| RFC 9098 | IPv6 Extension Headers | September 2021 |
| Gont, et al. | Informational | [Page] |
This document summarizes the operational implications of IPv6 extension headers specified in the IPv6 protocol specification (RFC 8200) and attempts to analyze reasons why packets with IPv6 extension headers are often dropped in the public Internet.¶
This document is not an Internet Standards Track specification; it is published for informational purposes.¶
This document is a product of the Internet Engineering Task Force (IETF). It represents the consensus of the IETF community. It has received public review and has been approved for publication by the Internet Engineering Steering Group (IESG). Not all documents approved by the IESG are candidates for any level of Internet Standard; see Section 2 of RFC 7841.¶
Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at https://www.rfc-editor.org/info/rfc9098.¶
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.¶
IPv6 extension headers (EHs) allow for the extension of the IPv6 protocol and provide support for core functionality such as IPv6 fragmentation. However, common implementation limitations suggest that EHs present a challenge for IPv6 packet routing equipment and middleboxes, and evidence exists that IPv6 packets with EHs are intentionally dropped in the public Internet in some circumstances.¶
This document has the following goals:¶
Section 4 of this document provides background information about the IPv6 packet structure and associated implications. Section 5 summarizes previous work that has been carried out in the area of IPv6 extension headers. Section 6 discusses packet-forwarding engine constraints in contemporary routers. Section 7 discusses why intermediate systems may need to access Layer 4 information to make a forwarding decision. Finally, Section 8 discusses operational implications of IPv6 EHs.¶
This document uses the term "intermediate system" to describe both routers and middleboxes when there is no need to distinguish between the two and where the important issue is that the device being discussed forwards packets.¶
This document analyzes the operational challenges represented by packets that employ IPv6 extension headers and documents some of the operational reasons why these packets are often dropped in the public Internet. This document is not a recommendation to drop such packets, but rather an analysis of why they are currently dropped.¶
It is useful to compare the basic structure of IPv6 packets against that of IPv4 packets and analyze the implications of the two different packet structures.¶
IPv4 packets have a variable-length header size that allows for the use of IPv4 "options" -- optional information that may be of use to nodes processing IPv4 packets. The IPv4 header length is specified in the "Internet Header Length" (IHL) field of the mandatory IPv4 header and must be in the range of 20 octets (the minimum IPv4 header size) to 60 octets, accommodating at most 40 octets of options. The upper-layer protocol type is specified via the "Protocol" field of the mandatory IPv4 header.¶
Protocol, IHL
+--------+
| |
| v
+------//-----+------------------------+
| | |
| IPv4 | Upper-Layer |
| Header | Protocol |
| | |
+-----//------+------------------------+
variable length
<------------->
IPv6 took a different approach to the IPv6 packet structure. Rather than employing a variable-length header as IPv4 does, IPv6 employs a packet structure similar to a linked list, where a mandatory fixed-length IPv6 header is followed by an arbitrary number of optional extension headers, with the upper-layer header being the last header in the IPv6 header chain. Each extension header typically specifies its length (unless it is implicit from the extension header type) and the "next header" (NH) type that follows in the IPv6 header chain.¶
NH NH, EH-length NH, EH-length
+-------+ +------+ +-------+
| | | | | |
| v | v | v
+-------------+-------------+-//-+---------------+--------------+
| | | | | |
| IPv6 | Ext. | | Ext. | Upper-Layer |
| header | Header | | Header | Protocol |
| | | | | |
+-------------+-------------+-//-+---------------+--------------+
fixed length variable number of EHs & length
<------------> <-------------------------------->
This packet structure has the following implications:¶
Some of the operational and security implications of IPv6 extension headers have been discussed in the IETF:¶
A number of recent RFCs have discussed issues related to IPv6 extension headers and have specified updates to RFC 2460 [RFC2460] (an earlier version of the IPv6 standard). Many of these updates have now been incorporated into the current IPv6 core standard [RFC8200] or the IPv6 node requirements [RFC8504]. Namely,¶
Additionally, [RFC8200] has relaxed the requirement that "all nodes must examine and process the Hop-by-Hop Options header" from [RFC2460], by specifying that only nodes that have been explicitly configured to process the Hop-by-Hop Options header are required to do so.¶
A number of studies have measured the extent to which packets employing IPv6 extension headers are dropped in the public Internet:¶
Most contemporary carrier-grade routers use dedicated hardware, e.g., Application-Specific Integrated Circuits (ASICs) or Network Processing Units (NPUs), to determine how to forward packets across their internal fabrics (see [IEPG94-Scudder] and [APNIC-Scudder] for details). One common method of handling next-hop lookups is to send a small portion of the ingress packet to a lookup engine with specialized hardware, e.g., ternary content-addressable memory (TCAM) or reduced latency dynamic random-access memory (RLDRAM), to determine the packet's next hop. Technical constraints mean that there is a trade-off between the amount of data sent to the lookup engine and the overall packet-forwarding rate of the lookup engine. If more data is sent, the lookup engine can inspect further into the packet, but the overall packet-forwarding rate of the system will be reduced. If less data is sent, the overall packet-forwarding rate of the router will be increased, but the packet lookup engine may not be able to inspect far enough into a packet to determine how it should be handled.¶
If a hardware-forwarding engine on a contemporary router cannot make a forwarding decision about a packet because critical information is not sent to the lookup engine, then the router will normally drop the packet. Section 7 discusses some of the reasons for which a contemporary router might need to access Layer 4 information to make a forwarding decision.¶
Historically, some packet-forwarding engines punted packets of this kind to the control plane for more in-depth analysis, but this is unfeasible on most contemporary router architectures as a result of the vast difference between the hardware-based forwarding capacity of the router and the processing capacity of the control plane and the size of the management link that connects the control plane to the forwarding plane. Other platforms may have a separate software-based forwarding plane that is distinct both from the hardware-based forwarding plane and the control plane. However, the limited CPU resources of this software-based forwarding plane, as well as the limited bandwidth of the associated link, results in similar throughput constraints.¶
If an IPv6 header chain is sufficiently long such that it exceeds the packet lookup capacity of the router, the router might be unable to determine how the packet should be handled and thus could resort to dropping the packet.¶
Although type-length-value (TLV) chains are amenable to iterative processing on architectures that have packet lookup engines with deep inspection capabilities, some packet-forwarding engines manage IPv6 header chains using recirculation. This approach processes extension headers one at a time: when processing on one extension header is completed, the packet is looped back through the processing engine again. This recirculation process continues repeatedly until there are no more extension headers left to be processed.¶
Recirculation is typically used on packet-forwarding engines with limited lookup capability, because it allows arbitrarily long header chains to be processed without the complexity and cost associated with packet-forwarding engines, which have deep lookup capabilities. However, recirculation can impact the forwarding capacity of hardware, as each packet will pass through the processing engine multiple times. Depending on configuration, the type of packets being processed, and the hardware capabilities of the packet-forwarding engine, the data-plane throughput performance on the router might be negatively affected.¶
The following subsections discuss some of the reasons for which intermediate systems may need to process Layer 3 / Layer 4 information to make a forwarding decision.¶
In the case of Equal Cost Multipath (ECMP) load sharing, the intermediate system needs to make a decision regarding which of its interfaces to use to forward a given packet. Since round-robin usage of the links is usually avoided to prevent packet reordering, forwarding engines need to use a mechanism that will consistently forward the same data streams down the same forwarding paths. Most forwarding engines achieve this by calculating a simple hash using an n-tuple gleaned from a combination of Layer 2 through to Layer 4 protocol header information. This n-tuple will typically use the src/dst Media Access Control (MAC) addresses, src/dst IP addresses, and, if possible, further Layer 4 src/dst port information.¶
In the IPv6 world, flows are expected to be identified by means of the IPv6 "Flow Label" [RFC6437]. Thus, ECMP and hash-based load sharing should be possible without the need to process the entire IPv6 header chain to obtain upper-layer information to identify flows. [RFC7098] discusses how the IPv6 Flow Label can be used to enhance Layer 3/4 load distribution and balancing for large server farms.¶
Historically, many IPv6 implementations failed to set the Flow Label, and hash-based ECMP/load-sharing devices also did not employ the Flow Label for performing their task. While support of [RFC6437] is currently widespread for current versions of all popular host implementations, there is still only marginal usage of the IPv6 Flow Label for ECMP and load balancing [Almeida-2020]. A contributing factor could be the issues that have been found in host implementations and middleboxes [Jaeggli-2018].¶
Clearly, widespread support of [RFC6437] would relieve intermediate systems from having to process the entire IPv6 header chain, making Flow Label-based ECMP and load sharing [RFC6438] feasible.¶
If an intermediate system cannot determine consistent n-tuples for calculating flow hashes, data streams are more likely to end up being distributed unequally across ECMP and load-shared links. This may lead to packet drops or reduced performance.¶
Infrastructure Access Control Lists (iACLs) drop unwanted packets destined to a network's infrastructure. Typically, iACLs are deployed because external direct access to a network's infrastructure addresses is operationally unnecessary and can be used for attacks of different sorts against router control planes. To this end, traffic usually needs to be differentiated on the basis of Layer 3 or Layer 4 criteria to achieve a useful balance of protection and functionality. For example, an infrastructure may be configured with the following policy:¶
If a forwarding router cannot determine consistent n-tuples for calculating flow hashes, data streams are more likely to end up being distributed unequally across ECMP and load-shared links. This may lead to packet drops or reduced performance.¶
If a network cannot deploy infrastructure ACLs, then the security of the network may be compromised as a result of the increased attack surface.¶
The case of customer Distributed Denial-of-Service (DDoS) protection and edge-to-core customer protection filters is similar in nature to the iACL protection. Similar to iACL protection, Layer 4 ACLs generally need to be applied as close to the edge of the network as possible, even though the intent is usually to protect the customer edge rather than the provider core. Application of Layer 4 DDoS protection to a network edge is often automated using BGP Flowspec [RFC8955] [RFC8956].¶
For example, a website that normally only handles traffic on TCP ports 80 and 443 could be subject to a volumetric DDoS attack using NTP and DNS packets with a randomized source IP address, thereby rendering source-based remote triggered black hole [RFC5635] mechanisms useless. In this situation, ACLs that provide DDoS protection could be configured to block all UDP traffic at the network edge without impairing the web server functionality in any way. Thus, being able to block arbitrary protocols at the network edge can avoid DDoS-related problems both in the provider network and on the customer edge link.¶
Network Intrusion Detection Systems (NIDS) examine network traffic and try to identify traffic patterns that can be correlated to network-based attacks. These systems generally attempt to inspect application-layer traffic (if possible) but, at the bare minimum, inspect Layer 4 flows. When attack activity is inferred, the operator is notified of the potential intrusion attempt.¶
Network Intrusion Prevention Systems (IPS) operate similarly to NIDSs, but they can also prevent intrusions by reacting to detected attack attempts by e.g., triggering packet filtering policies at firewalls and other devices.¶
Use of extension headers can be problematic for NIDS/IPS, since:¶
As a result, in order to increase the efficiency or effectiveness of these systems, packets employing IPv6 extension headers are often dropped at the network ingress point(s) of networks that deploy these systems.¶
Firewalls enforce security policies by means of packet filtering. These systems usually inspect Layer 3 and Layer 4 traffic but can often also examine application-layer traffic flows.¶
As with a NIDS or IPS (Section 7.4), use of IPv6 extension headers can represent a challenge to network firewalls, since:¶
Additionally, a common firewall filtering policy is the so-called "default deny", where all traffic is blocked (by default), and only expected traffic is added to an "allow/accept list".¶
As a result, packets employing IPv6 extension headers are often dropped by network firewalls, either because of the challenges represented by extension headers or because the use of IPv6 extension headers has not been explicitly allowed.¶
Note that although the data presented in [Zack-FW-Benchmark] was several years old at the time of publication of this document, many contemporary firewalls use comparable hardware and software architectures; consequently, the conclusions of this benchmark are still relevant, despite its age.¶
As discussed in Section 7, intermediate systems that need to find the Layer 4 header must process the entire IPv6 header chain. When such devices are unable to obtain the required information, the forwarding device has the option to drop the packet unconditionally, forward the packet unconditionally, or process the packet outside the normal forwarding path. Forwarding packets unconditionally will usually allow for the circumvention of security controls (see, e.g., Section 7.5), while processing packets outside of the normal forwarding path will usually open the door to Denial-of-Service (DoS) attacks (see, e.g., Section 6). Thus, in these scenarios, devices often simply resort to dropping such packets unconditionally.¶
Most contemporary carrier-grade routers have a fast hardware-assisted forwarding plane and a loosely coupled control plane, connected together with a link that has much less capacity than the forwarding plane could handle. Traffic differentiation cannot be performed by the control plane because this would overload the internal link connecting the forwarding plane to the control plane.¶
The Hop-by-Hop Options header has been particularly challenging since, in most circumstances, the corresponding packet is punted to the control plane for processing. As a result, many operators drop IPv6 packets containing this extension header [RFC7872]. [RFC6192] provides advice regarding protection of a router's control plane.¶
Some intermediate systems do not have support for fine-grained filtering of IPv6 extension headers. For example, an operator that wishes to drop packets containing RHT0 may only be able to filter on the extension header type (Routing Header). This could result in an operator enforcing a coarser filtering policy (e.g., "drop all packets containing a Routing Header" vs. "only drop packets that contain a Routing Header Type 0").¶
The security implications of IPv6 extension headers generally fall into one or more of these categories:¶
Unlike IPv4 packets where the upper-layer protocol can be trivially found by means of the IHL field of the IPv4 header, the structure of IPv6 packets is more flexible and complex. This can represent a challenge for devices that need to find this information, since locating upper-layer protocol information requires that all IPv6 extension headers be examined. In turn, this presents implementation difficulties, since some packet-filtering mechanisms that require upper-layer information (even if just the upper-layer protocol type) can be trivially circumvented by inserting IPv6 extension headers between the main IPv6 header and the upper-layer protocol header. [RFC7113] describes this issue for the RA-Guard case, but the same techniques could be employed to circumvent other IPv6 firewall and packet-filtering mechanisms. Additionally, implementation inconsistencies in packet-forwarding engines can result in evasion of security controls [PARSING] [Atlasis2014] [BH-EU-2014].¶
Sometimes, packets with IPv6 extension headers can impact throughput performance on intermediate systems. Unless appropriate mitigations are put in place (e.g., packet dropping and/or rate limiting), an attacker could simply send a large amount of IPv6 traffic employing IPv6 extension headers with the purpose of performing a DoS attack (see Sections 6.1 and 8 for further details). The extent to which performance is affected on these devices is implementation dependent.¶
IPv6 implementations, like all other software, tend to mature with time and wide-scale deployment. While the IPv6 protocol itself has existed for over 20 years, serious bugs related to IPv6 extension header processing continue to be discovered (see, e.g., [Cisco-Frag], [Microsoft-SA], and [FreeBSD-SA]). Because there is currently little operational reliance on IPv6 extension headers, the corresponding code paths are rarely exercised, and there is the potential for bugs that still remain to be discovered in some implementations.¶
The IPv6 Fragment Header is employed for the fragmentation and reassembly of IPv6 packets. While many of the security implications of the fragmentation/reassembly mechanism are known from the IPv4 world, several related issues have crept into IPv6 implementations. These range from DoS attacks to information leakages, as discussed in [RFC7739], [Bonica-NANOG58], and [Atlasis2012].¶
This document has no IANA actions.¶
The security implications of IPv6 extension headers are discussed in Section 8.4. This document does not introduce any new security issues.¶
The authors would like to thank (in alphabetical order) Mikael Abrahamsson, Fred Baker, Dale W. Carder, Brian Carpenter, Tim Chown, Owen DeLong, Gorry Fairhurst, Guillermo Gont, Tom Herbert, Lee Howard, Tom Petch, Sander Steffann, Eduard Vasilenko, Éric Vyncke, Rob Wilton, Jingrong Xie, and Andrew Yourtchenko for providing valuable comments on earlier draft versions of this document.¶
Fernando Gont would like to thank Jan Zorz / Go6 Lab <https://go6lab.si/>, Jared Mauch, and Sander Steffann <https://steffann.nl/> for providing access to systems and networks that were employed to perform experiments and measurements involving packets with IPv6 extension headers.¶