Proxying UDP in HTTP

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 19 December 2022.¶

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

While HTTP provides the CONNECT method (see Section 9.3.6 of [HTTP]) for creating a TCP [TCP] tunnel to a proxy, prior to this specification it lacked a method for doing so for UDP [UDP] traffic.¶

This document describes a protocol for tunnelling UDP to a server acting as a UDP-specific proxy over HTTP. UDP tunnels are commonly used to create an end-to-end virtual connection, which can then be secured using QUIC [QUIC] or another protocol running over UDP. Unlike CONNECT, the UDP proxy itself is identified with an absolute URL containing the traffic's destination. Clients generate those URLs using a URI Template [TEMPLATE], as described in Section 2.¶

This protocol supports all existing versions of HTTP by using HTTP Datagrams [HTTP-DGRAM]. When using HTTP/2 [HTTP/2] or HTTP/3 [HTTP/3], it uses HTTP Extended CONNECT as described in [EXT-CONNECT2] and [EXT-CONNECT3]. When using HTTP/1.x [HTTP/1.1], it uses HTTP Upgrade as defined in Section 7.8 of [HTTP].¶

2. Client Configuration

HTTP clients are configured to use a UDP proxy with a URI Template [TEMPLATE] that has the variables "target_host" and "target_port". Examples are shown below:¶

https://example.org/.well-known/masque/udp/{target_host}/{target_port}/
https://proxy.example.org:4443/masque?h={target_host}&p={target_port}
https://proxy.example.org:4443/masque{?target_host,target_port}

The following requirements apply to the URI Template:¶

• The URI Template MUST be a level 3 template or lower.¶
• The URI Template MUST be in absolute form, and MUST include non-empty scheme, authority and path components.¶
• The path component of the URI Template MUST start with a slash "/".¶
• All template variables MUST be within the path or query components of the URI.¶
• The URI template MUST contain the two variables "target_host" and "target_port" and MAY contain other variables.¶
• The URI Template MUST NOT contain any non-ASCII unicode characters and MUST only contain ASCII characters in the range 0x21-0x7E inclusive (note that percent-encoding is allowed; see Section 2.1 of [URI]).¶
• The URI Template MUST NOT use Reserved Expansion ("+" operator), Fragment Expansion ("#" operator), Label Expansion with Dot-Prefix, Path Segment Expansion with Slash-Prefix, nor Path-Style Parameter Expansion with Semicolon-Prefix.¶

Clients SHOULD validate the requirements above; however, clients MAY use a general-purpose URI Template implementation that lacks this specific validation. If a client detects that any of the requirements above are not met by a URI Template, the client MUST reject its configuration and abort the request without sending it to the UDP proxy.¶

Since the original HTTP CONNECT method allowed conveying the target host and port but not the scheme, proxy authority, path, nor query, there exist clients with proxy configuration interfaces that only allow the user to configure the proxy host and the proxy port. Client implementations of this specification that are constrained by such limitations MAY attempt to access UDP proxying capabilities using the default template, which is defined as: "https://$PROXY_HOST:$PROXY_PORT/.well-known/masque/udp/{target_host}/{target_port}/" where $PROXY_HOST and $PROXY_PORT are the configured host and port of the UDP proxy respectively. UDP proxy deployments SHOULD offer service at this location if they need to interoperate with such clients.¶

3. Tunnelling UDP over HTTP

To allow negotiation of a tunnel for UDP over HTTP, this document defines the "connect-udp" HTTP Upgrade Token. The resulting UDP tunnels use the Capsule Protocol (see Section 3.2 of [HTTP-DGRAM]) with HTTP Datagrams in the format defined in Section 5.¶

To initiate a UDP tunnel associated with a single HTTP stream, a client issues a request containing the "connect-udp" upgrade token. The target of the tunnel is indicated by the client to the UDP proxy via the "target_host" and "target_port" variables of the URI Template, see Section 2. If the request is successful, the UDP proxy commits to converting received HTTP Datagrams into UDP packets and vice versa until the tunnel is closed.¶

When sending its UDP proxying request, the client SHALL perform URI Template expansion to determine the path and query of its request. target_host supports using DNS names, IPv6 literals and IPv4 literals. Note that IPv6 scoped addressing zone identifiers are not supported. Also note that this URI Template expansion requires using percent-encoding, so for example if the target_host is "2001:db8::42", it will be encoded in the URI as "2001%3Adb8%3A%3A42".¶

By virtue of the definition of the Capsule Protocol (see Section 3.2 of [HTTP-DGRAM]), UDP proxying requests do not carry any message content. Similarly, successful UDP proxying responses also do not carry any message content.¶

GET https://example.org/.well-known/masque/udp/192.0.2.6/443/ HTTP/1.1
Host: example.org
Connection: Upgrade
Upgrade: connect-udp
Capsule-Protocol: ?1

HTTP/1.1 101 Switching Protocols
Connection: Upgrade
Upgrade: connect-udp
Capsule-Protocol: ?1

HEADERS
:method = CONNECT
:protocol = connect-udp
:scheme = https
:path = /.well-known/masque/udp/192.0.2.6/443/
:authority = example.org
capsule-protocol = ?1

HEADERS
:status = 200
capsule-protocol = ?1

4. Context Identifiers

The mechanism for proxying UDP in HTTP defined in this document allows future extensions to exchange HTTP Datagrams that carry different semantics from UDP payloads. Some of these extensions can augment UDP payloads with additional data, while others can exchange data that is completely separate from UDP payloads. In order to accomplish this, all HTTP Datagrams associated with UDP Proxying request streams start with a context ID, see Section 5.¶

Context IDs are 62-bit integers (0 to 2⁶²-1). Context IDs are encoded as variable-length integers, see Section 16 of [QUIC]. The context ID value of 0 is reserved for UDP payloads, while non-zero values are dynamically allocated: non-zero even-numbered context IDs are client-allocated, and odd-numbered context IDs are proxy-allocated. The context ID namespace is tied to a given HTTP request: it is possible for a context ID with the same numeric value to be simultaneously allocated in distinct requests, potentially with different semantics. Context IDs MUST NOT be re-allocated within a given HTTP namespace but MAY be allocated in any order. The context ID allocation restrictions to the use of even-numbered and odd-numbered context IDs exist in order to avoid the need for synchronisation between endpoints. However, once a context ID has been allocated, those restrictions do not apply to the use of the context ID: it can be used by any client or UDP proxy, independent of which endpoint initially allocated it.¶

Registration is the action by which an endpoint informs its peer of the semantics and format of a given context ID. This document does not define how registration occurs. Future extensions MAY use HTTP header fields or capsules to register contexts. Depending on the method being used, it is possible for datagrams to be received with Context IDs which have not yet been registered, for instance due to reordering of the packet containing the datagram and the packet containing the registration message during transmission.¶

5. HTTP Datagram Payload Format

When HTTP Datagrams (see Section 2 of [HTTP-DGRAM]) are associated with UDP proxying request streams, the HTTP Datagram Payload field has the format defined in Figure 6. Note that when HTTP Datagrams are encoded using QUIC DATAGRAM frames [DGRAM], the Context ID field defined below directly follows the Quarter Stream ID field which is at the start of the QUIC DATAGRAM frame payload (see Section 2.1 of [HTTP-DGRAM]).¶

UDP Proxying HTTP Datagram Payload {
  Context ID (i),
  Payload (..),
}

Context ID:

A variable-length integer (see Section 16 of [QUIC]) that contains the value of the Context ID. If an HTTP/3 Datagram which carries an unknown Context ID is received, the receiver SHALL either drop that datagram silently or buffer it temporarily (on the order of a round trip) while awaiting the registration of the corresponding Context ID.¶

Payload:

The payload of the datagram, whose semantics depend on value of the previous field. Note that this field can be empty.¶

UDP packets are encoded using HTTP Datagrams with the Context ID set to zero. When the Context ID is set to zero, the Payload field contains the unmodified payload of a UDP packet (referred to as "data octets" in [UDP]).¶

By virtue of the definition of the UDP header [UDP], it is not possible to encode UDP payloads longer than 65527 bytes. Therefore, endpoints MUST NOT send HTTP Datagrams with a Payload field longer than 65527 using Context ID zero. An endpoint that receives an HTTP Datagram using Context ID zero whose Payload field is longer than 65527 MUST abort the corresponding stream. If a UDP proxy knows it can only send out UDP packets of a certain length due to its underlying link MTU, it has no choice but to discard incoming HTTP Datagrams using Context ID zero whose Payload field is longer than that limit. If the discarded HTTP Datagram was transported by a DATAGRAM capsule, the receiver SHOULD discard that capsule without buffering the capsule contents.¶

If a UDP proxy receives an HTTP Datagram before it has received the corresponding request, it SHALL either drop that HTTP Datagram silently or buffer it temporarily (on the order of a round trip) while awaiting the corresponding request.¶

Note that buffering datagrams (either because the request was not yet received, or because the Context ID is not yet known) consumes resources. Receivers that buffer datagrams SHOULD apply buffering limits in order to reduce the risk of resource exhaustion occuring. For example, receivers can limit the total number of buffered datagrams, or the cumulative size of buffered datagrams, on a per-stream, per-context, or per-connection basis.¶

A client MAY optimistically start sending UDP packets in HTTP Datagrams before receiving the response to its UDP proxying request. However, implementers should note that such proxied packets may not be processed by the UDP proxy if it responds to the request with a failure, or if the proxied packets are received by the UDP proxy before the request and the UDP proxy chooses to not buffer them.¶

6. Performance Considerations

Bursty traffic can often lead to temporally correlated packet losses, which in turn can lead to suboptimal responses from congestion controllers in protocols running over UDP. To avoid this, UDP proxies SHOULD strive to avoid increasing burstiness of UDP traffic: they SHOULD NOT queue packets in order to increase batching.¶

When the protocol running over UDP that is being proxied uses congestion control (e.g., [QUIC]), the proxied traffic will incur at least two nested congestion controllers. The underlying HTTP connection MUST NOT disable congestion control unless it has an out-of-band way of knowing with absolute certainty that the inner traffic is congestion-controlled.¶

If a client or UDP proxy with a connection containing a UDP proxying request stream disables congestion control, it MUST NOT signal Explicit Congestion Notification (ECN) [ECN] support on that connection. That is, it MUST mark all IP headers with the Not-ECT codepoint. It MAY continue to report ECN feedback via QUIC ACK_ECN frames or the TCP "ECE" bit, as the peer may not have disabled congestion control.¶

When the protocol running over UDP that is being proxied uses loss recovery (e.g., [QUIC]), and the underlying HTTP connection runs over TCP, the proxied traffic will incur at least two nested loss recovery mechanisms. This can reduce performance as both can sometimes independently retransmit the same data. To avoid this, UDP proxying SHOULD be performed over HTTP/3 to allow leveraging the QUIC DATAGRAM frame.¶

7. Security Considerations

There are significant risks in allowing arbitrary clients to establish a tunnel to arbitrary targets, as that could allow bad actors to send traffic and have it attributed to the UDP proxy. HTTP servers that support UDP proxying ought to restrict its use to authenticated users.¶

There exist software and network deployments that perform access control checks based on the source IP address of incoming requests. For example, some software allows unauthenticated configuration changes if they originated from 127.0.0.1. Such software could be running on the same host as the UDP proxy, or in the same broadcast domain. Proxied UDP traffic would then be received with a source IP address belonging to the UDP proxy. If this source address is used for access control, UDP proxying clients could use the UDP proxy to escalate their access privileges beyond those they might otherwise have. This could lead to unauthorized access by UDP proxying clients unless the UDP proxy disallows UDP proxying requests to vulnerable targets, such as the UDP proxy's own addresses and localhost, link-local, multicast, and broadcast addresses. UDP proxies can use the destination_ip_prohibited Proxy Error Type from Section 2.3.5 of [PROXY-STATUS] when rejecting such requests.¶

UDP proxies share many similarities to TCP CONNECT proxies when considering them as infrastructure for abuse to enable denial of service attacks. Both can obfuscate the attacker's source address from the attack target. In the case of a stateless volumetric attack (e.g., a TCP SYN flood or a UDP flood), both types of proxies pass the traffic to the target host. With stateful volumetric attacks (e.g., HTTP flooding) being sent over a TCP CONNECT proxy, the proxy will only send data if the target has indicated its willingness to accept data by responding with a TCP SYN-ACK. Once the path to the target is flooded, the TCP CONNECT proxy will no longer receive replies from the target and will stop sending data. Since UDP does not establish shared state between the UDP proxy and the target, the UDP proxy could continue sending data to the target in such a situation. While a UDP proxy could potentially limit the number of UDP packets it is willing to forward until it has observed a response from the target, that provides limited protection against denial of service attacks when attacks target open UDP ports where the protocol running over UDP would respond, and that would be interpreted as willingness to accept UDP by the UDP proxy. Such a packet limit could also cause issues for valid traffic.¶

The security considerations described in Section 4 of [HTTP-DGRAM] also apply here. Since it is possible to tunnel IP packets over UDP, the guidance in [TUNNEL-SECURITY] can apply.¶

8. IANA Considerations

Acknowledgments

This document is a product of the MASQUE Working Group, and the author thanks all MASQUE enthusiasts for their contibutions. This proposal was inspired directly or indirectly by prior work from many people, in particular HELIUM by Ben Schwartz, HiNT by Lucas Pardue, and the original MASQUE Protocol by the author of this document.¶

The author would like to thank Eric Rescorla for suggesting the use of an HTTP method to proxy UDP. The author is indebted to Mark Nottingham and Lucas Pardue for the many improvements they contributed to this document. The extensibility design in this document came out of the HTTP Datagrams Design Team, whose members were Alan Frindell, Alex Chernyakhovsky, Ben Schwartz, Eric Rescorla, Lucas Pardue, Marcus Ihlar, Martin Thomson, Mike Bishop, Tommy Pauly, Victor Vasiliev, and the author of this document.¶

Author's Address