]> Verizon Federal Network Systems

weddy@grc.nasa.gov NASA Glenn Research Center 21000 Brookpark Rd, MS 54-5 ClevelandOH 44135 216-433-6682

Transport TCP SYN Flood TCP SYN Cookies This document describes TCP SYN flooding attacks, which have been well-known to the community for several years. Various countermeasures against these attacks, and the trade-offs of each, are described. This document archives explanations of the attack and common defense techniques for the benefit of TCP implementers and administrators of TCP servers or networks, but does not make any standards-level recommendations.

The SYN flooding attack is a denial-of-service method affecting hosts that run TCP server processes. The attack takes advantage of the state retention TCP performs for some time after receiving a SYN segment to a port that has been put into the LISTEN state. The basic idea is to exploit this behavior by causing a host to retain enough state for bogus half-connections that there are no resources left to establish new legitimate connections. This SYN flooding attack has been well-known to the community for many years, and has been observed in the wild by network operators and end hosts. A number of methods have been developed and deployed to make SYN flooding less effective. Despite the notoriety of the attack, and the widely available countermeasures, the RFC series only documented the vulnerability as an example motivation for ingress filtering , and has not suggested any mitigation techniques for TCP implementations. This document addresses both points, but does not define any standards. Formal specifications and requirements of defense mechanisms are outside the scope of this document. Many defenses only impact an end host's implementation without changing interoperability. These may not require standardization, but their side-effects should at least be well understood. This document intentionally focuses on SYN flooding attacks from an individual end host or application's perspective, as a means to deny service to that specific entity. High packet-rate attacks that target the network's packet-processing capability and capacity have been observed operationally. Since such attacks target the network, and not a TCP implementation, they are out of scope for this document, whether or not they happen to use TCP SYN segments as part of the attack, as the nature of the packets used is irrelevant in comparison to the packet-rate in such attacks. The majority of this document consists of three sections. explains the SYN flooding attack in greater detail. Several common mitigation techniques are described in . An analysis and discussion of these techniques and their use is presented in . Further information on SYN cookies is contained in .

This section describes both the history and the technical basis of the SYN flooding attack.

The TCP SYN flooding weakness was discovered as early as 1994 by Bill Cheswick and Steve Bellovin . They included, and then removed, a paragraph on the attack in their book "Firewalls and Internet Security: Repelling the Wily Hacker" . Unfortunately, no countermeasures were developed within the next two years. The SYN flooding attack was first publicized in 1996, with the release of a description and exploit tool in Phrack Magazine . Aside from some minor inaccuracies, this article is of high enough quality to be useful, and code from the article was widely distributed and used. By September of 1996, SYN flooding attacks had been observed in the wild. Particularly, an attack against one ISP's mail servers caused well-publicized outages. CERT quickly released an advisory on the attack . SYN flooding was particularly serious in comparison to other known denial-of-service attacks at the time. Rather than relying on the common brute-force tactic of simply exhausting the network's resources, SYN flooding targets end-host resources, which require fewer packets to deplete. The community quickly developed many widely differing techniques for preventing or limiting the impact of SYN flooding attacks. Many of these have been deployed to varying degrees on the Internet, in both end hosts and intervening routers. Some of these techniques have become important pieces of the TCP implementations in certain operating systems, although some significantly diverge from the TCP specification and none of these techniques have yet been standardized or sanctioned by the IETF process.

As described in RFC 793, a TCP implementation may allow the LISTEN state to be entered with either all, some, or none of the pair of IP addresses and port numbers specified by the application. In many common applications like web servers, none of the remote host's information is pre-known or preconfigured, so that a connection can be established with any client whose details are unknown to the server ahead of time. This type of "unbound" LISTEN is the target of SYN flooding attacks due to the way it is typically implemented by operating systems. For success, the SYN flooding attack relies on the victim host TCP implementation's behavior. In particular, it assumes that the victim allocates state for every TCP SYN segment when it is received, and that there is a limit on the amount of such state than can be kept at any time. The current base TCP specification, RFC 793 , describes the standard processing of incoming SYN segments. RFC 793 describes the concept of a Transmission Control Block (TCB) data structure to store all the state information for an individual connection. In practice, operating systems may implement this concept rather differently, but the key is that each TCP connection requires some memory space. Per RFC 793, when a SYN is received for a local TCP port where a connection is in the LISTEN state, then the state transitions to SYN-RECEIVED, and some of the TCB is initialized with information from the header fields of the received SYN segment. In practice, many operating systems do not alter the TCB in LISTEN, but instead make a copy of the TCB and perform the state transition and update on the copy. This is done so that the local TCP port may be shared amongst several distinct connections. This TCB-copying behavior is not actually essential for this purpose, but influences the way in which applications that wish to handle multiple simultaneous connections through a single TCP port are written. The crucial result of this behavior is that, instead of updating already-allocated memory, new (or unused) memory must be devoted to the copied TCB. As an example, in the Linux 2.6.10 networking code, a "sock" structure is used to implement the TCB concept. By examination, this structure takes over 1300 bytes to store in memory. In other systems that implement less-complex TCP algorithms and options, the overhead may be less, although it typically exceeds 280 bytes . To protect host memory from being exhausted by connection requests, the number of TCB structures that can be resident at any time is usually limited by operating system kernels. Systems vary on whether limits are globally applied or local to a particular port number. There is also variation on whether the limits apply to fully established connections as well as those in SYN-RECEIVED. Commonly, systems implement a parameter to the typical listen() system call that allows the application to suggest a value for this limit, called the backlog. When the backlog limit is reached, then either incoming SYN segments are ignored, or uncompleted connections in the backlog are replaced. The concept of using a backlog is not described in the standards documents, so the failure behavior when the backlog is reached might differ between stacks (for instance, TCP RSTs might be generated). The exact failure behavior will determine whether initiating hosts continue to retransmit SYN segments over time, or quickly cease. These differences in implementation are acceptable since they only affect the behavior of the local stack when its resources are constrained, and do not cause interoperability problems. The SYN flooding attack does not attempt to overload the network's resources or the end host's memory, but merely attempts to exhaust the backlog of half-open connections associated with a port number. The goal is to send a quick barrage of SYN segments from IP addresses (often spoofed) that will not generate replies to the SYN-ACKs that are produced. By keeping the backlog full of bogus half-opened connections, legitimate requests will be rejected. Three important attack parameters for success are the size of the barrage, the frequency with which barrages are generated, and the means of selecting IP addresses to spoof. To be effective, the size of the barrage must be made large enough to reach the backlog. Ideally, the barrage size is no larger than the backlog, minimizing the volume of traffic the attacker must source. Typical default backlog values vary from a half-dozen to several dozen, so the attack might be tailored to the particular value determined by the victim host and application. On machines intended to be servers, especially for a high volume of traffic, the backlogs are often administratively configured to higher values. To limit the lifetime of half-opened connection state, TCP implementations commonly reclaim memory from half-opened connections if they do not become fully opened after some time period. For instance, a timer of 75 seconds might be set when the first SYN-ACK is sent, and on expiration cause SYN-ACK retransmissions to cease and the TCB to be released. The TCP specifications do not include this behavior of giving up on connection establishment after an arbitrary time. Some purists have expressed that the TCP implementation should continue retransmitting SYN and SYN-ACK segments without artificial bounds (but with exponential backoff to some conservative rate) until the application gives up. Despite this, common operating systems today do implement some artificial limit on half-open TCB lifetime. For instance, backing off and stopping after a total of 511 seconds can be observed in 4.4 BSD-Lite , and is still practiced in some operating systems derived from this code. To remain effective, a SYN flooding attack needs to send new barrages of bogus connection requests as soon as the TCBs from the previous barrage begin to be reclaimed. The frequency of barrages are tailored to the victim TCP implementation's TCB reclamation timer. Frequencies higher than needed source more packets, potentially drawing more attention, and frequencies that are too low will allow windows of time where legitimate connections can be established. For an effective attack, it is important that the spoofed IP addresses be unresponsive to the SYN-ACK segments that the victim will generate. If addresses of normal connected hosts are used, then those hosts will send the victim a TCP reset segment that will immediately free the corresponding TCB and allow room in the backlog for legitimate connections to be made. The code distributed in the original Phrack article used a single source address for all spoofed SYN segments. This makes the attack segments somewhat easier to identify and filter. A strong attacker will have a list of unresponsive and unrelated addresses that it chooses spoofed source addresses from. It is important to note that this attack is directed at particular listening applications on a host, and not the host itself or the network. The attack also attempts to prevent only the establishment of new incoming connections to the victim port, and does not impact outgoing connection requests, nor previously established connections to the victim port. In practice, an attacker might choose not to use spoofed IP addresses, but instead to use a multitude of hosts to initiate a SYN flooding attack. For instance, a collection of compromised hosts under the attacker's control (i.e., a "botnet") could be used. In this case, each host utilized in the attack would have to suppress its operating system's native response to the SYN-ACKs coming from the target. It is also possible for the attack TCP segments to arrive in a more continuous fashion than the "barrage" terminology used here suggests; as long as the rate of new SYNs exceeds the rate at which TCBs are reaped, the attack will be successful.

This section discusses a number of defense techniques that are known to the community, many of which are available in off-the-shelf products.

Since in the absence of an army of controlled hosts, the ability to send packets with spoofed source IP addresses is required for this attack to work, removing an attacker's ability to send spoofed IP packets is an effective solution that requires no modifications to TCP. The filtering techniques described in RFCs 2827, 3013, and 3704 represent the best current practices for packet filtering based on IP addresses . While perfectly effective, end hosts should not rely on filtering policies to prevent attacks from spoofed segments, as global deployment of filters is neither guaranteed nor likely. An attacker with the ability to use a group of compromised hosts or to rapidly change between different access providers will also make filtering an impotent solution.

An obvious attempt at a defense is for end hosts to use a larger backlog. Lemon has shown that in FreeBSD 4.4, this tactic has some serious negative aspects as the size of the backlog grows . The implementation has not been designed to scale past backlogs of a few hundred, and the data structures and search algorithms that it uses are inefficient with larger backlogs. It is reasonable to assume that other TCP implementations have similar design factors that limit their performance with large backlogs, and there seems to be no compelling reason why stacks should be re-engineered to support extremely large backlogs, since other solutions are available. However, experiments with large backlogs using efficient data structures and search algorithms have not been conducted, to our knowledge.

Another quickly implementable defense is shortening the timeout period between receiving a SYN and reaping the created TCB for lack of progress. Decreasing the timer that limits the lifetime of TCBs in SYN-RECEIVED is also flawed. While a shorter timer will keep bogus connection attempts from persisting for as long in the backlog, and thus free up space for legitimate connections sooner, it can prevent some fraction of legitimate connections from becoming fully established. This tactic is also ineffective because it only requires the attacker to increase the barrage frequency by a linearly proportional amount. This timer reduction is sometimes implemented as a response to crossing some threshold in the backlog occupancy, or some rate of SYN reception.

Once the entire backlog is exhausted, some implementations allow incoming SYNs to overwrite the oldest half-open TCB entry. This works under the assumption that legitimate connections can be fully established in less time than the backlog can be filled by incoming attack SYNs. This can fail when the attacking packet rate is high and/or the backlog size is small, and is not a robust defense.

The SYN cache, best described by Lemon , is based on minimizing the amount of state that a SYN allocates, i.e., not immediately allocating a full TCB. The full state allocation is delayed until the connection has been fully established. Hosts implementing a SYN cache have some secret bits that they select from the incoming SYN segments. The secret bits are hashed along with the IP addresses and TCP ports of a segment, and the hash value determines the location in a global hash table where the incomplete TCB is stored. There is a bucket limit for each hash value, and when this limit is reached, the oldest entry is dropped. The SYN cache technique is effective because the secret bits prevent an attacker from being able to target specific hash values for overflowing the bucket limit, and it bounds both the CPU time and memory requirements. Lemon's evaluation of the SYN cache shows that even under conditions where a SYN flooding attack is not being performed, due to the modified processing path, connection establishment is slightly more expedient. Under active attack, SYN cache performance was observed to approximately linearly shift the distribution of times to establish legitimate connections to about 15% longer than when not under attack . If data accompanies the SYN segment, then this data is not acknowledged or stored by the receiver, and will require retransmission. This does not affect the reliability of TCP's data transfer service, but it does affect its performance to some small extent. SYNs carrying data are used by the T/TCP extensions . While T/TCP is implemented in a number of popular operating systems , it currently seems to be rarely used. Measurements at one site's border router logged 2,545,785 SYN segments (not SYN-ACKs), of which 36 carried the T/TCP CCNEW option (or 0.001%). These came from 26 unique hosts, and no other T/TCP options were seen. 2,287 SYN segments with data were seen (or 0.09% of all SYN segments), all of which had exactly 24 bytes of data. These observations indicate that issues with SYN caches and data on SYN segments may not be significant in deployment.

SYN cookies go a step further and allocate no state at all for connections in SYN-RECEIVED. Instead, they encode most of the state (and all of the strictly required) state that they would normally keep into the sequence number transmitted on the SYN-ACK. If the SYN was not spoofed, then the acknowledgement number (along with several other fields) in the ACK that completes the handshake can be used to reconstruct the state to be put into the TCB. To date, one of the best references on SYN cookies can be found on Dan Bernstein's web site . This technique exploits the long-understood low entropy in TCP header fields . In , we describe the SYN cookie technique, to avoid the possibility that the web page will become unavailable. The exact mechanism for encoding state into the SYN-ACK sequence number can be implementation dependent. A common consideration is that to prevent replay, some time-dependent random bits must be embedded in the sequence number. One technique used 7 bits for these bits and 25 bits for the other data . One way to encode these bits has been to XOR the initial sequence number received with a truncated cryptographic hash of the IP address and TCP port number pairs, and secret bits. In practice, this hash has been generated using MD5 . Any similar one-way hash could be used instead without impacting interoperability since the hash value is checked by the same host who generates it. The problem with SYN cookies is that commonly implemented schemes are incompatible with some TCP options, if the cookie generation scheme does not consider them. For example, an encoding of the Maximum Segment Size (MSS) advertised on the SYN has been accommodated by using 2 sequence number bits to represent 4 predefined common MSS values. Similar techniques would be required for some other TCP options, while negotiated use of other TCP options can be detected implicitly. A timestamp on the ACK, as an example, indicates that Timestamp use was successfully negotiated on the SYN and SYN-ACK, while the reception of a SACK option at some point during the connection implies that SACK was negotiated. Note that SACK blocks should normally not be sent by a host using TCP cookies unless they are first received. For the common unidirectional data flow in many TCP connections, this can be a problem, as it limits SACK usage. For this reason, SYN cookies typically are not used by default on systems that implement them, and are only enabled either under high-stress conditions indicative of an attack, or via administrative action. Recently, a new SYN cookie technique developed for release in FreeBSD 7.0 leverages the bits of the Timestamp option in addition to the sequence number bits for encoding state. Since the Timestamp value is echoed back in the Timestamp Echo field of the ACK packet, any state stored in the Timestamp option can be restored similarly to the way that it is from the sequence number / acknowledgement in a basic SYN cookie. Using the Timestamp bits, it is possible to explicitly store state bits for things like send and receive window scales, SACK-allowed, and TCP-MD5-enabled, for which there is no room in a typical SYN cookie. This use of Timestamps to improve the compromises inherent in SYN cookies is unique to the FreeBSD implementation, to our knowledge. A limitation is that the technique can only be used if the SYN itself contains a Timestamp option, but this option seems to be widely implemented today, and hosts that support window scaling and SACK typically support timestamps as well. Similarly to SYN caches, SYN cookies do not handle application data piggybacked on the SYN segment. Another problem with SYN cookies is for applications where the first application data is sent by the passive host. If this host is handling a large number of connections, then packet loss may be likely. When a handshake-completing ACK from the initiator is lost, the passive side's application layer never is notified of the connection's existence and never sends data, even though the initiator thinks that the connection has been successfully established. An example application where the first application-layer data is sent by the passive side is SMTP, if implemented according to RFC 2821, where a "service ready" message is sent by the passive side after the TCP handshake is completed. Although SYN cookie implementations exist and are deployed, the use of SYN cookies is often disabled in default configurations, so it is unclear how much operational experience actually exists with them or if using them opens up new vulnerabilities. Anecdotes of incidents where SYN cookies have been used on typical web servers seem to indicate that the added processing burden of computing MD5 sums for every SYN packet received is not significant in comparison to the loss of application availability when undefended. For some computationally constrained mobile or embedded devices, this situation might be different.

The SYN cache and SYN cookie techniques can be combined. For example, in the event that the cache becomes full, then SYN cookies can be sent instead of purging cache entries upon the arrival of new SYNs. Such hybrid approaches may provide a strong combination of the positive aspects of each approach. Lemon has demonstrated the utility of this hybrid .

Firewall-based tactics may also be used to defend end hosts from SYN flooding attacks. The basic concept is to offload the connection establishment procedures onto a firewall that screens connection attempts until they are completed and then proxies them back to protected end hosts. This moves the problem away from end hosts to become the firewall's or proxy's problem, and may introduce other problems related to altering TCP's expected end-to-end semantics. A common tactic used in these firewall and proxy products is to implement one of the end host based techniques discussed above, and screen incoming SYNs from the protected network until the connection is fully established. This is accomplished by spoofing the source addresses of several packets to the initiator and listener at various stages of the handshake .

Several of the defenses discussed in the previous section rely on changes to behavior inside the network; via router filtering, firewalls, and proxies. These may be highly effective, and often require no modification or configuration of end-host software. Given the mobile nature and dynamic connectivity of many end hosts, it is optimistic for TCP implementers to assume the presence of such protective devices. TCP implementers should provide some means of defense to SYN flooding attacks in end-host implementations. Among end-host modifications, the SYN cache and SYN cookie approaches seem to be the only viable techniques discovered to date. Increasing the backlog and reducing the SYN-RECEIVED timer are measurably problematic. The SYN cache implies a higher memory footprint than SYN cookies; however, SYN cookies may not be fully compatible with some TCP options, and may hamper development of future TCP extensions that require state. For these reasons, SYN cookies should not be enabled by default on systems that provide them. SYN caches do not have the same negative implications and may be enabled as a default mode of processing. In October of 1996, Dave Borman implemented a SYN cache at BSDi for BSD/OS, which was given to the community with no restrictions. This code seems to be the basis for the SYN cache implementations adopted later in other BSD variants. The cache was used when the backlog became full, rather than by default, as we have described. A note to the tcp-impl mailing list explains that this code does not retransmit SYN-ACKs . More recent implementations have chosen to reverse this decision and retransmit SYN-ACKs. It is known that loss of SYN-ACK packets is not uncommon and can severely slow the performance of connections when initial retransmission timers for SYNs are overly conservative (as in some operating systems) or retransmitted SYNs are lost. Furthermore, if a SYN flooding attacker has a high sending rate, loss of retransmitted SYNs is likely, so if SYN-ACKs are not retransmitted, the chance of efficiently establishing legitimate connections is reduced. In 1997, NetBSD incorporated a modified version of Borman's code. Two notable differences from the original code stem from the decision to use the cache by default (for all connections). This implied the need to perform retransmissions for SYN-ACKs, and to use larger structures to keep more complete data. The original structure was 32 bytes long for IPv4 connections and 56 bytes with IPv6 support, while the current FreeBSD structure is 196 bytes long. As previously cited, Lemon implemented the SYN cache and cookie techniques in FreeBSD 4.4 . Lemon notes that a SYN cache structure took up 160 bytes compared to 736 for the full TCB (now 196 bytes for the cache structure). We have examined the OpenBSD 3.6 code and determined that it includes a similar SYN cache. Linux 2.6.5 code, also by examination, contains a SYN cookie implementation that encodes 8 MSS values, and does not use SYN cookies by default. This functionality has been present in the Linux kernel for several years previous to 2.6.5. When a SYN cache and/or SYN cookies are implemented with IPv6, the IPv6 flow label value used on the SYN-ACK should be consistent with the flow label used for the rest of the packets within that flow. There have been implementation bugs that caused random flow labels to be used in SYN-ACKs generated by SYN cache and SYN cookie code . Beginning with Windows 2000, Microsoft's Windows operating systems have had a "TCP SYN attack protection" feature, which can be toggled on or off in the registry. This defaulted to off, until Windows 2003 SP1, in which it is on by default. With this feature enabled, when the number of half-open connections and half-open connections with retransmitted SYN-ACKs exceeds configurable thresholds, then the number of times that SYN-ACKs are retransmitted before giving up is reduced, and the "Route Cache Entry" creation is delayed, which prevents some features (e.g., window scaling) from being used . Several vendors of commercial firewall products sell devices that can mitigate SYN flooding's effects on end hosts by proxying connections. Discovery and exploitation of the SYN flooding vulnerability in TCP's design provided a valuable lesson for protocol designers. The Stream Control Transmission Protocol , which was designed more recently, incorporated a 4-way handshake with a stateless cookie-based component for the listening end. In this way, the passive-opening side has better evidence that the initiator really exists at the given address before it allocates any state. The Host Identity Protocol base exchange is similarly designed as a 4-way handshake, but also involves a puzzle sent to the initiator that must be solved before any state is reserved by the responder. The general concept of designing statelessness into protocol setup to avoid denial-of-service attacks has been discussed by Aura and Nikander .

The SYN flooding attack on TCP has been described in numerous other publications, and the details and code needed to perform the attack have been easily available for years. Describing the attack in this document does not pose any danger of further publicizing this weakness in unmodified TCP stacks. Several widely deployed operating systems implement the mitigation techniques that this document discusses for defeating SYN flooding attacks. In at least some cases, these operating systems do not enable these countermeasures by default; however, the mechanisms for defeating SYN flooding are well deployed, and easily enabled by end-users. The publication of this document should not influence the number of SYN flooding attacks observed, and might increase the robustness of the Internet to such attacks by encouraging use of the commonly available mitigations.

A conversation with Ted Faber was the impetus for writing this document. Comments and suggestions from Joe Touch, Dave Borman, Fernando Gont, Jean-Baptiste Marchand, Christian Huitema, Caitlin Bestler, Pekka Savola, Andre Oppermann, Alfred Hoenes, Mark Allman, Lars Eggert, Pasi Eronen, Warren Kumari, David Malone, Ron Bonica, and Lisa Dusseault were useful in strengthening this document. The original work on TCP SYN cookies presented in is due to D.J. Bernstein. Work on this document was performed at NASA's Glenn Research Center. Funding was partially provided by a combination of NASA's Advanced Communications, Navigation, and Surveillance Architectures and System Technologies (ACAST) project, the Sensis Corporation, NASA's Space Communications Architecture Working Group, and NASA's Earth Science Technology Office.

&RFC0793; &RFC2827; &RFC3013; &RFC3704; &RFC1321; &RFC1644; &RFC1144; &RFC4413; &RFC2960;

This information is taken from Bernstein's web page on SYN cookies . This is a rewriting of the technical information on that web page and not a full replacement. There are other slightly different ways of implementing the SYN cookie concept than the exact means described here, although the basic idea of encoding data into the SYN-ACK sequence number is constant. A SYN cookie is an initial sequence number sent in the SYN-ACK, that is chosen based on the connection initiator's initial sequence number, MSS, a time counter, and the relevant addresses and port numbers. The actual bits comprising the SYN cookie are chosen to be the bitwise difference (exclusive-or) between the SYN's sequence number and a 32 bit quantity computed so that the top five bits come from a 32-bit counter value modulo 32, where the counter increases every 64 seconds, the next 3 bits encode a usable MSS near to the one in the SYN, and the bottom 24 bits are a server-selected secret function of pair of IP addresses, the pair of port numbers, and the 32-bit counter used for the first 5 bits. This means of selecting an initial sequence number for use in the SYN-ACK complies with the rule that TCP sequence numbers increase slowly. When a connection in LISTEN receives a SYN segment, it can generate a SYN cookie and send it in the sequence number of a SYN-ACK, without allocating any other state. If an ACK comes back, the difference between the acknowledged sequence number and the sequence number of the ACK segment can be checked against recent values of the counter and the secret function's output given those counter values and the IP addresses and port numbers in the ACK segment. If there is a match, the connection can be accepted, since it is statistically very likely that the other side received the SYN cookie and did not simply guess a valid cookie value. If there is not a match, the connection can be rejected under the heuristic that it is probably not in response to a recently sent SYN-ACK. With SYN cookies enabled, a host will be able to remain responsive even when under a SYN flooding attack. The largest price to be paid for using SYN cookies is in the disabling of the window scaling option, which disables high performance. Bernstein's web page contains more information about the initial conceptualization and implementation of SYN cookies, and archives of emails documenting this history. It also lists some false negative claims that have been made about SYN cookies, and discusses reducing the vulnerability of SYN cookie implementations to blind connection forgery by an attacker guessing valid cookies. The best description of the exact SYN cookie algorithms is in a part of an email from Bernstein, that is archived on the web site (notice it does not set the top five bits from the counter modulo 32, as the previous description did, but instead uses 29 bits from the second MD5 operation and 3 bits for the index into the MSS table; establishing the secret values is also not discussed). The remainder of this section is excerpted from Bernstein's email : Here's what an implementation would involve: Maintain two (constant) secret keys, sec1 and sec2. Maintain a (constant) sorted table of 8 common MSS values, msstab[8]. Keep track of a ``last overflow time.'' Maintain a counter that increases slowly over time and never repeats, such as ``number of seconds since 1970, shifted right 6 bits.'' When a SYN comes in from (saddr,sport) to (daddr,dport) with ISN x, find the largest i for which msstab[i] <= the incoming MSS. Compute z = MD5(sec1,saddr,sport,daddr,dport,sec1) + x + (counter << 24) + (MD5(sec2,counter,saddr,sport,daddr,dport,sec2) % (1 << 24)) and then y = (i << 29) + (z % (1 << 29)) Create a TCB as usual, with y as our ISN. Send back a SYNACK. Exception: _If_ we're out of memory for TCBs, set the ``last overflow time'' to the current time. Send the SYNACK anyway, with all fancy options turned off. When an ACK comes back, follow this procedure to find a TCB: (1) Look for a (saddr,sport,daddr,dport) TCB. If it's there, done. (2) If the ``last overflow time'' is earlier than a few minutes ago, give up. (3) Figure out whether our alleged ISN makes sense. This means recomputing y as above, for each of the counters that could have been used in the last few minutes (say, the last four counters), and seeing whether any of the y's match the ISN in the bottom 29 bits. If none of them do, give up. (4) Create a new TCB. The top three bits of our ISN give a usable MSS. Turn off all fancy options.