| Network Working Group | M. Thomson |
| Internet-Draft | Mozilla |
| Intended status: Informational | January 02, 2019 |
| Expires: July 6, 2019 |
Long-term Viability of Protocol Extension Mechanisms
draft-thomson-use-it-or-lose-it-03
The ability to change protocols depends on exercising the extension and version negotiation mechanisms that support change. Protocols that don’t use these mechanisms can find that deploying changes can be difficult and costly.
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 July 6, 2019.
Copyright (c) 2019 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.
A successful protocol [SUCCESS] will change in ways that allow it to continue to fulfill the needs of its users. New use cases, conditions and constraints on the deployment of a protocol can render a protocol that does not change obsolete.
Usage patterns and requirements for a protocol shift over time. Protocols can react to these shifts in one of three ways: adjust usage patterns within the constraints of the protocol, extend the protocol, and replace the protocol. These reactions are progressively more disruptive, but are also dictated by the nature of the change in requirements over longer periods.
Experience with Internet-scale protocol deployment shows that changing protocols is not uniformly successful. [TRANSITIONS] examines the problem more broadly.
This document examines the specific conditions that determine whether protocol maintainers have the ability to design and deploy new or modified protocols. Section 4 outlines several strategies that might aid in ensuring that protocol changes remain possible over time.
A change to a protocol can be made extremely difficult to deploy if there are bugs in implementations with which the new deployment needs to interoperate. Bugs in the handling of new codepoints or extensions can mean that instead of handling the mechanism as designed, endpoints react poorly. This can manifest as abrupt termination of sessions, errors, crashes, or disappearances of endpoints and timeouts.
Interoperability with other implementations is usually highly valued, so deploying mechanisms that trigger adverse reactions like these can be untenable. Where interoperability is a competitive advantage, this is true even if the negative reactions happen infrequently or only under relatively rare conditions.
Deploying a change to a protocol could require fixing a substantial proportion of the bugs that the change exposes. This can involve a difficult process that includes identifying the cause of these errors, finding the responsible implementation, coordinating a bug fix and release plan, contacting the operator of affected services, and waiting for the fix to be deployed to those services.
Given the effort involved in fixing these problems, the existence of these sorts of bugs can outright prevent the deployment of some types of protocol changes. It could even be necessary to come up with a new protocol design that uses a different method to achieve the same result.
The set of interoperable features in a protocol is often the subset of its features that have some value to those implementing and deploying the protocol. It is not always the case that future extensibility is in that set.
It is often argued that the design of a protocol extension point or version negotiation capability is critical to the freedom that it ultimately offers.
RFC 6709 [EXTENSIBILITY] contains a great deal of well-considered advice on designing for extension. It includes the following advice:
This has proven to be insufficient in practice. Many protocols have evidence of imperfect implementation of these critical mechanisms. Mechanisms that aren’t used are the ones that fail most often. The same paragraph from RFC 6709 acknowledges the existence of this problem, but does not offer any remedy:
Indeed, basic interoperability is considered critical early in the deployment of a protocol, and any engineering practice that values simplicity will tend to make version negotiation and extension mechanisms optional for this basic interoperability. This leads to these mechanisms being uniquely affected by this problem.
Transport Layer Security (TLS) [TLS12] provides examples of where a design that is objectively sound fails when incorrectly implemented. TLS provides examples of failures in protocol version negotiation and extensibility.
Version negotiation in TLS 1.2 and earlier uses the “Highest mutually supported version (HMSV)” scheme exactly as it is described in [EXTENSIBILITY]. However, clients are unable to advertise a new version without causing a non-trivial proportions of sessions to fail due to bugs in server and middlebox implementations.
Intolerance to new TLS versions is so severe [INTOLERANCE] that TLS 1.3 [TLS13] has abandoned HMSV version negotiation for a new mechanism.
The server name indication (SNI) [TLS-EXT] in TLS is another excellent example of the failure of a well-designed extensibility point. SNI uses the same technique for extension that is used with considerable success in other parts of the TLS protocol. The original design of SNI includes the ability to include multiple names of different types.
What is telling in this case is that SNI was defined with just one type of name: a domain name. No other type has ever been standardized, though several have been proposed. Despite an otherwise exemplary design, SNI is so inconsistently implemented that any hope for using the extension point it defines has been abandoned [SNI].
Even the most superficially simple protocols can often involve more actors than is immediately apparent. A two-party protocol has two ends, but even at the endpoints of an interaction, protocol elements can be passed on to other entities in ways that can affect protocol operation.
One of the key challenges in deploying new features in a protocol is ensuring compatibility with all actors that could influence the outcome.
Protocols deployed without active measures against intermediation will tend to become intermediated over time, as network operators deploy middleboxes to perform some function on traffic. In particular, one of the consequences of an unencrypted protocol is that any element on path can interact with the protocol. For example, HTTP was specifically designed with intermediation in mind, transparent proxies [HTTP] are not only possible but sometimes advantageous, despite some significant downsides. Consequently, transparent proxies for cleartext HTTP are commonplace.
Middleboxes are also protocol participants, to the degree that they are able to observe and act in ways that affect the protocol. The degree to which a middlebox participates varies from the basic functions that a router performs to full participation. For example, a SIP back-to-back user agent (B2BUA) [B2BUA] can be very deeply involved in the SIP protocol.
This phenomenon appears at all layers of the protocol stack, even when protocols are not designed with middlebox participation in mind. TCP’s [TCP] extension points have been rendered difficult to use, largely due to middlebox interactions, as experience with Multipath TCP [MPTCP] has shown. IP’s version field was rendered useless when encapsulated over Ethernet, requring a new ethertype with IPv6 [RFC2462], due in part to layer 2 devices making version-independent assumptions about the structure of the IPv4 header.
By increasing the number of different actors involved in any single protocol exchange, the number of potential implementation bugs that a deployment needs to contend with also increases. In particular, incompatible changes to a protocol that might be negotiated between endpoints in ignorance of the presence of a middlebox can result in a middlebox acting badly.
Thus, middleboxes can increase the difficulty of deploying changes to a protocol considerably.
The design of a protocol for extensibility and eventual replacement [EXTENSIBILITY] does not guarantee the ability to exercise those options. The set of features that enable future evolution need to be interoperable in the first implementations and deployments of the protocol. Active use of mechanisms that support evolution is the only way to ensure that they remain available for new uses.
The conditions for retaining the ability to evolve a design is most clearly evident in the protocols that are known to have viable version negotiation or extension points. The definition of mechanisms alone is insufficient; it’s the active use of those mechanisms that determines the existence of freedom.
For example, header fields in email [SMTP], HTTP [HTTP] and SIP [SIP] all derive from the same basic design. There is no evidence of significant barriers to deploying header fields with new names and semantics in email and HTTP, though the widespread deployment of SIP B2BUAs means that new SIP header fields can be more difficult.
In another example, the attribute-value pairs (AVPs) in Diameter [DIAMETER] are fundamental to the design of the protocol. The definition of new uses of Diameter regularly exercise the ability to add new AVPs and do so with no fear that the new feature might not be successfully deployed.
These examples show extension points that are heavily used also being relatively unaffected by deployment issues preventing addition of new values for new use cases.
These examples also confirm the case that good design is not a prerequisite for success. On the contrary, success is often despite shortcomings in the design. For instance, the shortcomings of HTTP header fields are significant enough that there are ongoing efforts to improve the syntax [HTTP-HEADERS].
Only using a protocol capability is able to ensure availability of that capability. Protocols that fail to use a mechanism, or a protocol that only rarely uses a mechanism, suffer an inability to rely on that mechanism.
The best way to guarantee that a protocol mechanism is used is to make it critical to an endpoint participating in that protocol. This means that implementations rely on both the existence of the protocol mechanism and its use.
For example, the message format in SMTP relies on header fields for most of its functions, including the most basic functions. A deployment of SMTP cannot avoid including an implementation of header field handling. In addition to this, the regularity with which new header fields are defined and used ensures that deployments frequently encounter header fields that it does not understand. An SMTP implementation therefore needs to be able to both process header fields that it understands and ignore those that it does not.
In this way, implementing the extensibility mechanism is not merely mandated by the specification, it is crucial to the functioning of a protocol deployment. Should an implementation fail to correctly implement the mechanism, that failure would quickly become apparent.
Caution is advised to avoid assuming that building a dependency on an extension mechanism is sufficient to ensure availability of that mechanism in the long term. If the set of possible uses is narrowly constrained and deployments do not change over time, implementations might not see new variations or assume a narrower interpretation of what is possible. Those implementations might still exhibit errors when presented with a new variation.
In contrast, there are many examples of extension points in protocols that have been either completely unused, or their use was so infrequent that they could no longer be relied upon to function correctly.
HTTP has a number of very effective extension points in addition to the aforementioned header fields. It also has some examples of extension point that are so rarely used that it is possible that they are not at all usable. Extension points in HTTP that might be unwise to use include the extension point on each chunk in the chunked transfer coding [HTTP], the ability to use transfer codings other than the chunked coding, and the range unit in a range request [HTTP-RANGE].
Even where extension points have multiple valid values, if the set of permitted values does not change over time, there is still a risk that new values are not tolerated by existing implementations. If the set of values for a particular field remains fixed over a long period, some implementations might not correctly handle a new value when it is introduced. For example, implementations of TLS broke when new values of the signature_algorithms extension were introduced.
There are several potential approaches that can provide some measure of protection against a protocol deployment becoming resistant to change.
As discussed in Section 3, the most effective defense against misuse of protocol extension points is active use.
Implementations are most likely to be tolerant of new values if they depend on being able to use new values. Failing that, implementations that routinely see new values are more likely to correctly handle new values. More frequent changes will improve the likelihood that incorrect handling or intolerance is discovered and rectified. The longer an intolerant implementation is deployed, the more difficult it is to correct.
What active use means could depend greatly on the environment in which a protocol is deployed. The frequency of changes necessary to safeguard some mechanisms might be slow enough to attract ossification in another protocol deployment, while being excessive in others. There are currently no firm guidelines for new protocol development, as much is being learned about what techniques are most effective.
“Grease” [GREASE] identifies lack of use as an issue (protocol mechanisms “rusting” shut) and proposes a system of use that exercises extension points by using dummy values.
The primary feature of the grease design is aimed at the style of negotiation most used in TLS, where the client offers a set of options and the server chooses the one that it most prefers from those that it supports. A client that uses grease randomly offers options - usually just one - from a set of reserved values. These values are guaranteed to never be assigned real meaning, so the server will never have cause to genuinely select one of these values.
The principle that grease operates on is that an implementation that is regularly exposed to unknown values is not likely to become intolerant of new values when they appear. This depends largely on the assumption that the difficulty of implementing the protocol mechanism correctly is not significantly more effort than implementing code to specifically filter out the randomized grease values.
To avoid simple techniques for filtering greasing codepoints, grease values are not reserved from a single contiguous block of code points, but are distributed evenly across the entire space of code points. Reserving a randomly selected set of code points has a greater chance of avoiding this problem, though it might be more difficult to specify and implement, especially over larger code point spaces.
Without reserved greasing codepoints, an implementation can use code points from spaces used for private or experimental use if such a range exists. In addition to the risk of triggering participation in an unwanted experiment, this can be less effective. Incorrect implementations might still be able to correctly identify these code points and ignore them.
Grease is deployed with the intent of quickly detecting errors in implementing the mechanisms it safeguards. Though it has been effective at revealing problems in some cases with TLS, its efficacy isn’t proven more generally.
This style of defensive design has some limitations. It does not necessarily create the need for an implementation to rely on the mechanism it safeguards; that is determined by the underlying protocol itself. More critically, it does not easily translate to other forms of extension point. For instance, HMSV negotiation cannot be greased in this fashion. Other techniques might be necessary for protocols that don’t rely on the particular style of exchange that is predominant in TLS.
Cryptography can be used to reduce the number of entities that can participate in a protocol. Using tools like TLS ensures that only authorized participants are able to influence whether a new protocol feature is used.
Permitting fewer protocol participants reduces the number of implementations that can prevent a new mechanism from being deployed. As recommended in [PATH-SIGNALS], use of encryption and integrity protection can be used to limit participation.
For example, the QUIC protocol [QUIC] adopts both encryption and integrity protection. Encryption is used to carefully control what information is exposed to middleboxes. For those fields that are not encrypted, QUIC uses integrity protection to prevent modification.
While not a direct means of protecting extensibility mechanisms, feedback systems can be important to discovering problems.
Visibility of errors is critical to the success of the grease technique (see Section 4.2). The grease design is most effective if a deployment has a means of detecting and reporting errors. Ignoring errors could allow problems to become entrenched.
Feedback on errors is more important during the development and early deployment of a change. It might also be helpful to disable automatic error recovery methods during development.
Automated feedback systems are important for automated systems, or where error recovery is also automated. For instance, connection failures with HTTP alternative services [ALT-SVC] are not permitted to affect the outcome of transactions. An automated feedback system for capturing failures in alternative services is therefore necessary for failures to be detected.
The ability to design, implement, and deploy new protocol mechanisms can be critical to security. In particular, it is important to be able to replace cryptographic algorithms over time [AGILITY]. For example, preparing for replacement of weak hash algorithms was made more difficult through misuse [HASH].
This document makes no request of IANA.
Mirja Kühlewind, Mark Nottingham, and Brian Trammell made significant contributions to this document.