Controlling IP Fragmentation on Common Platforms

Introduction defines the Don't Fragment (DF) bit in the IPv4 header. When set, this bit prevents routers from fragmenting IP packets. If a router needs to fragment a packet with the DF bit set, it will instead drop the packet and send an ICMP "Fragmentation Needed" message back to the sender. The DF bit has historically been most relevant to TCP (), where the kernel handles Path MTU Discovery (PMTUD) internally. Applications using TCP sockets do not need to interact with the DF bit directly. In IPv6 (), fragmentation by intermediate nodes is not permitted. All IPv6 packets effectively have the DF bit set, however, the sender's kernel might still break up UDP datagrams that are too large to fit the MTU of the interface before sending a packet into the network. defines Datagram Packetization Layer Path MTU Discovery (DPLPMTUD), a mechanism that allows protocols running over UDP to determine the maximum packet size they can send. Setting the DF bit is crucial for DPLPMTUD, as it ensures that packets larger than the Path MTU are dropped, allowing the endpoint to detect the MTU limitation. QUIC is one such protocol that runs over UDP and makes use of DPLPMTUD. As QUIC implementations typically run in user space, they need to configure the DF bit on their UDP sockets to perform DPLPMTUD correctly. This document provides guidance for implementers on how to set the DF bit on UDP sockets across different platforms.

Setting the DF bit While routers don't fragment IPv6 packets in transit, the sender's kernel will still fragment UDP datagrams that are too large to fit the MTU of the interface before sending a packet into the network. Therefore, operating systems offer socket options to control the fragmentation behavior for IPv6 packets. For user-space implementations of DPLPMTUD, applications need to set the DF bit on IPv4 sockets and prevent fragmentation on IPv6 sockets.

Linux PMTUD behavior is controlled via the IP_MTU_DISCOVER and IPV6_MTU_DISCOVER socket options at the IPPROTO_IP and IPPROTO_IPV6 levels. There are two closely related values: For IPv4, both IP_PMTUDISC_DO and IP_PMTUDISC_PROBE enable setting the DF bit. For IPv6, both IPV6_PMTUDISC_DO and IPV6_PMTUDISC_PROBE prevent fragmentation by the sender. The difference between the former and the latter is that the former instructs the kernel to process ICMP "Fragmentation Needed" messages, while the latter does not. When the kernel processes an ICMP "Fragmentation Needed" message, it will prevent the transmission of larger datagrams. Consequently, IP_PMTUDISC_DO makes applications vulnerable to the UDP equivalent of the Blind Performance-Degrading ICMP attack described in . This vulnerability does not exist when using IP_PMTUDISC_PROBE and IPV6_PMTUDISC_PROBE. For dual-stack sockets, both IPv4 and IPv6 socket options can be set independently.

Apple For IPv4, Apple platforms use the socket option of level IPPROTO_IP with name IP_DONTFRAG with value 1. For IPv6, IPV6_DONTFRAG with value 1 is used for the IPPROTO_IPV6 level. However, dual-stack sockets are handled differently: To open a dual-stack socket, an IPv6 socket needs to be opened and the IPV6_V6ONLY option needs to be set to 0. This enables the socket to send both IPv4 and IPv6 packets. IPv4 packets must be sent using an IPv4-mapped IPv6 address. When using a dual-stack socket, it is only necessary (and possible) to set the IPV6_DONTFRAG socket option. This results in the DF bit being set when sending IPv4 packets, and prevents fragmentation of IPv6 packets. It is not possible to control the fragmentation behavior of IPv4 and IPv6 separately.

Windows For IPv4, Windows uses the socket option of level IPPROTO_IP with name IP_DONTFRAGMENT with value 1. For IPv6, IPV6_DONTFRAG with value 1 is used for the IPPROTO_IPV6 level. Similar to the Apple platforms, dual-stack sockets are IPv6 sockets with the IPV6_V6ONLY option set to 0. IPv4 packets must be sent using an IPv4-mapped IPv6 address. However, contrary to Apple platforms, the DF bit on IPv4 packets is controlled by the IP_DONTFRAGMENT socket option.