Infrastructure addressing: loopbacks, point-to-point and link-local-only links
Number router loopbacks with /128s and point-to-point links with /127s, reserving a /64 per link. Avoid the addresses that RFC 6164 says must not be used as unicast on /127 links. Weigh link-local-only infrastructure links (RFC 7404) against globally addressed links. Apply the forwarding-prefix-length rule of RFC 7608, and route a /64 per VPS on a hypervisor.
Take the quiz · pass mark 70% · 10 questions
Explanation
The network itself needs addresses too: router loopbacks, links between routers, management interfaces, and on a hosting platform, the next hops that customer prefixes are routed to. This "infrastructure" space is small compared with customer space, but mistakes in it hurt the most. Your routing protocols, monitoring and filters all depend on it.
Where infrastructure lives in the plan
In the ictNetworks scenario (L2.04), infrastructure gets 3fff:a00::/48:
| Block | Use | Rule |
|---|---|---|
3fff:a00::/64 | Router loopbacks | Router n gets 3fff:a00::n/128 (n in hex) |
3fff:a00:0:100::/56 | Management networks | One /64 per PoP or DC |
3fff:a00:0:1000::/52 | Point-to-point links | Link L reserves 3fff:a00:0:1LLL::/64 (L in three hex digits) and uses a /127 inside it |
| the rest | Reserved |
Keeping all infrastructure in one /48 means a single infrastructure ACL entry (3fff:a00::/48) can protect every router address from the Internet (Level 5).
Loopbacks: /128
A loopback address is a single address on a virtual interface that never goes down. Routing protocols use it as a stable identity, management tools and BGP sessions use it as the address to talk to, and ICMP errors may come from it. Give each router one /128 from a dedicated /64. Do not put a /64 on each loopback: that wastes nothing in address terms, but it adds a needless route per router and confuses filters.
Point-to-point links: /127, with a /64 reserved
On a link between two routers there are no hosts and no SLAAC, so the /64 rule (L2.03) does not apply. RFC 6164 recommends /127 on inter-router point-to-point links, like /31 in IPv4. Its reasons:
- Neighbor Discovery exhaustion: if a link is a /64, an attacker can send packets to millions of non-existent addresses in that /64. The router tries to resolve each one and fills its ND cache. With a /127 there are only two addresses, so this attack does not work.
- Ping-pong: on some point-to-point link types, a packet to an unused address in a /64 can bounce between the two routers until its Hop Limit expires. A /127 has no unused addresses.
RFC 6164 also requires routers to support /127 and to disable the Subnet-Router anycast address on /127 links. Otherwise the all-zero address would be anycast rather than an ordinary unicast address.
Which /127 inside the /64? RFC 6164 says that addresses with all-zero interface identifiers (the Subnet-Router anycast address of RFC 4291) and the highest 128 interface identifier values of the /64 (reserved subnet anycast, RFC 2526) should not be used as unicast addresses. So avoid the very first /127 (::0 and ::1) and the very top of the /64. This course uses ::a for the A side and ::b for the B side. 0xa = 1010 and 0xb = 1011 differ only in the last bit, so they form a valid /127 (…::a/127), and the letters are easy to remember.
Why reserve a /64 per link? It costs little (a /52 holds 4,096 links), keeps every link on a nibble boundary for documentation and reverse DNS, and allows you to switch a link to /64 later if a device cannot do /127.
Drawing diagram…
Link-local-only infrastructure links (RFC 7404)
A third option is to give router links no global addresses at all. Interior routing protocols such as OSPFv3 and IS-IS already work with link-local next hops. Only loopbacks are routed. RFC 7404 lists the trade-offs:
| Advantages | Caveats |
|---|---|
| Smaller routing tables: one /128 per router | You cannot ping or traceroute to a specific link interface from afar |
| Simpler plan: only loopbacks need addresses | Traceroute shows a loopback (or other global) address for each hop, not the interface |
| Less configuration and DNS | Link-local addresses change with hardware unless you configure them statically, and they need a zone ID (RFC 4007), e.g. fe80::1%eth0 |
| Smaller attack surface: nothing on the links is reachable from outside | Management tools that use interface addresses must use the loopback instead |
| MPLS RSVP-TE explicit paths that name interface addresses do not work |
RFC 5837 defines an ICMP extension that reports the incoming interface, which would solve the traceroute caveat, but RFC 7404 notes that few implementations support it.
Typical choice: many ISPs use global /127s on links where they need per-interface monitoring and traceroute clarity, and link-local-only in large data-centre fabrics where links are numerous and identical (Level 7). Both are valid. Choose deliberately and document it.
Forwarding works on any prefix length (RFC 7608)
RFC 7608 (BCP 198) says that IPv6 forwarding must use longest-match on prefixes of any length from 0 to 128. Implementations must not assume /64 or any other fixed boundary. That is what makes /127 links, /128 loopbacks, and summaries like /36 or /40 work. If a device cannot route a /127 or a /65 correctly, it does not meet this BCP, and that is worth raising with the vendor.
Hosting edge: a routed /64 per VPS
On a Proxmox host, the simplest design puts all VPSs on a shared bridge with one shared /64, but that brings back the problems of a shared LAN between customers: ND exhaustion, address squatting and harder abuse handling. The ictNetworks scenario instead routes one /64 to each VPS (L2.03, L2.04):
- The Proxmox node has a customer bridge (
vmbr1) with no global prefix on it, and a well-known link-local gateway addressfe80::1. - Each VPS gets a static link-local address derived from its VPS number, and the node routes the VPS's /64 to that link-local next hop.
# On the Proxmox node (VPS 300 = 0x12c, cluster /48 3fff:a00:1003::/48)
ip -6 addr add fe80::1/64 dev vmbr1
ip -6 route add 3fff:a00:1003:12c::/64 via fe80::12c dev vmbr1
# Inside VPS 300
ip -6 addr add fe80::12c/64 dev eth0
ip -6 addr add 3fff:a00:1003:12c::1/64 dev eth0
ip -6 route add default via fe80::1 dev eth0
The node's routing table now holds one route per VPS, all inside the cluster /48. The cluster announces only the /48 upstream. Level 5 adds source-address filtering so that a VPS cannot send from another customer's prefix. On a MikroTik edge router, the same design is a per-customer route such as /ipv6 route add dst-address=3fff:a00:1003:12c::/64 gateway=fe80::12c%bridge-vps, where bridge-vps is the customer-facing bridge.
Worked examples
Example 1: Number a link
Link number 37 (0x025) reserves 3fff:a00:0:1025::/64. The routers use 3fff:a00:0:1025::a/127 (A side) and 3fff:a00:0:1025::b/127 (B side).
Example 2: Find the partner address
A router has 3fff:a00:0:1025::b/127. The /127 is 3fff:a00:0:1025::a/127, and its partner is 3fff:a00:0:1025::a: flip the last bit.
Example 3: Capacity
The point-to-point block 3fff:a00:0:1000::/52 holds 2^12 = 4,096 links with one /64 each. Loopbacks come from one /64, so router count is never the limit.
Example 4: A bad /127
An engineer configures 3fff:a00:0:1025::/127 and ::1/127. The address 3fff:a00:0:1025:: has an all-zero interface identifier, which is the Subnet-Router anycast address. RFC 6164 says it should not be used as unicast. Use ::a and ::b instead.
Example 5: Loopback by router number
Router 26 (0x1a) has loopback 3fff:a00::1a/128.
Lab
Goal: Number a four-router ring with /127 links and /128 loopbacks, verify it with traceroute, then compare it with link-local-only links.
You need: Linux with Docker and containerlab (bash -c "$(curl -sL https://get.containerlab.dev)"), and an FRRouting image (for example quay.io/frrouting/frr:10.1.0). OSPFv3 is used here only to distribute the loopbacks. You will learn it properly in Level 4.
The lab uses ISP-A's infrastructure block from the course lab plan: 2001:db8:1000::/48, loopbacks from 2001:db8:1000::/64, and one /64 per link from 2001:db8:1000:1000::/52.
Step 1: Topology
ring.clab.yml:
name: l205
topology:
defaults:
kind: linux
image: quay.io/frrouting/frr:10.1.0
nodes:
r1: { binds: [r1.conf:/etc/frr/frr.conf, daemons:/etc/frr/daemons] }
r2: { binds: [r2.conf:/etc/frr/frr.conf, daemons:/etc/frr/daemons] }
r3: { binds: [r3.conf:/etc/frr/frr.conf, daemons:/etc/frr/daemons] }
r4: { binds: [r4.conf:/etc/frr/frr.conf, daemons:/etc/frr/daemons] }
links:
- endpoints: ["r1:eth1", "r2:eth1"] # link 1
- endpoints: ["r2:eth2", "r3:eth1"] # link 2
- endpoints: ["r3:eth2", "r4:eth1"] # link 3
- endpoints: ["r4:eth2", "r1:eth2"] # link 4
Copy FRR's default daemons file and set ospf6d=yes.
Step 2: Address plan
Fill in this table before you configure anything:
| Link | Reserved /64 | A side | B side |
|---|---|---|---|
| 1 (r1–r2) | 2001:db8:1000:1001::/64 | r1 ::a/127 | r2 ::b/127 |
| 2 (r2–r3) | ? | ? | ? |
| 3 (r3–r4) | ? | ? | ? |
| 4 (r4–r1) | ? | ? | ? |
Loopbacks: rn gets 2001:db8:1000::n/128.
Step 3: Configure
r1.conf (adapt it for the others):
frr defaults traditional
hostname r1
interface lo
ipv6 address 2001:db8:1000::1/128
ipv6 ospf6 area 0
interface eth1
ipv6 address 2001:db8:1000:1001::a/127
ipv6 ospf6 area 0
ipv6 ospf6 network point-to-point
interface eth2
ipv6 address 2001:db8:1000:1004::b/127
ipv6 ospf6 area 0
ipv6 ospf6 network point-to-point
router ospf6
ospf6 router-id 0.0.0.1
Deploy with sudo containerlab deploy -t ring.clab.yml.
Step 4: Verify
docker exec clab-l205-r1 vtysh -c "show ipv6 route ospf6"
docker exec clab-l205-r1 ping -c2 2001:db8:1000::3
docker exec clab-l205-r1 traceroute -6 -n 2001:db8:1000::3
Record which addresses appear as traceroute hops.
Step 5: Link-local only
Remove the global addresses from all eth interfaces (keep the loopbacks), wait for OSPFv3 to reconverge, and repeat step 4. Explain the difference in the traceroute output using RFC 7404, and say which design you would use for ISP-A and why.
What to submit
Your completed address table, one router config, the two traceroute outputs, and your explanation.
Summary
- Infrastructure goes in one dedicated block (here
3fff:a00::/48) so that one ACL protects it. - Loopbacks: one /128 each, from a single /64.
- Point-to-point router links: /127 (RFC 6164), with a /64 reserved per link. Disable Subnet-Router anycast, and do not use the all-zero IID or the top of the /64. This course uses
::aand::b. - Link-local-only links (RFC 7404) shrink the routing table and attack surface, but cost per-interface ping/traceroute and need static link-locals for explicit neighbours.
- Forwarding must work on any prefix length (RFC 7608).
- A routed /64 per VPS uses a link-local next hop on the hypervisor and keeps customers apart.
- Next lesson (L2.06): Unique Local Addresses, and hosts with more than one prefix.
Quiz imported · 10 questions · pass mark 70%. Lab questions are not auto-marked.