The Generalized TTL Security Mechanism (GTSM)

The Generalized TTL Security Mechanism (GTSM) is designed to protect a router's IP-based control plane from CPU-utilization based attacks. In particular, while cryptographic techniques can protect the router-based infrastructure (e.g., BGP , ) from a wide variety of attacks, many attacks based on CPU overload can be prevented by the simple mechanism described in this document. Note that the same technique protects against other scarce-resource attacks involving a router's CPU, such as attacks against processor-line card bandwidth. GTSM is based on the fact that the vast majority of protocol peerings are established between routers that are adjacent. Thus, most protocol peerings are either directly between connected interfaces or, in the worst case, are between loopback and loopback, with static routes to loopbacks. Since TTL spoofing is considered nearly impossible, a mechanism based on an expected TTL value can provide a simple and reasonably robust defense from infrastructure attacks based on forged protocol packets from outside the network. Note, however, that GTSM is not a substitute for authentication mechanisms. In particular, it does not secure against insider on-the-wire attacks, such as packet spoofing or replay. Finally, the GTSM mechanism is equally applicable to both TTL (IPv4) and Hop Limit (IPv6), and from the perspective of GTSM, TTL and Hop Limit have identical semantics. As a result, in the remainder of this document the term "TTL" is used to refer to both TTL or Hop Limit (as appropriate). The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 .

GTSM is predicated upon the following assumptions: The vast majority of protocol peerings are between adjacent routers. Service providers may or may not configure strict ingress filtering on non-trusted links. If maximal protection is desired, such filtering is necessary as described in . Use of GTSM is OPTIONAL, and can be configured on a per-peer (group) basis. The peer routers both implement GTSM. The router supports a method to use separate resource pools (e.g., queues, processing quotas) for differently classified traffic. Note that this document does not prescribe further restrictions that a router may apply to packets not matching the GTSM filtering rules, such as dropping packets that do not match any configured protocol session and rate-limiting the rest. This document also does not suggest the actual means of resource separation, as those are hardware and implementation-specific. However, the possibility of denial-of-service (DoS) attack prevention is based on the assumption that classification of packets and separation of their paths are done before the packets go through a scarce resource in the system. In practice, the closer GTSM processing is done to the line-rate hardware, the more resistant the system is to DoS attacks.

This document assumes that, when used with existing protocols, GTSM will be manually configured between protocol peers. That is, no automatic GTSM capability negotiation, such as is provided by RFC 3392 , is assumed or defined. If a new protocol is designed with built-in GTSM support, then it is recommended that procedures are always used for sending and validating received protocol packets (GTSM is always on, see for example ). If, however, dynamic negotiation of GTSM support is necessary, protocol messages used for such negotiation MUST be authenticated using other security mechanisms to prevent DoS attacks. Also note that this specification does not offer a generic GTSM capability negotiation mechanism, so messages of the protocol augmented with the GTSM behavior will need to be used if dynamic negotiation is deemed necessary.

Throughout this document, we assume that potential attackers have evolved in both sophistication and access to the point that they can send control traffic to a protocol session, and that this traffic appears to be valid control traffic (i.e., it has the source/destination of configured peer routers). We also assume that each router in the path between the attacker and the victim protocol speaker decrements TTL properly (clearly, if either the path or the adjacent peer is compromised, then there are worse problems to worry about). For maximal protection, ingress filtering should be applied before the packet goes through the scarce resource. Otherwise an attacker directly connected to one interface could disturb a GTSM-protected session on the same or another interface. Interfaces that aren't configured with this filtering (e.g., backbone links) are assumed to not have such attackers (i.e., are trusted). As a specific instance of such interfaces, we assume that tunnels are not a back-door for allowing TTL-spoofing on protocol packets to a GTSM-protected peering session with a directly connected neighbor. We assume that: 1) there are no tunneled packets terminating on the router, 2) tunnels terminating on the router are assumed to be secure and endpoints are trusted, 3) tunnel decapsulation includes source address spoofing prevention , or 4) the GTSM-enabled session does not allow protocol packets coming from a tunnel. Since the vast majority of peerings are between adjacent routers, we can set the TTL on the protocol packets to 255 (the maximum possible for IP) and then reject any protocol packets that come in from configured peers that do NOT have an inbound TTL of 255. GTSM can be disabled for applications such as route-servers and other multi-hop peerings. In the event that an attack comes in from a compromised multi-hop peering, that peering can be shut down.

If GTSM is not built into the protocol and is used as an additional feature (e.g., for BGP, LDP, or MSDP), it SHOULD NOT be enabled by default in order to remain backward-compatible with the unmodified protocol. However, if the protocol defines a built-in dynamic capability negotiation for GTSM, a protocol peer MAY suggest the use of GTSM provided that GTSM would only be enabled if both peers agree to use it. If GTSM is enabled for a protocol session, the following steps are added to the IP packet sending and reception procedures: Sending protocol packets: The TTL field in all IP packets used for transmission of messages associated with GTSM-enabled protocol sessions MUST be set to 255. This also applies to the related ICMP error handling messages. On some architectures, the TTL of control plane originated traffic is under some configurations decremented in the forwarding plane. The TTL of GTSM-enabled sessions MUST NOT be decremented. Receiving protocol packets: The GTSM packet identification step associates each received packet addressed to the router's control plane with one of the following three trustworthiness categories: Unknown: these are packets that cannot be associated with any registered GTSM-enabled session, and hence GTSM cannot make any judgment on the level of risk associated with them. Trusted: these are packets that have been identified as belonging to one of the GTSM-enabled sessions, and their TTL values are within the expected range. Dangerous: these are packets that have been identified as belonging to one of the GTSM-enabled sessions, but their TTL values are NOT within the expected range, and hence GTSM believes there is a risk that these packets have been spoofed. The exact policies applied to packets of different classifications are not postulated in this document and are expected to be configurable. Configurability is likely necessary in particular with the treatment of related messages (ICMP errors). It should be noted that fragmentation may restrict the amount of information available for classification. However, by default, the implementations: SHOULD ensure that packets classified as Dangerous do not compete for resources with packets classified as Trusted or Unknown. MUST NOT drop (as part of GTSM processing) packets classified as Trusted or Unknown. MAY drop packets classified as Dangerous.

The use of the TTL field to protect BGP originated with many different people, including Paul Traina and Jon Stewart. Ryan McDowell also suggested a similar idea. Steve Bellovin, Jay Borkenhagen, Randy Bush, Alfred Hoenes, Vern Paxon, Robert Raszuk, and Alex Zinin also provided useful feedback on earlier versions of this document. David Ward provided insight on the generalization of the original BGP-specific idea. Alex Zinin, Alia Atlas, and John Scudder provided a significant amount of feedback for the newer versions of the document. During and after the IETF Last Call, useful comments were provided by Francis Dupont, Sam Hartman, Lars Eggert, and Ross Callon.

GTSM is a simple procedure that protects single-hop protocol sessions, except in those cases in which the peer has been compromised. In particular, it does not protect against the wide range of on-the-wire attacks; protection from these attacks requires more rigorous security mechanisms.

The approach described here is based on the observation that a TTL (or Hop Limit) value of 255 is non-trivial to spoof, since as the packet passes through routers towards the destination, the TTL is decremented by one per router. As a result, when a router receives a packet, it may not be able to determine if the packet's IP address is valid, but it can determine how many router hops away it is (again, assuming none of the routers in the path are compromised in such a way that they would reset the packet's TTL). Note, however, that while engineering a packet's TTL such that it has a particular value when sourced from an arbitrary location is difficult (but not impossible), engineering a TTL value of 255 from non-directly connected locations is not possible (again, assuming none of the directly connected neighbors are compromised, the packet has not been tunneled to the decapsulator, and the intervening routers are operating in accordance with RFC 791 ).

The security of any tunneling technique depends heavily on authentication at the tunnel endpoints, as well as how the tunneled packets are protected in flight. Such mechanisms are, however, beyond the scope of this memo. An exception to the observation that a packet with TTL of 255 is difficult to spoof may occur when a protocol packet is tunneled and the tunnel is not integrity-protected (i.e., the lower layer is compromised). When the protocol packet is tunneled directly to the protocol peer (i.e., the protocol peer is the decapsulator), the GTSM provides some limited added protection as the security depends entirely on the integrity of the tunnel. For protocol adjacencies over a tunnel, if the tunnel itself is deemed secure (i.e., the underlying infrastructure is deemed secure, and the tunnel offers degrees of protection against spoofing such as keys or cryptographic security), the GTSM can serve as a check that the protocol packet did not originate beyond the head-end of the tunnel. In addition, if the protocol peer can receive packets for the GTSM-protected protocol session from outside the tunnel, the GTSM can help thwart attacks from beyond the adjacent router. When the tunnel tail-end decapsulates the protocol packet and then IP-forwards the packet to a directly connected protocol peer, the TTL is decremented as described below. This means that the tunnel decapsulator is the penultimate node from the GTSM-protected protocol peer's perspective. As a result, the GTSM check protects from attackers encapsulating packets to your peers. However, specific cases arise when the connection from the tunnel decapsulator node to the protocol peer is not an IP forwarding hop, where TTL-decrementing does not happen (e.g., layer-2 tunneling, bridging, etc). In the IPsec architecture , another example is the use of Bump-in-the-Wire (BITW) .

Protocol packets may be tunneled over IP directly to a protocol peer, or to a decapsulator (tunnel endpoint) that then forwards the packet to a directly connected protocol peer. Examples of tunneling IP over IP include IP-in-IP , GRE , and various forms of IPv6-in-IPv4 (e.g., ). These cases are depicted below.

Peer router ---------- Tunnel endpoint router and peer TTL=255 [tunnel] [TTL=255 at ingress] [TTL=255 at processing] Peer router -------- Tunnel endpoint router ----- On-link peer TTL=255 [tunnel] [TTL=255 at ingress] [TTL=254 at ingress] [TTL=254 at egress] In both cases, the encapsulator (origination tunnel endpoint) is the (supposed) sending protocol peer. The TTL in the inner IP datagram can be set to 255, since RFC 2003 specifies the following behavior:

When encapsulating a datagram, the TTL in the inner IP header is decremented by one if the tunneling is being done as part of forwarding the datagram; otherwise, the inner header TTL is not changed during encapsulation. In the first case, the encapsulated packet is tunneled directly to the protocol peer (also a tunnel endpoint), and therefore the encapsulated packet's TTL can be received by the protocol peer with an arbitrary value, including 255. In the second case, the encapsulated packet is tunneled to a decapsulator (tunnel endpoint), which then forwards it to a directly connected protocol peer. For IP-in-IP tunnels, RFC 2003 specifies the following decapsulator behavior:

The TTL in the inner IP header is not changed when decapsulating. If, after decapsulation, the inner datagram has TTL = 0, the decapsulator MUST discard the datagram. If, after decapsulation, the decapsulator forwards the datagram to one of its network interfaces, it will decrement the TTL as a result of doing normal IP forwarding. See also Section 4.4. And similarly, for GRE tunnels, RFC 2784 specifies the following decapsulator behavior:

When a tunnel endpoint decapsulates a GRE packet which has an IPv4 packet as the payload, the destination address in the IPv4 payload packet header MUST be used to forward the packet and the TTL of the payload packet MUST be decremented. Hence the inner IP packet header's TTL, as seen by the decapsulator, can be set to an arbitrary value (in particular, 255). If the decapsulator is also the protocol peer, it is possible to deliver the protocol packet to it with a TTL of 255 (first case). On the other hand, if the decapsulator needs to forward the protocol packet to a directly connected protocol peer, the TTL will be decremented (second case).

Protocol packets may also be tunneled over MPLS Label Switched Paths (LSPs) to a protocol peer. The following diagram depicts the topology.

Peer router -------- LSP Termination router and peer TTL=255 MPLS LSP [TTL=x at ingress] MPLS LSPs can operate in Uniform or Pipe tunneling models. The TTL handling for these models is described in RFC 3443 that updates RFC 3032 in regards to TTL processing in MPLS networks. RFC 3443 specifies the TTL processing in both Uniform and Pipe Models, which in turn can used with or without penultimate hop popping (PHP). The TTL processing in these cases results in different behaviors, and therefore are analyzed separately. Please refer to Section 3.1 through Section 3.3 of RFC 3443. The main difference from a TTL processing perspective between Uniform and Pipe Models at the LSP termination node resides in how the incoming TTL (iTTL) is determined. The tunneling model determines the iTTL: For Uniform Model LSPs, the iTTL is the value of the TTL field from the popped MPLS header (encapsulating header), whereas for Pipe Model LSPs, the iTTL is the value of the TTL field from the exposed header (encapsulated header). For Uniform Model LSPs, RFC 3443 states that at ingress:

For each pushed Uniform Model label, the TTL is copied from the label/IP-packet immediately underneath it. From this point, the inner TTL (i.e., the TTL of the tunneled IP datagram) represents non-meaningful information, and at the egress node or during PHP, the ingress TTL (iTTL) is equal to the TTL of the popped MPLS header (see Section 3.1 of RFC 3443). In consequence, for Uniform Model LSPs of more than one hop, the TTL at ingress (iTTL) will be less than 255 (x <= 254), and as a result the check described in of this document will fail. The TTL treatment is identical between Short Pipe Model LSPs without PHP and Pipe Model LSPs (without PHP only). For these cases, RFC 3443 states that:

For each pushed Pipe Model or Short Pipe Model label, the TTL field is set to a value configured by the network operator. In most implementations, this value is set to 255 by default. In these models, the forwarding treatment at egress is based on the tunneled packet as opposed to the encapsulation packet. The ingress TTL (iTTL) is the value of the TTL field of the header that is exposed, that is the tunneled IP datagram's TTL. The protocol packet's TTL as seen by the LSP termination can therefore be set to an arbitrary value (including 255). If the LSP termination router is also the protocol peer, it is possible to deliver the protocol packet with a TTL of 255 (x = 255). Finally, for Short Pipe Model LSPs with PHP, the TTL of the tunneled packet is unchanged after the PHP operation. Therefore, the same conclusions drawn regarding the Short Pipe Model LSPs without PHP and Pipe Model LSPs (without PHP only) apply to this case. For Short Pipe Model LSPs, the TTL at egress has the same value with or without PHP. In conclusion, GTSM checks are possible for IP tunneled over Pipe model LSPs, but not for IP tunneled over Uniform model LSPs. Additionally, for all tunneling modes, if the LSP termination router needs to forward the protocol packet to a directly connected protocol peer, it is not possible to deliver the protocol packet to the protocol peer with a TTL of 255. If the packet is further forwarded, the outgoing TTL (oTTL) is calculated by decrementing iTTL by one.

As described in , an attacker directly connected to one interface can disturb a GTSM-protected session on the same or another interface (by spoofing a GTSM peer's address) unless ingress filtering has been applied on the connecting interface. As a result, interfaces that do not include such protection need to be trusted not to originate attacks on the router.

As mentioned, fragmentation may restrict the amount of information available for classification. Since non-initial IP fragments do not contain Layer 4 information, it is highly likely that they cannot be associated with a registered GTSM-enabled session. Following the receiving protocol procedures described in , non-initial IP fragments would likely be classified with Unknown trustworthiness. And since the IP packet would need to be reassembled in order to be processed, the end result is that the initial-fragment of a GTSM-enabled session effectively receives the treatment of an Unknown-trustworthiness packet, and the complete reassembled packet receives the aggregate of the Unknowns. In principle, an implementation could remember the TTL of all received fragments. Then when reassembling the packet, verify that the TTL of all fragments match the required value for an associated GTSM-enabled session. In the likely common case that the implementation does not do this check on all fragments, then it is possible for a legitimate first fragment (which passes the GTSM check) to be combined with spoofed non-initial fragments, implying that the integrity of the received packet is unknown and unprotected. If this check is performed on all fragments at reassembly, and some fragment does not pass the GTSM check for a GTSM-enabled session, the reassembled packet is categorized as a Dangerous-trustworthiness packet and receives the corresponding treatment. Further, reassembly requires to wait for all the fragments and therefore likely invalidates or weakens the fifth assumption presented in : it may not be possible to classify non-initial fragments before going through a scarce resource in the system, when fragments need to be buffered for reassembly and later processed by a CPU. That is, when classification cannot be done with the required granularity, non-initial fragments of GTSM-enabled session packets would not use different resource pools. Consequently, to get practical protection from fragment attacks, operators may need to rate-limit or discard all received fragments. As such, it is highly RECOMMENDED for GTSM-protected protocols to avoid fragmentation and reassembly by manual MTU tuning, using adaptive measures such as Path MTU Discovery (PMTUD), or any other available method , , or .

GTSM could possibly offer some small, though difficult to quantify, degree of protection when used with multi-hop protocol sessions (see ). In order to avoid having to quantify the degree of protection and the resulting applicability of multi-hop, we only describe the single-hop case because its security properties are clearer.

GTSM is only applicable to environments with inherently limited topologies (and is most effective in those cases where protocol peers are directly connected). In particular, its application should be limited to those cases in which protocol peers are directly connected. GTSM will not protect against attackers who are as close to the protected station as its legitimate peer. For example, if the legitimate peer is one hop away, GTSM will not protect from attacks from directly connected devices on the same interface (see for more). Experimentation on GTSM's applicability and security properties is needed in multi-hop scenarios. The multi-hop scenarios where GTSM might be applicable is expected to have the following characteristics: the topology between peers is fairly static and well-known, and in which the intervening network (between the peers) is trusted.

RFC 3682 did not specify how to handle "related messages" (ICMP errors). This specification mandates setting and verifying TTL=255 of those as well as the main protocol packets. Setting TTL=255 in related messages does not cause issues for RFC 3682 implementations. Requiring TTL=255 in related messages may have impact with RFC 3682 implementations, depending on which default TTL the implementation uses for originated packets; some implementations are known to use 255, while 64 or other values are also used. Related messages from the latter category of RFC 3682 implementations would be classified as Dangerous and treated as described in . This is not believed to be a significant problem because protocols do not depend on related messages (e.g., typically having a protocol exchange for closing the session instead of doing a TCP-RST), and indeed the delivery of related messages is not reliable. As such, related messages typically provide an optimization to shorten a protocol keepalive timeout. Regardless of these issues, given that related messages provide a significant attack vector to e.g., reset protocol sessions, making this further restriction seems sensible.