has a name and a number. RFC 7084, Basic Requirements for IPv6 Customer Edge Routers, defines what a home router must do with prefix delegation. RFC 9096, Improving the Reaction of Customer Edge Routers to IPv6 Renumbering Events, tightens that requirement: when the ISP withdraws a prefix and hands out a new one, the router must advertise the old prefix with both preferred and valid lifetime set to zero. That zero-lifetime announcement is the explicit signal for every device on the network to stop using that address as the source for its connections (RFC 9096, RFC 7084).
Without that announcement, the device has no way of knowing the address is now orphaned. It keeps treating it as valid until it expires on its own, which can take minutes or hours depending on the original lifetime. In the meantime, any new connection that picks that old address as its source goes out with an IP the ISP no longer routes back to, and the packet gets lost along the way. Name resolution, on the other hand, doesn’t depend on which source address you use to reach the internet, only on reaching a DNS server: that’s why DNS keeps working while the actual connection silently fails.
What This Means for You if You Run Your Own Router
If you manage a home router on plain Linux (NixOS, Debian, Arch) or on pfSense or OPNsense behind an ISP that renegotiates the connection fairly often (PPPoE that drops, a short DHCPv6-PD lease, provider maintenance), there’s a concrete check you can run this week: force a redial of the WAN link and watch, from a LAN client, whether the old IPv6 addresses disappear or linger alongside the new ones.
If you use a systemd-based daemon for RA, you won’t see that withdrawal announcement today: full RFC 7084 support is still an open issue and the patch hasn’t been merged yet. If you use radvd or dnsmasq on OPNsense, you won’t see it either: the request was closed as not planned on the first and remains unresolved on the second. VyOS has the same request open for its dynamic prefix support. The only piece that already implements RFC 9096 is odhcpd, OpenWrt’s daemon, which for anyone outside OpenWrt means packaging it separately (Jha did this for nixpkgs) or migrating that specific role to a router that already ships with it (source).
odhcpd vs. networkd, radvd and dnsmasq
All four pieces solve the same problem (announcing prefixes and handing out addresses), but support for the renumbering case differs sharply:
| Daemon | Implements RFC 9096 | Dependencies | Status |
|---|---|---|---|
| odhcpd (OpenWrt) | Yes | Talks to the kernel over netlink; doesn’t need ubus, systemd or dbus | The reference implementation; runs without the rest of the OpenWrt stack |
| systemd-based daemon | No | systemd | Open issue, patch not merged |
| radvd (on OPNsense) | No | None specific | Request closed as not planned |
| dnsmasq (on OPNsense) | No | None specific | Request open, unresolved |
That last column is what matters for deciding: odhcpd is the only option on the list that doesn’t require the full OpenWrt ecosystem to work, because it can read interface state directly from the kernel over netlink. That’s what let Jha install it on NixOS without dragging in the rest of OpenWrt’s infrastructure (source).
flowchart TD
A["ISP renegotiates the WAN session"] --> B["Router receives a new IPv6 prefix via DHCPv6-PD"]
B --> C{"Does the router announce the old prefix with lifetime 0 (RFC 9096)?"}
C -->|"Yes, with odhcpd"| D["Clients drop the old IP and use the new one"]
C -->|"No, with current networkd, radvd or dnsmasq"| E["Clients keep the old IP and lose outgoing packets"]
What the Community Is Debating: Renumbering Less Matters More Than Announcing Well
In the Hacker News thread about the case, a comment from ebiederm qualifies the fix: the underlying problem isn’t just a missing announcement, but that IPv6 renumbering in general isn’t as robust as it should be, so the recommended practice is still to minimize it as much as possible. Renumbering without prior notice is documented as a problem even when the protocol works as intended (ebiederm’s comment on HN).
Another comment, from wernam, questions a configuration detail the author himself describes: it points out that NAT64 is a technique designed for IPv4, not for IPv6 routing, and that ULA addresses rank lower than IPv4 in the address selection order, so IPv4 will always be preferred over IPv6 in that kind of setup (wernam’s comment on HN). Take it as a warning: fixing the RFC 9096 announcement doesn’t automatically fix how your system prioritizes one address family over another, those are two separate layers of the same networking problem.
A third comment, from lmz, points to another inconsistency between implementations, though on a different problem: the fact that Windows and some Linux systems treat a search-domain value as a single domain instead of a list doesn’t come from a recent change in the RFCs, but from a decision not to preserve each stack’s previous behavior (lmz’s comment on HN). It shows that every operating system interprets the same DHCPv6 options with its own quirks, beyond this specific withdrawn-prefix case.
Limitations and When This Doesn’t Affect You
If your connection is a static IPv6 block (typical of a business plan or a dedicated tunnel), none of this applies: the prefix never changes and there’s nothing to withdraw. If your router already runs unmodified OpenWrt, it doesn’t apply either, because odhcpd is already the default RA daemon.
Where it does matter is in the exact profile of this case: a self-run router, outside OpenWrt, behind a residential ISP that renegotiates the WAN session fairly often. Packaging odhcpd outside OpenWrt isn’t trivial: as of the original post, the nixpkgs package was still under review, so this isn’t a one-command fix for every distribution (source).
📌 Note: the classic symptom (DNS resolves, the connection doesn’t arrive) isn’t unique to this case: any mechanism that leaves a stale source address in circulation, without announcing its withdrawal, can produce the same effect.
References
- IPv6 delegation, an unusually subtle failure mode (Savyasachee Jha): the real case, the step by step diagnosis, and the state of RFC 9096 support in networkd, radvd, dnsmasq and odhcpd.
- RFC 9096: the requirement to announce the withdrawn prefix with zero lifetime.
- RFC 7084: the baseline requirements for IPv6 edge routers that RFC 9096 tightens.
- Hacker News thread about the case: technical discussion around the diagnosis and the fix.
Featured image: Foto de Jonathan en Unsplash
How each article is made (sources, testing and editorial criteria): How we work.
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments