Secure Negotiation of Incompatible Protocols in TLS

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶

This Internet-Draft will expire on 26 February 2022.¶

Copyright Notice

Copyright (c) 2021 IETF Trust and the persons identified as the document authors. All rights reserved.¶

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.¶

Table of Contents

1. Introduction

With increased diversity in protocol choice, some applications are able to use one of several semantically-equivalent protocols to achieve their goals. This is particularly notable in HTTP where there are currently three distinct protocols: HTTP/1.1 [HTTP11], HTTP/2 [HTTP2], and HTTP/3 [HTTP3]. This is also true of protocols that support variants based on both TLS [TLS] and DTLS [DTLS].¶

For protocols that are mutually compatible, Application-Layer Protocol Negotiation (ALPN; [ALPN]) provides a secure way to negotiate protocol selection.¶

In ALPN, the client offers a list of options in a TLS ClientHello and the server chooses the option that it most prefers. A downgrade attack occurs where both client and server support a protocol that the server prefers more than than the selected protocol. ALPN protects against this attack by ensuring that the server is aware of all options the client supports and including those options and the server choice under the integrity protection provided by the TLS handshake.¶

This downgrade protection functions because protocol negotiation is part of the TLS handshake. The introduction of semantically-equivalent protocols that use incompatible handshakes introduces new opportunities for downgrade attack. For instance, it is not possible to negotiate the use of HTTP/2 based on an attempt to connect using HTTP/3. The former relies on TCP, whereas the latter uses UDP. These protocols are therefore mutually incompatible.¶

This document defines an extension to TLS that allows clients to discover when servers support alternative protocols that are incompatible with the currently-selected TLS version. This might be used to avoid downgrade attack caused by interference in protocol discovery mechanisms.¶

This extension is motivated by the addition of new mechanisms, such as [SVCB]. SVCB enables the discovery of servers that support multiple different protocols, some of which are incompatible. The extension can also be used to authenticate protocol choices that are discovered by other means.¶

2. Terminology

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.¶

Two protocols are consider "compatible" if it is possible to negotiate either using the same connection attempt. In comparison, protocols are "incompatible" if they require separate attempts to establish a connection.¶

3. Incompatible Protocols and SVCB

The SVCB record [SVCB] allows a client to learn about services associated with a domain name. This includes how to locate a server, along with supplementary information about the server, including protocols that the server supports. This allows a client to start using a protocol of their choice without added latency, as the lookup can be performed concurrently with other name resolution. The added cost of the additional DNS queries is minimal.¶

However, SVCB provides no protection against a downgrade attack between incompatible protocols. An attacker could remove DNS records for client-preferred protocols, leaving the client to believe that only less-preferred options are available. If those options are not compatible with the client-preferred option, the client will not know to attempt these. The client then only offers options compatible with the less-preferred options when attempting a TLS handshake. Even if a client were to inform the server that it supports a more preferred protocol, the server would not be able to act upon it.¶

Authenticating all of the information presented in SVCB records might provide clients with complete information about server support, but this is impractical for several reasons:¶

• it is not possible to ensure that all server instances in a deployment have the same protocol configuration, as deployments for a single name routinely include multiple providers that cannot coordinate closely;¶
• the ability to provide a subset of valid DNS records is integral to many strategies for managing servers; and¶
• it is difficult to ensure that cached DNS records are synchronized with server state.¶

Overall, an authenticated TLS handshake is a better source of authoritative information about the protocols that are supported by servers.¶

4. Authenticating Incompatible Protocols

The incompatible_protocols(TBD) TLS extension provides clients with information about the incompatible protocols that are supported by the same logical server; see Section 6 for a definition of a logical server.¶

enum {
    incompatible_protocols(TBD), (65535)
} ExtensionType;

A client that supports the extension advertises an empty extension. In response, a server that supports this extension includes a list of application protocol identifiers. The "extension_data" field of the value server extension uses the ProtocolName type defined in [ALPN], which is repeated here. This syntax is shown in Figure 1.¶

opaque ProtocolName<1..2^8-1>;
ProtocolName IncompatibleProtocol;

struct {
  select (Handshake.msg_type) {
    case client_hello:
      Empty;
    case encrypted_extensions:
      IncompatibleProtocol incompatible_protocols<3..2^16-1>;
  };
} IncompatibleProtocols;

This extension only applies to the ClientHello and EncryptedExtensions messages. An implementation that receives this extension in any other handshake message MUST send a fatal illegal_parameter alert.¶

A server deployment that supports multiple incompatible protocols MAY advertise all protocols that are supported by the same logical server. A server needs to ensure that protocols advertised in this fashion are available to the client.¶

A server MUST omit any compatible protocols from this extension. That is, any protocol that the server might be able to select, had the client offered the protocol in the application_layer_protocol_negotiation extension. In comparison, clients are expected to include all compatible protocols in the application_layer_protocol_negotiation extension.¶

5. Incompatible Protocol Selection

This document expands the definition of protocol negotiation to include both compatible and incompatible protocols and provide protection against downgrade for both types of selection. ALPN [ALPN] only considers compatible protocols: the client presents a set of compatible options and the server chooses its most preferred.¶

With an selection of protocols that includes incompatible options, the client makes a selection between incompatible options before making a connection attempt. Therefore, this design does not enable negotiation, it instead provides the client with information about other incompatible protocols that the server might support.¶

Detecting a potential downgrade between incompatible protocols does not automatically imply that a client abandon a connection attempt. It only provides the client with authenticated information about its options. What a client does with this information is left to client policy.¶

In brief:¶

• For compatible protocols, the client offers all acceptable options and the server selects its most preferred¶
• For incompatible protocols, information the server offers is authenticated and the client is able to act on that¶

For a protocol like HTTP/3, this might not result in the client choosing to use HTTP/3, even if HTTP/3 is preferred and the server indicates that a service endpoint supporting HTTP/3 is available. Blocking of UDP or QUIC is known to be widespread. As a result, clients might adopt a policy of tolerating a downgrade to a TCP-based protocol, even if HTTP/3 were preferred. However, as blocking of UDP is highly correlated by access network, clients that are able to establish HTTP/3 connections to some servers might choose to apply a stricter policy when a server that indicates HTTP/3 support is unreachable.¶

6. Logical Servers

The set of endpoints over which clients can assume availability of incompatible protocols is the set of endpoints that share an IP version, IP address, and port number with the TLS server that provides the incompatible_protocols extension.¶

This definition includes a port number that is independent of the protocol that is used. Any protocol that defines a port number is considered to be equivalent. In particular, incompatible protocols can be deployed to TCP, UDP, SCTP, or DCCP ports as long as the IP address and port number is the same.¶

This determination is made from the perspective of a client. This means that servers need to be aware of all instances that might answer to the same IP address and port; see Section 7.¶

7. Operational Considerations

By listing incompatible protocols a server needs reliable knowledge of the existence of these alternatives. This depends on some coordination of deployments. In particular, coordination is important if a load balancer distributes load for a single IP address to multiple server instances. Ensuring consistent configuration of servers could present operational difficulties as it requires that incompatible protocols are only listed when those protocols are deployed across all server instances.¶

Server deployments can choose not to provide information about incompatible protocols, which denies clients information about downgrade attacks but might avoid the operational complexity of providing accurate information.¶

During rollout of a new, incompatible protocol, until the deployment is stable and not at risk of being disabled, servers SHOULD NOT advertise the existence of the new protocol. Protocol deployments that are disabled, first need to be removed from the incompatible_protocols extension or there could be some loss of service. Though the incompatible_protocols extension only applies at the time of the TLS handshake, clients might take some time to act on the information. If an incompatible protocol is removed from deployment between when the client completes a handshake and when it acts, this could be treated as an error by the client.¶

If a server does not list incompatible protocols, clients cannot learn about other services and so cannot detect downgrade attacks against those protocols.¶

8. Security Considerations

This design depends on the integrity of the TLS handshake across all forms, including TLS [RFC8446], DTLS [DTLS], and QUIC [QUIC-TLS]. An attacker that can modify a TLS handshake in any one of these protocols can cause a client to believe that other options do not exist.¶

9. IANA Considerations

TODO: register the extension¶

Appendix A. Acknowledgments

Benjamin Schwartz provided significant input into the design of the mechanism and helped simplify the overall design.¶

Appendix B. Defining Logical Servers

As incompatible protocols use different protocol stacks, they also use different endpoints. In other words, it is in many cases impossible for the exactly same endpoint to support multiple incompatible protocols. Thus, it is necessary to understand the set of endpoints at a server that offer the incompatible protocols.¶

A number of choices are possible here:¶

• The set of endpoints that are authoritative for the same domain name.¶
• The set of endpoints that are authoritative for the same "authority" as defined in RFC 3986 [URI], which is in effect domain name plus port number.¶
• The set of endpoints that are referenced by the same SVCB ServiceMode record.¶
• The set of endpoints that share an IP address.¶
• The set of endpoints that share an IP address and port number.¶

The challenge with options based on domain name is that it might prevent the use of multiple service providers. This is a common practice for HTTP, where the same domain name can be operated by multiple CDN operators.¶

Having multiple service operators also rules out using SVCB ServiceMode records also as different records might be used to identify different operators.¶

Hosts on the same IP address might work, but common deployment practices include use of different ports for entirely different services, which can have different operational constraints such as deployment schedules. Including different ports in the same scope could force all services on the same host to support a consistent set of protocols.¶

This leaves IP and port. There is always a risk that the same port number is used for completely different purposes depending on the choice of protocol, but this practice is sufficiently rare that it is not anticipated to be a problem.¶

Author's Address