Internet Architecture Board (IAB) D. Thaler Request for Comments: 7288 Microsoft Category: Informational June 2014 ISSN: 2070-1721
Reflections on Host Firewalls
In today's Internet, the need for firewalls is generally accepted in the industry, and indeed firewalls are widely deployed in practice. Unlike traditional firewalls that protect network links, host firewalls run in end-user systems. Often the result is that software may be running and potentially consuming resources, but then communication is blocked by a host firewall. It's taken for granted that this end state is either desirable or the best that can be achieved in practice, rather than (for example) an end state where the relevant software is not running or is running in a way that would not result in unwanted communication. In this document, we explore the issues behind these assumptions and provide suggestions on improving the architecture going forward.
Status of This Memo
This document is not an Internet Standards Track specification; it is published for informational purposes.
This document is a product of the Internet Architecture Board (IAB) and represents information that the IAB has deemed valuable to provide for permanent record. It represents the consensus of the Internet Architecture Board (IAB). Documents approved for publication by the IAB are not a candidate for any level of Internet Standard; see Section 2 of RFC 5741.
Copyright (c) 2014 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 (http://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.
[BLOCK-FILTER] discusses the issue of blocking or filtering abusive or objectionable content and communications, and the effects on the overall Internet architecture. This document complements that discussion by focusing on the architectural effects of host firewalls on hosts and applications.
"Behavior of and Requirements for Internet Firewalls" [RFC2979] provides an introduction to firewalls and the requirement for transparency in particular, stating:
The introduction of a firewall and any associated tunneling or access negotiation facilities MUST NOT cause unintended failures of legitimate and standards-compliant usage that would work were the firewall not present.
Many firewalls today do not follow that guidance, such as by blocking traffic containing IP options or IPv6 extension headers (see [RFC7045] for more discussion).
In Section 2.1 of "Reflections on Internet Transparency" [RFC4924], the IAB provided additional thoughts on firewalls and their impact on the Internet architecture, including issues around disclosure obligations and complexity as applications evolve to circumvent firewalls. This document extends that discussion with additional considerations.
Traditionally, firewalls have been about arming customers to protect against bugs in applications and services. This document discusses a number of fundamental questions, such as who a firewall is meant to protect from what. It does so primarily, though not exclusively, from an end system perspective with a focus on host firewalls in particular.
While the Internet Security Glossary [RFC4949] contains an extended definition of a firewall, informally, most people would tend to think of a firewall as simply "something that blocks unwanted traffic" (see [RFC4948] for a discussion on many types of unwanted traffic). A fundamental question is, however: "unwanted by whom?"
Possible answers include end users, application developers, network administrators, host administrators, firewall vendors, and content providers. We will exclude by definition the sender of the traffic in question, since the sender would generally want such traffic to be delivered. Still, the other entities have different, and often conflicting, desires which means that a type of traffic might be
Thaler Informational [Page 3]
RFC 7288 Host Firewalls June 2014
wanted by one entity and unwanted by another entity. Thus, not surprisingly, there exist various types of firewalls, and various types of "arms race" as we will discuss in Section 4.1.2.
In this document we distinguish between a "host firewall", which simply intends to protect the single computer on which it runs, and a "network firewall", which is located in the network and intends to protect the network and any hosts behind it.
A Network Address Translator (NAT) is also often viewed, or even marketed, as a type of network firewall; Section 2.2 of [RFC4864] addresses this misconception, but nevertheless some of the same observations in the present document may also apply to NATs.
Sandboxed environments, such as those provided by browsers, can also be thought of as a type of host firewall in the more general sense. For example, a cross-site check in a browser can be thought of as a mechanism to block unwanted outbound traffic per a "same origin policy" where a script can only communicate with the "site" from which the script was obtained, for some definition of site such as the scheme and authority in a URI.
The term "application" is used in this document generically to apply to any component that can receive traffic. In this sense, it could refer to a process running on a computer (including a system service) or even to a portion of a TCP/IP stack itself, such as a component that responds to pings.
Desires for wanted or unwanted traffic can be expressed in terms of "allow" vs. "block" rules, with some way to resolve conflicting rules. Many firewalls are actually implemented in terms of such rules. Figure 1 shows some typical sources of such rules.
Figure 1 assumes that network firewalls are administered by network administrators (if any), and host firewalls are administered by host administrators (if any). A firewall may also have rules provided by the firewall vendor itself.
End users typically cannot directly provide rules to firewalls that affect other users, unless the end user is a host or network administrator. Application developers can, however, provide such rules to some firewalls, such as providing rules at installation time. They can do this, for example, by invoking an API provided by a host firewall included with the operating system, or by providing metadata to the operating system for use by firewalls, or by using a protocol such as Universal Plug and Play (UPnP) [UPNPWANIP] or the Port Control Protocol (PCP) [RFC6887] to communicate with a network firewall (see Section 4.1.3 for a longer discussion).
Thaler Informational [Page 5]
RFC 7288 Host Firewalls June 2014
Firewall rules generally fall into two categories:
1. Attack surface reduction: Rules intended to prevent an application from doing things the developer does not want it to do.
2. Security policy: Rules intended to prevent an application from doing things the developer might want it to do, but an administrator does not.
A firewall is unnecessary if both categories are empty. We will now treat each category in turn, focusing specifically on host firewalls (although some points might be equally applicable to network firewalls).
As noted above, this category of firewall rule typically attempts to prevent applications from doing things the developer did not intend.
One might ask whether this category of rules is typically empty, and the answer is that it is not, at present. One reason stems from mitigating the threat of vulnerability exploitation by putting a security barrier in a separate process, isolated from the potentially compromised process. Furthermore, there is also some desire for a "stealth mode" (see Section 5 below).
Hence, typically a firewall will have rules to block everything by default. A one-time, privileged, application-install step adds one or more Allow rules, and then normal (unprivileged) application execution is then constrained by the resulting rules.
A second reason this category of rules is non-empty is where they are used as workarounds for cases the application developer found too onerous to implement. These cases include:
1. Simple policies that the developer would want but that are difficult to implement. One example might be a policy that an application should communicate only within the local network (e.g., a home or enterprise), but not be reachable from the global Internet or while the device is moved to some public network such as a hotspot. A second example might be the reverse, i.e., a policy to communicate over the Internet but not with local entities. The need for this category would be reduced by better platform support for such policies, making them easier for developers to implement and use.
Thaler Informational [Page 6]
RFC 7288 Host Firewalls June 2014
2. Complex policies where the developer cannot possibly be aware of specifics. One example might be a policy to communicate only during, or only outside of, normal business hours, where the exact hours may vary by location and time of year. Another example might be a policy to avoid communication over links that cost too much, where the definition of "too much" may vary by customer, and indeed, the end host and application might not even be aware of the costs. The need for this category would be reduced by better platform support for such policies, allowing the application to communicate through some simple API with some other library or service that can deal with the specifics.
Firewall vendors point out that one can more quickly and reliably update firewall rules than application software. Indeed, some applications might have no way to update them, and support for other applications might no longer be available (e.g., if the developers are no longer around). Most modern operating systems (and any applications that come with them) have automatic updates, as do some independent applications. But until all applications have automatic updates, and automatic updates are actually used, it will still be the case that firewall rules can be updated more quickly than software patches. Furthermore, in some contexts (e.g., within some enterprises), a possibly lengthy retesting and recertification process must be employed before applications can be updated.
In short, mechanisms to encourage and ease the use of secure automatic software updates are important and would greatly reduce overall complexity. Such mechanisms should allow scheduling updates at appropriate times, taking into account operational considerations such as dependencies, compatibility, testing and maintenance windows.
A key question to ask is whether the application could still do something useful when firewalled. If the answer is yes, then not using the software is probably unrealistic. For example, a game with both single-player and multi-player capabilities could still be useful in single-player mode when firewalled. If instead the answer is no, it is better to not allow the application to run in the first place, and some host firewalls can indeed block applications from running.
As noted earlier, one disadvantage of this approach is that resources still get consumed. For example, the application can still consume memory, CPU, bandwidth (up to the point of blockage), ports in the transport layer protocol, and possibly other resources depending on the application, for operations that provide no benefit while firewalled.
A second important disadvantage of this approach is the bad user experience. Typically the application and the end-user won't know why the application doesn't work. A poorly designed application might not cope well and consume even more resources (e.g., retrying an operation that continually fails).
A third disadvantage is that it is common for a firewall rule to block more that is appropriate for attack surface reduction, impacting protocol operation and even having adverse effects on other endpoints. For example, some firewalls that cannot perform full deep packet inspection at line speed have adopted a block by default approach to anything they don't understand from the first few bytes; this is very harmful to innovation as it interferes with the ability to deploy new protocols and features.
As another example, blocking ICMP adversely affects path MTU discovery which can cause problems for other entities (see [RFC4890] and Section 3.1.1 of [RFC2979] for further discussion). This can happen due to lack of understanding all the details of application behavior, or due to accidental misconfiguration. Section 2.1 of [RFC5505] states, "Anything that can be configured can be misconfigured," and discusses this in more detail.
In short, it is important to make applications more aware of the constraints of their environment, and hence better able to behave well when constrained.
As noted in Section 2, this category of firewall rule typically attempts to prevent an application from doing things an administrator does not want them to do, even if the application developer did.
One might ask whether this category of rules is typically empty, and the answer varies somewhat. For enterprise-scenario firewalls, it is almost never empty, and hence the problems discussed in Section 3.1.3 can be common here too. Similarly, for consumer-scenario firewalls, it is generally not empty but there are some notable exceptions. For example, for the host firewall in some operation systems, this category always starts empty and most users never change this.
In this option, each application must implement support for potentially complex security policies, along with ways for administrators to configure them. Although the explicit interaction with applications avoids the problems discussed in Section 3.1.3, this approach is impractical for a number of reasons. First, the complexity makes it difficult to implement and is error-prone, especially for application developers whose primary expertise is not networking. Second, the potentially large number of applications (and application developers) results in an inconsistent experience that makes it difficult for an administrator to manage common policies across applications, thus driving up training and operational costs.
Putting security policies in firewalls without explicit interaction with the applications results in the problems discussed in Section 3.1.3. In addition, this leads to "arms races" where the applications are incented to evolve to get around the security policies, since the desires of the end user or developer can conflict with the desires of an administrator. As stated in Section 2.1 of [RFC4924]:
In practice, filtering intended to block or restrict application usage is difficult to successfully implement without customer consent, since over time developers will tend to re-engineer
Thaler Informational [Page 9]
RFC 7288 Host Firewalls June 2014
filtered protocols so as to avoid the filters. Thus, over time, filtering is likely to result in interoperability issues or unnecessary complexity. These costs come without the benefit of effective filtering since many application protocols began to use HTTP as a transport protocol after application developers observed that firewalls allow HTTP traffic while dropping packets for unknown protocols.
Such arms races stem from inherent tussles between the desires of different entities. For example, the tussle between end-user desires and administrator desires leads to an arms race between firewalls and deep packet inspection on the one hand, vs. the use of obfuscation or tunnels on the other.
Although such arms races are often thought of in the context of network firewalls, they also occur with host firewalls. It is, however, generally easier for a host firewall to overcome, since it may be more practical for a host firewall to establish some form of trust between the policy-desiring entities, and have a trusted arbiter.
In this approach, applications use a library or other external service whereby the applications have explicit knowledge of the impact of the security policies in order to avoid the problems in Section 3.1.3, and in a sandboxed environment, this might be the application's only way to interact with the network.
Thus, in this opt-in approach, applications provide a description of the network access requested, and the library/service can ensure that applications and/or users are informed in a way they can understand and that administrators can craft policy that affects the applications.
This approach is very difficult to do in a firewall-vendor-specific library/service when there can be multiple firewall implementations (including ones in the middle of the network), since it is usually impractical for an application developer to know about and develop for many different firewall APIs. It is, however, possible to employ this approach with a firewall-vendor-agnostic library/service that can communicate with both applications and firewalls. Thus, application developers and firewall developers can use a common platform.
Thaler Informational [Page 10]
RFC 7288 Host Firewalls June 2014
We observe that this approach is very different from the classic firewall approach. It is, however the approach used by some host operating system firewalls, and it is also the approach used by PCP in the IETF. As such, we encourage the deployment and use of this model.
Furthermore, while this approach lessens the incentive for arms races as discussed above, one important issue still remains. Namely, there is no standard mechanism today for a library/service to learn complex policies from the network. Further work in this area is needed.
There is often a desire to hide from address and port scans on a public network. However, compliance to many RFCs requires responding to various messages. For example, TCP [RFC0793] compliance requires sending a RST in response to a SYN when there is no listener, and ICMPv6 [RFC4443] compliance requires sending an Echo Reply in response to an Echo Request.
Firewall rules can allow such stealth without changing the statement of compliance of the basic protocols. However, stealth mode could instead be implemented as a configurable option used by the applications themselves. For example, rather than a firewall rule to drop a certain outbound message after an application generates it, fewer resources would be consumed if the application knew not to generate it in the first place.
There is a common misconception that firewalls protect users from malware on their computer, when in fact firewalls protect users from buggy software. There is some concern that firewalls give users a false sense of security; firewalls are not invulnerable and will not prevent malware from running if the user allows it.
This document has focused primarily on host firewalls. For additional discussion (focused more on network firewalls) see [RFC2979] and [BLOCK-FILTER].
Stuart Cheshire provided the motivation for this document by asking the thought-provoking question of why one would want to firewall an application rather than simply stop running it. The ensuing discussion, and subsequent IAB tech chat in November 2011, led to this document. Dan Simon, Stephen Bensley, Gerardo Diaz Cuellar, Brian Carpenter, and Paul Hoffman also provided helpful suggestions.