IPv6 address types

Every IPv6 address is either unicast, multicast or anycast. Broadcast does not exist in IPv6 — its functions were folded into multicast, which is a deliberate change because multicast can be scoped and directed rather than always interrupting every host on the segment.

The distinction between anycast and unicast is not visible in the address itself. An anycast address is a unicast address that happens to be configured on several nodes, with routers delivering to whichever is nearest. This is how IPv6 root DNS servers work.

Most of the space is still unassigned. These are the ranges that matter in practice, with what each one is actually for.

A global unicast address is divided into three parts: a routing prefix from the provider, a subnet ID you control, and a 64-bit interface identifier. That gives you 65,536 subnets of 18 quintillion addresses each, which is why IPv6 subnet planning is about the number of /64s rather than the number of hosts.

|       2001:db8:abcd      |   12    |        0000:0000:0000:0001        |
|<--- routing prefix /48 --->|<-subnet->|<---- interface identifier /64 ---->|
       (from the provider)     (yours)          (device identifier)

Network  /48  = 2001:db8:abcd::
Subnet   /64  = 2001:db8:abcd:12::
Host           = 2001:db8:abcd:12::1
Addresses in one /64 = 2^64 = 18,446,744,073,709,551,616

The same structure applies at other sizes. An ISP holding a /32 can give each customer a /48, which is 65,536 customer allocations, and each customer then has 65,536 /64s to use internally. Nothing is sized tightly, and that is intentional — the design removes address scarcity from the planning problem.

Every IPv6 interface has a link-local address whether or not it has a global one. It is generated from the interface's MAC address or a random value, and it is what makes IPv6 work before any global configuration exists.

• It is the next hop in routing protocols. OSPFv3 and other protocols form adjacencies over link-local addresses, which removes the need for a globally routable transit subnet on point-to-point links.
• It carries neighbour discovery. Router advertisements, neighbour solicitations and duplicate address detection all ride on link-local and multicast addresses.
• It needs a zone identifier when you use it. Since the same fe80:: prefix exists on every interface, you must specify which interface you mean. On Linux that is fe80::1%eth0; on Windows it is fe80::1%12 using the interface index.
• It is never routed. Routers drop link-local traffic at the interface boundary, which is what makes the scope meaningful.

• Forgetting the zone identifier on a link-local address. Pinging fe80::1 without %eth0 fails with an ambiguous address error, which is confusing until you remember that every interface has one.
• Using a prefix other than /64 for a subnet with SLAAC hosts. Stateless autoconfiguration requires a 64-bit interface identifier, so a /72 or /80 subnet will not autoconfigure correctly even though it is technically valid.
• Assuming fd00::/8 is the whole ULA range. The allocation is fc00::/7; the fd00::/8 half is the part defined for local assignment. Rules intended to cover all ULA traffic should match fc00::/7.
• Treating 2001:db8:: like a real address. It is documentation space and no router will forward it. Seeing it in a live system means a template was applied without substitution.
• Expecting broadcast to exist. There is no broadcast in IPv6. Code that tries to send to a subnet broadcast address has no equivalent and needs to be rewritten to use a multicast group, usually ff02::1 for all nodes on the link.
• Planning subnets to conserve addresses. Sizing a subnet to fit the hosts is an IPv4 habit. In IPv6 every subnet is a /64; the planning effort goes into the subnet ID instead.