| TOC |
|
By submitting this Internet-Draft, each author represents that any applicable patent or other IPR claims of which he or she is aware have been or will be disclosed, and any of which he or she becomes aware will be disclosed, in accordance with Section 6 of 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 August 20, 2007.
Copyright © The IETF Trust (2007).
The honesty of TCP senders and receivers has become a major concern to the Internet community. Currently, TCP senders rely on receiver honesty so they can correctly react to network congestion. Such honesty cannot be taken for granted. Receivers may conceal dropped packets to prevent their flow being subject to a congestion response or may acknowledge data optimistically to get a higher bandwidth. This document introduces a simple two-stage test of receiver honesty. Once a receiver fails the first stage it can be subjected to the second stage test that conclusively proves cheating. The performance hit of the first stage is very slight compared to the second. So, although the first stage is not decisive, it selects which receivers are acting suspiciously enough to warrant the second stage. This specification does not modify the TCP protocol - the tests only require a change to sender implementations.
By submitting this Internet-Draft, each author represents that any applicable patent or other IPR claims of which he or she is aware have been or will be disclosed, and any of which he or she becomes aware will be disclosed, in accordance with Section 6 of 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.
1.
Introduction
2.
Requirements notation
3.
The Problems
3.1.
Concealing Lost Segments
3.2.
Optimistic Acknowledgements
4.
Requirements for a robust solution
5.
Existing Proposals
5.1.
Randomly Skipped Segments
5.2.
The ECN nonce
5.3.
A transport layer nonce
6.
The Test for Receiver Cheating
6.1.
Solution Overview
6.2.
Probabilistic Testing
6.2.1.
Performing the Probabilistic Test
6.2.2.
Assessing the Probabilistic Test
6.2.3.
RTT Measurement Considerations
6.2.4.
Protocol Details for the Probabilistic Test
6.3.
Deterministic Testing
6.3.1.
Performing the Deterministic Test
6.3.2.
Assessing the Deterministic Test
6.3.3.
Protocol Details for the Deterministic Test
6.4.
Responding to Cheating
7.
IANA Considerations
8.
Security Considerations
9.
Conclusions
10.
Acknowledgements
11.
Comments Solicited
12.
References
12.1.
Normative References
12.2.
Informative References
§
Authors' Addresses
§
Intellectual Property and Copyright Statements
| TOC |
This document specifies how a TCP sender implementation can be modified to detect a cheating receiver. It uses the standard wire protocol and protocol semantics of basic TCP [RFC0793] (Postel, J., “Transmission Control Protocol,” September 1981.) without modification.
When any network resource (e.g. a link) becomes congested, the congestion control protocol [RFC2581] (Allman, M., Paxson, V., and W. Stevens, “TCP Congestion Control,” April 1999.) within TCP/IP relies on the voluntary compliance of all senders and all receivers that are using paths through the resource. The protocol expects all receivers to correctly feed back congestion information and it expects each sender to respond by backing off its rate.
Over the past several years the Internet has become increasingly adversarial. Self-interested or malicious parties may produce non-compliant protocol implementations if it is to their advantage, or to the disadvantage of their chosen victims. To enforce congestion control when no-one can be trusted is extremely hard within the current Internet architecture. This specification deals with one specific case: where a TCP sender is well-behaved and wants to ensure its receivers behave as well.
Simple attacks have been published showing that TCP receivers can manipulate feedback to fool TCP senders into massively exceeding the compliant rate [Savage] (Savage, S., Wetherall, D., and T. Anderson, “TCP Congestion Control with a Misbehaving Receiver,” 1999.). Such receivers might want to make senders unwittingly launch a denial of service attack on other flows sharing part of the path between them [Sherwood] (Sherwood, R., Bhattacharjee, B., and R. Braud, “Misbehaving TCP Receivers Can Cause Internet-Wide Congestion Collapse,” 2005.). But a more likely motivation is simple self-interest---a receiver can hugely improve its own download speed, without any need for the sender to be a willing accomplice.
To be clear, the measures in this specification are intended for senders that can be trusted to behave. Unless all senders are trusted, this scheme alone cannot prevent other misbehaving senders from causing congestion collapse of the Internet. But the more trustworthy senders that deploy these measures, the less likely that misbehaving receivers will be able to find senders that can be fooled into causing congestion collapse.
However, senders do not have to be motivated solely by "the common good" to deploy these changes. It is directly in their own interest for senders serving multiple receivers (e.g. large file servers and certain file-sharing peers) to detect cheating receivers. A large server relies on honest network congestion feedback to efficiently apportion its own resources between receivers. If such a large server devotes an excessive fraction of its own resources to misbehaving receivers, it may well hit its own resource limits and have to starve other half-connections even if their network path has spare capacity.
In order for a sender to test a receiver, we avoid requiring the receiver to have deployed any new or optional protocol features, as any misbehaving receiver could simply circumvent the test by claiming it did not support the optional feature. Instead, the sender emulates network re-ordering then loss to test that the receiver behaves as it should within the basic TCP protocol.
This document specifies a two-stage test in which the sender deliberately re-orders some data segments so as to check if the destination honestly acknowledges out-of-order segments. The first stage test introduces a small reordering which will have a related very minor performance hit. It is not a conclusive test. However, failing it raises sufficient suspicion to warrant the more intrusive but conclusive second stage. The second stage proves beyond doubt whether the receiver is cheating but it also requires significant re-ordering, which harms performance. Therefore it should not be used unless a receiver is already strongly suspected of cheating (through failing the first stage).
The technique is designed to work with all known variants of TCP, with or without ECN [RFC3168] (Ramakrishnan, K., Floyd, S., and D. Black, “The Addition of Explicit Congestion Notification (ECN) to IP,” September 2001.), with or without SACK [RFC2018] (Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, “TCP Selective Acknowledgment Options,” October 1996.), and so on. The technique is probably transferable to derivatives of TCP, such as SCTP [RFC2960] (Stewart, R., Xie, Q., Morneault, K., Sharp, C., Schwarzbauer, H., Taylor, T., Rytina, I., Kalla, M., Zhang, L., and V. Paxson, “Stream Control Transmission Protocol,” October 2000.), but separate specifications will be required for such related transports. The requirements for a robust solution in Section 4 serve as guidelines for these separate specifications.
The document is structured as follows. It begins with a detailed description of the problems outlined above. It cites some published results that show how damaging these problems potentially are. It sets out some simple requirements that have to be met by any robust solution. It examines three existing proposed solutions in more detail, compares them against the list of requirements and demonstrates why they are not suitably robust. It then details the proposed two-stage re-ordering test, directly utilising one of the solutions already proposed as its second stage and modifying it slightly for the first stage.
| TOC |
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.).
| TOC |
TCP is widely used as the end-to-end transport in the Internet. In order to avoid the congestion collapses that plagued the Internet in the mid 1980's, TCP utilises a number of mechanisms to avoid congestion [RFC2581] (Allman, M., Paxson, V., and W. Stevens, “TCP Congestion Control,” April 1999.). These mechanisms all rely on knowing that data has been received (through acknowledgments of that data) and knowing when congestion has happened (either through knowing that a segment was lost in flight or through being notified of an Explicit Congestion Notification (ECN) [RFC3168] (Ramakrishnan, K., Floyd, S., and D. Black, “The Addition of Explicit Congestion Notification (ECN) to IP,” September 2001.)). TCP also uses a flow control mechanism to control the rate at which data is sent [RFC0813] (Clark, D., “Window and Acknowledgement Strategy in TCP,” July 1982.). Both the flow control and congestion avoidance mechanisms utilise a transmission window that limits the number of unacknowledged segments that are allowed to be sent at any given time. In order to work out the size of the transmission window, TCP monitors the average round trip time (RTT) for each flow and the number of unacknowledged segments still in flight.
A strategising receiver can take advantage of the congestion and flow control mechanisms to increase its data throughput. The three known ways in which it can do this are: optimistic acknowledgements, concealing segment losses and dividing acknowledgements into smaller parts. The first two are examined in more detail below and details of the third can be found in [Savage] (Savage, S., Wetherall, D., and T. Anderson, “TCP Congestion Control with a Misbehaving Receiver,” 1999.).
| TOC |
TCP is designed to view a lost segment as an indication of congestion on the channel. This is because TCP makes the assumption that packets are most likely to be lost through deliberately being dropped by a congested node rather than through transmission losses or errors.
In order to avoid congestion collapse [RFC3714] (Floyd, S. and J. Kempf, “IAB Concerns Regarding Congestion Control for Voice Traffic in the Internet,” March 2004.), whichever TCP connection detects the congestion (through detecting that a packet has been dropped) is expected to respond to it either by reducing its congestion window to 1 segment after a timeout or by halving it on receipt of three duplicate acks (the precise rules are in [RFC2581] (Allman, M., Paxson, V., and W. Stevens, “TCP Congestion Control,” April 1999.)).
For applications where missing data is not an issue, it is in the interest of a receiver to maximise the data rate it gets from the sender. If it conceals lost segments by falsely generating acknowledgements for them it will not suffer a reduction in data rate. There are a number of ways to make an application loss-insensitive. Some applications such as streaming media are inherently insensitive anyway, as a loss will just be seen as a transient error. TCP is widely used to transmit media files, either audio or video, which are relatively insensitive to data loss (depending on the encoding used). Also senders may be serving data containing redundant parity to allow the application to recreate lost data. A cheating receiver can also exploit application layer protocols such as the partial GET in HTTP 1.1 [RFC2616] (Fielding, R., Gettys, J., Mogul, J., Frystyk, H., Masinter, L., Leach, P., and T. Berners-Lee, “Hypertext Transfer Protocol -- HTTP/1.1,” June 1999.) to recover missing data over a secondary connection.
| Figure 1: Concealing lost segments |
| TOC |
Optimistic acknowledgements were identified as a possible attack in [Savage] (Savage, S., Wetherall, D., and T. Anderson, “TCP Congestion Control with a Misbehaving Receiver,” 1999.). If a receiver is downloading a file from a server, it is probably in its interest to acquire as high a bandwidth as possible for this. One way of increasing the bandwidth is to encourage the sender to believe the round trip time is shorter than it actually is. This means the sender will open up its transmission window faster and thus will send data faster. Of course any lost segments will also be concealed during such an attack.
The receiver can achieve this by sending acknowledgements for data it hasn't actually received yet. As long as the acknowledgement is for a packet that has already been transmitted, the sender will assume the RTT has become shorter. This will cause it to increase its transmission window more rapidly and thus send more data. Optimistic acknowledgements are particularly damaging since they can also be used to significantly amplify the effect of a denial of service (DoS) attack on a network. This form of attack is explained in more detail in [Sherwood] (Sherwood, R., Bhattacharjee, B., and R. Braud, “Misbehaving TCP Receivers Can Cause Internet-Wide Congestion Collapse,” 2005.).
The flow on the left acknowledges data only once it is received. The flow on the right acknowledges data before it is received and consequently the apparent RTT is reduced.
| Figure 2: Optimistic acknowledgements |
In 2005 US-CERT (the United States Computer Emergency Readiness Team) issued a vulnerability notice [VU102014] (Doherty, “Optimistic TCP Acknowledgements Can Cause Denial of Service,” .) specifically addressed to 80 major network equipment manufacturers and vendors who could be affected if someone maliciously exploited optimistic acknowledgements to cause a denial of service. This highlights the potential severity of such an attack were one to be launched.
| TOC |
Since the above problems come about through the inherent behaviour of the TCP protocol, there is no gain in introducing a new protocol as misbehaving receivers can claim to only support the old protocol. The best approach is to provide a mechanism within the existing protocol to test whether a receiver is cheating. The following requirements should be met by any such test in TCP and are likely to be applicable for similar tests in other transport protocols:
| TOC |
| TOC |
[Sherwood] (Sherwood, R., Bhattacharjee, B., and R. Braud, “Misbehaving TCP Receivers Can Cause Internet-Wide Congestion Collapse,” 2005.) suggests a simple approach to test a receiver's honesty. The test involves randomly dropping segments at the sender before they are transmitted. All TCP "flavours" require that a receiver should generate duplicate acknowledgements for all subsequent segments until a missing segment is received. This system requires that SACK be enabled so the sender can reliably tell that the duplicate acknowledgements are generated by the segment that is meant to be missing and are not concealing other congestion. Once the first duplicate acknowledgement arrives, the missing segment can then be re-transmitted. Because this loss has been deliberately introduced, the sender doesn't treat it as a sign of congestion. If a receiver sends an acknowledgement for a segment that was sent after the gap, it proves it is behaving dishonestly and can thus be sanctioned. As soon as the first duplicate acknowledgement is received the missing segment is re-transmitted. This will introduce a 1 RTT delay for some segments which could adversely affect some low-latency applications.
This scheme does work perfectly well in principle and does allow the sender to clearly identify dishonest behaviour. However it fails to meet requirement 4 in Section 4 (Requirements for a robust solution) above since it requires SACK to be used. If SACK were not used then it would fail to meet requirement 1 as it would be impossible to differentiate between the loss introduced on purpose and any additional loss introduced by the network.
It might be possible to incentivise the use of SACK by receivers by stating that senders are entitled to discriminate against receivers that don't support it. Given that SACK is now widely implemented across the Internet this might be a feasible, but controversial, deployment strategy. However the solution in Section 6 (The Test for Receiver Cheating) builds on Sherwood's scheme but avoids the need for SACK.
| TOC |
The authors of the ECN scheme [RFC3168] (Ramakrishnan, K., Floyd, S., and D. Black, “The Addition of Explicit Congestion Notification (ECN) to IP,” September 2001.) identified the failure to echo ECN marks as a potential attack on ECN. The ECN nonce was proposed as a possible solution to this in [RFC3540] (Spring, N., Wetherall, D., and D. Ely, “Robust Explicit Congestion Notification (ECN) Signaling with Nonces,” June 2003.). It uses a 1 bit nonce in every IP header. The nonce works by randomly setting the ECN field to ECN(0) or ECN(1). It then maintains the least significant bit of the sum of this value and stores the expected sum for each segment boundary. At the receiver end, the same 1-bit sum is calculated and is echoed back in the NS (nonce sum) flag added to the TCP header. If a packet has been congestion marked then it loses the information of which ECT codepoint it was carrying. A receiver wishing to conceal the ECN mark will have to guess whether to increment NS or not. Once congestion has been echoed back and the source has started a congestion response the nonce sum in the TCP header is not checked. Once congestion recovery is over the source resets its NS to that of the destination and starts checking again.
On the face of it this solution also fully covers the two problems identified in Section 3 (The Problems). If a receiver conceals a lost segment it has to guess what mark was there and, over several guesses, is very likely to be found out. If a receiver tries to use optimistic acknowledgements it has to guess what nonce was set on all the packets it acknowledges but hasn't received yet. However there are some key weaknesses to this system. Firstly, it assumes that ECN will be widely deployed (not currently true). Secondly, it relies on the receiver honestly declaring support for both ECN and the ECN nonce - a strategising receiver can simply declare it is neither ECN nor ECN nonce capable and thus avoid the nonce. Thirdly, the mechanism is suspended during any congestion response. Comparing it against the requirements in Section 4 (Requirements for a robust solution) above, it is clear that the ECN nonce fails to meet requirements 3 and 4 and arguably fails to meet requirement 2 as [RFC3540] (Spring, N., Wetherall, D., and D. Ely, “Robust Explicit Congestion Notification (ECN) Signaling with Nonces,” June 2003.) is experimental. The authors do state that any sender that implements the ECN nonce is entitled to discriminate against any receiver that doesn't support it. Given there are currently no implementations of the ECN nonce, discriminating against the large majority of receivers that don't support it is not a feasible deployment strategy.
| TOC |
One possible solution to the above issues is a multi-bit transport layer nonce. Two versions of this are proposed in [Savage] (Savage, S., Wetherall, D., and T. Anderson, “TCP Congestion Control with a Misbehaving Receiver,” 1999.). The first is the so called "Singular Nonce" where each segment is assigned a unique random number. This value is then echoed back to the receiver with the ack for that segment. The second version is the "Cumulative Nonce" where the nonce is set as before, but the cumulative sum of all nonces is echoed back. Whilst such a system is robust and allows a sender to correctly identify a misbehaving receiver, it has the key drawback that it requires either the creation of a new TCP option to carry the nonce and nonce reply or it requires the TCP header to be extended to include both these fields.
This proposal clearly breaches several of the requirements listed in Section 4 (Requirements for a robust solution). It breaches requirement 2 in that it needs a completely new TCP option or a change to the TCP header. It breaches requirement 3 because it needs the receiver to actively echo the nonce (as does the ECN nonce scheme) and if it uses a TCP option it breaches requirement 4. On the face of it there is no obvious route by which this sort of system can be widely implemented.
| TOC |
| TOC |
The ideal solution to the above problems should fully meet the requirements set out in Section 4 (Requirements for a robust solution). The most important of these is that the solution should leverage existing TCP behaviours rather than mandating new behaviours and options. The proposed solution utilises TCP's receiver behaviour on detecting missing data. To test a receiver the sender delays a segment during transmission by D segments. There is a trade off because increasing D increases the probability of detecting cheating but also increases the probability of masking a congestion event during the test. The completely safe strategy for the sender would be to reduce its rate pessimistically as if there were congestion during the test however this will hurt honest receivers thus breaching requirement 6. To overcome this dilemma, the test consists of two stages. In the first stage, the sender uses small displacements without the pessimistic congestion response to determine which receivers are probably cheating. The sender can then prove the dishonesty of these receivers by subjecting them to a deterministic test. This test uses a longer displacement but given the receiver is already under suspicion, it can risk harming performance by pessimistically reducing its rate as if the segment it held back was really lost by the network.
| TOC |
The first requirement for a sender is to decide when to test a receiver. This document doesn't specify when the test should be performed but the following guidance may be helpful. The simplest option is for a sender to perform the test at frequent random intervals for all its half-connections. There are also some heuristic triggers that might indicate the need for a test. Firstly, if a sender is itself too busy, it would be sensible for it to test all its receivers. Secondly, if the sender has many half-connections that are within a RTT of a congestion response, it would be sensible to test all the half-connections that aren't in a congestion response. Thirdly, the sender could aim to test all its half-connections at least once. Finally it is to be expected that there is a certain degree of existing segment reordering and thus a sender should be suspicious of any receiver that isn't generating as many duplicate acknowledgements as other receivers. [Piratla] (Piratla, N., Jayasumana, A., and T. Banka, “On Reorder Density and its Application to Characterization of Packet Reordering,” 2005.) explores how prevalent reordering might be in the Internet though it is unclear whether the figures given are more widely applicable.
The proposed solution depends, like the skipped segments solution, on the strict requirement that all TCP receivers have to send a duplicate acknowledgement as soon as they receive an out-of-order segment. This acknowledges that some data has been received, however the acknowledgement is for the last in order segment that was received (hence duplicating an acknowledgment already made). SACK extends this behaviour to allow the sender to infer exactly which segments are missing. This leads to a simple statement: if a receiver is behaving honestly it must respond to an out-of-order packet by generating a duplicate acknowledgement.
Following from the above statement, a sender can test the honesty of a given receiver by simply delaying transmission of a given segment by several places. An honest receiver will respond to this by generating a number of duplicate acknowledgements. The sender would strongly suspect a receiver of cheating if it received no duplicate acknowledgements as a result of the test. A dishonest receiver can only conceal its actions by waiting until the delayed segment arrives and then generating an appropriate stream of duplicate acknowledgements to appear to be honest.
| TOC |
The actual mechanism for conducting the test is extremely simple. Having decided to conduct a test the sender selects a segment, N. It then chooses a displacement, D (in segments) for this segment where strictly 2 < D < K - 2 where K is the current window size. In practice only low values of D should be chosen to conceal the test among the background reordering and limit the chance of masking congestion. Typically D might be less than 6. If K is less than 5, the sender can proceed straight to the deterministic test. To conduct the probabilistic test, instead of transmitting segment N, it transmits N+1, N+2, etc. as shown in the figure below. Once it has transmitted N+D it can transmit segment N. The sender needs to record the sequence number, N as well as the displacement, D.
According to data in [Piratla] (Piratla, N., Jayasumana, A., and T. Banka, “On Reorder Density and its Application to Characterization of Packet Reordering,” 2005.), as much as 15% of segments in the Internet arrive out of order though this claim may not be accurate. Whatever the actual degree of re-ordering, receivers always expect occasional losses of packets which they cannot distinguish from re-ordering without waiting for the re-ordered packet to arrive. Consequently a misbehaving receiver is unsure how to react to any out-of-order packets it receives. It should be noted that the natural reordering may reduce the displacement deliberately introduced by the test so the sender should conduct the test more than once.
| Figure 3: A receiver reacting honestly to a probabilistic test |
During testing, loss of segment L in the range from N+1 to N+D inclusive will be temporarily masked by the duplicate acknowledgements from the intentional gap that was introduced. In this case the sender's congestion response will be delayed by at most the offset D. If there is an actual loss during the test then, once the receiver receives segment N, it will generate an acknowledgement for L-1. This will lie between N and N+D. Thus it is reasonable to treat receipt of any acknowledgement between N and N+D inclusive as an indication of congestion and react accordingly. This will also discourage the receiver from sending optimistic acknowledgements in case these prove to lie in the middle of a testing sequence, in which case it will trigger a congestion response by the sender. It also means a dishonest receiver has to wait for a full K segments after any genuine lost segment to be sure it isn't a test as it will otherwise trigger a congestion response. Delaying by that long will quickly increase the RTT estimate and will soon reduce the transmission rate by as much as if the receiver had reacted honestly to the congestion.
As an additional safety measure, if the sender is performing slow start when it decides to test the receiver, it should change to congestion avoidance. The reason for this is in case there is any congestion that is concealed during the test. If there is congestion, and the sender's window is still increasing exponentially, this might significantly exacerbate the situation. This does mean that any receiver being tested during this period will suffer reduced throughput, but such testing should only be triggered by the sender being overloaded.
| TOC |
This approach to testing receiver honesty appears to meet all the requirements set out in Section 4 (Requirements for a robust solution). The most attractive feature is that it enforces equivalence with honest behaviour. That is to say, a receiver can either honestly report the missing packets or it can suffer a reduced throughput by delaying segments and increasing the RTT. The only significant drawback is that during a test it introduces some delay to the reporting of actual congestion. Given that TCP only reacts once to congestion in each RTT the delay doesn't significantly adversely affect the overall response to severe congestion.
Some receivers may choose to behave dishonestly despite this. These can be quickly identified by looking at their acknowledgements. A receiver that never sends duplicate acknowledgements in response to being tested is likely to be misbehaving. Equally, a receiver that delays transmission of the duplicate acknowledgements until it is sure it is being tested will leave an obvious pattern of acknowledgements that the sender can identify. Because a receiver is unlikely to be able to differentiate this test from actual re-ordering events, the receiver will be forced to behave in the same fashion for any re-ordered packet even in the absence of a test, making it continually appear to have longer RTT.
| TOC |
Clearly, if the sender has re-ordered segment N, it cannot use it to take an accurate RTT measurement. However it is desirable to ensure that, during a test, the sender still measures the RTT of the flow. One of the key aspects of this test is that the only way to cheat is for a dishonest receiver to delay sending acknowledgements until it is certain a test is happening. If accurate RTTs can be measured during a test, this delay will cause a dishonest receiver to suffer an increase in RTT and thus a reduction in data throughput. This will help act as a disincentive to cheating.
Measurement of the RTT usually depends on receiving an acknowledgement for a segment and measuring the delay between when the segment was sent and when the acknowledgement arrives. The TCP timestamp option is often used to provide accurate RTT measurement but again, this is not going to function correctly during a test phase. During a test therefore, the RTT has to be estimated using the arrival of duplicate acknowledgements. Figure 4 (Measuring the RTT during a test) shows how one can measure the RTT in this way, and also demonstrates how this will increase if a dishonest sender chooses to cheat. However it is not sufficient simply to measure a single RTT during the test. A clever receiver might decide that the safe reaction to any missing segment is to immediately send one or two duplicate acknowledgements in order to disrupt this RTT measurement without running the risk of triggering a fast retransmit if the segment is genuinely missing.
| Figure 4: Measuring the RTT during a test |
| TOC |
| TOC |
If after one or more probabilistic tests the sender deems that a receiver is acting suspiciously, the sender can perform a deterministic test similar to the skipped segment scheme in Section 5.1 (Randomly Skipped Segments) above.
| TOC |
In order to perform the deterministic test the sender again needs to choose a segment, M to use for testing. This time the sender holds back the segment until the receiver indicates that it is missing. Once the receiver sends a duplicate acknowledgement for segment M-1 then the sender transmits segment M. In the meantime data transmission should proceed as usual. If SACK is not in use, this test clearly increases the delay in reporting of genuine segment losses by up to a RTT. This is because it is only once segment M reaches the receiver that it will be able to acknowledge the later loss. Therefore, unless SACK is in use, the sender MUST pessimistically perform a congestion response following the arrival of 3 duplicate acknowledgements for segment M-1 as mandated in [RFC2581] (Allman, M., Paxson, V., and W. Stevens, “TCP Congestion Control,” April 1999.).
| TOC |
A dishonest receiver that is concealing segment losses will establish that this isn't a probabilistic test once the missing segment fails to arrive within the space of 1 congestion window. In order to conceal the loss the receiver will simply carry on acknowledging all subsequent data. The sender can therefore state that if it receives an acknowledgement for a segment with a sequence number greater than M before it has actually sent segment M then the receiver must be cheating. A sender would be expected to close a connection with any receiver that had failed the deterministic test, but this draft was only written to document a test procedure to establish dishonesty, not what the sender should or must do if the receiver fails the test.
It is important to be aware that a third party who is able to correctly guess the initial sequence number of a connection might be able to masquerade as a receiver and send acknowledgements on their behalf to make them appear dishonest. Such an attack can be identified because an honest receiver will also be generating a stream of duplicate acknowledgements until such time as it receives the missing segment.
| TOC |
| TOC |
Having identified that a receiver is actually being dishonest, the appropriate response is to terminate the connection with that receiver. If a sender is under severe attack it might also choose to ignore all subsequent requests to connect by that receiver. However this is a risky strategy as it might give an increased incentive to launch an attack against someone by making them appear to be behaving dishonestly. It also is risky in the current network where many users might share a quite small bank of IP addresses assigned dynamically to them by their ISP's DHCP server. A safer alternative to blacklisting a given IP address might be to simply test future connections more rigorously.
| TOC |
This memo includes no request to IANA.
| TOC |
The two tests described in this document provide a solution to two of the significant security problems that were outlined in [Savage] (Savage, S., Wetherall, D., and T. Anderson, “TCP Congestion Control with a Misbehaving Receiver,” 1999.). Both these attacks could potentially cause major congestion of senders own resources (by making them transmit at too high a rate) and could lead to network congestion collapse through subverting the correct reporting of congestion or by amplifying any DoS attack [Sherwood] (Sherwood, R., Bhattacharjee, B., and R. Braud, “Misbehaving TCP Receivers Can Cause Internet-Wide Congestion Collapse,” 2005.). The proposed solution cannot alone prevent misbehaving senders from causing congestion collapse of the Internet. However, the more widely it is deployed by trustworthy senders, the more these particular attacks would be mitigated through ensuring accurate reporting of segment losses. The more senders that deploy these measures, the less likely it is that a misbehaving receiver will be able to find a sender to fool into causing congestion collapse.
It should be noted that if a third party is able to correctly guess the initial sequence number of a connection, they might be able to masquerade as a receiver and send acknowledgements on their behalf to make them appear dishonest during a deterministic test.
[Savage] (Savage, S., Wetherall, D., and T. Anderson, “TCP Congestion Control with a Misbehaving Receiver,” 1999.) also identified a further attack involving splitting acknowledgements into smaller parts. TCP is designed such that increases in the congestion window are driven by the arrival of a valid acknowledgement. It doesn't matter if this acknowledgement covers all of a transmitted segment or not. This means a receiver that divides all its acknowledgements into two will cause the congestion window to open at twice the rate it would do otherwise. The tests described above can't protect against that attack. However there is a straightforward solution to this - every time the sender transmits a new segment it increments a counter; every acknowledgment it receives decrements that counter; if the counter reaches zero, the sender won't increase its congestion window in response to a new acknowledgement arriving. To comply with this document, senders MUST implement a solution to this problem.
| TOC |
The issue of mutual trust between TCP senders and receivers is a significant one in the current Internet. This document has introduced a mechanism by which honest senders can verify that their receivers are not cheating. The whole process is robust, lightweight, elegant and efficient. The probabilistic test might delay a congestion notification by a fraction of a RTT, however this is compensated for by the protocol reacting more rapidly to any such indication. The deterministic test carries a greater risk of delaying congestion notification and consequently the protocol mandates that a congestion response should happen whilst performing the test. The two tests combine to provide a mechanism to allow the sender to judge the honesty of a receiver in a manner that both encourages honest behaviour and proves dishonesty in a robust manner. The most attractive feature of this scheme is that it requires no active participation by the receiver as it utilises the standard behaviour of TCP in the presence of missing data. The only changes required are at the sender.
| TOC |
{ToDo:}
| TOC |
Comments and questions are encouraged and very welcome. They can be addressed to the IETF Transport Area working group mailing list <tsvwg@ietf.org>, and/or to the authors.
| TOC |
| TOC |
| [RFC0793] | Postel, J., “Transmission Control Protocol,” STD 7, RFC 793, September 1981. |
| [RFC0813] | Clark, D., “Window and Acknowledgement Strategy in TCP,” RFC 813, July 1982. |
| [RFC2018] | Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, “TCP Selective Acknowledgment Options,” RFC 2018, October 1996 (TXT, HTML, XML). |
| [RFC2119] | Bradner, S., “Key words for use in RFCs to Indicate Requirement Levels,” BCP 14, RFC 2119, March 1997 (TXT, HTML, XML). |
| [RFC2581] | Allman, M., Paxson, V., and W. Stevens, “TCP Congestion Control,” RFC 2581, April 1999. |
| TOC |
| [Piratla] | Piratla, N., Jayasumana, A., and T. Banka, “On Reorder Density and its Application to Characterization of Packet Reordering,” 2005. |
| [RFC2616] | Fielding, R., Gettys, J., Mogul, J., Frystyk, H., Masinter, L., Leach, P., and T. Berners-Lee, “Hypertext Transfer Protocol -- HTTP/1.1,” RFC 2616, June 1999 (TXT, PS, PDF, HTML, XML). |
| [RFC2960] | Stewart, R., Xie, Q., Morneault, K., Sharp, C., Schwarzbauer, H., Taylor, T., Rytina, I., Kalla, M., Zhang, L., and V. Paxson, “Stream Control Transmission Protocol,” RFC 2960, October 2000. |
| [RFC3168] | Ramakrishnan, K., Floyd, S., and D. Black, “The Addition of Explicit Congestion Notification (ECN) to IP,” RFC 3168, September 2001. |
| [RFC3540] | Spring, N., Wetherall, D., and D. Ely, “Robust Explicit Congestion Notification (ECN) Signaling with Nonces,” RFC 3540, June 2003. |
| [RFC3714] | Floyd, S. and J. Kempf, “IAB Concerns Regarding Congestion Control for Voice Traffic in the Internet,” RFC 3714, March 2004. |
| [Savage] | Savage, S., Wetherall, D., and T. Anderson, “TCP Congestion Control with a Misbehaving Receiver,” 1999. |
| [Sherwood] | Sherwood, R., Bhattacharjee, B., and R. Braud, “Misbehaving TCP Receivers Can Cause Internet-Wide Congestion Collapse,” 2005. |
| [VU102014] | Doherty, “Optimistic TCP Acknowledgements Can Cause Denial of Service.” |
| TOC |
| Toby Moncaster | |
| BT | |
| B54/70, Adastral Park | |
| Martlesham Heath | |
| Ipswich IP5 3RE | |
| UK | |
| Phone: | +44 1473 648734 |
| Email: | toby.moncaster@bt.com |
| Bob Briscoe | |
| BT & UCL | |
| B54/77, Adastral Park | |
| Martlesham Heath | |
| Ipswich IP5 3RE | |
| UK | |
| Phone: | +44 1473 645196 |
| Email: | bob.briscoe@bt.com |
| Arnaud Jacquet | |
| BT | |
| B54/70, Adastral Park | |
| Martlesham Heath | |
| Ipswich IP5 3RE | |
| UK | |
| Phone: | +44 1473 647284 |
| Email: | arnaud.jacquet@bt.com |
| TOC |
Copyright © The IETF Trust (2007).
This document is subject to the rights, licenses and restrictions contained in BCP 78, and except as set forth therein, the authors retain all their rights.
This document and the information contained herein are provided on an “AS IS” basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY, THE IETF TRUST AND THE INTERNET ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
The IETF takes no position regarding the validity or scope of any Intellectual Property Rights or other rights that might be claimed to pertain to the implementation or use of the technology described in this document or the extent to which any license under such rights might or might not be available; nor does it represent that it has made any independent effort to identify any such rights. Information on the procedures with respect to rights in RFC documents can be found in BCP 78 and BCP 79.
Copies of IPR disclosures made to the IETF Secretariat and any assurances of licenses to be made available, or the result of an attempt made to obtain a general license or permission for the use of such proprietary rights by implementers or users of this specification can be obtained from the IETF on-line IPR repository at http://www.ietf.org/ipr.
The IETF invites any interested party to bring to its attention any copyrights, patents or patent applications, or other proprietary rights that may cover technology that may be required to implement this standard. Please address the information to the IETF at ietf-ipr@ietf.org.
Funding for the RFC Editor function is provided by the IETF Administrative Support Activity (IASA). This document was produced using xml2rfc v1.32 (of http://xml.resource.org/) from a source in RFC-2629 XML format.