Internet Engineering Task Force (IETF) A. Ramaiah Request for Comments: 5961 Cisco Category: Standards Track R. Stewart ISSN: 2070-1721 Huawei M. Dalal Cisco August 2010
Improving TCP's Robustness to Blind In-Window Attacks
Abstract
TCP has historically been considered to be protected against spoofed off-path packet injection attacks by relying on the fact that it is difficult to guess the 4-tuple (the source and destination IP addresses and the source and destination ports) in combination with the 32-bit sequence number(s). A combination of increasing window sizes and applications using longer-term connections (e.g., H-323 or Border Gateway Protocol (BGP) [RFC4271]) have left modern TCP implementations more vulnerable to these types of spoofed packet injection attacks.
Many of these long-term TCP applications tend to have predictable IP addresses and ports that makes it far easier for the 4-tuple (4-tuple is the same as the socket pair mentioned in RFC 793) to be guessed. Having guessed the 4-tuple correctly, an attacker can inject a TCP segment with the RST bit set, the SYN bit set or data into a TCP connection by systematically guessing the sequence number of the spoofed segment to be in the current receive window. This can cause the connection to abort or cause data corruption. This document specifies small modifications to the way TCP handles inbound segments that can reduce the chances of a successful attack.
Status of This Memo
This is an Internet Standards Track document.
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). Further information on Internet Standards is available in Section 2 of RFC 5741.
Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at http://www.rfc-editor.org/info/rfc5961.
Ramaiah, et al. Standards Track [Page 1]
RFC 5961 TCP Security August 2010
Copyright Notice
Copyright (c) 2010 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. 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.
Table of Contents
1. Introduction ....................................................3 1.1. Applicability Statement ....................................3 1.2. Basic Attack Methodology ...................................4 1.3. Attack probabilities .......................................5 2. Terminology .....................................................7 3. Blind Reset Attack Using the RST Bit ............................7 3.1. Description of the Attack ..................................7 3.2. Mitigation .................................................7 4. Blind Reset Attack Using the SYN Bit ............................9 4.1. Description of the Attack ..................................9 4.2. Mitigation .................................................9 5. Blind Data Injection Attack ....................................10 5.1. Description of the Attack .................................10 5.2. Mitigation ................................................11 6. Suggested Mitigation Strengths .................................12 7. ACK Throttling .................................................12 8. Backward Compatibility and Other Considerations ................13 9. Middlebox Considerations .......................................14 9.1. Middlebox That Resend RSTs ................................14 9.2. Middleboxes That Advance Sequence Numbers .................15 9.3. Middleboxes That Drop the Challenge ACK ...................15 10. Security Considerations .......................................16 11. Contributors ..................................................17 12. Acknowledgments ...............................................17 13. References ....................................................17 13.1. Normative References .....................................17 13.2. Informative References ...................................17
TCP [RFC0793] is widely deployed and the most common reliable end-to- end transport protocol used for data communication in today's Internet. Yet, when it was standardized over 20 years ago, the Internet was a different place, lacking many of the threats that are now common. The off-path TCP spoofing attacks, which are seen in the Internet today, fall into this category.
In a TCP spoofing attack, an off-path attacker crafts TCP packets by forging the IP source and destination addresses as well as the source and destination ports (referred to as a 4-tuple value in this document). The targeted TCP endpoint will then associate such a packet with an existing TCP connection. It needs to be noted that, guessing this 4-tuple value is not always easy for an attacker. But there are some applications (e.g., BGP [RFC4271]) that have a tendency to use the same set(s) of ports on either endpoint, making the odds of correctly guessing the 4-tuple value much easier. When an attacker is successful in guessing the 4-tuple value, one of three types of injection attacks may be waged against a long-lived connection. RST - Where an attacker injects a RST segment hoping to cause the connection to be torn down. "RST segment" here refers to a TCP segment with the RST bit set. SYN - Where an attacker injects a SYN hoping to cause the receiver to believe the peer has restarted and therefore tear down the connection state. "SYN segment" here refers to a TCP segment with SYN bit set. DATA - Where an attacker tries to inject a DATA segment to corrupt the contents of the transmission. "DATA segment" here refers to any TCP segment containing data.
This document talks about some known in-window attacks and suitable defenses against these. The mitigations suggested in this document SHOULD be implemented in devices that regularly need to maintain TCP connections of the kind most vulnerable to the attacks described in this document. Examples of such TCP connections are the ones that tend to be long-lived and where the connection endpoints can be determined, in cases where no auxiliary anti-spoofing protection mechanisms like TCP MD5 [RFC2385] can be deployed. These mitigations MAY be implemented in other cases.
Focusing upon the RST attack, we examine this attack in more detail to get an overview as to how it works and how this document addresses the issue. For this attack, the goal is for the attacker to cause one of the two endpoints of the connection to incorrectly tear down the connection state, effectively aborting the connection. One of the important things to note is that for the attack to succeed the RST needs to be in the valid receive window. It also needs to be emphasized that the receive window is independent of the current congestion window of the TCP connection. The attacker would try to forge many RST segments to try to cover the space of possible windows by putting out a packet in each potential window. To do this, the attacker needs to have or guess several pieces of information namely:
1) The 4-tuple value containing the IP address and TCP port number of both ends of the connection. For one side (usually the server), guessing the port number is a trivial exercise. The client side may or may not be easy for an attacker to guess depending on a number of factors, most notably the operating system and application involved.
2) A sequence number that will be used in the RST. This sequence number will be a starting point for a series of guesses to attempt to present a RST segment to a connection endpoint that would be acceptable to it. Any random value may be used to guess the starting sequence number.
3) The window size that the two endpoints are using. This value does NOT have to be the exact window size since a smaller value used in lieu of the correct one will just cause the attacker to generate more segments before succeeding in his mischief. Most modern operating systems have a default window size that usually is applied to most connections. Some applications however may change the window size to better suit the needs of the application. So often times the attacker, with a fair degree of certainty (knowing the application that is under attack), can come up with a very close approximation as to the actual window size in use on the connection.
After assembling the above set of information, the attacker begins sending spoofed TCP segments with the RST bit set and a guessed TCP sequence number. Each time a new RST segment is sent, the sequence number guess is incremented by the window size. The feasibility of this methodology (without mitigations) was first shown in [SITW]. This is because [RFC0793] specifies that any RST within the current window is acceptable. Also, [RFC4953] talks about the probability of a successful attack with varying window sizes and bandwidth.
Ramaiah, et al. Standards Track [Page 4]
RFC 5961 TCP Security August 2010
A slight enhancement to TCP's segment processing rules can be made, which makes such an attack much more difficult to accomplish. If the receiver examines the incoming RST segment and validates that the sequence number exactly matches the sequence number that is next expected, then such an attack becomes much more difficult than outlined in [SITW] (i.e., the attacker would have to generate 1/2 the entire sequence space, on average). This document will discuss the exact details of what needs to be changed within TCP's segment processing rules to mitigate all three types of attacks (RST, SYN, and DATA).
Every application has control of a number of factors that drastically affect the probability of a successful spoofing attack. These factors include such things as:
Window Size - Normally settable by the application but often times defaulting to 32,768 or 65,535 depending upon the operating system (see Figure 6 of [Medina05]).
Server Port number - This value is normally a fixed value so that a client will know where to connect to the peer. Thus, this value normally provides no additional protection.
Client Port number - This value may be a random ephemeral value, if so, this makes a spoofing attack more difficult. There are some clients, however, that for whatever reason either pick a fixed client port or have a very guessable one (due to the range of ephemeral ports available with their operating system or other application considerations) for such applications a spoofing attack becomes less difficult.
For the purposes of the rest of this discussion we will assume that the attacker knows the 4-tuple values. This assumption will help us focus on the effects of the window size versus the number of TCP packets an attacker must generate. This assumption will rarely be true in the real Internet since at least the client port number will provide us with some amount of randomness (depending on the operating system).
To successfully inject a spoofed packet (RST, SYN, or DATA), in the past, the entire sequence space (i.e., 2^32) was often considered available to make such an attack unlikely. [SITW] demonstrated that this assumption was incorrect and that instead of (1/2 * 2^32) packets (assuming a random distribution), (1/2 * (2^32/window))
Ramaiah, et al. Standards Track [Page 5]
RFC 5961 TCP Security August 2010
packets are required. In other words, the mean number of tries needed to inject a RST segment is (2^31/window) rather than the 2^31 assumed before.
Substituting numbers into this formula, we see that for a window size of 32,768, an average of 65,536 packets would need to be transmitted in order to "spoof" a TCP segment that would be acceptable to a TCP receiver. A window size of 65,535 reduces this even further to 32,768 packets. At today's access bandwidths, an attack of that size is feasible.
With rises in bandwidth to both the home and office, it can only be expected that the values for default window sizes will continue to rise in order to better take advantage of the newly available bandwidth. It also needs to be noted that this attack can be performed in a distributed fashion in order potentially gain access to more bandwidth.
As we can see from the above discussion this weakness lowers the bar quite considerably for likely attacks. But there is one additional dependency that is the duration of the TCP connection. A TCP connection that lasts only a few brief packets, as often is the case for web traffic, would not be subject to such an attack since the connection may not be established long enough for an attacker to generate enough traffic. However, there is a set of applications, such as BGP [RFC4271], that is judged to be potentially most affected by this vulnerability. BGP relies on a persistent TCP session between BGP peers. Resetting the connection can result in term- medium unavailability due to the need to rebuild routing tables and route flapping; see [NISCC] for further details.
For applications that can use the TCP MD5 option [RFC2385], such as BGP, that option makes the attacks described in this specification effectively impossible. However, some applications or implementations may find that option expensive to implement.
There are alternative protections against the threats that this document addresses. For further details regarding the attacks and the existing techniques, please refer to [RFC4953]. It also needs to be emphasized that, as suggested in [TSVWG-PORT] and [RFC1948], port randomization and initial sequence number (ISN) randomization would help improve the robustness of the TCP connection against off-path attacks.
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]. TCP terminology should be interpreted as described in [RFC0793].
As described in the introduction, it is possible for an attacker to generate a RST segment that would be acceptable to a TCP receiver by guessing in-window sequence numbers. In particular [RFC0793], page 37, states the following:
In all states except SYN-SENT, all reset (RST) segments are validated by checking their SEQ-fields [sequence numbers]. A reset is valid if its sequence number is in the window. In the SYN-SENT state (a RST received in response to an initial SYN), the RST is acceptable if the ACK field acknowledges the SYN.
[RFC0793] currently requires handling of a segment with the RST bit when in a synchronized state to be processed as follows:
1) If the RST bit is set and the sequence number is outside the current receive window (SEG.SEQ <= RCV.NXT || SEG.SEQ > RCV.NXT+ RCV.WND), silently drop the segment.
2) If the RST bit is set and the sequence number is acceptable, i.e., (RCV.NXT <= SEG.SEQ < RCV.NXT+RCV.WND), then reset the connection.
Instead, implementations SHOULD implement the following steps in place of those specified in [RFC0793] (as listed above).
1) If the RST bit is set and the sequence number is outside the current receive window, silently drop the segment.
2) If the RST bit is set and the sequence number exactly matches the next expected sequence number (RCV.NXT), then TCP MUST reset the connection.
Ramaiah, et al. Standards Track [Page 7]
RFC 5961 TCP Security August 2010
3) If the RST bit is set and the sequence number does not exactly match the next expected sequence value, yet is within the current receive window (RCV.NXT < SEG.SEQ < RCV.NXT+RCV.WND), TCP MUST send an acknowledgment (challenge ACK):
<SEQ=SND.NXT><ACK=RCV.NXT><CTL=ACK>
After sending the challenge ACK, TCP MUST drop the unacceptable segment and stop processing the incoming packet further. Further segments destined to this connection will be processed as normal.
The modified RST segment processing would thus become:
In all states except SYN-SENT, all reset (RST) segments are validated by checking their SEQ-fields [sequence numbers]. A reset is valid if its sequence number exactly matches the next expected sequence number. If the RST arrives and its sequence number field does NOT match the next expected sequence number but is within the window, then the receiver should generate an ACK. In all other cases, where the SEQ-field does not match and is outside the window, the receiver MUST silently discard the segment.
In the SYN-SENT state (a RST received in response to an initial SYN), the RST is acceptable if the ACK field acknowledges the SYN. In all other cases the receiver MUST silently discard the segment.
With the above slight change to the TCP state machine, it becomes much harder for an attacker to generate an acceptable reset segment.
In cases where the remote peer did generate a RST, but it fails to meet the above criteria (the RST sequence number was within the window but NOT the exact expected sequence number), when the challenge ACK is sent back, it will no longer have the transmission control block (TCB) related to this connection and hence as per [RFC0793], the remote peer will send a second RST back. The sequence number of the second RST is derived from the acknowledgment number of the incoming ACK. This second RST, if it reaches the sender, will cause the connection to be aborted since the sequence number would now be an exact match.
A valid RST received out of order would still generate a challenge ACK in response. If this RST happens to be a genuine one, the other end would send an RST with an exact sequence number match that would cause the connection to be dropped.
Note that the above mitigation may cause a non-amplification ACK exchange. This concern is discussed in Section 10.
The analysis of the reset attack using the RST bit highlights another possible avenue for a blind attacker using a similar set of sequence number guessing. Instead of using the RST bit, an attacker can use the SYN bit with the exact same semantics to tear down a connection.
[RFC0793] currently requires handling of a segment with the SYN bit set in the synchronized state to be as follows:
1) If the SYN bit is set and the sequence number is outside the expected window, send an ACK back to the sender.
2) If the SYN bit is set and the sequence number is acceptable, i.e., (RCV.NXT <= SEG.SEQ < RCV.NXT+RCV.WND), then send a RST segment to the sender.
Instead, the handling of the SYN in the synchronized state SHOULD be performed as follows:
1) If the SYN bit is set, irrespective of the sequence number, TCP MUST send an ACK (also referred to as challenge ACK) to the remote peer:
<SEQ=SND.NXT><ACK=RCV.NXT><CTL=ACK>
After sending the acknowledgment, TCP MUST drop the unacceptable segment and stop processing further.
By sending an ACK, the remote peer is challenged to confirm the loss of the previous connection and the request to start a new connection. A legitimate peer, after restart, would not have a TCB in the synchronized state. Thus, when the ACK arrives, the peer should send a RST segment back with the sequence number derived from the ACK field that caused the RST.
This RST will confirm that the remote peer has indeed closed the previous connection. Upon receipt of a valid RST, the local TCP endpoint MUST terminate its connection. The local TCP endpoint should then rely on SYN retransmission from the remote end to re-establish the connection.
Ramaiah, et al. Standards Track [Page 9]
RFC 5961 TCP Security August 2010
A spoofed SYN, on the other hand, will then have generated an additional ACK that the peer will discard as a duplicate ACK and will not affect the established connection.
Note that this mitigation does leave one corner case un-handled, which will prevent the reset of a connection when it should be reset (i.e., it is a non-spoofed SYN wherein a peer really did restart). This problem occurs when the restarting host chooses the exact same IP address and port number that it was using prior to its restart. By chance, the restarted host must also choose an initial sequence number of exactly (RCV.NXT - 1) of the remote peer that is still in the established state. Such a case would cause the receiver to generate a "challenge" ACK as described above. But since the ACK would be within the outgoing connections window, the inbound ACK would be acceptable, and the sender of the SYN will do nothing with the response ACK. This sequence will continue as the SYN sender continually times out and retransmits the SYN until such time as the connection attempt fails.
This corner case is a result of the [RFC0793] specification and is not introduced by these new requirements.
Note that the above mitigation may cause a non-amplification ACK exchange. This concern is discussed in Section 10.
A third type of attack is also highlighted by both the RST and SYN attacks. It is also possible to inject data into a TCP connection by simply guessing a sequence number within the current receive window of the victim. The ACK value of any data segment is considered valid as long as it does not acknowledge data ahead of the next segment to send. In other words, an ACK value is acceptable if it is ((SND.UNA-(2^31-1)) <= SEG.ACK <= SND.NXT). The (2^31 - 1) in the above inequality takes into account the fact that comparisons on TCP sequence and acknowledgment numbers is done using the modulo 32-bit arithmetic to accommodate the number wraparound. This means that an attacker has to guess two ACK values with every guessed sequence number so that the chances of successfully injecting data into a connection are 1 in ( 1/2 (2^32 / RCV.WND) * 2). Thus, the mean number of tries needed to inject data successfully is 1/2 (2*2^32/RWND) = 2^32/RCV.WND.
Ramaiah, et al. Standards Track [Page 10]
RFC 5961 TCP Security August 2010
When an attacker successfully injects data into a connection, the data will sit in the receiver's re-assembly queue until the peer sends enough data to bridge the gap between the RCV.NXT value and the injected data. At that point, one of two things will occur:
1) A packet war will ensue with the receiver indicating that it has received data up until RCV.NXT (which includes the attacker's data) and the sender sending an ACK with an acknowledgment number less than RCV.NXT.
2) The sender will send enough data to the peer that will move RCV.NXT even further along past the injected data.
Depending upon the TCP implementation in question and the TCP traffic characteristics at that time, data corruption may result. In case (a), the connection will eventually be reset by one of the sides unless the sender produces more data that will transform the ACK war into case (b). The reset will usually occur via User Time Out (UTO) (see section 4.2.3.5 of [RFC1122]).
Note that the protections illustrated in this section neither cause an ACK war nor prevent one from occurring if data is actually injected into a connection. The ACK war is a product of the attack itself and cannot be prevented (other than by preventing the data from being injected).
All TCP stacks MAY implement the following mitigation. TCP stacks that implement this mitigation MUST add an additional input check to any incoming segment. The ACK value is considered acceptable only if it is in the range of ((SND.UNA - MAX.SND.WND) <= SEG.ACK <= SND.NXT). All incoming segments whose ACK value doesn't satisfy the above condition MUST be discarded and an ACK sent back. It needs to be noted that RFC 793 on page 72 (fifth check) says: "If the ACK is a duplicate (SEG.ACK < SND.UNA), it can be ignored. If the ACK acknowledges something not yet sent (SEG.ACK > SND.NXT) then send an ACK, drop the segment, and return". The "ignored" above implies that the processing of the incoming data segment continues, which means the ACK value is treated as acceptable. This mitigation makes the ACK check more stringent since any ACK < SND.UNA wouldn't be accepted, instead only ACKs that are in the range ((SND.UNA - MAX.SND.WND) <= SEG.ACK <= SND.NXT) get through.
A new state variable MAX.SND.WND is defined as the largest window that the local sender has ever received from its peer. This window may be scaled to a value larger than 65,535 bytes ([RFC1323]). This small check will reduce the vulnerability to an attacker guessing a
Ramaiah, et al. Standards Track [Page 11]
RFC 5961 TCP Security August 2010
valid sequence number, since, not only one must guess the in-window sequence number, but also guess a proper ACK value within a scoped range. This mitigation reduces, but does not eliminate, the ability to generate false segments. It does however reduce the probability that invalid data will be injected.
Implementations can also chose to hard code the MAX.SND.WND value to the maximum permissible window size, i.e., 65535 in the absence of window scaling. In the presence of the window scaling option, the value becomes (MAX.SND.WND << Snd.Wind.Scale).
This mitigation also helps in improving robustness on accepting spoofed FIN segments (FIN attacks). Among other things, this mitigation requires that the attacker also needs to get the acknowledgment number to fall in the range mentioned above in order to successfully spoof a FIN segment leading to the closure of the connection. Thus, this mitigation greatly improves the robustness to spoofed FIN segments.
Note that the above mitigation may cause a non-amplification ACK exchange. This concern is discussed in Section 10.
As described in the above sections, recommendation levels for RST, SYN, and DATA are tagged as SHOULD, SHOULD, and MAY, respectively. The reason that DATA mitigation is tagged as MAY, even though it increased the TCP robustness in general is because, the DATA injection is perceived to be more difficult (twice as unlikely) when compared to RST and SYN counterparts. However, it needs to be noted that all the suggested mitigations improve TCP's robustness in general and hence the choice of implementing some or all mitigations recommended in the document is purely left to the implementer.
In order to alleviate multiple RSTs/SYNs from triggering multiple challenge ACKs, an ACK throttling mechanism is suggested as follows:
1) The system administrator can configure the number of challenge ACKs that can be sent out in a given interval. For example, in any 5 second window, no more than 10 challenge ACKs should be sent.
2) The values for both the time and number of ACKs SHOULD be tunable by the system administrator to accommodate different perceived levels of threat and/or system resources.
Ramaiah, et al. Standards Track [Page 12]
RFC 5961 TCP Security August 2010
It should be noted that these numbers are empirical in nature and have been obtained from the RST throttling mechanisms existing in some implementations. Also, note that no timer is needed to implement the above mechanism, instead a timestamp and a counter can be used.
An implementation SHOULD include an ACK throttling mechanism to be conservative. While we have not encountered a case where the lack of ACK throttling can be exploited, as a fail-safe mechanism we recommend its use. An implementation may take an excessive number of invocations of the throttling mechanism as an indication that network conditions are unusual or hostile.
An administrator who is more concerned about protecting his bandwidth and CPU utilization may set smaller ACK throttling values whereas an administrator who is more interested in faster cleanup of stale connections (i.e., concerned about excess TCP state) may decide to set a higher value thus allowing more RST's to be processed in any given time period.
The time limit SHOULD be tunable to help timeout brute force attacks faster than a potential legitimate flood of RSTs.
8. Backward Compatibility and Other Considerations
All of the new required mitigation techniques in this document are totally compatible with existing ([RFC0793]) compliant TCP implementations as this document introduces no new assumptions or conditions.
There is a corner scenario in the above mitigations that will require more than one round-trip time to successfully abort the connection as per the figure below. This scenario is similar to the one in which the original RST was lost in the network.
Ramaiah, et al. Standards Track [Page 13]
RFC 5961 TCP Security August 2010
TCP A TCP B 1.a. ESTAB <-- <SEQ=300><ACK=101><CTL=ACK><DATA> <-- ESTAB b. (delayed) ... <SEQ=400><ACK=101><CTL=ACK><DATA> <-- ESTAB c. (in flight) ... <SEQ=500><ACK=101><CTL=RST> <-- CLOSED 2. ESTAB --> <SEQ=101><ACK=400><CTL=ACK> --> CLOSED (ACK for 1.a) ... <SEQ=400><ACK=0><CTL=RST> <-- CLOSED 3. CHALLENGE --> <SEQ=101><ACK=400><CTL=ACK> --> CLOSED (for 1.c) ... <SEQ=400><ACK=0><CTL=RST> <-- RESPONSE 4.a. ESTAB <-- <SEQ=400><ACK=101><CTL=ACK><DATA> 1.b reaches A b. ESTAB --> <SEQ=101><ACK=500><CTL=ACK> c. (in flight) ... <SEQ=500><ACK=0><CTL=RST> <-- CLOSED 5. RESPONSE arrives at A, but dropped since its outside of window. 6. ESTAB <-- <SEQ=500><ACK=0><CTL=RST> 4.c reaches A 7. CLOSED CLOSED
For the mitigation to be maximally effective against the vulnerabilities discussed in this document, both ends of the TCP connection need to have the fix. Although, having the mitigations at one end might prevent that end from being exposed to the attack, the connection is still vulnerable at the other end.
Consider a middlebox M-B tracking connections between two TCP end hosts E-A and E-C. If E-C sends a RST with a sequence number that is within the window but not an exact match to reset the connection and M-B does not have the fix recommended in this document, it may clear the connection and forward the RST to E-A saving an incorrect sequence number. If E-A does not have the fix, the connection would get cleared as required. However, if E-A does have the required fix, it will send a challenge ACK to E-C. M-B, being a middlebox, may intercept this ACK and resend the RST on behalf of E-C with the old sequence number. This RST will, again, not be acceptable and may trigger a challenge ACK.
The above situation may result in a RST/ACK war. However, we believe that if such a case exists in the Internet, the middlebox is generating packets a conformant TCP endpoint would not generate. [RFC0793] dictates that the sequence number of a RST has to be derived from the acknowledgment number of the incoming ACK segment. It is outside the scope of this document to suggest mitigations to the ill-behaved middleboxes.
Ramaiah, et al. Standards Track [Page 14]
RFC 5961 TCP Security August 2010
Consider a similar scenario where the RST from M-B to E-A gets lost, E-A will continue to hold the connection and E-A might send an ACK an arbitrary time later after the connection state was destroyed at M-B. For this case, M-B will have to cache the RST for an arbitrary amount of time until it is confirmed that the connection has been cleared at E-A.
Some middleboxes may compute RST sequence numbers at the higher end of the acceptable window. The scenario is the same as the earlier case, but in this case instead of sending the cached RST, the middlebox (M-B) sends a RST that computes its sequence number as the sum of the acknowledgment field in the ACK and the window advertised by the ACK that was sent by E-A to challenge the RST as depicted below. The difference in the sequence numbers between step 1 and 2 below is due to data lost in the network.
TCP A Middlebox
1. ESTABLISHED <-- <SEQ=500><ACK=100><CTL=RST> <-- CLOSED
2. ESTABLISHED --> <SEQ=100><ACK=300><WND=500><CTL=ACK> --> CLOSED
3. ESTABLISHED <-- <SEQ=800><ACK=100><CTL=RST> <-- CLOSED
4. ESTABLISHED --> <SEQ=100><ACK=300><WND=500><CTL=ACK> --> CLOSED
5. ESTABLISHED <-- <SEQ=800><ACK=100><CTL=RST> <-- CLOSED
Although the authors are not aware of an implementation that does the above, it could be mitigated by implementing the ACK throttling mechanism described earlier.
It also needs to be noted that, some middleboxes (Firewalls/NATs) that don't have the fix recommended in the document, may drop the challenge ACK. This can happen because, the original RST segment that was in window had already cleared the flow state pertaining to the TCP connection in the middlebox. In such cases, the end hosts that have implemented the RST mitigation described in this document, will have the TCP connection left open. This is a corner case and can go away if the middlebox is conformant with the changes proposed in this document.
These changes to the TCP state machine do NOT protect an implementation from on-path attacks. It also needs to be emphasized that while mitigations within this document make it harder for off- path attackers to inject segments, it does NOT make it impossible. The only way to fully protect a TCP connection from both on- and off- path attacks is by using either IPsec Authentication Header (AH) [RFC4302] or IPsec Encapsulating Security Payload (ESP) [RFC4303].
Implementers also should be aware that the attacks detailed in this specification are not the only attacks available to an off-path attacker and that the counter measures described herein are not a comprehensive defense against such attacks.
In particular, administrators should be aware that forged ICMP messages provide off-path attackers the opportunity to disrupt connections or degrade service. Such attacks may be subject to even less scrutiny than the TCP attacks addressed here, especially in stacks not tuned for hostile environments. It is important to note that some ICMP messages, validated or not, are key to the proper function of TCP. Those ICMP messages used to properly set the path maximum transmission unit are the most obvious example. There are a variety of ways to choose which, if any, ICMP messages to trust in the presence of off-path attackers and choosing between them depends on the assumptions and guarantees developers and administrators can make about their network. This specification does not attempt to do more than note this and related issues. Unless implementers address spoofed ICMP messages [RFC5927], the mitigations specified in this document may not provide the desired protection level.
In any case, this RFC details only part of a complete strategy to prevent off-path attackers from disrupting services that use TCP. Administrators and implementers should consider the other attack vectors and determine appropriate mitigations in securing their systems.
Another notable consideration is that a reflector attack is possible with the required RST/SYN mitigation techniques. In this attack, an off-path attacker can cause a victim to send an ACK segment for each spoofed RST/SYN segment that lies within the current receive window of the victim. It should be noted, however, that this does not cause any amplification since the attacker must generate a segment for each one that the victim will generate.
Mitesh Dalal and Amol Khare of Cisco Systems came up with the solution for the RST/SYN attacks. Anantha Ramaiah and Randall Stewart of Cisco Systems discovered the data injection vulnerability and together with Patrick Mahan and Peter Lei of Cisco Systems found solutions for the same. Paul Goyette, Mark Baushke, Frank Kastenholz, Art Stine, and David Wang of Juniper Networks provided the insight that apart from RSTs, SYNs could also result in formidable attacks. Shrirang Bage of Cisco Systems, Qing Li and Preety Puri of Wind River Systems, and Xiaodan Tang of QNX Software along with the folks above helped in ratifying and testing the interoperability of the suggested solutions.
Special thanks to Mark Allman, Ted Faber, Steve Bellovin, Vern Paxson, Allison Mankin, Sharad Ahlawat, Damir Rajnovic, John Wong, Joe Touch, Alfred Hoenes, Andre Oppermann, Fernando Gont, Sandra Murphy, Brian Carpenter, Cullen Jennings, and other members of the tcpm WG for suggestions and comments. ACK throttling was introduced to this document by combining the suggestions from the tcpm working group.
[Medina05] Medina, A., Allman, M., and S. Floyd, "Measuring the Evolution of Transport Protocols in the Internet", ACM Computer Communication Review, 35(2), April 2005, <http://www.icir.org/mallman/papers/tcp-evo-ccr05.ps>.