This Internet-Draft is submitted to IETF in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts.
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.”
The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt.
The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html.
This Internet-Draft will expire on March 7, 2010.
Copyright (c) 2009 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 in effect on the date of publication of this document (http://trustee.ietf.org/license-info). Please review these documents carefully, as they describe your rights and restrictions with respect to this document.
This document describes a heuristic approach for distinguishing ESP-NULL (Encapsulating Security Payload without encryption) packets from encrypted ESP packets. The reason for using heuristics instead of modifying ESP is to provide a solution that can be used now without updating all end nodes. With heuristic methods, only the intermediate devices wanting to find ESP-NULL packets need to be updated.
1.1. Applicability: Heuristic Traffic Inspection and Wrapped ESP
1.2. Requirements notation
2. Other Options
2.2. Mandating by Policy
2.3. Modifying ESP
3. Description of Heuristics
4. IPsec flows
5. Deep Inspection Engine
6. Special and Error Cases
7. UDP encapsulation
8. Heuristic Checks
8.1. ESP-NULL format
8.2. Self Describing Padding Check
8.3. Protocol Checks
8.3.1. TCP checks
8.3.2. UDP checks
8.3.3. ICMP checks
8.3.4. SCTP checks
8.3.5. IPv4 and IPv6 Tunnel checks
9. Security Considerations
10.1. Normative References
10.2. Informative References
Appendix A. Example Pseudocode
§ Authors' Addresses
The ESP (Encapsulated Security Payload [RFC4303] (Kent, S., “IP Encapsulating Security Payload (ESP),” December 2005.)) protocol can be used with NULL encryption [RFC2410] (Glenn, R. and S. Kent, “The NULL Encryption Algorithm and Its Use With IPsec,” November 1998.) to provide authentication and integrity protection, but not confidentiality. This offers similar properties to IPsec's AH (Authentication Header [RFC4302] (Kent, S., “IP Authentication Header,” December 2005.)). The reason to use ESP-NULL instead of AH is that AH cannot be used if there are NATs (Network Address Translation devices) on the path. With AH it would be easy to detect packets which have only authentication and integrity protection, as AH has its own protocol number and deterministic packet length. With ESP-NULL such detection is nondeterministic, in spite of the base ESP packet format being fixed.
In some cases intermediate devices would like to detect ESP-NULL packets so they could perform deep inspection or enforce access control. This kind of deep inspection includes virus detection, spam filtering, and intrusion detection. As end nodes might be able to bypass those checks by using encrypted ESP instead of ESP-NULL, these kinds of scenarios also require very specific policies to forbid such circumvention.
These sorts of policy requirements usually mean that the whole network needs to be controlled, i.e. under the same adminstrative domain. Such setups are usually limited to inside the network of one enterprise or organization, and encryption is not used as the network is considered safe enough from eavesdroppers.
Because the traffic inspected is usually host to host traffic inside one organization, that usually means transport mode IPsec is used.
It should also be noted that even if new protocol modifications for ESP support easier detection of ESP-NULL in the future, this document will aid in transition of older end-systems. That way, a solution can be implemented immediately, and not after a 5-10 year upgrade-and-deployment time frame. Even with protocol modification for end nodes, the intermediate devices will need heuristics until they can assume that those protocol modifications can be found from all the end devices. To make sure that any solution does not break in the future it would be best if such heuristics are documented, i.e. we need to publish an RFC for what to do now even when there might be a new protocol coming in the future that will solve the same problem better.
There are two ways to enable intermediate security devices to distinguish between encrypted and unencrypted ESP traffic:
- The heuristics approach has the intermediate node inspect the unchanged ESP traffic, to determine with extremely high probability whether or not the traffic stream is encrypted.
- The Wrapped ESP approach [I‑D.ietf‑ipsecme‑traffic‑visibility] (Grewal, K., Montenegro, G., and M. Bhatia, “Wrapped ESP for Traffic Visibility,” January 2010.), in contrast, requires the ESP endpoints to be modified to support the new protocol. WESP allows the intermediate node to distinguish encrypted and unencrypted traffic deterministically, using a simpler implementation for the intermediate node.
Both approaches are being documented simultaneously by the IPsecME Working Group, with WESP being put on Standards Track while the heuristics approach is being published as an Informational RFC. While endpoints are being modified to adopt WESP, we expect both approaches to coexist for years, because the heuristic approach is needed to inspect traffic where at least one of the endpoints has not been modified. In other words, intermediate nodes are expected to support both approaches in order to achieve good security and performance during the transition period.
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 [RFC2119] (Bradner, S., “Key words for use in RFCs to Indicate Requirement Levels,” March 1997.).
This document will discuss the heuristic approach of detecting ESP-NULL packets. There are some other options which can be used, and this section will briefly discuss those.
The most logical approach would use the already defined protocol which offers authentication and integrity protection, but not confidentiality, namely AH. AH traffic is clearly marked as not encrypted, and can always be inspected by intermediate devices.
Using AH has two problems. First is that, as it also protects the IP headers, it will also protect against NATs on the path, thus it will not work if there is NAT on the path between end nodes. In some environments this might not be a problem, but some environments include heavy use of NATs even inside the internal network of the enterprise or organization. NAT-Traversal (NAT-T, [RFC3948] (Huttunen, A., Swander, B., Volpe, V., DiBurro, L., and M. Stenberg, “UDP Encapsulation of IPsec ESP Packets,” January 2005.)) could be extended to support AH also, and the early versions of the NAT-T proposals did include that, but it was left out as it was not seen as necessary.
The another problem is that in the new IPsec Architecture [RFC4301] (Kent, S. and K. Seo, “Security Architecture for the Internet Protocol,” December 2005.) the support for AH is now optional, meaning not all implementations support it. ESP-NULL has been defined to be mandatory to implement by Cryptographic Algorithm Implementation Requirements for Encapsulating Security Payload (ESP) [RFC4835] (Manral, V., “Cryptographic Algorithm Implementation Requirements for Encapsulating Security Payload (ESP) and Authentication Header (AH),” April 2007.).
AH has also quite complex processing rules compared to ESP when calculating the ICV, including things like zeroing out mutable fields. As AH is not as widely used than ESP, the AH support is not as well tested in the interoperability events, meaning it might have more bugs than ESP implementations.
Another easy way to solve this problem is to mandate the use of ESP-NULL with common parameters within an entire organization. This either removes the need for heuristics (if no ESP encrypted traffic is allowed at all) or simplifies them considerably (only one set of parameters needs to be inspected, e.g. everybody in the organization who is using ESP-NULL must use HMAC-SHA-1-96 as their integrity algorithm). This does not work if the machines are not under the same administrative domain. Also, such a solution might require some kind of centralized policy management to make sure everybody uses the same policy.
Several internet drafts discuss ways of modifying ESP to offer intermediate devices information about an ESP packet's use of NULL encryption. The following methods have been discussed: adding an IP-option, adding a new IP-protocol number plus an extra header [I‑D.ietf‑ipsecme‑traffic‑visibility] (Grewal, K., Montenegro, G., and M. Bhatia, “Wrapped ESP for Traffic Visibility,” January 2010.), adding a new IP-protocol numbers which tell the ESP-NULL parameters [I‑D.hoffman‑esp‑null‑protocol] (Hoffman, P. and D. McGrew, “An Authentication-only Profile for ESP with an IP Protocol Identifier,” August 2007.), reserving an SPI range for ESP-NULL [I‑D.bhatia‑ipsecme‑esp‑null] (Bhatia, M., “Identifying ESP-NULL Packets,” December 2008.), and using UDP encapsulation with a different format and ports.
All of the aforementioned drafts require modification to ESP, which requires that all end nodes needs to be modified before intermediate devices can assume that this new ESP format is in use. Updating end nodes will require lots of time. An example of the slowness of endpoint migration vs. intermediate migration can be seen from the IPv6 vs NAT case. IPv6 required updating all of the end nodes (and routers too) before it could be effectively used. This has taken a very long time, and IPv6 deployment is not yet widespread. NAT, on the other hand, only required modifying an existing intermediate device or adding a new one, and has spread out much faster. Another example of slow end-node deployment is IKEv2. Considering an implementation that requires both IKEv2 and a new ESP format, it would take several years, possibly as long as a decade, before widespread deployment.
The heuristics to detect ESP-NULL packets will only require changes to the those intermediate devices which do deep inspection or other operations which require detecting ESP-NULL. As those nodes require changes regardless of any ESP-NULL method, updating intermediate nodes is unavoidable. Heuristics do not require updating or modifying any other devices on the rest of the network, including (and especially) end-nodes.
In this document it is assumed that an affected intermediate node will act as a stateful interception device, meaning it will keep state of the flows - where flows are defined by the ESP SPI and IP addresses forming an IPsec SA - going through it. The heuristics can also be used without storing any state, but performance will be worse in that case, as heuristic checks will need to be done for each packet, not only once per flow. This will also affect the reliability of the heuristics.
Generally, an intermediate node runs heuristics only for the first few packets of the new flow (i.e. the new IPsec SA). After those few packets, the node detects parameters of the IPsec flow, it skips detection heuristics, and it can perform direct packet-inspecting action based on its own policy. Once detected, ESP-NULL packets will never be detected as encrypted ESP packets, meaning that valid ESP-NULL packets will never bypass the deep inspection. The only failure mode of these heuristics is to assume encrypted ESP packets are ESP-NULL packet, thus causing completely random packet data to be deeply inspected. An attacker can easily send random-looking ESP-NULL packets which will cause heuristics to detect packets as encrypted ESP, but that is no worse than sending non-ESP fuzz through an intermediate node.
For hardware implementations all the flow lookup based on the ESP next header number (50), source address, destination address, and SPI can be done by the hardware (there is usually already similar functionality there, for TCP/UDP flows). The heuristics can be implemented by the hardware, but using software will allow faster updates when new protocol modifications come out or new protocols need support.
ESP is a stateful protocol, meaning there is state stored in the both end nodes of the ESP IPsec SA, and the state is identified by the pair of destination IP and SPI. End nodes also often fix the source IP address in an SA unless the destination is a multicast group. As most (if not all) flows of interest to an intermediate device are unicast, it is safer to assume the receiving node also uses a source address, and the intermediate device should do the same. In some cases this might cause extraneous cached ESP IPsec SA flows, but by using the source address two distinct flows will never be mixed.
When the intermediate device sees an new ESP IPsec flow, i.e. a flow of ESP packets where the source address, destination address, and SPI number forms a triplet which has not been cached, it will start the heuristics to detect whether this flow is ESP-NULL or not. These heuristics appear in the Section 8 (Heuristic Checks).
When the heuristics finish, they will label the flow as either encrypted (which tells that packets in this flow are encrypted, and cannot be ESP-NULL packets) or as ESP-NULL. This information, along with the ESP-NULL parameters detected by the heuristics, is stored to a flow cache, which will be used in the future when processing packets of the same flow.
Both encrypted ESP and ESP-NULL flows are processed based on the local policy. In normal operation encrypted ESP flows are passed through or dropped per local policy, and ESP-NULL flows are passed to the deep inspection engine. Local policy will also be used to determine other packet-processing parameters. Local policy issues will be clearly marked in this document to ease implementation.
In some cases the heuristics cannot determine the type of flow from a single packet, and in that case it might need multiple packets before it can finish the process. In those cases the heuristics return "unsure" status. In that case the packet processed based on the local policy and flow cache is updated with "unsure" status. Local policy for "unsure" packets could range from dropping (which encourages end-node retransmission) to queuing (which may preserve delivery, at the cost of artificially inflating round-trip times if they are measured). When the next packet to the flow arrives, it is heuristically processed again, and the cached flow may continue to be "unsure", marked as ESP, or marked as an ESP-NULL flow.
There are several reasons why a single packet might not be enough to detect type of flow. One of them is that the next header number was unknown, i.e. if heuristics do not know about the protocol for the packet, it cannot verify it has properly detected ESP-NULL parameters, even when the packet otherwise looks like ESP-NULL. If the packet does not look like ESP-NULL at all, then encrypted ESP status can be returned quickly. As ESP-NULL heuristics should know the same protocols as a deep inspection device, an unknown protocol should not be handled any differently than a cleartext instance of an unknown protocol if possible.
A deep inspection engine running on an intermediate node usually checks deeply into the packet and performs policy decisions based on the contents of the packet. The deep inspection engine should be able to tell the difference between success, failure, and garbage. Success means that a packet was successfully checked with the deep inspection engine, and it passed the checks and is allowed to be forwarded. Failure means that a packet was successfully checked but the actual checks done indicated that packets should be dropped, i.e. the packet contained a virus, was a known attack, or something similar.
Garbage means that the packet's protocol headers or other portions were unparseable. For the heuristics, it would be useful if the deep inspection engine can differentiate the garbage and failure cases, as garbage cases can be used to detect certain error cases (e.g. where the ESP-NULL parameters are incorrect, or the flow is really an encrypted ESP flow, not an ESP-NULL flow).
If the deep inspection engine will only return failure for all garbage packets in addition to real failure cases, then a system implementing the ESP-NULL heuristics cannot recover from error situations quickly.
There is a small probability that an encrypted ESP packet (which looks like contain completely random bytes) will have correct bytes in correct locations, such that heuristics will detect the packet as an ESP-NULL packet instead of detecting that it is encrypted ESP packet. The actual probabilities will be computed later in this document. Such a packet will not cause problems, as the deep inspection engine will most likely reject the packet and return that it is garbage. If the deep inspection engine is rejecting a high number of packets as garbage, it might indicate an original ESP-NULL detection for the flow was wrong (i.e. an encrypted ESP flow was improperly detected as ESP-NULL). In that case, the cached flow should be invalidated and discovery should happen again.
Each ESP-NULL flow should also keep statistics about how many packets have been detected as garbage by deep inspection, how many have passed checks, or how many have failed checks with policy violations (i.e. failed because actual inspection policy failures, not because the packet looked like garbage). If the number of garbage packets suddenly increases (e.g. most of the packets start to be look like garbage according to the deep inspection engine), it is possible the old ESP-NULL SA was replaced by an identical-SPI encrypting ESP SA. If both ends use random SPI generation, this is a very unlikely situation (1 in 2^32), but it is possible that some nodes reuse SPI numbers (e.g. a 32-bit memory address of the SA descriptor), thus this situation needs to be handled.
Actual limits for cache invalidation are local policy decisions. Sample invalidation policies include: 50% of packets marked as garbage within a second; or if a deep inspection engine cannot differentiate between garbage and failure, failing more than 95% of packets in last 10 seconds. For implementations that do not distinguish between garbage and failure, failures should not be treated too quickly as indication of SA reuse. Often, single packets cause state-related errors that block otherwise normal packets from passing.
The flow lookup code needs to detect UDP packets to or from port 4500 in addition to the ESP packets, and perform similar processing to them after skipping the UDP header. Each unique port pair makes separate IPsec flow, i.e. UDP encapsulated IPsec flows are identified by the source and destination IP, source and destination port number and SPI number. As devices might be using MOBIKE, that means that the flow cache should be shared between the UDP encapsulated IPsec flows and non encapsulated IPsec flows. As previously mentioned, differentiating between garbage and actual policy failures will help in proper detection immensely.
Because the checks are also run for packets having source port 4500 in addition to those having destination port 4500, this might cause the checks to be run for non-ESP traffic too. The UDP encapsulation processing should also be avare of that. We cannot limit the checks for only UDP packets having destination port 4500, as return packets from the SGW going towards the NAT box do have source port 4500, and some other port as destination port.
Normally, HMAC-SHA1-96 or HMAC-MD5-96 gives 1 out of 2^96 probability that a random packet will pass the HMAC test. This yields a 99.999999999999999999999999998% probability that an end node will correctly detect a random packet as being invalid. This means that it should be enough for an intermediate device to check around 96 bits from the input packet. By comparing them against known values for the packet we get more or less same probability as an end node is using. This gives an upper limit of how many bits heuristics need to check - there is no point of checking much more than that many bits (since that same probability is acceptable for the end node). In most of the cases the intermediate device does not need that high probability, perhaps something around 32-64 bits is enough.
IPsec's ESP has a well-understood packet layout, but its variable-length fields reduce the ability of pure algorithmic matching to one requiring heuristics and assigning probabilities.
The ESP-NULL format is as follows:
0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Security Parameters Index (SPI) | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Sequence Number | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | IV (optional) | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Payload Data (variable) | ~ ~ | | + +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | | Padding (0-255 bytes) | +-+-+-+-+-+-+-+-+ +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | | Pad Length | Next Header | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Integrity Check Value-ICV (variable) | ~ ~ | | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Figure 1 |
The output of the heuristics should provide us information whether the packet is encrypted ESP or ESP-NULL. In case it is ESP-NULL we also need to know the Integrity Check Value (ICV) field length and the Initialization Vector (IV) length.
The currently defined ESP authentication algorithms have 5 different lengths for the ICV field. Most commonly used is 96 bits for AUTH_HMAC_MD5_96, AUTH_HMAC_SHA1_96, AUTH_AES_XCBC_96, and AUTH_AES_CMAC_96 algorithms. After that comes 128 bit ICV lengths for AUTH_HMAC_MD5_128, AUTH_HMAC_SHA2_256_128 AUTH_AES_128_GMAC, AUTH_AES_192_GMAC, AUTH_AES_256_GMAC algorithms. 160, 192, and 256 bit field lengths are used by AUTH_HMAC_SHA1_160, AUTH_HMAC_SHA2_384_192, and AUTH_HMAC_SHA2_512_256 algorithms, respectively.
In addition to the ICV length, there are also two possible values for IV lengths: zero bytes (default) and eight bytes (for AUTH_AES_*_GMAC). Detecting the IV length requires understanding the payload, i.e. the actual protocol data (meaning TCP, UDP, etc). This is required to distinguish the optional IV from the actual protocol data. How well IV can be distinguished from the actual protocol data depends how the IV is generated. If IV is generated using method that generates random looking data (i.e. encrypted counter etc) then disginguishing protocol data from IV is quite easy. If IV is counter or similar non-random value, then there are bit more possibilities for error. If the protocol (also known as the, "next header") of the packet is something that heuristics doesn't have protocol checking support, then detecting the IV length is impossible, thus the heuristics cannot finish. In that case heuristics returns "unsure" and requires further packets.
Before obtaining the next header field, the ICV length must be measured. Five different ICV lengths leads to five possible places for the pad length and padding. Implementations must be careful when trying larger sizes of ICV such that the inspected bytes do not belong to data that is not payload data. For example, a ten-byte ICMP echo request will have zero-length padding, but any checks for 256-bit ICVs will inspect sequence number or SPI data if the packet actually contains a 96-bit or 128-bit ICV.
ICV lengths should always be checked from shorted to longest. It is much more likely to obtain valid-looking padding bytes in the cleartext part of the payload than from the ICV field of a longer ICV than what is currently inspected. For example, if a packet has a 96-bit ICV and implementation starts first checking for a 256-bit ICV, it is possible that the cleartext part of the payload contains valid looking bytes. If done in the other order, i.e. a packet having a 256-bit ICV and the implementation checks for a 96-bit ICV first, the inspected bytes are part of the longer ICV field, and should be indistinguishable from random noise.
Each ESP packet always has between 0-255 bytes of padding, and payload, pad length, and next header are always right aligned within a 4-byte boundary. Normally implementations use minimal amount of padding, but heuristics method would be even more reliable if some extra padding is added. The actual padding data has bytes starting from 01 and ending to the pad length, i.e. exact padding and pad length bytes for 4 bytes of padding would be 01 02 03 04 04.
Two cases of ESP-NULL padding are matched bytes (like the 04 04 shown above), or the zero-byte padding case. In cases where there is one or more bytes of padding, a node can perform a very simple and fast test -- a sequence of N N in any of those five locations. Given five two-byte locations (assuming the packet size allows all five possible ICV lengths), the upper-bound probability of finding a random encrypted packet that exhibits non-zero length ESP-NULL properties is 1-(1-255/65536)^5 == 0.019 == 1.9%. In the cases where there is 0 bytes of padding, a random encrypted ESP packet has 1-(1-1/256)^5 == 0.019 == 1.9%. Together, both cases yields a 3.8% upper-bound chance of misclassifying an encrypted packet as an ESP-NULL packet.
In the matched bytes case, further inspection (counting the pad bytes backward and downward from the pad-length match) can reduce the number of misclassified packets further. A padding length of 255 means a specific 256^254 sequence of bytes must occur. This virtually eliminates pairs of 'FF FF' as viable ESP-NULL padding. Every one of the 255 pairs for padding length N has only a 1 / 256^N probability of being correct ESP-NULL padding. This shrinks the aforementioned 1.9% of matched-pairs to virtually nothing.
At this point a maximum of 2% of packets remain, so the next header number is inspected. If the next header number is known (and supported) then the packet can be inspected based on the next header number. If the next header number is unknown (i.e. not any of those with protocol checking support) the packet is marked "unsure", because there is no way to detect the IV length without inspecting the inner protocol payload.
There are six different next header fields which are in common use (TCP (6), UDP (17), ICMP (1), SCTP (132), IPv4 (4) and IPv6 (41)), and if IPv6 is in heavy use, that number increases to nine (Fragment (44), ICMPv6 (58), and IPv6 options (60)). To insure that no packet is misinterpreted as an encrypted ESP packet even when it is ESP-NULL packet, a packet cannot be marked as a failure even when the next header number is one of those which is not known and supported. In those cases the packet must be marked "unsure".
An intermediate node's policy, however, can aid in detecting an ESP-NULL flow even when the protocol is not a common-case one. By counting how many "unsure" returns obtained via heuristics, and after the receipt of a consistent, but unknown, next-header number of those with same ICV length (i.e. in same location) that means that flow is ESP-NULL with more probability than what it is to fool a given hash algorithm with a random packet. The flow can be classified as ESP-NULL with a known ICV length, but an unknown IV length.
Fortunately, in unknown protocol cases the IV length does not matter, as the protocol is unknown to the heuristics, it will most likely be unknown by the deep inspection engine also. It is therefore important that heuristics should support at least those same protocols as the deep inspection engine does. Upon receipt of any inner next header number that is known by the heuristics (and deep inspection engine), the heuristics can detect the IV length properly.
Generic protocol checking is much easier with pre-existing state. For example, when many TCP / UDP flows are established over one SA, a rekey with a new SA which needs heuristics. Then it is enough to just check that if the flow is already known by the deep inspection engine, it will give a strong leaning that the new SA is really ESP-NULL.
The worst case scenario is when an end node starts up communcation, i.e. it does not have any previous flows through the device. Heuristics will run on the first few packets received from the end node. The later subsections mainly cover these bringup cases, as they are the most difficult.
In the protocol checks there are two different types of checks. The first check is for packet validity, i.e. certain locations must contain specific values. For example, an inner IPv4 header of an IPv4 tunnel packet must have its 4-bit version number set to 4. If it does not, the packet is not valid, and can be marked as a failure. Other positions depending on ICV and IV lengths must also be checked, and if all of them are failures, then the packet is a failure. If any of the checks are "unsure" the packet is marked as such.
The second type of check is for variable, but easy-to-parse values. For example, the 4-bit header length field of an inner IPv4 packet. It has a fixed value (5) as long as there are no inner IPv4 options. If the header-length has that specific value, the number of known "good" bits increases. If it has some other value, the known "good" bit count stays the same. A local policy might include reaching a bit count that is over a threshold (for example 96 bits), causing a packet to be marked as valid.
When the first TCP packet is feed to the heuristics, it is most likely going to be the SYN packet of the new connection, thus it will have less useful information than what other later packets might have. Best valid packet checks include: checking that header length and reserved and other bits have valid values; checking source and destination port numbers, which in some cases can be used for heuristics (but in general they cannot be used to reliably distinguished from random numbers apart from some well-known ports like 25/80/110/143).
The most obvious field, TCP checksum, might not be usable, as it is possible that the packet has already transitted a NAT box, thus the IP numbers used in the checksum are wrong, thus the checksum is wrong. If the checksum is correct that can again be used to increase valid bit count, but verifying checksums is a costly operation, thus skipping that check might be best unless there is hardware to help the calculation. Window size, urgent pointer, sequence number, and acknowledgement numbers can be used, but there is not one specific known value for them.
One good method of detection is if the a packet is dropped then the next packet will most likely be a retransmission of the previous packet. Thus if two packets are received with the same source, and destination port numbers, and where sequence numbers are either same or right after each other, then it's likely a TCP packet has been correctly detected.
The deep inspection engines usually do very good TCP flow checking already, including flow tracking, verification of sequence numbers, and reconstruction of the whole TCP flow. Similar methods can be used here, but they are implementation-dependent and not described here.
UDP header has even more problems than the TCP header, as UDP has even less known data. The checksum has the same problem as the TCP checksum, due to NATs. The UDP length field might not match the overall packet length, as the sender is allowed to include TFC (traffic flow confidentiality, see section 2.7 of IP Encapsulating Security Payload document [RFC4303] (Kent, S., “IP Encapsulating Security Payload (ESP),” December 2005.)) padding.
With UDP packets similar multiple packet methods can be used as with TCP, as UDP protocols usually include several packets using same port numbers going from one end node to another, thus receiving multiple packets having a known pair of UDP port numbers is good indication that the heuristics have passed.
Some UDP protocols also use identical source and destination port numbers, thus that is also a good check.
As ICMP messages are usually sent as return packets for other packets, they are not very common packets to get as first packets for the SA, the ICMP Echo message being a noteworthy exception. ICMP ECHO has known type and code, identifier, and sequence number. The checksum, however, might be incorrect again because of NATs.
For error ICMP messages the ICMP message contains part of the original IP packet inside, and then same rules which are used to detect IPv4/IPv6 tunnel checks can be used.
SCTP [RFC4960] (Stewart, R., “Stream Control Transmission Protocol,” September 2007.) has a self-contained checksum, which is computed over the SCTP payload and is not affected by NATs unless the NAT is SCTP-aware. Even more than the TCP and UDP checksums, the SCTP checksum is expensive, and may be prohibitive even for deep-packet inspections.
SCTP chunks can be inspected to see if their lengths are consistent across the total length of the IP datagram, so long as TFC padding is not present.
In cases of tunneled traffic the packet inside contains a full IPv4 or IPv6 packet. Many fields are useable. For IPv4 those fields include version, header length, total length (again TFC padding might confuse things there), protocol number, and 16-bit header checksum. In those cases the intermediate device should give the decapsulated IP packet to the deep inspection engine. IPv6 has less usable fields, but the version number, packet length (modulo TFC confusion) and next-header all can be used by deep-packet inspection.
The whole deep inspection for ESP-NULL flows only has the problem that attacker can always bypass it by using encrypted ESP (or some other encryption or tunneling method) instead of ESP-NULL unless that is not forbidden by policy. This means that in the end the responsibility whether end node can bypass deep inspection is for the policy enforcement of the both end nodes, i.e. if a server allows encrypted connections also, then attacker who wants to attack the server and wants to bypass deep inspection device in the middle, will use encrypted traffic. This means that the protection of the whole network is only as good as the policy enforcement and protection of the end node. One way to enforce deep inspection for all traffic, is to forbid encrypted ESP completely, but in that case ESP-NULL detection is easier, as all packets must be ESP-NULL based on the policy, and further restriction can eliminate ambiguities in ICV and IV sizes.
|[RFC2119]||Bradner, S., “Key words for use in RFCs to Indicate Requirement Levels,” BCP 14, RFC 2119, March 1997 (TXT, HTML, XML).|
|[RFC2410]||Glenn, R. and S. Kent, “The NULL Encryption Algorithm and Its Use With IPsec,” RFC 2410, November 1998 (TXT, HTML, XML).|
|[RFC4301]||Kent, S. and K. Seo, “Security Architecture for the Internet Protocol,” RFC 4301, December 2005 (TXT).|
|[RFC4302]||Kent, S., “IP Authentication Header,” RFC 4302, December 2005 (TXT).|
|[RFC4303]||Kent, S., “IP Encapsulating Security Payload (ESP),” RFC 4303, December 2005 (TXT).|
|[I-D.bhatia-ipsecme-esp-null]||Bhatia, M., “Identifying ESP-NULL Packets,” draft-bhatia-ipsecme-esp-null-00 (work in progress), December 2008 (TXT).|
|[I-D.hoffman-esp-null-protocol]||Hoffman, P. and D. McGrew, “An Authentication-only Profile for ESP with an IP Protocol Identifier,” draft-hoffman-esp-null-protocol-00 (work in progress), August 2007 (TXT).|
|[I-D.ietf-ipsecme-traffic-visibility]||Grewal, K., Montenegro, G., and M. Bhatia, “Wrapped ESP for Traffic Visibility,” draft-ietf-ipsecme-traffic-visibility-12 (work in progress), January 2010 (TXT).|
|[RFC3948]||Huttunen, A., Swander, B., Volpe, V., DiBurro, L., and M. Stenberg, “UDP Encapsulation of IPsec ESP Packets,” RFC 3948, January 2005 (TXT).|
|[RFC4835]||Manral, V., “Cryptographic Algorithm Implementation Requirements for Encapsulating Security Payload (ESP) and Authentication Header (AH),” RFC 4835, April 2007 (TXT).|
|[RFC4960]||Stewart, R., “Stream Control Transmission Protocol,” RFC 4960, September 2007 (TXT).|
This appendix is meant for the implementors. It does not include all the required checks, and this is just example pseudocode, so final implementation can be very different. It mostly lists things what needs to be done, but implementations can optimize things depending on their other parts. For example implementation might combine heuristics and deep inspection tightly together.
The following example pseudocode show the fastpath part of the packet processing engine. This part is usually implemented in hardware.
//////////////////////////////////////////////////////////// // This pseudocode uses following variables: // // SPI_offset: Number of bytes between start of protocol // data and SPI. This is 0 for ESP, and // 8 for UDP encapsulated ESP (i.e skipping // UDP header). // // IV_len: Length of the IV of the ESP-NULL packet. // // ICV_len: Length of the ICV of the ESP-NULL packet. // // State: State of the packet, i.e. ESP-NULL, ESP, or // unsure. // // Also following data is taken from the packet: // // IP_total_len: Total IP packet length // IP_hdr_len: Header length of IP packet in bytes // IP_Src_IP: Source address of IP packet // IP_Dst_IP: Destination address of IP packet // // UDP_len: Length of the UDP packet taken from UDP header. // UDP_src_port: Source port of UDP packet. // UDP_dst_port: Destination port of UDP packet. // // SPI: SPI number from ESP packet. // // Protocol: Actual protocol number of the protocol inside // ESP-NULL packet. // Protocol_off: Calculated offset to the protocol payload data // inside ESP-NULL packet. //////////////////////////////////////////////////////////// // This is the main processing code for the packet // This will check if the packet requires ESP processing, // Process packet: * If IP protocol is ESP * Set SPI_offset to 0 bytes * Goto Process ESP * If IP protocol is UDP * Goto Process UDP * Continue Non-ESP processing //////////////////////////////////////////////////////////// // This code is run for UDP packets, and it checks if the // packet is UDP encapsulated UDP packet, or UDP // encapsulated IKE packet, or keepalive packet. // Process UDP: // Reassembly is not mandatory here, we could // do reassembly also only after detecting the // packet being UDP encapsulated ESP packet, but // that would complicated the pseudocode here // a lot, as then we would need to add code // for checking if the UDP header is in this // packet or not. // Reassembly is to simplify things * If packet is fragment * Do full reassembly before processing * If UDP_src_port != 4500 and UDP_dst_port != 4500 * Continue Non-ESP processing * Set SPI_offset to 8 bytes * If UDP_len > 4 and first 4 bytes of UDP packet are 0x000000 * Continue Non-ESP processing (pass IKE-packet) * If UDP_len == 1 and first byte is 0xff * Continue Non-ESP processing (pass NAT-Keepalive Packet) * Goto Process ESP //////////////////////////////////////////////////////////// // This code is run for ESP packets (or UDP encapsulated ESP // packets). This checks if IPsec flow is known, and // if not calls heuristics. If IPsec flow is known // then it continues processing based on the policy. // Process ESP: * If packet is fragment * Do full reassembly before processing * If IP_total_len < IP_hdr_len + SPI_offset + 4 * Drop invalid packet * Load SPI from IP_hdr_len + SPI_offset * Initialize State to ESP // In case this was UDP encapsulated ESP then use UDP_src_port and // UDP_dst_port also when finding data from SPI cache. * Find IP_Src_IP + IP_Dst_IP + SPI from SPI cache * If SPI found * Load State, IV_len, ICV_len from cache * If SPI not found or State is unsure * Call Autodetect ESP parameters (drop to slowpath) * If State is ESP * Continue Non-ESP-NULL processing * Goto Check ESP-NULL packet //////////////////////////////////////////////////////////// // This code is run for ESP-NULL packets, and this // finds out the data required for deep inspection // engine (protocol number, and offset to data) // and calls the deep inspection engine. // Check ESP-NULL packet: * If IP_total_len < IP_hdr_len + SPI_offset + IV_len + ICV_len + 4 (spi) + 4 (seq no) + 4 (protocol + padding) * Drop invalid packet * Load Protocol from IP_total_len - ICV_len - 1 * Set Protocol_off to IP_hdr_len + SPI_offset + IV_len + 4 (spi) + 4 (seq no) * Do normal deep inspection on packet.
| Figure 2 |
The following example pseudocode show the actual heuristics part of the packet processing engine. This part is usually implemented in software.
//////////////////////////////////////////////////////////// // This pseudocode uses following variables: // // SPI_offset, IV_len, ICV_len, State, SPI, // IP_total_len, IP_hdr_len, IP_Src_IP, IP_Dst_IP // as defined in fastpath pseudocode. // // Stored_Check_Bits:Number of bits we have successfully // checked to contain acceptable values // in the actual payload data. This value // is stored / retrieved from SPI cache. // // Check_Bits: Number of bits we have successfully // checked to contain acceptable values // in the actual payload data. This value // is updated during the packet // verification. // // Last_Packet_Data: Contains selected pieces from the // last packet. This is used to compare // certain fields of this packet to // same fields in previous packet. // // Packet_Data: Selected pieces of this packet, same // fields as Last_Packet_Data, and this // is stored as new Last_Packet_Data to // SPI cache after this packet is processed. // // Test_ICV_len: Temporary ICV length used during tests. // This is stored to ICV_len when // padding checks for the packet succeed // and the packet didn't yet have unsure // status. // // Test_IV_len: Temporary IV length used during tests. // // Pad_len: Padding length from the ESP packet. // // Protocol: Protocol number of the packet inside ESP // packet. // // TCP.*: Fields from TCP header (from inside ESP) // UDP.*: Fields from UDP header (from inside ESP) //////////////////////////////////////////////////////////// // This code starts the actual heuristics. // During this the fastpath has already loaded // State, ICV_len and IV_len in case they were // found from the SPI cache (i.e. in case the flow // had unsure status). // Autodetect ESP parameters: // First we check if this is unsure flow, and // if so, we check next packet against the // already set IV/ICV_len combination. * If State is unsure * Call Verify next packet * If State is ESP-NULL * Goto Store ESP-NULL SPI cache info * If State is unsure * Goto Verify unsure // If we failed the test, i.e. State // was changed to ESP, we check other // ICV/IV_len values, i.e. fall through // ICV lengths are tested in order of ICV lengths, // from shortest to longest. * Call Try standard algorithms * If State is ESP-NULL * Goto Store ESP-NULL SPI cache info * Call Try 128bit algorithms * If State is ESP-NULL * Goto Store ESP-NULL SPI cache info * Call Try 160bit algorithms * If State is ESP-NULL * Goto Store ESP-NULL SPI cache info * Call Try 192bit algorithms * If State is ESP-NULL * Goto Store ESP-NULL SPI cache info * Call Try 256bit algorithms * If State is ESP-NULL * Goto Store ESP-NULL SPI cache info // No idea how to test AUTH_DES_MAC and AUTH_KPDK_MD5 as they are // not defined anywhere // If any of those test above set state to unsure // we mark IPsec flow as unsure. * If State is unsure * Goto Store unsure SPI cache info // All of the test failed, meaning the packet cannot // be ESP-NULL packet, thus we mark IPsec flow as ESP * Goto Store ESP SPI cache info //////////////////////////////////////////////////////////// // Store ESP-NULL status to the IPsec flow cache. // Store ESP-NULL SPI cache info: * Store State, IV_len, ICV_len to SPI cache using IP_Src_IP + IP_Dst_IP + SPI as key * Continue Check ESP-NULL packet //////////////////////////////////////////////////////////// // Store encrypted ESP status to the IPsec flow cache. // Store ESP SPI cache info: * Store State, IV_len, ICV_len to SPI cache using IP_Src_IP + IP_Dst_IP + SPI as key * Continue Check non-ESP-NULL packet //////////////////////////////////////////////////////////// // Store unsure flow status to IPsec flow cache. // Here we also store the Check_Bits. // Store unsure SPI cache info: * Store State, IV_len, ICV_len, Stored_Check_Bits to SPI cache using IP_Src_IP + IP_Dst_IP + SPI as key * Contine Check unknown packet //////////////////////////////////////////////////////////// // Verify this packet against the previously selected // ICV_len and IV_len values. This will either // fail (and set state to ESP to mark we do not yet // know what type of flow this is), or it will // increment Check_Bits. // Verify next packet: // We already have IV_len, ICV_len and State loaded * Load Stored_Check_Bits, Last_Packet_Data from SPI Cache * Set Test_ICV_len to ICV_len, Test_IV_len to IV_len * Initialize Check_Bits to 0 * Call Verify padding * If verify padding returned Failure // Initial guess was wrong, restart * Set State to ESP * Clear IV_len, ICV_len, State, Stored_Check_Bits, Last_Packet_Data from SPI Cache * Return // Ok, padding check succeeded again * Call Verify packet * If verify packet returned Failure // Guess was wrong, restart * Set State to ESP * Clear IV_len, ICV_len, State, Stored_Check_Bits, Last_Packet_Data from SPI Cache * Return // It succeeded and updated Check_Bits and Last_Packet_Data store // them to SPI cache * Increment Stored_Check_Bits by Check_Bits * Store Stored_Check_Bits to SPI Cache * Store Packet_Data as Last_Packet_Data to SPI cache * Return //////////////////////////////////////////////////////////// // This will check if we have already seen enough bits // acceptable from the payload data, so we can decide // that this IPsec flow is ESP-NULL flow. // Verify unsure: // Check if we have enough check bits * If Stored_Check_Bits > configured limit // We have checked enough bits, return ESP-NULL * Set State ESP-NULL * Goto Store ESP-NULL SPI cache info // Not yet enough bits, continue * Continue Check unknown packet //////////////////////////////////////////////////////////// // Check for standard 96-bit algorithms. // Try standard algorithms: // AUTH_HMAC_MD5_96, AUTH_HMAC_SHA1_96, AUTH_AES_XCBC_96, // AUTH_AES_CMAC_96 * Set Test_ICV_len to 12, Test_IV_len to 0 * Goto Check packet //////////////////////////////////////////////////////////// // Check for 128-bit algorithms, this is only one that // can have IV, so we need to check different IV_len values // here too. // Try 128bit algorithms: // AUTH_HMAC_MD5_128, AUTH_HMAC_SHA2_256_128 // AUTH_AES_128_GMAC, AUTH_AES_192_GMAC, AUTH_AES_256_GMAC * Set Test_ICV_len to 16, Test_IV_len to 0 * If IP_total_len < IP_hdr_len + SPI_offset + Test_IV_len + Test_ICV_len + 4 (spi) + 4 (seq no) + 4 (protocol + padding) * Return * Call Verify padding * If verify padding returned Failure * Return * Initialize Check_Bits to 0 * Call Verify packet * If verify packet returned Failure * Goto Try GMAC // Ok, packet seemed ok, but go now and check if we have enough // data bits so we can assume it is ESP-NULL * Goto Check if done for unsure //////////////////////////////////////////////////////////// // Check for GMAC macs, i.e. macs having 8 byte IV. // Try GMAC: // AUTH_AES_128_GMAC, AUTH_AES_192_GMAC, AUTH_AES_256_GMAC * Set Test_IV_len to 8 * If IP_total_len < IP_hdr_len + SPI_offset + Test_IV_len + Test_ICV_len + 4 (spi) + 4 (seq no) + 4 (protocol + padding) * Return * Initialize Check_Bits to 0 * Call Verify packet * If verify packet returned Failure // Guess was wrong, continue * Return // Ok, packet seemed ok, but go now and check if we have enough // data bits so we can assume it is ESP-NULL * Goto Check if done for unsure //////////////////////////////////////////////////////////// // Check for 160-bit algorithms. // Try 160bit algorithms: // AUTH_HMAC_SHA1_160 * Set Test_ICV_len to 20, Test_IV_len to 0 * Goto Check packet //////////////////////////////////////////////////////////// // Check for 192-bit algorithms. // Try 192bit algorithms: // AUTH_HMAC_SHA2_384_192 * Set Test_ICV_len to 24, Test_IV_len to 0 * Goto Check packet //////////////////////////////////////////////////////////// // Check for 256-bit algorithms. // Try 256bit algorithms: // AUTH_HMAC_SHA2_512_256 * Set Test_ICV_len to 32, Test_IV_len to 0 * Goto Check packet //////////////////////////////////////////////////////////// // This actually does the checking for the packet, by // first verifying the length, and then self describing // padding, and if that succeeds, then checks the actual // payload content. // Check packet: * If IP_total_len < IP_hdr_len + SPI_offset + Test_IV_len + Test_ICV_len + 4 (spi) + 4 (seq no) + 4 (protocol + padding) * Return * Call Verify padding * If verify padding returned Failure * Return * Initialize Check_Bits to 0 * Call Verify packet * If verify packet returned Failure // Guess was wrong, continue * Return // Ok, packet seemed ok, but go now and check if we have enough // data bits so we can assume it is ESP-NULL * Goto Check if done for unsure //////////////////////////////////////////////////////////// // This code checks if we have seen enough acceptable // values in the payload data, so we can decide that this // IPsec flow is ESP-NULL flow. // Check if done for unsure: * If Stored_Check_Bits > configured limit // We have checked enough bits, return ESP-NULL * Set State ESP-NULL * Set IV_len to Test_IV_len, ICV_len to Test_ICV_len * Clear Stored_Check_Bits, Last_Packet_Data from SPI Cache * Return // Not yet enough bits, check this is first unsure, if so // store information. In case there is multiple // tests succeeding, we always assume the first one // (the wone using shortest MAC) is the one we want to // check in the future. * If State is not unsure * Set State unsure // These values will be stored to SPI cache if // the final state will be unsure * Set IV_len to Test_IV_len, ICV_len to Test_ICV_len * Set Stored_Check_Bits as Check_Bits * Return //////////////////////////////////////////////////////////// // Verify self describing padding // Verify padding: * Load Pad_len from IP_total_len - Test_ICV_len - 2 * Verify padding bytes at IP_total_len - Test_ICV_len - 1 - Pad_len .. IP_total_len - Test_ICV_len - 2 are 1, 2, ..., Pad_len * If Verify of padding bytes succeeded * Return Success * Return Failure //////////////////////////////////////////////////////////// // This will verify the actual protocol content inside ESP // packet. // Verify packet: // We need to first check things that cannot be set, i.e if any of // those are incorrect, then we return Failure. For any / fields which might be correct, we increment the Check_Bits // for a suitable amount of bits. If all checks pass, then // we just return Success, and the upper layer will then // later check if we have enough bits checked already. * Load Protocol From IP_total_len - Test_ICV_len - 1 * If Protocol TCP * Goto Verify TCP * If Protocol UDP * Goto Verify UDP // Other protocols can be added here as needed, most likely same // protocols as deep inspection does // Tunnel mode checks is also left out from here to make the // document shorter. * Return Failure //////////////////////////////////////////////////////////// // Verify TCP protocol headers // Verify TCP: // First we check things that must be set correctly. * Check TCP.reserved_bits are non-zero * Return Failure * If TCP.Data_Offset field < 5 // TCP head length too small * Return Failure // After that we start to check things that does not // have one definitive value, but can have multiple possible // valid values * If TCP.ACK bit is not set, then check that TCP.Acknowledgment_number field contains 0 // If ACK bit is not set then the acknowledgment // field usually contains 0, but I do not think // RFCs mandate it being zero, so we cannot make // this a failure if it is not so. * Increment Check_Bits by 32 * If TCP.URG bit is not set, then check that TCP.Urgent_Pointer field contains 0 // If URG bit is not set then urgent pointer // field usually contains 0, but I do not think // RFCs mandate it being zero, so we cannot make // this failure if it is not so. * Increment Check_Bits by 16 * If TCP.Data_Offset field == 5 * Increment Check_Bits by 4 * If TCP.Data_Offset field > 5 * If TCP options format is valid and it is padded correctly * Increment Check_Bits accordingly * If TCP options format was garbage * Return Failure * If TCP.checksum is correct // This might be wrong because packet passed NAT, so // we cannot make this failure case * Increment Check_Bits by 16 // We can also do normal deeper TCP inspection here, i.e. // check that SYN/ACK/FIN/RST bits are correct and state // matches the state of existing flow if this is packet // to existing flow etc. // If there is anything clearly wrong in the packet (i.e. // some data is set to something that it cannot be), then // this can return Failure, otherwise it should just // increment Check_Bits matching the number of bits checked. // // We can also check things here compared to the last packet * If Last_Packet_Data.TCP.source port = Packet_Data.TCP.source_port and Last_Packet_Data.TCP.destination port = Packet_Data.TCP.destination port * Increment Check_Bits by 32 * If Last_Packet_Data.TCP.acknowledgement_number = Packet_Data.TCP.acknowledgement_number * Increment Check_Bits by 32 * If Last_Packet_Data.TCP.sequence_number = Packet_Data.TCP.sequence_number * Increment Check_Bits by 32 // We can do other similar checks here * Return Success //////////////////////////////////////////////////////////// // Verify UDP protocol headers // Verify UDP: // First we check things that must be set correctly. * If UDP.UDP_length > IP_total_len - IP_hdr_len - SPI_offset - Test_IV_len - Test_ICV_len - 4 (spi) - 4 (seq no) - 1 (protocol) - Pad_len - 1 (Pad_len) * Return Failure * If UDP.UDP_length < 8 * Return Failure // After that we start to check things that does not // have one definitive value, but can have multiple possible // valid values * If UDP.UDP_checksum is correct // This might be wrong because packet passed NAT, so // we cannot make this failure case * Increment Check_Bits by 16 * If UDP.UDP_length = IP_total_len - IP_hdr_len - SPI_offset - Test_IV_len - Test_ICV_len - 4 (spi) - 4 (seq no) - 1 (protocol) - Pad_len - 1 (Pad_len) // If there is no TFC padding then UDP_length // will be matching the full packet length * Increment Check_Bits by 16 // We can also do normal deeper UDP inspection here. // If there is anything clearly wrong in the packet (i.e. // some data is set to something that it cannot be), then // this can return Failure, otherwise it should just // increment Check_Bits matching the number of bits checked. // // We can also check things here compared to the last packet * If Last_Packet_Data.UDP.source_port = Packet_Data.UDP.source_port and Last_Packet_Data.destination_port = Packet_Data.UDP.destination_port * Increment Check_Bits by 32 * Return Success
| Figure 3 |
|Daniel L. McDonald|
|Sun Microsystems, Inc.|
|35 Network Drive|
|Burlington, MA 01803|