A review of implementation DNS over port 80/443

Name servers use port 53, on both UDP and TCP . However, users or operators occasionally find it useful to use an alternative way to deliver DNS information, and often pick port 80 (the default HTTP port) or 443 (the default HTTPS port) for this purpose. There are several use cases: Case 1: Firewalls or other middleboxes may interfere with normal DNS traffic . In addition, some ISPs and hotels block external DNS and perform DNS rewriting to send users to advertising pages, or may use IP addresses which cause misleading geographic location for the user . Users may want DNSSEC support which is not deployed locally in such a case, and so on. Case 2: Users may use DNS over TLS or HTTPS to protect privacy. This also allows the client to authenticate the server. Case 3: Developers may want a higher level DNS API. Web developers may prefer different abstractions or familiar tools like JSON or XML, transmitted using HTTP or HTTPS. This memo does not aim to develop standards or tools. The purpose is to review various implementation options as a reference for developers. However, it may be helpful for anyone hoping to develop specifications for DNS over 80/443 in the future. Note that most of the implementations described in this memo are on port 80/443 and combined with TCP/TLS/HTTP(S). The main focus is between stub resolvers and recursive servers, and the discussion is about the stub resolver to recursive server communication.

The simplest approach is just moving the DNS traffic to port 80 or 443 from 53. This approach serves the requirement use case 1. In this way, the whole protocol is the same as current DNS transport in TCP, except the transport port is moved to port 80 or 443. The difference between port 80 and 443 is that the traffic of port 80 is usually intercepted as HTTP traffic for purposes of deeper inspection, while the traffic of port 443 is usually considered to be encrypted, and typically ignored by middle-boxes. One example where DNS is transported through port 80/443 is one of the fallback cases of NLnetLabs' DNSSEC trigger . Transporting DNS through port 80/443 is an easy implementation. Developers can simply run an existing DNS server and configure the DNS software to listen on ports 80/443. The client can also apply this change without any significant changes. One drawback of this approach is that it might mislead the client because of the port used. For example, clients might think DNS over 443 as a secure protocol because normally the session would be encrypted. In this case, however, it is not.

Another approach is DNS over TLS on port 443, which is also implemented in DNSSEC trigger. DNS over TLS is documented in , which uses the well-known port 853. Using port 443 to carry the traffic instead still serves the purpose in use case 1, as some middle boxes may block traffic on the port 853. also discusses authentication and privacy profiles. TLS provide many benefits for DNS. First, it significantly reduces the DNS conversation's vulnerability to being hijacked. Second, it prevents resolvers from becoming amplifiers of reflection attack. Additionally, it provides privacy by encrypting the conversation between client and server. One concern of DNS over TLS is its cost. Compared to UDP, DNS-over-TCP requires an additional round-trip-time (RTT) of latency to establish a TCP connection. Use of TLS encryption algorithms adds an additional RTT, and results in slightly higher CPU usage. It should also be considered that the DNS packet over TLS on a new port might be dropped by some middle boxes. Another concern of TLS is the deployment difficulty when authenticating the server. If servers are authenticated, certificate management is required.

Different from raw DNS over TCP using port 80/443, another option is encapsulating DNS wire-format data into an HTTP body and sending it as HTTP(S) traffic. It is quiet useful in use cases 1 & 2 described in the introduction. This approach has the benefit that HTTP usually makes it through even the worst coffee shop or hotel room firewalls, as this expected by Internet users. Using HTTP also benefits from HTTP's persistent TCP connection pool concept (see section 6.3 in ), which DNS on TCP port 53 does not have. Note that if HTTP/2 (see ) is used then there may be concurrent streams, and answers on different streams may arrive out-of-order. Finally, as with DNS over TLS, HTTPS provides data integrity and privacy. Use of such encryption is recommended. The basic methodology works as follows: The client creates a DNS query message. The client encapsulates the DNS message in a HTTP(S) message body and assigns parameters with the HTTP header. The client connects to the server and issues an HTTP(S) POST request method. The server decapsulates the HTTP package to DNS query, and resolves the DNS query. The server encapsulates the DNS response in HTTP(S) and sends it back via the HTTP(S) session. Note that if the original DNS query is sent by TCP, first two bits of the package is the message length and should be removed. (This is only true if some software is translating from the DNS protocol to DNS over HTTP, for example via a proxy. Native implementations will of course not need this.) There is an implementation of this methodology in the Go Programming Language (https://github.com/BII-Lab/DNSoverHTTPinGO) as well as C (https://github.com/BII-Lab/DNSoverHTTP), maintained by BII lab. In addition to the benefits mentioned before, the HTTP header makes DNS wire-format over HTTP(S) easy to extend. Compared to creating a new option in EDNS0, using new parameters in HTTP header is far easier to deploy, since DNS messages with EDNS0 may not pass some middle boxes. The DNS wire-format approach has the advantage that any future changes to the DNS protocol will be transparently supported by both client and server, even while continuing to use HTTP. One disadvantage of packaging DNS into HTTP is its cost. Packing and unpacking uses CPU and may result in higher response time. The DNS over HTTP messages also have a risk of being dropped by firewalls which intercepts HTTP packets. And it should be noted that if HTTPS is used, then all the discussion of TLS in previous section is also applicable here.

As mentioned in use case 3, one motivation of a REST HTTP API is for web developers who need to get DNS information but cannot create raw requests (for example JavaScript developers). They can work by creating HTTP requests other than real DNS queries. In this style of implementation DNS data is exchanged in other formats than wire format, like JSON , or XML . There are also lots of REST DNS API developed by DNS service providers. Most of these APIs are developed in the scope of their own system with different specification. But a typical query is a client will requesting a special formatted URI. Usually there is a HTTP(S) server listening to port 80/443, which will parse the request and create a DNS query or DNS operation command towards the real DNS. Unlike wire-format DNS over HTTP(S), once the HTTP(S) server receives the response, it create the response by putting DNS data into various structured formats like JSON, XML, YAML, or even plain text. However, such an approach may have issues, because it is not based on traditional DNS protocol. So there is no guarantee of the protocol's completeness and correctness. Support for DNSSEC might also be a problem because the response usually do not contain RR records with the answer, making it impossible for a client to validate the reply. As with DNS over DNS Wire-format over HTTP, use of encryption is encouraged.