| Internet-Draft | avoid-fragmentation | December 2021 |
| Fujiwara & Vixie | Expires 27 June 2022 | [Page] |
EDNS0 enables a DNS server to send large responses using UDP and is widely deployed. Path MTU discovery remains widely undeployed due to security issues, and IP fragmentation has exposed weaknesses in application protocols. Currently, DNS is known to be the largest user of IP fragmentation. It is possible to avoid IP fragmentation in DNS by limiting response size where possible, and signaling the need to upgrade from UDP to TCP transport where necessary. This document proposes to avoid IP fragmentation in DNS.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 27 June 2022.¶
Copyright (c) 2021 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
DNS has EDNS0 [RFC6891] mechanism. It enables a DNS server to send large responses using UDP. EDNS0 is now widely deployed, and DNS (over UDP) is said to be the biggest user of IP fragmentation.¶
Fragmented DNS UDP responses have systemic weaknesses, which expose the requestor to DNS cache poisoning from off-path attackers. (See Appendix A for references and details.)¶
[RFC8900] summarized that IP fragmentation introduces fragility to Internet communication. The transport of DNS messages over UDP should take account of the observations stated in that document.¶
TCP avoids fragmentation using its Maximum Segment Size (MSS) parameter, but each transmitted segment is header-size aware such that the size of the IP and TCP headers is known, as well as the far end's MSS parameter and the interface or path MTU, so that the segment size can be chosen so as to keep the each IP datagram below a target size. This takes advantage of the elasticity of TCP's packetizing process as to how much queued data will fit into the next segment. In contrast, DNS over UDP has little datagram size elasticity and lacks insight into IP header and option size, and so must make more conservative estimates about available UDP payload space.¶
This document proposes to set IP_DONTFRAG / IPV6_DONTFRAG in DNS/UDP messages in order to avoid IP fragmentation, and describes how to avoid packet losses due to IP_DONTFRAG / IPV6_DONTFRAG.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
"Requestor" refers to the side that sends a request. "Responder" refers to an authoritative, recursive resolver or other DNS component that responds to questions. (Quoted from EDNS0 [RFC6891])¶
"Path MTU" is the minimum link MTU of all the links in a path between a source node and a destination node. (Quoted from [RFC8201])¶
"Path MTU discovery" is defined by [RFC1191], [RFC8201] and [RFC8899].¶
IP_DONTFRAG option is not defined by any RFCs. It is similar to IPV6_DONTFRAG option defined in [RFC3542]. IP_DONTFRAG option is used on BSD systems to set the Don't Fragment bit [RFC0791] when sending IPv4 packets. On Linux systems this is done via IP_MTU_DISCOVER and IP_PMTUDISC_DO.¶
Many of the specialized terms used in this document are defined in DNS Terminology [RFC8499].¶
The methods to avoid IP fragmentation in DNS are described below:¶
UDP responders SHOULD compose UDP responses that result in IP packets that do not exceed the path MTU to the requestor. If the path MTU discovery failed or is impossible, UDP responders SHOULD compose UDP responses that result in IP packets that do not exceed the default maximum DNS/UDP payload size described in Section 3.3.¶
The cause and effect of the TC bit is unchanged from EDNS0 [RFC6891].¶
Fragmentation avoidance is achieved with the IP(V6)_DONTFRAG option. The purpose of packet size limitation is to decrease packet loss due to the effects of the IP(V6)_DONTFRAG option.¶
Default maximum DNS/UDP payload size depends on the connectivity of each node, it cannot be determined unconditionally. However, there are good proposed values.¶
Operators MAY select a good number from Table 1. Details of proposed values are described in Appendix B.¶
| Source | IPv4 | IPv6 |
|---|---|---|
| Minimal: RFC 4035 MUST | 1220 | 1220 |
| Software developers / DNSFlagDay2020 propose | 1232 | 1232 (1280-40-8) |
| Authors' recommendation | 1400 | 1400 (1500 -40 -8 - some headers) |
| Maximum: Ethernet MTU 1500 [Huston2021] | 1472 (1500-20-8) | 1452 (1500-40-8) |
| Measured | MTU -20-8 | MTU -40-8 |
However, operators of DNS servers SHOULD measure their path MTU to the Internet at setting up DNS servers (and when network configuration changes).¶
How to measure path MTU is described in Appendix D.¶
Operators of authoritative servers (that offer global DNS zones) and full-service resolvers (that access authoritative servers of the global DNS) SHOULD measure their path MTU to well-known locations on the Internet, such as [a-m].root-servers.net or [a-m].gtld-servers.net.¶
Operators of full-service resolvers would be well advised to measure their path MTU to several authority name servers and to a random sample of their expected stub resolver client networks, to find the upper boundary on IP/UDP packet size in the average case. Or, operators of ISPs know their customers' connectivity and customers' MTU to ISPs' servers. This limit should not be exceeded by most messages received or transmitted by a full resolver, or else fallback to TCP will occur too often.¶
DNS clients (stub resolvers) need to specify an appropriate requestor's payload size when supporting EDNS0. In case of CPEs, embedded devices, and user devices, network operators can not control them, developers may choose small values such as 1220 and 1232.¶
Other DNS servers are out-of-scope of this document. (For example, Forwarding only resolvers, or private DNS).¶
The proposed method supports incremental deployment.¶
When a full-service resolver implements the proposed method, its stub resolvers (clients) and the authority server network will no longer observe IP fragmentation or reassembly from that server, and will fall back to TCP when necessary.¶
When an authoritative server implements the proposed method, its full service resolvers (clients) will no longer observe IP fragmentation or reassembly from that server, and will fall back to TCP when necessary.¶
Large DNS responses are the result of zone configuration. Zone operators SHOULD seek configurations resulting in small responses. For example,¶
In prior research ([Fujiwara2018] and dns-operations mailing list discussions), there are some authoritative servers that ignore EDNS0 requestor's UDP payload size, and return large UDP responses.¶
It is also well known that there are some authoritative servers that do not support TCP transport.¶
Such non-compliant behavior cannot become implementation or configuration constraints for the rest of the DNS. If failure is the result, then that failure must be localized to the non-compliant servers.¶
This document has no IANA actions.¶
The author would like to specifically thank Paul Wouters, Mukund Sivaraman, Tony Finch, Hugo Salgado, Peter van Dijk, Brian Dickson, Puneet Sood and Jim Reid for extensive review and comments.¶
"Fragmentation Considered Poisonous" [Herzberg2013] proposed effective off-path DNS cache poisoning attack vectors using IP fragmentation. "IP fragmentation attack on DNS" [Hlavacek2013] and "Domain Validation++ For MitM-Resilient PKI" [Brandt2018] proposed that off-path attackers can intervene in path MTU discovery [RFC1191] to perform intentionally fragmented responses from authoritative servers. [RFC7739] stated the security implications of predictable fragment identification values.¶
DNSSEC is a countermeasure against cache poisoning attacks that use IP fragmentation. However, DNS delegation responses are not signed with DNSSEC, and DNSSEC does not have a mechanism to get the correct response if an incorrect delegation is injected. This is a denial-of-service vulnerability that can yield failed name resolutions. If cache poisoning attacks can be avoided, DNSSEC validation failures will be avoided.¶
In Section 3.2 (Message Side Guidelines) of UDP Usage Guidelines [RFC8085] we are told that an application SHOULD NOT send UDP datagrams that result in IP packets that exceed the Maximum Transmission Unit (MTU) along the path to the destination.¶
A DNS message receiver cannot trust fragmented UDP datagrams primarily due to the small amount of entropy provided by UDP port numbers and DNS message identifiers, each of which being only 16 bits in size, and both likely being in the first fragment of a packet, if fragmentation occurs. By comparison, TCP protocol stack controls packet size and avoid IP fragmentation under ICMP NEEDFRAG attacks. In TCP, fragmentation should be avoided for performance reasons, whereas for UDP, fragmentation should be avoided for resiliency and authenticity reasons.¶
There are many discussions for default path MTU size and maximum DNS/UDP payload size.¶
Socket options: "IP_MTU (since Linux 2.2) Retrieve the current known path MTU of the current socket. Valid only when the socket has been connected. Returns an integer. Only valid as a getsockopt(2)." (Quoted from Debian GNU Linux manual: ip(7))¶
"IPV6_MTU getsockopt(): Retrieve the current known path MTU of the current socket. Only valid when the socket has been connected. Returns an integer." (Quoted from Debian GNU Linux manual: ipv6(7))¶
Section 3.4 of [RFC1122] specifies FIND_MAXSIZES() as one of "INTERNET/TRANSPORT LAYER INTERFACEs".¶
The Linux tool "tracepath" can be used to measure the path MTU to a destination.¶
Or, "ping/ping6" command with "-D" Don't Fragment bit set / Disable IPv6 fragmentation options.¶
Some implementations have 'minimal responses' configuration that causes a DNS server to make response packets smaller, containing only mandatory and required data.¶
Under the minimal-responses configuration, DNS servers compose response messages using only RRSets corresponding to queries. In case of delegation, DNS servers compose response packets with delegation NS RRSet in authority section and in-domain (in-zone and below-zone) glue in the additional data section. In case of non-existent domain name or non-existent type, the start of authority (SOA RR) will be placed in the Authority Section.¶
In addition, if the zone is DNSSEC signed and a query has the DNSSEC OK bit, signatures are added in answer section, or the corresponding DS RRSet and signatures are added in authority section. Details are defined in [RFC4035] and [RFC5155].¶