TLS for DNS: Initiation and Performance Considerations

Today, nearly all DNS queries ( and ) are sent unencrypted, which makes them vulnerable to eavesdropping by an attacker that has access to the network channel, reducing the privacy of the querier. Recent news reports have elevated these concerns, and ongoing efforts are beginning to identify privacy concerns about DNS (). Prior work has addressed some aspects of DNS security, but until recently there has been little work on privacy between a DNS client and server. DNS Security Extensions (DNSSEC, ) provide response integrity by defining mechanisms to cryptographically sign zones, allowing end-users (or their first-hop resolver) to verify replies are correct. By intention, DNSSEC does not protect request and response privacy. Traditionally, either privacy was not considered a requirement for DNS traffic, or it was assumed that network traffic was sufficiently private, however these perceptions are evolving due to recent events . DNSCurve defines a method to add confidentiality to the link between DNS clients and servers; however, it does so with a new cryptographic protocol and does not take advantage of an existing standard protocol such as TLS. ConfidentialDNS and IPSECA use opportunistic encryption to offer privacy for DNS queries and responses. Finally, others have suggested DNS-over-TLS. Unbound DNS software includes a DNS-over-TLS implementation. The present document goes beyond past DNS-over-TLS discussions by providing two modes of initiation for DNS-over-TLS: use of a well-known port, and use of a negotiation mechanism in an established connection. Protocol changes proposed here must consider potential interactions with middle boxes. The port-based initiation of TLS is very straightforward, but might be blocked by firewalls or be unwelcome to some DNS client or server implementations. If port-based initiation of TLS fails, the negotiation mechanism allows DNS clients and servers to upgrade an existing DNS-over-TCP connection to a DNS-over-TLS connection, analogous to upgrade mechanisms in other uses of TLS, such as STARTTLS used in SMTP , IMAP and POP , to name just a few of many. Adding TLS to DNS-over-TCP avoids port blocking, but maybe interact poorly with middle boxes that inspect DNS traffic. As is generally the case with TLS, both approaches are subject to downgrade attacks, as discussed in . The protocol described here works for any DNS client to server communication using DNS-over-TCP. There can be different profiles providing different levels of privacy, as discussed in . The protocol may be used for any DNS communication both from stub to recursive, and from recursive to authoritative servers, but different protocols may be preferable for different environments. This document describes two profiles providing different levels of assurance of privacy: an opportunistic privacy profile and a pre-deployed profile.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

The only changes required for port-based DNS-over-TLS are those optimizing TCP and TLS performance discussed in the following. The DNS protocol itself is unchanged. Clients and servers negotiate upgrade-based DNS-over-TLS by setting a bit in the Flags field of the EDNS0 OPT meta-RR. The "TLS OK" (TO) bit is defined as the second bit of the third and fourth bytes of the "extended RCODE and flags" portion of the EDNS0 OPT meta-RR, immediately adjacent to the "DNSSEC OK" (DO) bit :

DNS clients first try port-based DNS-over-TLS. If that connection fails, they try upgrade-based DNS-over-TLS.

DNS clients SHOULD first try using port-based DNS-over-TLS by establishing the TCP connection to the dedicated port TBD (number to be defined in ). Clients MAY try STARTTLS upgrade before the dedicated port if there is information that this ordering is preferred. It SHOULD be an implementation and/or local determination as to whether to attempt TLS via the dedicated port first and then fall back to STARTTLS use, or to choose some other order of attempts and fallbacks.

Setting the TO bit in queries sent using UDP transport has no protocol meaning. However, the client MAY set the TO bit when using UDP transport. The server MUST ignore the TO bit when receiving UDP transport. DISCUSSION: community advice is sought on this. The advantage of allowing a client to send UDP on TO is that servers can collect information on deployment (as happened with the DO bit). The disadvantage is that a meaningless bit (TO over UDP) might cause confusion, and some middleboxes might not pass a UDP query with the TO bit set. DNS clients set the TO bit in the initial query sent to a server using TCP transport to signal their desire that the TCP connection be upgraded to TLS. DNS clients SHOULD NOT set the TO bit on queries when using TLS transport because doing so has no meaning in this protocol. Since the motivation for upgrade-based DNS-over-TLS is to preserve privacy, DNS clients SHOULD use an initial (unprotected) query that reveals no private information in the initial TO=1 query to a server. To provide a standard "dummy" query, it is RECOMMENDED to send the initial query with RD=0, QNAME="STARTTLS", QCLASS=CH, and QTYPE=TXT ("STARTTLS/CH/TXT") analogous to administrative queries already in widespread use . (For some profiles, the client MUST use a dummy query for the initial query.) After sending the initial TO=1 query using TCP transport, DNS clients MUST wait for the initial response before sending any subsequent queries over the same TCP connection.

A DNS client that receives a response using UDP transport that has the TO bit set handles that response as usual. It MAY record the server's support for DNS-over-TLS and use that information as part of its server selection algorithm in the case where multiple servers are available to service a particular query. A DNS client that has sent the TO bit using TCP transport and receives a response to its initial query that has the TO bit set MUST immediately initiate a TLS handshake using the procedure described in . (Note that this document does not yet deal with what happens when the TLS handshake does not succeed.) DISCUSSION: are there any cases in which a DNS client that sent TO on DNS-over-TCP and receives TO in the initial response from the server would not initiate the TLS handshake? Is there any reason for this to be SHOULD rather than MUST? A DNS client that receives a response to its initial query using TCP transport that has the TO bit clear MUST not initiate a TLS handshake and SHOULD utilize the existing TCP connection for subsequent queries. DNS clients SHOULD remember server IP addresses that don't support upgrade-based DNS-over-TLS, including TLS handshake failures, and not request DNS-over-TLS from them for reasonable period (such as one hour per server).

A DNS server that supports DNS-over-TLS SHOULD support port-based DNS-over-TLS, and SHOULD support upgrade-based DNS-over-TLS.

A DNS server receiving a query over UDP with the TO bit ignores that bit. A DNS server receiving a query over an existing TLS connection with the TO bit ignores that bit. A DNS server receiving an initial query over TCP that has the TO bit set MAY inform the client it is willing to establish a TLS session, as described in the next section. A DNS server receiving subsequent queries over TCP MUST ignore the TO bit. (A client wishing to start TLS after the initial query MUST open a new TCP connection to do so.)

A DNS server sending a response over UDP to a query that had an OPT meta-RR SHOULD set the TO bit to indicate its general support for DNS-over-TLS, as long as it is willing and able to support a TLS connection with the particular client. A DNS server receiving an initial query over TCP that has the TO bit set MAY set the TO bit in its response. The server MUST then proceed with the TLS handshake protocol. A DNS server receiving a "dummy" STARTTLS/CH/TXT query over TCP MUST respond with RCODE=0 and a TXT RR in the Answer section. Contents of the TXT RR are strictly informative (for humans) and MUST NOT be interpreted by the client software. Recommended TXT RDATA values are "STARTTLS" or "NO_TLS".

After TLS negotiation completes, the connection will be encrypted and is now protected from eavesdropping and normal DNS queries SHOULD take place, following DNS-over-TCP framing (, section 4.2.2). It is expected that multiple DNS queries will be made over the same TLS connection instead of tearing down the TLS connection after each response. A user of DNS-over-TLS SHOULD follow best practices for DNS-over-TCP, as described in . (For DNS clients that use library functions such as "gethostbyname()", current clients may open and close UDP connections each DNS call. We recommend they reuse a single TCP connection to the recursive resolver or use UDP to a caching resolver that uses a system-wide TCP connection to the recursive resolver.) Both clients and servers SHOULD follow existing DNS-over-TCP timeout rules, which are often implementation- and situation-dependent. In the absence of any other advice, the RECOMMENDED timeout values are 30 seconds for recursive name servers, 60 seconds for clients of recursive name servers, 10 seconds for authoritative name servers, and 20 seconds for clients of authoritative name servers. Current work in this area may assist DNS-over-TLS clients and servers select useful timeout values . As with current DNS-over-TCP, DNS servers MAY close the connection at any time (e.g., due to resource constraints). As with current DNS-over-TCP, clients MUST handle abrupt closes and be prepared to reestablish connections and/or retry queries. DNS servers SHOULD use the TLS close-notify request to shift TCP TIME-WAIT state to the clients. Additional requirements and guidance for optimizing DNS-over-TCP are provided by , . As discussed in , TCP Fast Open is of benefit. DNS servers SHOULD enable fast TLS session resumption to avoid keeping per-client session state.

Middleboxes may be present in some networks and have been known to interfere with normal DNS resolution and create problems for DNS-over-TLS. Remarkably, downgrade attacks can affect plaintext protocols that utilize "STARTTLS" signaling in a similar way. A DNS client attempting upgrade-based DNS-over-TLS through a middlebox, or in the presence of a downgrade attack, could have one of the following outcomes. (These outcomes are similar to those discussed in prior RFCs, such as .) The DNS client sends a TO=1 query and receives a TO=0 response. In this case there is no upgrade to TLS and DNS resolution occurs normally, without encryption. The DNS client sends a TO=1 query and receives a TO=1 response, but the middlebox does not understand the TLS negotiation and does not allow those packets to pass through. Clients SHOULD retry DNS without TO set if negotiation fails, and then retry with TLS after a reasonable period (see ). The DNS client sends a TO=1 query but receives no response at all. The middlebox might be silently dropping the query due to the presence of the TO bit, when it should, in fact, ignore and pass through unknown flag bits . The client SHOULD fall back to normal (unencrypted) DNS for a reasonable period (as discussed in ). In general, clients that attempt TLS and fail can either fall back on unencrypted DNS, or wait and retry later, depending on their privacy requirements.

This protocol provides flexibility to accommodate several different use cases. Two usage profiles are defined here to identify specific design points in performance and privacy. Other profiles are possible but are outside the scope of this document.

For opportunistic privacy, analogous to SMTP opportunistic encryption one desires privacy when possible, but does not require it. With opportunistic privacy, a client might acquire a recursive DNS resolver from an untrusted source (such as DHCP while roaming), it might or might not validate the TLS certificate, and it might not use a dummy value for the initial query. These choices maximize availability and performance, but they are vulnerable to on-path attacks. Opportunistic privacy can be used by any current client, but it only provides privacy when there are no on-path attackers.

For pre-deployed privacy, the DNS client has one or more trusted recursive DNS providers. This profile provides strong privacy guarantees to the user. With pre-deployed privacy, a client retains a copy of the TLS certificate and IP address of each provider. The client will only use one of those DNS providers. Because it has a pre-deployed TLS certificate, it may detect person-in-the-middle and downgrade attacks. With pre-deployed privacy, the DNS client MUST signal to the user when none of the designated DNS servers are available, and MUST NOT provide DNS service until one of the designated DNS servers becomes available. The designated DNS provider may be temporarily unavailable when configuring a network. For example, for clients on networks that require authentication through web-based login, such authentication may require DNS interception and spoofing. Techniques such as those used by DNSSEC-trigger MAY be used during network configuration, with the intent to transition to the designated DNS provider after authentication. The user MUST be alerted that the DNS is not private during such bootstrap. Methods for pre-deployment of the designated DNS provider are outside the scope of this document. In corporate settings, such information may be provided at system installation. Use of multiple public DNS providers suggests that end users are able to configure DNS by hand.

DNS-over-TLS incurs additional latency at session startup. It also requires additional state (memory) and increased processing (CPU). Latency: Compared to UDP, DNS-over-TCP requires an additional round-trip-time (RTT) of latency to establish the connection. The TLS handshake adds another two RTTs of latency. Clients and servers should support connection keepalive (reuse) and out-of-order processing to amortize connection setup costs. Moreover, TLS connection resumption can further reduce the setup delay. State: The use of connection-oriented TCP requires keeping additional state in both kernels and applications. TLS has marginal increases in state over TCP alone. The state requirements are of particular concerns on servers with many clients. Smaller timeout values will reduce the number of concurrent connections, and servers can preemptively close connections when resources limits are exceeded. Processing: Use of TLS encryption algorithms results in slightly higher CPU usage. Servers can choose to refuse new DNS-over-TCP clients if processing limits are exceeded. Number of connections: To minimize state on DNS servers and connection startup time, clients SHOULD minimize creation of new TCP connections. Use of a local DNS forwarder allows a single active DNS-over-TLS connection allows a single active TCP connection for DNS per client computer. Additional guidance can be found in . A full performance evaluation is outside the scope of this specification. A more detailed analysis of the performance implications of DNS-over-TLS (and DNS-over-TCP) is discussed in a technical report and .

This document defines a new bit ("TO") in the Flags field of the EDNS0 OPT meta-RR. At the time of approval of this draft in the standards track, as per the IANA Considerations of RFC 6891, IANA is requested to reserve the second leftmost bit of the flags as the TO bit, immediately adjacent to the DNSSEC DO bit, as shown in . IANA is requested add the following value to the "Service Name and Transport Protocol Port Number Registry" registry. That registry is populated by expert review , and such a review will be requested if this document progresses.

The goal of this proposal is to address the security risks that arise because DNS queries may be eavesdropped upon, as described above. There are a number of residual risks that may impact this goal. There are known attacks on TLS, such as person-in-the-middle and protocol downgrade. These are general attacks on TLS and not specific to DNS-over-TLS; please refer to the TLS RFCs for discussion of these security issues. Any protocol interactions prior to the TLS handshake are performed in the clear and can be modified by a man-in-the-middle attacker. For this reason, clients MAY discard cached information about server capabilities advertised prior to the start of the TLS handshake. As with other uses of STARTTLS-upgrade to TLS, the mechanism specified here is susceptible to downgrade attacks, where a person-in-the-middle prevents a successful TLS upgrade. Keeping track of servers known to support TLS (i.e., "pinning") enables clients to detect downgrade attacks. For servers with no connection history, clients may choose to refuse non-TLS DNS, or they may continue without TLS, depending on their privacy requirements. This document does not propose new ideas for certificate authentication for TLS in the context of DNS. Several external methods are possible, although each has weaknesses. The current Certificate Authority infrastructure is used by HTTP/TLS . With many trusted CAs, this approach has recognized weaknesses . Some work is underway to partially address these concerns (for example, with certificate pinning , but more work is needed. DANE provides mechanisms to root certificate trust with DNSSEC. That use here must be carefully evaluated to address potential issues in trust recursion. For stub-to-recursive resolver use, certificate authentication is sometimes either easy or nearly impossible. If the recursive resolver is manually configured, its certificate can be authenticated when it is configured. If the recursive resolver is automatically configured (such as with DHCP ), it could use DHCP authentication mechanisms ). Ongoing discussion and development of opportunistic TLS (connections without CA validation, ) may be relevant to DNS-over-TLS.

The authors would like to thank Stephane Bortzmeyer, Brian Haberman, Kim-Minh Kaplan, Bill Manning, George Michaelson, Eric Osterweil, Glen Wiley, John Dickinson, and Sara Dickinson for reviewing this Internet-draft, and Nikita Somaiya for early work on this idea. Work by Zi Hu, Liang Zhu, and John Heidemann in this paper is partially sponsored by the U.S. Dept. of Homeland Security (DHS) Science and Technology Directorate, HSARPA, Cyber Security Division, BAA 11-01-RIKA and Air Force Research Laboratory, Information Directorate under agreement number FA8750-12-2-0344, and contract number D08PC75599.