MTU Configuration and Path MTU Discovery

The gif interface uses the fixed length, 1280, to determine whether the outgoing IPv6 packets are split. This means the MTU value configured on the interface will be ignored when the outer protocol is IPv6. When the NOCLAMP interface flag is set, gif uses the same configured value as IPv4 communications. This behavior prevents potential issues when the path MTU is smaller than the interface MTU. This section describes the reason why the default behavior is different. The NOCLAMP interface flag can be set using the following command:

and clear the flag using the following:

where gif0 is the actual interface name.

A tunnel interface always has an implicit smaller MTU for the inner protocol than the outer protocol because of the additional header. Note that the interface MTU on a gif interface, the default value is 1280, is used as MTU for the outer protocol. This means that the MTU for the inner protocol varies depending on the outer protocol header length. If an outgoing packet bigger than the inner protocol MTU arrives at a gif interface for encapsulation, it will be split into fragments. Specifically, if IPv4 is used as the outer protocol, the inner is 20 octets smaller than the interface MTU. In the case of the default interface MTU, 1280, inner packets bigger than 1260 will be fragmented. In the case of IPv6, the inner is 40 octets smaller than the outer.

This fragmentation is not harmful though it can degrade the performance. Note that while an increased MTU on gif interface helps to mitigate this reduced performance issue, it can also cause packet losses on the intermediate narrowest path between the two communication endpoints in IPv6. IPv6 allows fragmentation only on the sender, not on the routers in the communication path. A big outgoing packet will be dropped on a router with a smaller MTU.

In normal IPv6 communication, an ICMPv6 Packet Too Big error will be sent back to the sender, who can adjust the packet length and re-send it. This process is performed in the upper protocols than L3, such as TCP, and makes the packet length shorter so that packets go through the path without fragmentation. This behavior is known as path MTU discovery.

When using a gif interface, the Packet Too Big message is generated for the outer protocol. Since the gif interface does not translate this error to the inner protocol, the inner protocol sees it just as a packet loss with no useful information to adjust the length of the next packets. In this situation, path MTU discovery does not work, and communications of the inner protocol become stalled.

In order to avoid this, a gif interface silently splits a packet of over 1240 octets into fragments to make the outer protocol packets equal or shorter than 1280 octets, even when the interface MTU is configured as larger than 1280. Note that this occurs only when the outer protocol is IPv6. 1280 is the smallest MTU in IPv6 and guarantees no packet loss occurs on intermediate routers.

As mentioned earlier, the performance is sub-optimal if the actual path MTU is larger than 1280. A typical confusing scenario is as follows. The gif interface can have Ethernet, whose MTU is usually 1500, as the inner protocol. It is called an EtherIP tunnel, and can be configured by adding the gif interface as a member of if_bridge(4) interface. The if_bridge(4) interface forcibly changes the MTU of the gif interface with those for the other member interfaces, which are likely 1500. In this case, a situation in which the MTU of the gif interface is 1500 but fragmentation in 1280 octets always occurs.

The default behavior is most conservative to prevent confusing packet loss. Depending on the network configuration, enabling the NOCLAMP interface flag might be helpful for better performance. It is crucial to ensure that the path MTU is equal to or larger than the interface MTU when enabling this flag.

ECN friendly behavior

The gif device can be configured to be ECN friendly, as described in draft-ietf-ipsec-ecn-02.txt. This is turned off by default, and can be turned on by the IFF_LINK1 interface flag.

Without IFF_LINK1, gif will show normal behavior, as described in RFC2893. This can be summarized as follows:

With IFF_LINK1, gif will copy ECN bits (0x02 and 0x01 on IPv4 TOS byte or IPv6 traffic class byte) on egress and ingress, as follows:

Note that the ECN friendly behavior violates RFC2893. This should be used in mutual agreement with the peer.

Security

A malicious party may try to circumvent security filters by using tunnelled packets. For better protection, gif performs both martian and ingress filtering against the outer source address on egress. Note that martian/ingress filters are in no way complete. You may want to secure your node by using packet filters. Ingress filtering can break tunnel operation in an asymmetrically routed network. It can be turned off by IFF_LINK2 bit.

Miscellaneous

By default, gif tunnels may not be nested. This behavior may be modified at runtime by setting the sysctl(8) variable net.link.gif.max_nesting to the desired level of nesting.