RFC 9210




Internet Engineering Task Force (IETF)                       J. Kristoff
Request for Comments: 9210                                 Dataplane.org
BCP: 235                                                      D. Wessels
Updates: 1123, 1536                                             Verisign
Category: Best Current Practice                               March 2022
ISSN: 2070-1721


           DNS Transport over TCP - Operational Requirements

Abstract



   This document updates RFCs 1123 and 1536.  This document requires the
   operational practice of permitting DNS messages to be carried over
   TCP on the Internet as a Best Current Practice.  This operational
   requirement is aligned with the implementation requirements in RFC
   7766.  The use of TCP includes both DNS over unencrypted TCP as well
   as over an encrypted TLS session.  The document also considers the
   consequences of this form of DNS communication and the potential
   operational issues that can arise when this Best Current Practice is
   not upheld.

Status of This Memo



   This memo documents an Internet Best Current Practice.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Further information on
   BCPs is available in Section 2 of RFC 7841.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   https://www.rfc-editor.org/info/rfc9210.

Copyright Notice



   Copyright (c) 2022 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.

Table of Contents



   1.  Introduction
     1.1.  Requirements Language
   2.  History of DNS over TCP
     2.1.  Uneven Transport Usage and Preference
     2.2.  Waiting for Large Messages and Reliability
     2.3.  EDNS(0)
     2.4.  Fragmentation and Truncation
     2.5.  "Only Zone Transfers Use TCP"
     2.6.  Reuse, Pipelining, and Out-of-Order Processing
   3.  DNS-over-TCP Requirements
   4.  Network and System Considerations
     4.1.  Connection Establishment and Admission
     4.2.  Connection Management
     4.3.  Connection Termination
     4.4.  DNS over TLS
     4.5.  Defaults and Recommended Limits
   5.  DNS-over-TCP Filtering Risks
     5.1.  Truncation, Retries, and Timeouts
     5.2.  DNS Root Zone KSK Rollover
   6.  Logging and Monitoring
   7.  IANA Considerations
   8.  Security Considerations
   9.  Privacy Considerations
   10. References
     10.1.  Normative References
     10.2.  Informative References
   Appendix A.  RFCs Related to DNS Transport over TCP
     A.1.  RFC 1035 - DOMAIN NAMES - IMPLEMENTATION AND SPECIFICATION
     A.2.  RFC 1536 - Common DNS Implementation Errors and Suggested
            Fixes
     A.3.  RFC 1995 - Incremental Zone Transfer in DNS
     A.4.  RFC 1996 - A Mechanism for Prompt Notification of Zone
            Changes (DNS NOTIFY)
     A.5.  RFC 2181 - Clarifications to the DNS Specification
     A.6.  RFC 2694 - DNS extensions to Network Address Translators
            (DNS_ALG)
     A.7.  RFC 3225 - Indicating Resolver Support of DNSSEC
     A.8.  RFC 3226 - DNSSEC and IPv6 A6 aware server/resolver message
            size requirements
     A.9.  RFC 4472 - Operational Considerations and Issues with IPv6
            DNS
     A.10. RFC 5452 - Measures for Making DNS More Resilient against
            Forged Answers
     A.11. RFC 5507 - Design Choices When Expanding the DNS
     A.12. RFC 5625 - DNS Proxy Implementation Guidelines
     A.13. RFC 5936 - DNS Zone Transfer Protocol (AXFR)
     A.14. RFC 7534 - AS112 Nameserver Operations
     A.15. RFC 6762 - Multicast DNS
     A.16. RFC 6891 - Extension Mechanisms for DNS (EDNS(0))
     A.17. IAB RFC 6950 - Architectural Considerations on Application
            Features in the DNS
     A.18. RFC 7477 - Child-to-Parent Synchronization in DNS
     A.19. RFC 7720 - DNS Root Name Service Protocol and Deployment
            Requirements
     A.20. RFC 7766 - DNS Transport over TCP - Implementation
            Requirements
     A.21. RFC 7828 - The edns-tcp-keepalive EDNS(0) Option
     A.22. RFC 7858 - Specification for DNS over Transport Layer
            Security (TLS)
     A.23. RFC 7873 - Domain Name System (DNS) Cookies
     A.24. RFC 7901 - CHAIN Query Requests in DNS
     A.25. RFC 8027 - DNSSEC Roadblock Avoidance
     A.26. RFC 8094 - DNS over Datagram Transport Layer Security
            (DTLS)
     A.27. RFC 8162 - Using Secure DNS to Associate Certificates with
            Domain Names for S/MIME
     A.28. RFC 8324 - DNS Privacy, Authorization, Special Uses,
            Encoding, Characters, Matching, and Root Structure: Time for
            Another Look?
     A.29. RFC 8467 - Padding Policies for Extension Mechanisms for
            DNS (EDNS(0))
     A.30. RFC 8482 - Providing Minimal-Sized Responses to DNS Queries
            That Have QTYPE=ANY
     A.31. RFC 8483 - Yeti DNS Testbed
     A.32. RFC 8484 - DNS Queries over HTTPS (DoH)
     A.33. RFC 8490 - DNS Stateful Operations
     A.34. RFC 8501 - Reverse DNS in IPv6 for Internet Service
            Providers
     A.35. RFC 8806 - Running a Root Server Local to a Resolver
     A.36. RFC 8906 - A Common Operational Problem in DNS Servers:
            Failure to Communicate
     A.37. RFC 8932 - Recommendations for DNS Privacy Service
            Operators
     A.38. RFC 8945 - Secret Key Transaction Authentication for DNS
            (TSIG)
   Acknowledgments

   Authors' Addresses



1.  Introduction



   DNS messages are delivered using UDP or TCP communications.  While
   most DNS transactions are carried over UDP, some operators have been
   led to believe that any DNS-over-TCP traffic is unwanted or
   unnecessary for general DNS operation.  When DNS over TCP has been
   restricted, a variety of communication failures and debugging
   challenges often arise.  As DNS and new naming system features have
   evolved, TCP as a transport has become increasingly important for the
   correct and safe operation of an Internet DNS.  Reflecting modern
   usage, the DNS standards declare that support for TCP is a required
   part of the DNS implementation specifications [RFC7766].  This
   document is the equivalent of formal requirements for the operational
   community, encouraging system administrators, network engineers, and
   security staff to ensure DNS-over-TCP communications support is on
   par with DNS-over-UDP communications.  It updates [RFC1123],
   Section 6.1.3.2 to clarify that all DNS resolvers and recursive
   servers MUST support and service both TCP and UDP queries and also
   updates [RFC1536] to remove the misconception that TCP is only useful
   for zone transfers.

1.1.  Requirements Language



   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
   BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
   capitals, as shown here.

2.  History of DNS over TCP



   The curious state of disagreement between operational best practices
   and guidance for DNS transport protocols derives from conflicting
   messages operators have received from other operators, implementors,
   and even the IETF.  Sometimes these mixed signals have been explicit;
   on other occasions, conflicting messages have been implicit.  This
   section presents an interpretation of the storied and conflicting
   history that led to this document.  This section is included for
   informational purposes only.

2.1.  Uneven Transport Usage and Preference



   In the original suite of DNS specifications, [RFC1034] and [RFC1035]
   clearly specify that DNS messages could be carried in either UDP or
   TCP, but they also state that there is a preference for UDP as the
   best transport for queries in the general case.  As stated in
   [RFC1035]:

   |  While virtual circuits can be used for any DNS activity, datagrams
   |  are preferred for queries due to their lower overhead and better
   |  performance.

   Another early, important, and influential document, [RFC1123], marks
   the preference for a transport protocol more explicitly:

   |  DNS resolvers and recursive servers MUST support UDP, and SHOULD
   |  support TCP, for sending (non-zone-transfer) queries.

   and it further stipulates that:

   |  A name server MAY limit the resources it devotes to TCP queries,
   |  but it SHOULD NOT refuse to service a TCP query just because it
   |  would have succeeded with UDP.

   Culminating in [RFC1536], DNS over TCP came to be associated
   primarily with the zone transfer mechanism, while most DNS queries
   and responses were seen as the dominion of UDP.

2.2.  Waiting for Large Messages and Reliability



   In the original specifications, the maximum DNS-over-UDP message size
   was enshrined at 512 bytes.  However, even while [RFC1123] prefers
   UDP for non-zone transfer queries, it foresaw that DNS over TCP would
   become more popular in the future to overcome this limitation:

   |  [...] it is also clear that some new DNS record types defined in
   |  the future will contain information exceeding the 512 byte limit
   |  that applies to UDP, and hence will require TCP.

   At least two new, widely anticipated developments were set to elevate
   the need for DNS-over-TCP transactions.  The first was dynamic
   updates defined in [RFC2136], and the second was the set of
   extensions collectively known as "DNSSEC", whose operational
   considerations were originally given in [RFC2541] (note that
   [RFC2541] has been obsoleted by [RFC6781]).  The former suggests that

   |  ...requestors who require an accurate response code must use TCP.

   while the latter warns that

   |  ... larger keys increase the size of the KEY and SIG RRs.  This
   |  increases the chance of DNS UDP packet overflow and the possible
   |  necessity for using higher overhead TCP in responses.

   Yet, defying some expectations, DNS over TCP remained little used in
   real traffic across the Internet in the late 1990s.  Dynamic updates
   saw little deployment between autonomous networks.  Around the time
   DNSSEC was first defined, another new feature helped solidify UDP
   transport dominance for message transactions.

2.3.  EDNS(0)



   In 1999, the IETF published the Extension Mechanisms for DNS
   (EDNS(0)) in [RFC2671] (which was obsoleted by [RFC6891] in 2013).
   That document standardized a way for communicating DNS nodes to
   perform rudimentary capabilities negotiation.  One such capability
   written into the base specification and present in every EDNS(0)-
   compatible message is the value of the maximum UDP payload size the
   sender can support.  This unsigned 16-bit field specifies, in bytes,
   the maximum (possibly fragmented) DNS message size a node is capable
   of receiving over UDP.  In practice, typical values are a subset of
   the 512- to 4096-byte range.  EDNS(0) became widely deployed over the
   next several years, and numerous surveys (see [CASTRO2010] and
   [NETALYZR]) have shown that many systems support larger UDP MTUs with
   EDNS(0).

   The natural effect of EDNS(0) deployment meant DNS messages larger
   than 512 bytes would be less reliant on TCP than they might otherwise
   have been.  While a non-negligible population of DNS systems lacked
   EDNS(0) or fell back to TCP when necessary, DNS clients still
   strongly prefer UDP to TCP.  For example, as of 2014, DNS-over-TCP
   transactions remained a very small fraction of overall DNS traffic
   received by root name servers [VERISIGN].

2.4.  Fragmentation and Truncation



   Although EDNS(0) provides a way for endpoints to signal support for
   DNS messages exceeding 512 bytes, the realities of a diverse and
   inconsistently deployed Internet may result in some large messages
   being unable to reach their destination.  Any IP datagram whose size
   exceeds the MTU of a link it transits will be fragmented and then
   reassembled by the receiving host.  Unfortunately, it is not uncommon
   for middleboxes and firewalls to block IP fragments.  If one or more
   fragments do not arrive, the application does not receive the
   message, and the request times out.

   For IPv4-connected hosts, the MTU is often an Ethernet payload size
   of 1500 bytes.  This means that the largest unfragmented UDP DNS
   message that can be sent over IPv4 is likely 1472 bytes, although
   tunnel encapsulation may reduce that maximum message size in some
   cases.

   For IPv6, the situation is a little more complicated.  First, IPv6
   headers are 40 bytes (versus 20 without options in IPv4).  Second,
   approximately one-third of DNS recursive resolvers use the minimum
   MTU of 1280 bytes [APNIC].  Third, fragmentation in IPv6 can only be
   done by the host originating the datagram.  The need to fragment is
   conveyed in an ICMPv6 "Packet Too Big" message.  The originating host
   indicates a fragmented datagram with IPv6 extension headers.
   Unfortunately, it is quite common for both ICMPv6 and IPv6 extension
   headers to be blocked by middleboxes.  According to [HUSTON], some
   35% of IPv6-capable recursive resolvers were unable to receive a
   fragmented IPv6 packet.  When the originating host receives a signal
   that fragmentation is required, it is expected to populate its path
   MTU cache for that destination.  The application will then retry the
   query after a timeout since the host does not generally retain copies
   of messages sent over UDP for potential retransmission.

   The practical consequence of all this is that DNS requestors must be
   prepared to retry queries with different EDNS(0) maximum message size
   values.  Administrators of [BIND] are likely to be familiar with
   seeing the following message in their system logs: "success resolving
   ... after reducing the advertised EDNS(0) UDP packet size to 512
   octets".

   Often, reducing the EDNS(0) UDP packet size leads to a successful
   response.  That is, the necessary data fits within the smaller
   message size.  However, when the data does not fit, the server sets
   the truncated flag in its response, indicating the client should
   retry over TCP to receive the whole response.  This is undesirable
   from the client's point of view because it adds more latency and is
   potentially undesirable from the server's point of view due to the
   increased resource requirements of TCP.

   Note that a receiver is unable to differentiate between packets lost
   due to congestion and packets (fragments) intentionally dropped by
   firewalls or middleboxes.  Over network paths with non-trivial
   amounts of packet loss, larger, fragmented DNS responses are more
   likely to never arrive and time out compared to smaller, unfragmented
   responses.  Clients might be misled into retrying queries with
   different EDNS(0) UDP packet size values for the wrong reason.

   The issues around fragmentation, truncation, and TCP are driving
   certain implementation and policy decisions in the DNS.  Notably,
   Cloudflare implemented a technique that minimizes the number of
   DNSSEC denial-of-existence records (for its online signing platform)
   [CLOUDFLARE] and uses an Elliptic Curve Digital Signature Algorithm
   (ECDSA) such that its signed responses fit easily in 512 bytes.  The
   Key Signing Key (KSK) Rollover Design Team [DESIGNTEAM] spent a lot
   of time thinking and worrying about response sizes.  There is growing
   sentiment in the DNSSEC community that RSA key sizes beyond 2048 bits
   are impractical and that critical infrastructure zones should
   transition to elliptic curve algorithms to keep response sizes
   manageable [ECDSA].

   More recently, renewed security concerns about fragmented DNS
   messages (see [AVOID_FRAGS] and [FRAG_POISON]) are leading
   implementors to consider smaller responses and lower default EDNS(0)
   UDP payload size values for both queriers and responders
   [FLAGDAY2020].

2.5.  "Only Zone Transfers Use TCP"



   Today, the majority of the DNS community expects, or at least has a
   desire, to see DNS-over-TCP transactions occur without interference
   [FLAGDAY2020].  However, there has also been a long-held belief by
   some operators, particularly for security-related reasons, that DNS-
   over-TCP services should be purposely limited or not provided at all
   [CHES94] [DJBDNS].  A popular meme is that DNS over TCP is only ever
   used for zone transfers and is generally unnecessary otherwise, with
   filtering all DNS-over-TCP traffic even described as a best practice.

   The position on restricting DNS over TCP had some justification given
   that historical implementations of DNS name servers provided very
   little in the way of TCP connection management (for example, see
   Section 6.1.2 of [RFC7766] for more details).  However, modern
   standards and implementations are nearing parity with the more
   sophisticated TCP management techniques employed by, for example,
   HTTP(S) servers and load balancers.

2.6.  Reuse, Pipelining, and Out-of-Order Processing



   The idea that a TCP connection can support multiple transactions goes
   back as far as [RFC0883], which states: "Multiple messages may be
   sent over a virtual circuit."  Although [RFC1035], which updates the
   former, omits this particular detail, it has been generally accepted
   that a TCP connection can be used for more than one query and
   response.

   [RFC5966] clarifies that servers are not required to preserve the
   order of queries and responses over any transport.  [RFC7766], which
   updates the former, further encourages query pipelining over TCP to
   achieve performance on par with UDP.  A server that sends out-of-
   order responses to pipelined queries avoids head-of-line blocking
   when the response for a later query is ready before the response to
   an earlier query.

   However, TCP can potentially suffer from a different head-of-line
   blocking problem due to packet loss.  Since TCP itself enforces
   ordering, a single lost segment delays delivery of data in any
   following segments until the lost segment is retransmitted and
   successfully received.

3.  DNS-over-TCP Requirements



   An average increase in DNS message size (e.g., due to DNSSEC), the
   continued development of new DNS features (Appendix A), and a denial-
   of-service mitigation technique (Section 8) all show that DNS-over-
   TCP transactions are as important to the correct and safe operation
   of the Internet DNS as ever, if not more so.  Furthermore, there has
   been research that argues connection-oriented DNS transactions may
   provide security and privacy advantages over UDP transport [TDNS].
   In fact, the standard for DNS over TLS [RFC7858] is just this sort of
   specification.  Therefore, this document makes explicit that it is
   undesirable for network operators to artificially inhibit DNS-over-
   TCP transport.

   Section 6.1.3.2 of [RFC1123] is updated as follows:

   OLD:

   |  DNS resolvers and recursive servers MUST support UDP, and SHOULD
   |  support TCP, for sending (non-zone-transfer) queries.

   NEW:

   |  All DNS resolvers and servers MUST support and service both UDP
   |  and TCP queries.

   Note that:

   *  DNS servers (including forwarders) MUST support and service TCP
      for receiving queries so that clients can reliably receive
      responses that are larger than what either side considers too
      large for UDP.

   *  DNS clients MUST support TCP for sending queries so that they can
      retry truncated UDP responses as necessary.

   Furthermore, the requirement in Section 6.1.3.2 of [RFC1123] around
   limiting the resources a server devotes to queries is hereby updated:

   OLD:

   |  A name server MAY limit the resources it devotes to TCP queries,
   |  but it SHOULD NOT refuse to service a TCP query just because it
   |  would have succeeded with UDP.

   NEW:

   |  A name server MAY limit the resources it devotes to queries, but
   |  it MUST NOT refuse to service a query just because it would have
   |  succeeded with another transport protocol.

   Lastly, Section 1 of [RFC1536] is updated to eliminate the
   misconception that TCP is only useful for zone transfers:

   OLD:

   |  DNS implements the classic request-response scheme of client-
   |  server interaction.  UDP is, therefore, the chosen protocol for
   |  communication though TCP is used for zone transfers.

   NEW:

   |  DNS implements the classic request-response scheme of client-
   |  server interaction.

   The filtering of DNS over TCP is harmful in the general case.  DNS
   resolver and server operators MUST support and provide DNS service
   over both UDP and TCP transports.  Likewise, network operators MUST
   allow DNS service over both UDP and TCP transports.  It is
   acknowledged that DNS-over-TCP service can pose operational
   challenges that are not present when running DNS over UDP alone, and
   vice versa.  However, the potential damage incurred by prohibiting
   DNS-over-TCP service is more detrimental to the continued utility and
   success of the DNS than when its usage is allowed.

4.  Network and System Considerations



   This section describes measures that systems and applications can
   take to optimize performance over TCP and to protect themselves from
   TCP-based resource exhaustion and attacks.

4.1.  Connection Establishment and Admission



   Resolvers and other DNS clients should be aware that some servers
   might not be reachable over TCP.  For this reason, clients MAY track
   and limit the number of TCP connections and connection attempts to a
   single server.  Reachability problems can be caused by network
   elements close to the server, close to the client, or anywhere along
   the path between them.  Mobile clients that cache connection failures
   MAY do so on a per-network basis or MAY clear such a cache upon
   change of network.

   Additionally, DNS clients MAY enforce a short timeout on
   unestablished connections rather than rely on the host operating
   system's TCP connection timeout, which is often around 60-120 seconds
   (i.e., due to an initial retransmission timeout of 1 second, the
   exponential back-off rules of [RFC6298], and a limit of six retries
   as is the default in Linux).

   The SYN flooding attack is a denial-of-service method affecting hosts
   that run TCP server processes [RFC4987].  This attack can be very
   effective if not mitigated.  One of the most effective mitigation
   techniques is SYN cookies, described in Section 3.6 of [RFC4987],
   which allows the server to avoid allocating any state until the
   successful completion of the three-way handshake.

   Services not intended for use by the public Internet, such as most
   recursive name servers, SHOULD be protected with access controls.
   Ideally, these controls are placed in the network, well before any
   unwanted TCP packets can reach the DNS server host or application.
   If this is not possible, the controls can be placed in the
   application itself.  In some situations (e.g., attacks), it may be
   necessary to deploy access controls for DNS services that should
   otherwise be globally reachable.  See also [RFC5358].

   The FreeBSD and NetBSD operating systems have an "accept filter"
   feature ([accept_filter]) that postpones delivery of TCP connections
   to applications until a complete, valid request has been received.
   The dns_accf(9) filter ensures that a valid DNS message is received.
   If not, the bogus connection never reaches the application.  The
   Linux TCP_DEFER_ACCEPT feature, while more limited in scope, can
   provide some of the same benefits as the BSD accept filter feature.
   These features are implemented as low-level socket options and are
   not activated automatically.  If applications wish to use these
   features, they need to make specific calls to set the right options,
   and administrators may also need to configure the applications to
   appropriately use the features.

   Per [RFC7766], applications and administrators are advised to
   remember that TCP MAY be used before sending any UDP queries.
   Networks and applications MUST NOT be configured to refuse TCP
   queries that were not preceded by a UDP query.

   TCP Fast Open (TFO) [RFC7413] allows TCP clients to shorten the
   handshake for subsequent connections to the same server.  TFO saves
   one round-trip time in the connection setup.  DNS servers SHOULD
   enable TFO when possible.  Furthermore, DNS servers clustered behind
   a single service address (e.g., anycast or load balancing) SHOULD
   either use the same TFO server key on all instances or disable TFO
   for all members of the cluster.

   DNS clients MAY also enable TFO.  At the time of this writing, it is
   not implemented or is disabled by default on some operating systems.
   [WIKIPEDIA_TFO] describes applications and operating systems that
   support TFO.

4.2.  Connection Management



   Since host memory for TCP state is a finite resource, DNS clients and
   servers SHOULD actively manage their connections.  Applications that
   do not actively manage their connections can encounter resource
   exhaustion leading to denial of service.  For DNS, as in other
   protocols, there is a trade-off between keeping connections open for
   potential future use and the need to free up resources for new
   connections that will arrive.

   Operators of DNS server software SHOULD be aware that operating
   system and application vendors MAY impose a limit on the total number
   of established connections.  These limits may be designed to protect
   against DDoS attacks or performance degradation.  Operators SHOULD
   understand how to increase these limits if necessary and the
   consequences of doing so.  Limits imposed by the application SHOULD
   be lower than limits imposed by the operating system so that the
   application can apply its own policy to connection management, such
   as closing the oldest idle connections first.

   DNS server software MAY provide a configurable limit on the number of
   established connections per source IP address or subnet.  This can be
   used to ensure that a single or small set of users cannot consume all
   TCP resources and deny service to other users.  Note, however, that
   if this limit is enabled, it possibly limits client performance while
   leaving some TCP resources unutilized.  Operators SHOULD be aware of
   these trade-offs and ensure this limit, if configured, is set
   appropriately based on the number and diversity of their users and
   whether users connect from unique IP addresses or through a shared
   Network Address Translator (NAT) [RFC3022].

   DNS server software SHOULD provide a configurable timeout for idle
   TCP connections.  This can be used to free up resources for new
   connections and to ensure that idle connections are eventually
   closed.  At the same time, it possibly limits client performance
   while leaving some TCP resources unutilized.  For very busy name
   servers, this might be set to a low value, such as a few seconds.
   For less busy servers, it might be set to a higher value, such as
   tens of seconds.  DNS clients and servers SHOULD signal their timeout
   values using the edns-tcp-keepalive EDNS(0) option [RFC7828].

   DNS server software MAY provide a configurable limit on the number of
   transactions per TCP connection.  This can help protect against
   unfair connection use (e.g., not releasing connection slots to other
   clients) and network evasion attacks.

   Similarly, DNS server software MAY provide a configurable limit on
   the total duration of a TCP connection.  This can help protect
   against unfair connection use, slow read attacks, and network evasion
   attacks.

   Since clients may not be aware of server-imposed limits, clients
   utilizing TCP for DNS need to always be prepared to re-establish
   connections or otherwise retry outstanding queries.

4.3.  Connection Termination



   The TCP peer that initiates a connection close retains the socket in
   the TIME_WAIT state for some amount of time, possibly a few minutes.
   It is generally preferable for clients to initiate the close of a TCP
   connection so that busy servers do not accumulate many sockets in the
   TIME_WAIT state, which can cause performance problems or even denial
   of service.  The edns-tcp-keepalive EDNS(0) option [RFC7828] can be
   used to encourage clients to close connections.

   On systems where large numbers of sockets in TIME_WAIT are observed
   (as either a client or a server) and are affecting an application's
   performance, it may be tempting to tune local TCP parameters.  For
   example, the Linux kernel has a "sysctl" parameter named
   net.ipv4.tcp_tw_reuse, which allows connections in the TIME_WAIT
   state to be reused in specific circumstances.  Note, however, that
   this affects only outgoing (client) connections and has no impact on
   servers.  In most cases, it is NOT RECOMMENDED to change parameters
   related to the TIME_WAIT state.  It should only be done by those with
   detailed knowledge of both TCP and the affected application.

4.4.  DNS over TLS



   DNS messages may be sent over TLS to provide privacy between stubs
   and recursive resolvers.  [RFC7858] is a Standards Track document
   describing how this works.  Although DNS over TLS utilizes TCP port
   853 instead of port 53, this document applies equally well to DNS
   over TLS.  Note, however, that DNS over TLS is only defined between
   stubs and recursives at the time of this writing.

   The use of TLS places even stronger operational burdens on DNS
   clients and servers.  Cryptographic functions for authentication and
   encryption require additional processing.  Unoptimized connection
   setup with TLS 1.3 [RFC8446] takes one additional round trip compared
   to TCP.  Connection setup times can be reduced with TCP Fast Open and
   TLS False Start [RFC7918] for TLS 1.2.  TLS 1.3 session resumption
   does not reduce round-trip latency because no application profile for
   use of TLS 0-RTT data with DNS has been published at the time of this
   writing.  However, TLS session resumption can reduce the number of
   cryptographic operations, and in TLS 1.2, session resumption does
   reduce the number of additional round trips from two to one.

4.5.  Defaults and Recommended Limits



   A survey of features and defaults was conducted for popular open-
   source DNS server implementations at the time of writing.  This
   section documents those defaults and makes recommendations for
   configurable limits that can be used in the absence of any other
   information.  Any recommended values in this document are only
   intended as a starting point for administrators that are unsure of
   what sorts of limits might be reasonable.  Operators SHOULD use
   application-specific monitoring, system logs, and system monitoring
   tools to gauge whether their service is operating within or exceeding
   these limits and adjust accordingly.

   Most open-source DNS server implementations provide a configurable
   limit on the total number of established connections.  Default values
   range from 20 to 150.  In most cases, where the majority of queries
   take place over UDP, 150 is a reasonable limit.  For services or
   environments where most queries take place over TCP or TLS, 5000 is a
   more appropriate limit.

   Only some open-source implementations provide a way to limit the
   number of connections per source IP address or subnet, but the
   default is to have no limit.  For environments or situations where it
   may be necessary to enable this limit, 25 connections per source IP
   address is a reasonable starting point.  The limit should be
   increased when aggregated by subnet or for services where most
   queries take place over TCP or TLS.

   Most open-source implementations provide a configurable idle timeout
   on connections.  Default values range from 2 to 30 seconds.  In most
   cases, 10 seconds is a reasonable default for this limit.  Longer
   timeouts improve connection reuse, but busy servers may need to use a
   lower limit.

   Only some open-source implementations provide a way to limit the
   number of transactions per connection, but the default is to have no
   limit.  This document does not offer advice on particular values for
   such a limit.

   Only some open-source implementations provide a way to limit the
   duration of connection, but the default is to have no limit.  This
   document does not offer advice on particular values for such a limit.

5.  DNS-over-TCP Filtering Risks



   Networks that filter DNS over TCP risk losing access to significant
   or important pieces of the DNS namespace.  For a variety of reasons,
   a DNS answer may require a DNS-over-TCP query.  This may include
   large message sizes, lack of EDNS(0) support, or DDoS mitigation
   techniques (including Response Rate Limiting [RRL]); additionally,
   perhaps some future capability that is as yet unforeseen will also
   demand TCP transport.

   For example, [RFC7901] describes a latency-avoiding technique that
   sends extra data in DNS responses.  This makes responses larger and
   potentially increases the effectiveness of DDoS reflection attacks.
   The specification mandates the use of TCP or DNS cookies [RFC7873].

   Even if any or all particular answers have consistently been returned
   successfully with UDP in the past, this continued behavior cannot be
   guaranteed when DNS messages are exchanged between autonomous
   systems.  Therefore, filtering of DNS over TCP is considered harmful
   and contrary to the safe and successful operation of the Internet.
   This section enumerates some of the known risks at the time of this
   writing when networks filter DNS over TCP.

5.1.  Truncation, Retries, and Timeouts



   Networks that filter DNS over TCP may inadvertently cause problems
   for third-party resolvers as experienced by [TOYAMA].  For example, a
   resolver receives queries for a moderately popular domain.  The
   resolver forwards the queries to the domain's authoritative name
   servers, but those servers respond with the TC bit set.  The resolver
   retries over TCP, but the authoritative server blocks DNS over TCP.
   The pending connections consume resources on the resolver until they
   time out.  If the number and frequency of these truncated-and-then-
   blocked queries are sufficiently high, the resolver wastes valuable
   resources on queries that can never be answered.  This condition is
   generally not easily or completely mitigated by the affected DNS
   resolver operator.

5.2.  DNS Root Zone KSK Rollover



   The plans for deploying DNSSEC KSK for the root zone highlighted a
   potential problem in retrieving the root zone key set [LEWIS].
   During some phases of the KSK rollover process, root zone DNSKEY
   responses were larger than 1280 bytes, the IPv6 minimum MTU for links
   carrying IPv6 traffic [RFC8200].  There was some concern that any DNS
   server unable to receive large DNS messages over UDP, or any DNS
   message over TCP, would experience disruption while performing DNSSEC
   validation [KSK_ROLLOVER_ARCHIVES].

   However, during the year-long postponement of the KSK rollover, there
   were no reported problems that could be attributed to the 1414 octet
   DNSKEY response when both the old and new keys were published in the
   zone.  Additionally, there were no reported problems during the two-
   month period when the old key was published as revoked and the DNSKEY
   response was 1425 octets in size [ROLL_YOUR_ROOT].

6.  Logging and Monitoring



   Developers of applications that log or monitor DNS SHOULD NOT ignore
   TCP due to the perception that it is rarely used or is hard to
   process.  Operators SHOULD ensure that their monitoring and logging
   applications properly capture DNS messages over TCP.  Otherwise,
   attacks, exfiltration attempts, and normal traffic may go undetected.

   DNS messages over TCP are in no way guaranteed to arrive in single
   segments.  In fact, a clever attacker might attempt to hide certain
   messages by forcing them over very small TCP segments.  Applications
   that capture network packets (e.g., with libpcap [libpcap]) SHOULD
   implement and perform full TCP stream reassembly and analyze the
   reassembled stream instead of the individual packets.  Otherwise,
   they are vulnerable to network evasion attacks [phrack].
   Furthermore, such applications need to protect themselves from
   resource exhaustion attacks by limiting the amount of memory
   allocated to tracking unacknowledged connection state data.  dnscap
   [dnscap] is an open-source example of a DNS logging program that
   implements TCP stream reassembly.

   Developers SHOULD also keep in mind connection reuse, query
   pipelining, and out-of-order responses when building and testing DNS
   monitoring applications.

   As an alternative to packet capture, some DNS server software
   supports dnstap [dnstap] as an integrated monitoring protocol
   intended to facilitate wide-scale DNS monitoring.

7.  IANA Considerations



   This document has no IANA actions.

8.  Security Considerations



   This document, providing operational requirements, is the companion
   to the implementation requirements of DNS over TCP provided in
   [RFC7766].  The security considerations from [RFC7766] still apply.

   Ironically, returning truncated DNS-over-UDP answers in order to
   induce a client query to switch to DNS over TCP has become a common
   response to source-address-spoofed, DNS denial-of-service attacks
   [RRL].  Historically, operators have been wary of TCP-based attacks,
   but in recent years, UDP-based flooding attacks have proven to be the
   most common protocol attack on the DNS.  Nevertheless, a high rate of
   short-lived DNS transactions over TCP may pose challenges.  In fact,
   [DAI21] details a class of IP fragmentation attacks on DNS
   transactions if the IP Identifier field (16 bits in IPv4 and 32 bits
   in IPv6) can be predicted and a system is coerced to fragment rather
   than retransmit messages.  While many operators have provided DNS-
   over-TCP service for many years without duress, past experience is no
   guarantee of future success.

   DNS over TCP is similar to many other Internet TCP services.  TCP
   threats and many mitigation strategies have been well documented in a
   series of documents such as [RFC4953], [RFC4987], [RFC5927], and
   [RFC5961].

   As mentioned in Section 6, applications that implement TCP stream
   reassembly need to limit the amount of memory allocated to connection
   tracking.  A failure to do so could lead to a total failure of the
   logging or monitoring application.  Imposition of resource limits
   creates a trade-off between allowing some stream reassembly to
   continue and allowing some evasion attacks to succeed.

   This document recommends that DNS servers enable TFO when possible.
   [RFC7413] recommends that a pool of servers behind a load balancer
   with a shared server IP address also share the key used to generate
   Fast Open cookies to prevent inordinate fallback to the three-way
   handshake (3WHS).  This guidance remains accurate but comes with a
   caveat: compromise of one server would reveal this group-shared key
   and allow for attacks involving the other servers in the pool by
   forging invalid Fast Open cookies.

9.  Privacy Considerations



   Since DNS over both UDP and TCP uses the same underlying message
   format, the use of one transport instead of the other does not change
   the privacy characteristics of the message content (i.e., the name
   being queried).  A number of protocols have recently been developed
   to provide DNS privacy, including DNS over TLS [RFC7858], DNS over
   DTLS [RFC8094], DNS over HTTPS [RFC8484], with even more on the way.

   Because TCP is somewhat more complex than UDP, some characteristics
   of a TCP conversation may enable DNS client fingerprinting and
   tracking that is not possible with UDP.  For example, the choice of
   initial sequence numbers, window size, and options might be able to
   identify a particular TCP implementation or even individual hosts
   behind shared resources such as NATs.

10.  References



10.1.  Normative References



   [RFC1035]  Mockapetris, P., "Domain names - implementation and
              specification", STD 13, RFC 1035, DOI 10.17487/RFC1035,
              November 1987, <https://www.rfc-editor.org/info/rfc1035>.

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,
              <https://www.rfc-editor.org/info/rfc2119>.

   [RFC2181]  Elz, R. and R. Bush, "Clarifications to the DNS
              Specification", RFC 2181, DOI 10.17487/RFC2181, July 1997,
              <https://www.rfc-editor.org/info/rfc2181>.

   [RFC6891]  Damas, J., Graff, M., and P. Vixie, "Extension Mechanisms
              for DNS (EDNS(0))", STD 75, RFC 6891,
              DOI 10.17487/RFC6891, April 2013,
              <https://www.rfc-editor.org/info/rfc6891>.

   [RFC7766]  Dickinson, J., Dickinson, S., Bellis, R., Mankin, A., and
              D. Wessels, "DNS Transport over TCP - Implementation
              Requirements", RFC 7766, DOI 10.17487/RFC7766, March 2016,
              <https://www.rfc-editor.org/info/rfc7766>.

   [RFC7828]  Wouters, P., Abley, J., Dickinson, S., and R. Bellis, "The
              edns-tcp-keepalive EDNS0 Option", RFC 7828,
              DOI 10.17487/RFC7828, April 2016,
              <https://www.rfc-editor.org/info/rfc7828>.

   [RFC7873]  Eastlake 3rd, D. and M. Andrews, "Domain Name System (DNS)
              Cookies", RFC 7873, DOI 10.17487/RFC7873, May 2016,
              <https://www.rfc-editor.org/info/rfc7873>.

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
              2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
              May 2017, <https://www.rfc-editor.org/info/rfc8174>.

10.2.  Informative References



   [accept_filter]
              FreeBSD, "FreeBSD accept_filter(9)", June 2000,
              <https://www.freebsd.org/cgi/man.cgi?query=accept_filter>.

   [APNIC]    Huston, G., "DNS XL", October 2020,
              <https://labs.apnic.net/?p=1380>.

   [AVOID_FRAGS]
              Fujiwara, K. and P. Vixie, "Fragmentation Avoidance in
              DNS", Work in Progress, Internet-Draft, draft-ietf-dnsop-
              avoid-fragmentation-06, 23 December 2021,
              <https://datatracker.ietf.org/doc/html/draft-ietf-dnsop-
              avoid-fragmentation-06>.

   [BIND]     Internet Systems Consortium, "BIND 9",
              <https://www.isc.org/bind/>.

   [CASTRO2010]
              Castro, S., Zhang, M., John, W., Wessels, D., and K.
              claffy, "Understanding and Preparing for DNS Evolution",
              DOI 10.1007/978-3-642-12365-8_1, April 2010,
              <https://doi.org/10.1007/978-3-642-12365-8_1>.

   [CHES94]   Cheswick, W. and S. Bellovin, "Firewalls and Internet
              Security: Repelling the Wily Hacker", First Edition, 1994.

   [CLOUDFLARE]
              Grant, D., "Economical With The Truth: Making DNSSEC
              Answers Cheap", June 2016,
              <https://blog.cloudflare.com/black-lies/>.

   [DAI21]    Dai, T., Shulman, H., and M. Waidner, "DNS-over-TCP
              Considered Vulnerable", DOI 10.1145/3472305.3472884, July
              2021, <https://doi.org/10.1145/3472305.3472884>.

   [DESIGNTEAM]
              ICANN, "Root Zone KSK Rollover Plan", March 2016,
              <https://www.iana.org/reports/2016/root-ksk-rollover-
              design-20160307.pdf>.

   [DJBDNS]   Bernstein, D., "When are TCP queries sent?", November
              2002, <https://cr.yp.to/djbdns/tcp.html#why>.

   [dnscap]   DNS-OARC, "DNSCAP", February 2014,
              <https://www.dns-oarc.net/tools/dnscap>.

   [dnstap]   "dnstap", <https://dnstap.info>.

   [ECDSA]    van Rijswijk-Deij, R., Sperotto, A., and A. Pras, "Making
              the Case for Elliptic Curves in DNSSEC",
              DOI 10.1145/2831347.2831350, October 2015,
              <https://dl.acm.org/doi/10.1145/2831347.2831350>.

   [FLAGDAY2020]
              DNS Software and Service Providers, "DNS Flag Day 2020",
              October 2020, <https://dnsflagday.net/2020/>.

   [FRAG_POISON]
              Herzberg, A. and H. Shulman, "Fragmentation Considered
              Poisonous", May 2012,
              <https://arxiv.org/pdf/1205.4011.pdf>.

   [HUSTON]   Huston, G., "Dealing with IPv6 fragmentation in the DNS",
              August 2017, <https://blog.apnic.net/2017/08/22/dealing-
              ipv6-fragmentation-dns/>.

   [KSK_ROLLOVER_ARCHIVES]
              ICANN, "KSK Rollover List Archives", January 2019,
              <https://mm.icann.org/pipermail/ksk-rollover/2019-January/
              date.html>.

   [LEWIS]    Lewis, E., "2017 DNSSEC KSK Rollover", RIPE 74, May 2017,
              <https://ripe74.ripe.net/presentations/25-RIPE74-lewis-
              submission.pdf>.

   [libpcap]  The Tcpdump Group, "Tcpdump and Libpcap",
              <https://www.tcpdump.org>.

   [NETALYZR] Kreibich, C., Weaver, N., Nechaev, B., and V. Paxson,
              "Netalyzr: Illuminating The Edge Network",
              DOI 10.1145/1879141.1879173, November 2010,
              <https://doi.org/10.1145/1879141.1879173>.

   [phrack]   horizon, "Defeating Sniffers and Intrusion Detection
              Systems", Phrack Magazine, December 1998,
              <http://phrack.org/issues/54/10.html>.

   [RFC0768]  Postel, J., "User Datagram Protocol", STD 6, RFC 768,
              DOI 10.17487/RFC0768, August 1980,
              <https://www.rfc-editor.org/info/rfc768>.

   [RFC0793]  Postel, J., "Transmission Control Protocol", STD 7,
              RFC 793, DOI 10.17487/RFC0793, September 1981,
              <https://www.rfc-editor.org/info/rfc793>.

   [RFC0883]  Mockapetris, P., "Domain names: Implementation
              specification", RFC 883, DOI 10.17487/RFC0883, November
              1983, <https://www.rfc-editor.org/info/rfc883>.

   [RFC1034]  Mockapetris, P., "Domain names - concepts and facilities",
              STD 13, RFC 1034, DOI 10.17487/RFC1034, November 1987,
              <https://www.rfc-editor.org/info/rfc1034>.

   [RFC1123]  Braden, R., Ed., "Requirements for Internet Hosts -
              Application and Support", STD 3, RFC 1123,
              DOI 10.17487/RFC1123, October 1989,
              <https://www.rfc-editor.org/info/rfc1123>.

   [RFC1536]  Kumar, A., Postel, J., Neuman, C., Danzig, P., and S.
              Miller, "Common DNS Implementation Errors and Suggested
              Fixes", RFC 1536, DOI 10.17487/RFC1536, October 1993,
              <https://www.rfc-editor.org/info/rfc1536>.

   [RFC1995]  Ohta, M., "Incremental Zone Transfer in DNS", RFC 1995,
              DOI 10.17487/RFC1995, August 1996,
              <https://www.rfc-editor.org/info/rfc1995>.

   [RFC1996]  Vixie, P., "A Mechanism for Prompt Notification of Zone
              Changes (DNS NOTIFY)", RFC 1996, DOI 10.17487/RFC1996,
              August 1996, <https://www.rfc-editor.org/info/rfc1996>.

   [RFC2136]  Vixie, P., Ed., Thomson, S., Rekhter, Y., and J. Bound,
              "Dynamic Updates in the Domain Name System (DNS UPDATE)",
              RFC 2136, DOI 10.17487/RFC2136, April 1997,
              <https://www.rfc-editor.org/info/rfc2136>.

   [RFC2541]  Eastlake 3rd, D., "DNS Security Operational
              Considerations", RFC 2541, DOI 10.17487/RFC2541, March
              1999, <https://www.rfc-editor.org/info/rfc2541>.

   [RFC2671]  Vixie, P., "Extension Mechanisms for DNS (EDNS0)",
              RFC 2671, DOI 10.17487/RFC2671, August 1999,
              <https://www.rfc-editor.org/info/rfc2671>.

   [RFC2694]  Srisuresh, P., Tsirtsis, G., Akkiraju, P., and A.
              Heffernan, "DNS extensions to Network Address Translators
              (DNS_ALG)", RFC 2694, DOI 10.17487/RFC2694, September
              1999, <https://www.rfc-editor.org/info/rfc2694>.

   [RFC3022]  Srisuresh, P. and K. Egevang, "Traditional IP Network
              Address Translator (Traditional NAT)", RFC 3022,
              DOI 10.17487/RFC3022, January 2001,
              <https://www.rfc-editor.org/info/rfc3022>.

   [RFC3225]  Conrad, D., "Indicating Resolver Support of DNSSEC",
              RFC 3225, DOI 10.17487/RFC3225, December 2001,
              <https://www.rfc-editor.org/info/rfc3225>.

   [RFC3226]  Gudmundsson, O., "DNSSEC and IPv6 A6 aware server/resolver
              message size requirements", RFC 3226,
              DOI 10.17487/RFC3226, December 2001,
              <https://www.rfc-editor.org/info/rfc3226>.

   [RFC4472]  Durand, A., Ihren, J., and P. Savola, "Operational
              Considerations and Issues with IPv6 DNS", RFC 4472,
              DOI 10.17487/RFC4472, April 2006,
              <https://www.rfc-editor.org/info/rfc4472>.

   [RFC4953]  Touch, J., "Defending TCP Against Spoofing Attacks",
              RFC 4953, DOI 10.17487/RFC4953, July 2007,
              <https://www.rfc-editor.org/info/rfc4953>.

   [RFC4987]  Eddy, W., "TCP SYN Flooding Attacks and Common
              Mitigations", RFC 4987, DOI 10.17487/RFC4987, August 2007,
              <https://www.rfc-editor.org/info/rfc4987>.

   [RFC5358]  Damas, J. and F. Neves, "Preventing Use of Recursive
              Nameservers in Reflector Attacks", BCP 140, RFC 5358,
              DOI 10.17487/RFC5358, October 2008,
              <https://www.rfc-editor.org/info/rfc5358>.

   [RFC5452]  Hubert, A. and R. van Mook, "Measures for Making DNS More
              Resilient against Forged Answers", RFC 5452,
              DOI 10.17487/RFC5452, January 2009,
              <https://www.rfc-editor.org/info/rfc5452>.

   [RFC5507]  IAB, Faltstrom, P., Ed., Austein, R., Ed., and P. Koch,
              Ed., "Design Choices When Expanding the DNS", RFC 5507,
              DOI 10.17487/RFC5507, April 2009,
              <https://www.rfc-editor.org/info/rfc5507>.

   [RFC5625]  Bellis, R., "DNS Proxy Implementation Guidelines",
              BCP 152, RFC 5625, DOI 10.17487/RFC5625, August 2009,
              <https://www.rfc-editor.org/info/rfc5625>.

   [RFC5927]  Gont, F., "ICMP Attacks against TCP", RFC 5927,
              DOI 10.17487/RFC5927, July 2010,
              <https://www.rfc-editor.org/info/rfc5927>.

   [RFC5936]  Lewis, E. and A. Hoenes, Ed., "DNS Zone Transfer Protocol
              (AXFR)", RFC 5936, DOI 10.17487/RFC5936, June 2010,
              <https://www.rfc-editor.org/info/rfc5936>.

   [RFC5961]  Ramaiah, A., Stewart, R., and M. Dalal, "Improving TCP's
              Robustness to Blind In-Window Attacks", RFC 5961,
              DOI 10.17487/RFC5961, August 2010,
              <https://www.rfc-editor.org/info/rfc5961>.

   [RFC5966]  Bellis, R., "DNS Transport over TCP - Implementation
              Requirements", RFC 5966, DOI 10.17487/RFC5966, August
              2010, <https://www.rfc-editor.org/info/rfc5966>.

   [RFC6298]  Paxson, V., Allman, M., Chu, J., and M. Sargent,
              "Computing TCP's Retransmission Timer", RFC 6298,
              DOI 10.17487/RFC6298, June 2011,
              <https://www.rfc-editor.org/info/rfc6298>.

   [RFC6762]  Cheshire, S. and M. Krochmal, "Multicast DNS", RFC 6762,
              DOI 10.17487/RFC6762, February 2013,
              <https://www.rfc-editor.org/info/rfc6762>.

   [RFC6781]  Kolkman, O., Mekking, W., and R. Gieben, "DNSSEC
              Operational Practices, Version 2", RFC 6781,
              DOI 10.17487/RFC6781, December 2012,
              <https://www.rfc-editor.org/info/rfc6781>.

   [RFC6950]  Peterson, J., Kolkman, O., Tschofenig, H., and B. Aboba,
              "Architectural Considerations on Application Features in
              the DNS", RFC 6950, DOI 10.17487/RFC6950, October 2013,
              <https://www.rfc-editor.org/info/rfc6950>.

   [RFC7413]  Cheng, Y., Chu, J., Radhakrishnan, S., and A. Jain, "TCP
              Fast Open", RFC 7413, DOI 10.17487/RFC7413, December 2014,
              <https://www.rfc-editor.org/info/rfc7413>.

   [RFC7477]  Hardaker, W., "Child-to-Parent Synchronization in DNS",
              RFC 7477, DOI 10.17487/RFC7477, March 2015,
              <https://www.rfc-editor.org/info/rfc7477>.

   [RFC7534]  Abley, J. and W. Sotomayor, "AS112 Nameserver Operations",
              RFC 7534, DOI 10.17487/RFC7534, May 2015,
              <https://www.rfc-editor.org/info/rfc7534>.

   [RFC7720]  Blanchet, M. and L-J. Liman, "DNS Root Name Service
              Protocol and Deployment Requirements", BCP 40, RFC 7720,
              DOI 10.17487/RFC7720, December 2015,
              <https://www.rfc-editor.org/info/rfc7720>.

   [RFC7858]  Hu, Z., Zhu, L., Heidemann, J., Mankin, A., Wessels, D.,
              and P. Hoffman, "Specification for DNS over Transport
              Layer Security (TLS)", RFC 7858, DOI 10.17487/RFC7858, May
              2016, <https://www.rfc-editor.org/info/rfc7858>.

   [RFC7901]  Wouters, P., "CHAIN Query Requests in DNS", RFC 7901,
              DOI 10.17487/RFC7901, June 2016,
              <https://www.rfc-editor.org/info/rfc7901>.

   [RFC7918]  Langley, A., Modadugu, N., and B. Moeller, "Transport
              Layer Security (TLS) False Start", RFC 7918,
              DOI 10.17487/RFC7918, August 2016,
              <https://www.rfc-editor.org/info/rfc7918>.

   [RFC8027]  Hardaker, W., Gudmundsson, O., and S. Krishnaswamy,
              "DNSSEC Roadblock Avoidance", BCP 207, RFC 8027,
              DOI 10.17487/RFC8027, November 2016,
              <https://www.rfc-editor.org/info/rfc8027>.

   [RFC8094]  Reddy, T., Wing, D., and P. Patil, "DNS over Datagram
              Transport Layer Security (DTLS)", RFC 8094,
              DOI 10.17487/RFC8094, February 2017,
              <https://www.rfc-editor.org/info/rfc8094>.

   [RFC8162]  Hoffman, P. and J. Schlyter, "Using Secure DNS to
              Associate Certificates with Domain Names for S/MIME",
              RFC 8162, DOI 10.17487/RFC8162, May 2017,
              <https://www.rfc-editor.org/info/rfc8162>.

   [RFC8200]  Deering, S. and R. Hinden, "Internet Protocol, Version 6
              (IPv6) Specification", STD 86, RFC 8200,
              DOI 10.17487/RFC8200, July 2017,
              <https://www.rfc-editor.org/info/rfc8200>.

   [RFC8324]  Klensin, J., "DNS Privacy, Authorization, Special Uses,
              Encoding, Characters, Matching, and Root Structure: Time
              for Another Look?", RFC 8324, DOI 10.17487/RFC8324,
              February 2018, <https://www.rfc-editor.org/info/rfc8324>.

   [RFC8446]  Rescorla, E., "The Transport Layer Security (TLS) Protocol
              Version 1.3", RFC 8446, DOI 10.17487/RFC8446, August 2018,
              <https://www.rfc-editor.org/info/rfc8446>.

   [RFC8467]  Mayrhofer, A., "Padding Policies for Extension Mechanisms
              for DNS (EDNS(0))", RFC 8467, DOI 10.17487/RFC8467,
              October 2018, <https://www.rfc-editor.org/info/rfc8467>.

   [RFC8482]  Abley, J., Gudmundsson, O., Majkowski, M., and E. Hunt,
              "Providing Minimal-Sized Responses to DNS Queries That
              Have QTYPE=ANY", RFC 8482, DOI 10.17487/RFC8482, January
              2019, <https://www.rfc-editor.org/info/rfc8482>.

   [RFC8483]  Song, L., Ed., Liu, D., Vixie, P., Kato, A., and S. Kerr,
              "Yeti DNS Testbed", RFC 8483, DOI 10.17487/RFC8483,
              October 2018, <https://www.rfc-editor.org/info/rfc8483>.

   [RFC8484]  Hoffman, P. and P. McManus, "DNS Queries over HTTPS
              (DoH)", RFC 8484, DOI 10.17487/RFC8484, October 2018,
              <https://www.rfc-editor.org/info/rfc8484>.

   [RFC8490]  Bellis, R., Cheshire, S., Dickinson, J., Dickinson, S.,
              Lemon, T., and T. Pusateri, "DNS Stateful Operations",
              RFC 8490, DOI 10.17487/RFC8490, March 2019,
              <https://www.rfc-editor.org/info/rfc8490>.

   [RFC8501]  Howard, L., "Reverse DNS in IPv6 for Internet Service
              Providers", RFC 8501, DOI 10.17487/RFC8501, November 2018,
              <https://www.rfc-editor.org/info/rfc8501>.

   [RFC8806]  Kumari, W. and P. Hoffman, "Running a Root Server Local to
              a Resolver", RFC 8806, DOI 10.17487/RFC8806, June 2020,
              <https://www.rfc-editor.org/info/rfc8806>.

   [RFC8906]  Andrews, M. and R. Bellis, "A Common Operational Problem
              in DNS Servers: Failure to Communicate", BCP 231,
              RFC 8906, DOI 10.17487/RFC8906, September 2020,
              <https://www.rfc-editor.org/info/rfc8906>.

   [RFC8932]  Dickinson, S., Overeinder, B., van Rijswijk-Deij, R., and
              A. Mankin, "Recommendations for DNS Privacy Service
              Operators", BCP 232, RFC 8932, DOI 10.17487/RFC8932,
              October 2020, <https://www.rfc-editor.org/info/rfc8932>.

   [RFC8945]  Dupont, F., Morris, S., Vixie, P., Eastlake 3rd, D.,
              Gudmundsson, O., and B. Wellington, "Secret Key
              Transaction Authentication for DNS (TSIG)", STD 93,
              RFC 8945, DOI 10.17487/RFC8945, November 2020,
              <https://www.rfc-editor.org/info/rfc8945>.

   [ROLL_YOUR_ROOT]
              Müller, M., Thomas, M., Wessels, D., Hardaker, W., Chung,
              T., Toorop, W., and R. van Rijswijk-Deij, "Roll, Roll,
              Roll Your Root: A Comprehensive Analysis of the First Ever
              DNSSEC Root KSK Rollover", DOI 10.1145/3355369.3355570,
              October 2019,
              <https://dl.acm.org/doi/10.1145/3355369.3355570>.

   [RRL]      Vixie, P. and V. Schryver, "DNS Response Rate Limiting
              (DNS RRL)", ISC-TN-2012-1-Draft1, April 2012.

   [TDNS]     Zhu, L., Heidemann, J., Wessels, D., Mankin, A., and N.
              Somaiya, "Connection-Oriented DNS to Improve Privacy and
              Security", DOI 10.1109/SP.2015.18, May 2015,
              <https://doi.org/10.1109/SP.2015.18>.

   [TOYAMA]   Toyama, K., Ishibashi, K., Toyono, T., Ishino, M.,
              Yoshimura, C., and K. Fujiwara, "DNS Anomalies and Their
              Impacts on DNS Cache Servers", NANOG 32, October 2004.

   [VERISIGN] Thomas, M. and D. Wessels, "An Analysis of TCP Traffic in
              Root Server DITL Data", DNS-OARC 2014 Fall Workshop,
              October 2014.

   [WIKIPEDIA_TFO]
              Wikipedia, "TCP Fast Open", February 2022,
              <https://en.wikipedia.org/w/
              index.php?title=TCP_Fast_Open&oldid=1071397204>.

Appendix A.  RFCs Related to DNS Transport over TCP



   This section enumerates all known RFCs with a status of Internet
   Standard, Proposed Standard, Informational, Best Current Practice, or
   Experimental that either implicitly or explicitly make assumptions or
   statements about the use of TCP as a transport for the DNS germane to
   this document.

A.1.  RFC 1035 - DOMAIN NAMES - IMPLEMENTATION AND SPECIFICATION



   The Internet Standard [RFC1035] is the base DNS specification that
   explicitly defines support for DNS over TCP.

A.2.  RFC 1536 - Common DNS Implementation Errors and Suggested Fixes



   The Informational document [RFC1536] states that UDP is "the chosen
   protocol for communication though TCP is used for zone transfers."
   That statement should now be considered in its historical context and
   is no longer a proper reflection of modern expectations.

A.3.  RFC 1995 - Incremental Zone Transfer in DNS



   The Proposed Standard [RFC1995] documents the use of TCP as the
   fallback transport when Incremental Zone Transfer (IXFR) responses do
   not fit into a single UDP response.  As with Authoritative Transfer
   (AXFR), IXFR messages are typically delivered over TCP by default in
   practice.

A.4.  RFC 1996 - A Mechanism for Prompt Notification of Zone Changes
      (DNS NOTIFY)



   The Proposed Standard [RFC1996] suggests that a primary server may
   decide to issue NOTIFY messages over TCP.  In practice, NOTIFY
   messages are generally sent over UDP, but this specification leaves
   open the possibility that the choice of transport protocol is up to
   the primary server; therefore, a secondary server ought to be able to
   operate over both UDP and TCP.

A.5.  RFC 2181 - Clarifications to the DNS Specification



   The Proposed Standard [RFC2181] includes clarifying text on how a
   client should react to the TC bit set on responses.  It is advised
   that the response be discarded and the query resent using TCP.

A.6.  RFC 2694 - DNS extensions to Network Address Translators (DNS_ALG)



   The Informational document [RFC2694] enumerates considerations for
   NAT devices to properly handle DNS traffic.  This document is
   noteworthy in its suggestion that "[t]ypically, TCP is used for AXFR
   requests," as further evidence that helps explain why DNS over TCP
   may have often been treated very differently than DNS over UDP in
   operational networks.

A.7.  RFC 3225 - Indicating Resolver Support of DNSSEC



   The Proposed Standard [RFC3225] makes statements indicating that DNS
   over TCP is "detrimental" as a result of increased traffic, latency,
   and server load.  This document is a companion to the next document
   in the RFC Series that describes the requirement for EDNS(0) support
   for DNSSEC.

A.8.  RFC 3226 - DNSSEC and IPv6 A6 aware server/resolver message size
      requirements



   Although updated by later DNSSEC RFCs, the Proposed Standard
   [RFC3226] strongly argues in favor of UDP messages instead of TCP,
   largely for performance reasons.  The document declares EDNS(0) a
   requirement for DNSSEC servers and advocates that packet
   fragmentation may be preferable to TCP in certain situations.

A.9.  RFC 4472 - Operational Considerations and Issues with IPv6 DNS



   The Informational document [RFC4472] notes that IPv6 data may
   increase DNS responses beyond what would fit in a UDP message.  What
   is particularly noteworthy, but perhaps less common today than when
   this document was written, is that it refers to implementations that
   truncate data without setting the TC bit to encourage the client to
   resend the query using TCP.

A.10.  RFC 5452 - Measures for Making DNS More Resilient against Forged
       Answers



   The Proposed Standard [RFC5452] arose as public DNS systems began to
   experience widespread abuse from spoofed queries, resulting in
   amplification and reflection attacks against unwitting victims.  One
   of the leading justifications for supporting DNS over TCP to thwart
   these attacks is briefly described in Section 9.3 of [RFC5452]
   ("Spoof Detection and Countermeasure").

A.11.  RFC 5507 - Design Choices When Expanding the DNS



   The Informational document [RFC5507] was largely an attempt to
   dissuade new DNS data types from overloading the TXT resource record
   type.  In so doing, it summarizes the conventional wisdom of DNS
   design and implementation practices.  The authors suggest TCP
   overhead and stateful properties pose challenges compared to UDP and
   imply that UDP is generally preferred for performance and robustness.

A.12.  RFC 5625 - DNS Proxy Implementation Guidelines



   The Best Current Practice document [RFC5625] provides DNS proxy
   implementation guidance including the mandate that a proxy "MUST
   [...] be prepared to receive and forward queries over TCP" even
   though it suggests that, historically, TCP transport has not been
   strictly mandatory in stub resolvers or recursive servers.

A.13.  RFC 5936 - DNS Zone Transfer Protocol (AXFR)



   The Proposed Standard [RFC5936] provides a detailed specification for
   the zone transfer protocol, as originally outlined in the early DNS
   standards.  AXFR operation is limited to TCP and not specified for
   UDP.  This document discusses TCP usage at length.

A.14.  RFC 7534 - AS112 Nameserver Operations



   The Informational document [RFC7534] enumerates the requirements for
   operation of AS112 project DNS servers.  New AS112 nodes are tested
   for their ability to provide service on both UDP and TCP transports,
   with the implication that TCP service is an expected part of normal
   operations.

A.15.  RFC 6762 - Multicast DNS



   In the Proposed Standard [RFC6762], the TC bit is deemed to have
   essentially the same meaning as described in the original DNS
   specifications.  That is, if a response with the TC bit set is
   received, "[...] the querier SHOULD reissue its query using TCP in
   order to receive the larger response."

A.16.  RFC 6891 - Extension Mechanisms for DNS (EDNS(0))



   The Internet Standard [RFC6891] helped slow the use of and need for
   DNS-over-TCP messages.  This document highlights concerns over server
   load and scalability in widespread use of DNS over TCP.

A.17.  IAB RFC 6950 - Architectural Considerations on Application
       Features in the DNS



   The Informational document [RFC6950] draws attention to large data in
   the DNS.  TCP is referenced in the context as a common fallback
   mechanism and counter to some spoofing attacks.

A.18.  RFC 7477 - Child-to-Parent Synchronization in DNS



   The Proposed Standard [RFC7477] specifies an RRType and a protocol to
   signal and synchronize NS, A, and AAAA resource record changes from a
   child-to-parent zone.  Since this protocol may require multiple
   requests and responses, it recommends utilizing DNS over TCP to
   ensure the conversation takes place between a consistent pair of end
   nodes.

A.19.  RFC 7720 - DNS Root Name Service Protocol and Deployment
       Requirements



   The Best Current Practice document [RFC7720] declares that root name
   service "MUST support UDP [RFC0768] and TCP [RFC0793] transport of
   DNS queries and responses."

A.20.  RFC 7766 - DNS Transport over TCP - Implementation Requirements



   The Proposed Standard [RFC7766] instructs DNS implementors to provide
   support for carrying DNS-over-TCP messages in their software and
   might be considered the direct ancestor of this operational
   requirements document.  The implementation requirements document
   codifies mandatory support for DNS-over-TCP in compliant DNS software
   but makes no recommendations to operators, which we seek to address
   here.

A.21.  RFC 7828 - The edns-tcp-keepalive EDNS(0) Option



   The Proposed Standard [RFC7828] defines an EDNS(0) option to
   negotiate an idle timeout value for long-lived DNS-over-TCP
   connections.  Consequently, this document is only applicable and
   relevant to DNS-over-TCP sessions and between implementations that
   support this option.

A.22.  RFC 7858 - Specification for DNS over Transport Layer Security
       (TLS)



   The Proposed Standard [RFC7858] defines a method for putting DNS
   messages into a TCP-based encrypted channel using TLS.  This
   specification is noteworthy for explicitly targeting the stub-to-
   recursive traffic but does not preclude its application from
   recursive-to-authoritative traffic.

A.23.  RFC 7873 - Domain Name System (DNS) Cookies



   The Proposed Standard [RFC7873] describes an EDNS(0) option to
   provide additional protection against query and answer forgery.  This
   specification mentions DNS over TCP as an alternative mechanism when
   DNS cookies are not available.  The specification does make mention
   of DNS-over-TCP processing in two specific situations.  In one, when
   a server receives only a client cookie in a request, the server
   should consider whether the request arrived over TCP, and if so, it
   should consider accepting TCP as sufficient to authenticate the
   request and respond accordingly.  In another, when a client receives
   a BADCOOKIE reply using a fresh server cookie, the client should
   retry using TCP as the transport.

A.24.  RFC 7901 - CHAIN Query Requests in DNS



   The Experimental specification [RFC7901] describes an EDNS(0) option
   that can be used by a security-aware validating resolver to request
   and obtain a complete DNSSEC validation path for any single query.
   This document requires the use of DNS over TCP or a transport
   mechanism verified by a source IP address such as EDNS-COOKIE
   [RFC7873].

A.25.  RFC 8027 - DNSSEC Roadblock Avoidance



   The Best Current Practice document [RFC8027] details observed
   problems with DNSSEC deployment and mitigation techniques.  Network
   traffic blocking and restrictions, including DNS-over-TCP messages,
   are highlighted as one reason for DNSSEC deployment issues.  While
   this document suggests these sorts of problems are due to "non-
   compliant infrastructure", the scope of the document is limited to
   detection and mitigation techniques to avoid so-called DNSSEC
   roadblocks.

A.26.  RFC 8094 - DNS over Datagram Transport Layer Security (DTLS)



   The Experimental specification [RFC8094] details a protocol that uses
   a datagram transport (UDP) but stipulates that "DNS clients and
   servers that implement DNS over DTLS MUST also implement DNS over TLS
   in order to provide privacy for clients that desire Strict Privacy
   [...]."  This requirement implies DNS over TCP must be supported in
   case the message size is larger than the path MTU.

A.27.  RFC 8162 - Using Secure DNS to Associate Certificates with Domain
       Names for S/MIME



   The Experimental specification [RFC8162] describes a technique to
   authenticate user X.509 certificates in an S/MIME system via the DNS.
   The document points out that the new experimental resource record
   types are expected to carry large payloads, resulting in the
   suggestion that "applications SHOULD use TCP -- not UDP -- to perform
   queries for the SMIMEA resource record."

A.28.  RFC 8324 - DNS Privacy, Authorization, Special Uses, Encoding,
       Characters, Matching, and Root Structure: Time for Another Look?



   The Informational document [RFC8324] briefly discusses the common
   role and challenges of DNS over TCP throughout the history of DNS.

A.29.  RFC 8467 - Padding Policies for Extension Mechanisms for DNS
       (EDNS(0))



   The Experimental document [RFC8467] reminds implementors to consider
   the underlying transport protocol (e.g., TCP) when calculating the
   padding length when artificially increasing the DNS message size with
   an EDNS(0) padding option.

A.30.  RFC 8482 - Providing Minimal-Sized Responses to DNS Queries That
       Have QTYPE=ANY



   The Proposed Standard [RFC8482] describes alternative ways that DNS
   servers can respond to queries of type ANY, which are sometimes used
   to provide amplification in DDoS attacks.  The specification notes
   that responders may behave differently, depending on the transport.
   For example, minimal-sized responses may be used over UDP transport,
   while full responses may be given over TCP.

A.31.  RFC 8483 - Yeti DNS Testbed



   The Informational document [RFC8483] describes a testbed environment
   that highlights some DNS-over-TCP behaviors, including issues
   involving packet fragmentation and operational requirements for TCP
   stream assembly in order to conduct DNS measurement and analysis.

A.32.  RFC 8484 - DNS Queries over HTTPS (DoH)



   The Proposed Standard [RFC8484] defines a protocol for sending DNS
   queries and responses over HTTPS.  This specification assumes TLS and
   TCP for the underlying security and transport layers, respectively.
   Self-described as a technique that more closely resembles a tunneling
   mechanism, DoH nevertheless likely implies DNS over TCP in some
   sense, if not directly.

A.33.  RFC 8490 - DNS Stateful Operations



   The Proposed Standard [RFC8490] updates the base protocol
   specification with a new OPCODE to help manage stateful operations in
   persistent sessions, such as those that might be used by DNS over
   TCP.

A.34.  RFC 8501 - Reverse DNS in IPv6 for Internet Service Providers



   The Informational document [RFC8501] identifies potential operational
   challenges with dynamic DNS, including denial-of-service threats.
   The document suggests TCP may provide some advantages but that
   updating hosts would need to be explicitly configured to use TCP
   instead of UDP.

A.35.  RFC 8806 - Running a Root Server Local to a Resolver



   The Informational document [RFC8806] describes how to obtain and
   operate a local copy of the root zone with examples showing how to
   pull from authoritative sources using a DNS-over-TCP zone transfer.

A.36.  RFC 8906 - A Common Operational Problem in DNS Servers: Failure
       to Communicate



   The Best Current Practice document [RFC8906] discusses a number of
   DNS operational failure scenarios and how to avoid them.  This
   includes discussions involving DNS-over-TCP queries, EDNS over TCP,
   and a testing methodology that includes a section on verifying DNS-
   over-TCP functionality.

A.37.  RFC 8932 - Recommendations for DNS Privacy Service Operators



   The Best Current Practice document [RFC8932] presents privacy
   considerations to DNS privacy service operators.  These mechanisms
   sometimes include the use of TCP and are therefore susceptible to
   information leakage such as TCP-based fingerprinting.  This document
   also references an earlier draft version of this document.

A.38.  RFC 8945 - Secret Key Transaction Authentication for DNS (TSIG)



   The Internet Standard [RFC8945] recommends that a client use TCP if
   truncated TSIG messages are received.

Acknowledgments

   This document was initially motivated by feedback from students who
   pointed out that they were hearing contradictory information about
   filtering DNS-over-TCP messages.  Thanks in particular to a teaching
   colleague, JPL, who perhaps unknowingly encouraged the initial
   research into the differences between what the community has
   historically said and did.  Thanks to all the NANOG 63 attendees who
   provided feedback for an early talk on this subject.

   The following individuals provided an array of feedback to help
   improve this document: Joe Abley, Piet Barber, Sara Dickinson, Tony
   Finch, Bob Harold, Paul Hoffman, Geoff Huston, Tatuya Jinmei, Puneet
   Sood, and Richard Wilhelm.  The authors are also indebted to the
   contributions stemming from discussion in the TCPM Working Group
   meeting at IETF 104.  Any remaining errors or imperfections are the
   sole responsibility of the document authors.

Authors' Addresses



   John Kristoff
   Dataplane.org
   Chicago, IL 60605
   United States of America
   Phone: +1 312 493 0305
   Email: jtk@dataplane.org
   URI:   https://dataplane.org/jtk/


   Duane Wessels
   Verisign
   12061 Bluemont Way
   Reston, VA 20190
   United States of America
   Phone: +1 703 948 3200
   Email: dwessels@verisign.com
   URI:   https://verisign.com