| Internet-Draft | ECN++ | January 2022 |
| Bagnulo & Briscoe | Expires 4 August 2022 | [Page] |
This document describes an experimental modification to ECN when used with TCP. It allows the use of ECN on the following TCP packets: SYNs, pure ACKs, Window probes, FINs, RSTs and retransmissions.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
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."¶
This Internet-Draft will expire on 4 August 2022.¶
Copyright (c) 2022 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 (https://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 Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
This document may contain material from IETF Documents or IETF Contributions published or made publicly available before November 10, 2008. The person(s) controlling the copyright in some of this material may not have granted the IETF Trust the right to allow modifications of such material outside the IETF Standards Process. Without obtaining an adequate license from the person(s) controlling the copyright in such materials, this document may not be modified outside the IETF Standards Process, and derivative works of it may not be created outside the IETF Standards Process, except to format it for publication as an RFC or to translate it into languages other than English.¶
RFC 3168 [RFC3168] specifies support of Explicit Congestion Notification (ECN) in IP (v4 and v6). By using the ECN capability, network elements (e.g. routers, switches) performing Active Queue Management (AQM) can use ECN marks instead of packet drops to signal congestion to the endpoints of a communication. This results in lower packet loss and increased performance. RFC 3168 also specifies support for ECN in TCP, but solely on data packets. For various reasons it precludes the use of ECN on TCP control packets (TCP SYN, TCP SYN-ACK, pure ACKs, Window probes) and on retransmitted packets. RFC 3168 is silent about the use of ECN on RST and FIN packets. RFC 5562 [RFC5562] is an experimental modification to ECN that enables ECN support for TCP SYN-ACK packets.¶
This document defines an experimental modification to ECN [RFC3168] that shall be called ECN++. It enables ECN support on all the aforementioned types of TCP packet. RFC 5562 (which was called ECN+) is obsoleted by the present specification, because it has the same goal of enabling ECT, but on only one type of control packet. The mechanisms proposed in this document have been defined conservatively and with safety in mind, possibly in some cases at the expense of performance.¶
ECN++ uses a sender-only deployment model. It works whether the two ends of the TCP connection use classic ECN feedback [RFC3168] or Accurate ECN feedback (AccECN [I-D.ietf-tcpm-accurate-ecn]), the two ECN feedback mechanisms for TCP being standardized at the time of writing.¶
Using ECN on initial SYN packets provides significant benefits, as we describe in the next subsection. However, only AccECN provides a way to feed back whether the SYN was CE marked, and RFC 3168 does not. Therefore, implementers of ECN++ are RECOMMENDED to also implement AccECN. Conversely, if AccECN (or an equivalent safety mechanism) is not implemented with ECN++, this specification rules out ECN on the SYN.¶
ECN++ is designed for compatibility with a number of latency improvements to TCP such as TCP Fast Open (TFO [RFC7413]), initial window of 10 SMSS (IW10 [RFC6928]) and Low latency Low Loss Scalable Transport (L4S [I-D.ietf-tsvwg-l4s-arch]), but they can all be implemented and deployed independently. [RFC8311] is a standards track procedural device that relaxes requirements in RFC 3168 and other standards track RFCs that would otherwise preclude the experimental modifications needed for ECN++ and other ECN experiments.¶
The absence of ECN support on TCP control packets and retransmissions has a potential harmful effect. In any ECN deployment, non-ECN-capable packets suffer a penalty when they traverse a congested bottleneck. For instance, with a drop probability of 1%, 1% of connection attempts suffer a timeout of about 1 second before the SYN is retransmitted, which is highly detrimental to the performance of short flows. TCP control packets, particularly TCP SYNs and SYN-ACKs, are important for performance, so dropping them is best avoided.¶
Not using ECN on control packets can be particularly detrimental to performance in environments where the ECN marking level is high. For example, [judd-nsdi] shows that in a controlled private data centre (DC) environment where ECN is used (in conjunction with DCTCP [RFC8257]), the probability of being able to establish a new connection using a non-ECN SYN packet drops to close to zero even when there are only 16 ongoing TCP flows transmitting at full speed. The issue is that DCTCP exhibits a much more aggressive response to packet marking (which is why it is only applicable in controlled environments). This leads to a high marking probability for ECN-capable packets, and in turn a high drop probability for non-ECN packets. Therefore non-ECN SYNs are dropped aggressively, rendering it nearly impossible to establish a new connection in the presence of even mild traffic load.¶
Finally, there are ongoing experimental efforts to promote the adoption of a slightly modified variant of DCTCP (and similar congestion controls) over the Internet to achieve low latency, low loss and scalable throughput (L4S) for all communications [I-D.ietf-tsvwg-l4s-arch]. In such an approach, L4S packets identify themselves using an ECN codepoint [I-D.ietf-tsvwg-ecn-l4s-id]. With L4S, preventing TCP control packets from obtaining the benefits of ECN would not only expose them to the prevailing level of congestion loss, but it would also classify them into a different queue. Then only L4S data packets would be classified into the L4S queue that is expected to have lower latency, while the packets controlling and retransmitting these data packets would still get stuck behind the queue induced by non-L4S-enabled TCP traffic.¶
The goal of the experimental modifications defined in this document is to allow the use of ECN on all TCP packets. Experiments are expected in the public Internet as well as in controlled environments to understand the following issues:¶
The data gathered through the experiments described in this document, particularly under the first 2 bullets above, will help in the redesign of the final mechanism (if needed) for adding ECN support to the different packet types considered in this document.¶
Success criteria: The experiment will be a success if we obtain enough data to have a clearer view of the deployability and benefits of enabling ECN on all TCP packets, as well as any issues. If the results of the experiment show that it is feasible to deploy such changes; that there are gains to be achieved through the changes described in this specification; and that no other major issues may interfere with the deployment of the proposed changes; then it would be reasonable to adopt the proposed changes in a standards track specification that would update RFC 3168.¶
The remainder of this document is structured as follows. In Section 2, we present the terminology used in the rest of the document. In Section 3, we specify the modifications to provide ECN support to TCP SYNs, pure ACKs, Window probes, FINs, RSTs and retransmissions. We describe both the network behaviour and the endpoint behaviour. Section 5 discusses variations of the specification that will be necessary to interwork with a number of popular variants or derivatives of TCP. RFC 3168 provides a number of specific reasons why ECN support is not appropriate for each packet type. In Section 4, we revisit each of these arguments for each packet type to justify why it is reasonable to conduct this experiment.¶
The keywords MUST, MUST NOT, REQUIRED, SHALL, SHALL NOT, SHOULD, SHOULD NOT, RECOMMENDED, NOT RECOMMENDED, MAY, and OPTIONAL in this document, are to be interpreted as described in BCP 14 [RFC2119] when and only when they appear in all capitals [RFC8174].¶
Pure ACK: A TCP segment with the ACK flag set and no data payload.¶
SYN: A TCP segment with the SYN (synchronize) flag set.¶
Window probe: Defined in [RFC0793], a window probe is a TCP segment with only one byte of data sent to learn if the receive window is still zero.¶
FIN: A TCP segment with the FIN (finish) flag set.¶
RST: A TCP segment with the RST (reset) flag set.¶
Retransmission: A TCP segment that has been retransmitted by the TCP sender.¶
TCP client: The initiating end of a TCP connection. Also called the initiator.¶
TCP server: The responding end of a TCP connection. Also called the responder.¶
ECT: ECN-Capable Transport. One of the two codepoints ECT(0) or ECT(1) in the ECN field [RFC3168] of the IP header (v4 or v6). An ECN-capable sender sets one of these to indicate that both transport end-points support ECN. When this specification says the sender sets an ECT codepoint, by default it means ECT(0). Optionally, it could mean ECT(1), which is in the process of being redefined for use by L4S experiments [RFC8311] [I-D.ietf-tsvwg-ecn-l4s-id].¶
Not-ECT: The ECN codepoint set by senders that indicates that the transport is not ECN-capable.¶
CE: Congestion Experienced. The ECN codepoint that an intermediate node sets to indicate congestion [RFC3168]. A node sets an increasing proportion of ECT packets to CE as the level of congestion increases.¶
The experimental ECN++ changes to the specification of TCP over ECN [RFC3168] defined here primarily alter the behaviour of the sending host for each half-connection. However, there are subsections for forwarding elements and receivers below, which recommend that they accept the new packets - they should do already, but might not. This will allow implementers to check the receive side code while they are altering the send-side code. All changes can be deployed at each end-point independently of others and independent of any network behaviour.¶
The feedback behaviour at the receiver depends on whether classic ECN TCP feedback [RFC3168] or Accurate ECN (AccECN) TCP feedback [I-D.ietf-tcpm-accurate-ecn] has been negotiated. Nonetheless, neither receiver feedback behaviour is altered by the present specification.¶
Previously the specification of ECN for TCP [RFC3168] required the sender to set not-ECT on TCP control packets and retransmissions. Some readers of RFC 3168 might have erroneously interpreted this as a requirement for firewalls, intrusion detection systems, etc. to check and enforce this behaviour. Section 4.3 of [RFC8311] updates RFC 3168 to remove this ambiguity. It requires firewalls or any intermediate nodes not to treat certain types of ECN-capable TCP segment differently (except potentially in one attack scenario). This is likely to only involve a firewall rule change in a fraction of cases (at most 0.4% of paths according to the tests reported in Section 4.2.2).¶
In case a TCP sender encounters a middlebox blocking ECT on certain TCP segments, the specification below includes behaviour to fall back to non-ECN. However, this loses the benefit of ECN on control packets. So operators are RECOMMENDED to alter their firewall rules to comply with the requirement referred to above (section 4.3 of [RFC8311]).¶
For each type of control packet or retransmission, the following sections detail changes to the sender's behaviour in two respects: i) whether it sets ECT; and ii) its response to congestion feedback. Table 1 summarises these two behaviours for each type of packet, but the relevant subsection below should be referred to for the detailed behaviour. The subsection on the SYN is more complex than the others, because it has to include fall-back behaviour if the ECT packet appears not to have got through, and caching of the outcome to detect persistent failures.¶
| TCP packet type | ECN field if AccECN f/b negotiated* | ECN field if RFC3168 f/b negotiated* | Congestion Response |
|---|---|---|---|
| SYN | ECT | not-ECT | If AccECN, reduce IW |
| SYN-ACK | ECT | ECT | Reduce IW |
| Pure ACK | ECT | not-ECT | If AccECN, usual cwnd response and optionally [RFC5690] |
| W Probe | ECT | ECT | Usual cwnd response |
| FIN | ECT | ECT | None or optionally [RFC5690] |
| RST | ECT | ECT | N/A |
| Re-XMT | ECT | ECT | Usual cwnd response |
Window probe and retransmission are abbreviated to W Probe an Re-XMT. * For a SYN, "negotiated" means "requested".¶
It can be seen that we recommend against the sender setting ECT on the SYN if it is not requesting AccECN feedback. Therefore it is RECOMMENDED that the AccECN specification [I-D.ietf-tcpm-accurate-ecn] is implemented, along with the ECN++ experiment, because it is expected that ECT on the SYN will give the most significant performance gain, particularly for short flows.¶
Nonetheless, this specification also caters for the case where an ECN++ TCP sender is not using AccECN. This could be because it does not support AccECN or because the other end of the TCP connection does not (AccECN can only be used for a connection if both ends support it).¶
Note that Table 1 does not imply any obligation to set any packet to ECT. ECN++ removes the restrictions that RFC 3168 places against setting ECT on these types of packets, and an implementation would normally be expected to take advantage of this, but it does not have to. Therefore, an implementation of the ECN++ experiment would be compliant if, for instance, it set ECT on some types of control packets but not others.¶
With classic [RFC3168] ECN feedback, the SYN was not expected to be ECN-capable, so the flag provided to feed back congestion was put to another use (it is used in combination with other flags to indicate that the responder supports ECN). In contrast, Accurate ECN (AccECN) feedback [I-D.ietf-tcpm-accurate-ecn] provides a codepoint in the SYN-ACK for the responder to feed back whether the SYN arrived marked CE. Therefore the setting of the IP/ECN field on the SYN is specified separately for each case in the following two subsections.¶
For the ECN++ experiment, if the SYN is requesting AccECN feedback, the TCP sender will also set ECT on the SYN. It can ignore the prohibition in section 6.1.1 of RFC 3168 against setting ECT on such a SYN, as per Section 4.3 of [RFC8311].¶
If the SYN sent by a TCP initiator does not attempt to negotiate Accurate ECN feedback, or does not use an equivalent safety mechanism, it MUST still comply with RFC 3168, which says that a TCP initiator "MUST NOT set ECT on a SYN".¶
The only envisaged examples of "equivalent safety mechanisms" are: a) some future TCP ECN feedback protocol, perhaps evolved from AccECN, that feeds back CE marking on a SYN; b) setting the initial window to 1 SMSS. IW=1 is NOT RECOMMENDED because it could degrade performance, but might be appropriate for certain lightweight TCP implementations.¶
See Section 4.2 for discussion and rationale.¶
If the TCP initiator does not set ECT on the SYN, the rest of Section 3.2.1 does not apply.¶
This subsection only applies if the ECN++ TCP client set ECTs on the SYN and supports AccECN.¶
Until AccECN servers become widely deployed, a TCP initiator that sets ECT on a SYN (which typically implies the same SYN also requests AccECN, as above) SHOULD also maintain a cache entry per server to record servers that it is not worth sending an ECT SYN to, e.g. because they do not support AccECN and therefore have no logic for congestion markings on the SYN. Mobile hosts MAY maintain a cache entry per access network to record 'non-ECT SYN' entries against proxies (see Section 4.2.3). This cache can be implemented as part of the shared state across multiple TCP connections, following [RFC2140].¶
Subsequently the initiator will not set ECT on a SYN to such a server or proxy, but it can still always request AccECN support (because the response will state any earlier stage of ECN evolution that the server supports with no performance penalty). If a server subsequently upgrades to support AccECN, the initiator will discover this as soon as it next connects, then it can remove the server from its cache and subsequently always set ECT for that server.¶
The client can limit the size of its cache of 'non-ECT SYN' servers. Then, while AccECN is not widely deployed, it will only cache the 'non-ECT SYN' servers that are most used and most recently used by the client. As the client accesses servers that have been expelled from its cache, it will simply use ECT on the SYN by default.¶
Servers that do not support ECN as a whole do not need to be recorded separately from non-support of AccECN because the response to a request for AccECN immediately states which stage in the evolution of ECN the server supports (AccECN [I-D.ietf-tcpm-accurate-ecn], classic ECN [RFC3168] or no ECN).¶
The above strategy is named "optimistic ECT and cache failures". It is believed to be sufficient based on three measurement studies and assumptions detailed in Section 4.2.3. However, Section 4.2.3 gives two other strategies and the choice between them depends on the implementer's goals and the deployment prevalence of ECN variants in the network and on servers, not to mention the prevalence of some significant bugs.¶
If the initiator times out without seeing a SYN-ACK, it will separately cache this fact (see fall-back in Section 3.2.1.4 for details).¶
As explained above, this subsection only applies if the ECN++ TCP client sets ECT on the initial SYN.¶
If the SYN-ACK returned to the TCP initiator confirms that the server supports AccECN, it will also be able to indicate whether or not the SYN was CE-marked. If the SYN was CE-marked, and if the initial window is greater than 1 MSS, then, the initiator MUST reduce its Initial Window (IW) and SHOULD reduce it to 1 SMSS (sender maximum segment size). The rationale is the same as that for the response to CE on a SYN-ACK (Section 4.3.2).¶
If the initiator has set ECT on the SYN and if the SYN-ACK shows that the server does not support feedback of a CE on the SYN (e.g. it does not support AccECN) and if the initial congestion window of the initiator is greater than 1 MSS, then the TCP initiator MUST conservatively reduce its Initial Window and SHOULD reduce it to 1 SMSS. A reduction to greater than 1 SMSS MAY be appropriate (see Section 4.2.1). Conservatism is necessary because the SYN-ACK cannot show whether the SYN was CE-marked.¶
If the TCP initiator (host A) receives a SYN from the remote end (host B) after it has sent a SYN to B, it indicates the (unusual) case of a simultaneous open. Host A will respond with a SYN-ACK. Host A will probably then receive a SYN-ACK in response to its own SYN, after which it can follow the appropriate one of the two paragraphs above.¶
In all the above cases, the initiator does not have to back off its retransmission timer as it would in response to a timeout following no response to its SYN [RFC6298], because both the SYN and the SYN-ACK have been successfully delivered through the network. Also, the initiator does not need to exit slow start or reduce ssthresh, which is not even required when a SYN is lost [RFC5681].¶
If an initial window of more than 3 segments is implemented (e.g. IW10 [RFC6928]), Section 5 gives additional recommendations.¶
As explained above, this subsection only applies if the ECN++ TCP client also sets ECT on the initial SYN.¶
An ECT SYN might be lost due to an over-zealous path element (or server) blocking ECT packets that do not conform to RFC 3168. Some evidence of this was found in a 2014 study [ecn-pam], but in a more recent study using 2017 data [Mandalari18] extensive measurements found no case where ECT on TCP control packets was treated any differently from ECT on TCP data packets. Loss is commonplace for numerous other reasons, e.g. congestion loss at a non-ECN queue on the forward or reverse path, transmission errors, etc. Alternatively, the cause of the loss might be the associated attempt to negotiate AccECN, or possibly other unrelated options on the SYN.¶
Therefore, if the timer expires after the TCP initiator has sent the first ECT SYN, it SHOULD make one more attempt to retransmit the SYN with ECT set (backing off the timer as usual). If the retransmission timer expires again, it SHOULD retransmit the SYN with the not-ECT codepoint in the IP header, to expedite connection set-up. If other experimental fields or options were on the SYN, it will also be necessary to follow their specifications for fall-back too. It would make sense to coordinate all the strategies for fall-back in order to isolate the specific cause of the problem.¶
If the TCP initiator is caching failed connection attempts, it SHOULD NOT give up using ECT on the first SYN of subsequent connection attempts until it is clear that a blockage persistently and specifically affects ECT on SYNs. This is because loss is so commonplace for other reasons. Even if it does eventually decide to give up setting ECT on the SYN, it will probably not need to give up on AccECN on the SYN. In any case, if a cache is used, it SHOULD be arranged to expire so that the initiator will infrequently attempt to check whether the problem has been resolved.¶
Other fall-back strategies MAY be adopted where applicable (see Section 4.2.2 for suggestions, and the conditions under which they would apply).¶
For the ECN++ experiment, the TCP implementation will set ECT on SYN-ACKs. It can ignore the requirement in section 6.1.1 of RFC 3168 to set not-ECT on a SYN-ACK, as per Section 4.3 of [RFC8311].¶
A host that sets ECT on SYN-ACKs MUST reduce its initial window in response to any congestion feedback, whether using classic ECN or AccECN (see Section 4.3.1). It SHOULD reduce it to 1 SMSS. This is different to the behaviour specified in an earlier experiment that set ECT on the SYN-ACK [RFC5562]. This is justified in Section 4.3.2.¶
The responder does not have to back off its retransmission timer because the ECN feedback proves that the network is delivering packets successfully and is not severely overloaded. Also the responder does not have to leave slow start or reduce ssthresh, which is not even required when a SYN-ACK has been lost.¶
The congestion response to CE-marking on a SYN-ACK for a server that implements either the TCP Fast Open experiment (TFO [RFC7413]) or experimentation with an initial window of more than 3 segments (e.g. IW10 [RFC6928]) is discussed in Section 5.¶
After the responder sends a SYN-ACK with ECT set, if its retransmission timer expires it SHOULD retransmit one more SYN-ACK with ECT set (and back-off its timer as usual). If the timer expires again, it SHOULD retransmit the SYN-ACK with not-ECT in the IP header. If other experimental fields or options were on the initial SYN-ACK, it will also be necessary to follow their specifications for fall-back. It would make sense to co-ordinate all the strategies for fall-back in order to isolate the specific cause of the problem.¶
This fall-back strategy attempts to use ECT one more time than the strategy for ECT SYN-ACKs in [RFC5562] (which is made obsolete, being superseded by the present specification). Other fall-back strategies MAY be adopted if found to be more effective, e.g. fall-back to not-ECT on the first retransmission attempt.¶
The server MAY cache failed connection attempts, e.g. per client access network. A client-based alternative to caching at the server is given in Section 4.3.3. If the TCP server is caching failed connection attempts, it SHOULD NOT give up using ECT on the first SYN-ACK of subsequent connection attempts until it is clear that the blockage persistently and specifically affects ECT on SYN-ACKs. This is because loss is so commonplace for other reasons (see Section 3.2.1.4). If a cache is used, it SHOULD be arranged to expire so that the server will infrequently attempt to check whether the problem has been resolved.¶
A Pure ACK is an ACK packet that does not carry data, which includes the Pure ACK at the end of TCP's 3-way handshake.¶
For the ECN++ experiment, whether a TCP implementation sets ECT on a Pure ACK depends on whether or not Accurate ECN TCP feedback [I-D.ietf-tcpm-accurate-ecn] has been successfully negotiated for a particular TCP connection, as specified in the following two subsections.¶
If AccECN has not been successfully negotiated for a connection, ECT MUST NOT be set on Pure ACKs by either end.¶
For the ECN++ experiment, if AccECN has been successfully negotiated, either end of the connection will set ECT on Pure ACKs. They can ignore the requirement in section 6.1.4 of RFC 3168 to set not-ECT on a pure ACK, as per Section 4.3 of [RFC8311].¶
See Section 3.3.3 for the implications if a host receives a CE-marked Pure ACK.¶
As explained above, this subsection only applies if AccECN has been successfully negotiated for the TCP connection.¶
A host that sets ECT on pure ACKs SHOULD respond to the congestion signal resulting from pure ACKs being marked with the CE codepoint. The specific response will need to be defined as an update to each congestion control specification. Possible responses to congestion feedback include reducing the congestion window (CWND) and/or regulating the pure ACK rate (see Section 4.4.1.1).¶
Note that, in comparison, TCP Congestion Control [RFC5681] does not require a TCP to detect or respond to loss of pure ACKs at all; it requires no reduction in congestion window or ACK rate.¶
For the ECN++ experiment, the TCP sender will set ECT on window probes. It can ignore the prohibition in section 6.1.6 of RFC 3168 against setting ECT on a window probe, as per Section 4.3 of [RFC8311].¶
A window probe contains a single octet, so it is no different from a regular TCP data segment. Therefore a TCP receiver will feed back any CE marking on a window probe as normal (either using classic ECN feedback or AccECN feedback). The sender of the probe will then reduce its congestion window as normal.¶
A receive window of zero indicates that the application is not consuming data fast enough and does not imply anything about network congestion. Once the receive window opens, the congestion window might become the limiting factor, so it is correct that CE-marked probes reduce the congestion window. This complements cwnd validation [RFC7661], which reduces cwnd as more time elapses without having used available capacity. However, CE-marking on window probes does not reduce the rate of the probes themselves. This is unlikely to present a problem, given the duration between window probes doubles [RFC1122] as long as the receiver is advertising a zero window (currently minimum 1 second, maximum at least 1 minute [RFC6298]).¶
A TCP implementation can set ECT on a FIN.¶
See Section 3.3.4 for the implications if a host receives a CE-marked FIN.¶
A congestion response to a CE-marking on a FIN is not required.¶
After sending a FIN, the endpoint will not send any more data in the connection. Therefore, even if the FIN-ACK indicates that the FIN was CE-marked (whether using classic or AccECN feedback), reducing the congestion window will not affect anything.¶
After sending a FIN, a host might send one or more pure ACKs. If it is using one of the techniques in Section 3.2.3 to regulate the delayed ACK ratio for pure ACKs, it could equally be applied after a FIN. But this is not required.¶
A TCP implementation can set ECT on a RST.¶
See Section 3.3.5 for the implications if a host receives a CE-marked RST.¶
A congestion response to a CE-marking on a RST is not required (and actually not possible).¶
Implementers SHOULD ensure that RST packets (and control packets generally) are always sent out with the same ECN field regardless of the TCP state machine. Otherwise the ECN field could reveal internal TCP state. For instance, the ECN field on a RST ought not to reveal any distinction between a non-listening port, a recently in-use port, and a closed session port.¶
For the ECN++ experiment, the TCP sender will set ECT on retransmitted segments. It can ignore the prohibition in section 6.1.5 of RFC 3168 against setting ECT on retransmissions, as per Section 4.3 of [RFC8311].¶
See Section 3.3.6 for the implications if a host receives a CE-marked retransmission.¶
If the TCP sender receives feedback that a retransmitted packet was CE-marked, it will react as it would to any feedback of CE-marking on a data packet.¶
Extensive measurements in fixed and mobile networks [Mandalari18] have found no evidence of blockages due to ECT being set on any type of TCP control packet.¶
In case traversal problems arise in future, fall-back measures have been specified above, but only for the cases where ECT on the initial packet of a half-connection (SYN or SYN-ACK) is persistently failing to get through.¶
Fall-back measures for blockage of ECT on other TCP control packets MAY be implemented. However they are not specified here given the lack of any evidence they will be needed. Section 4.9 justifies this advice in more detail.¶
The present ECN++ specification primarily concerns the behaviour for sending TCP control packets or retransmissions. Below are a few changes to the receive side of an implementation that are recommended while updating its send side. Nonetheless, where deployment is concerned, ECN++ is still a sender-only deployment, because it does not depend on receivers complying with any of these recommendations.¶
RFC8311 is a standards track update to RFC 3168 in order to (amongst other things) "...allow the use of ECT codepoints on SYN packets, pure acknowledgement packets, window probe packets, and retransmissions of packets..., provided that the changes from RFC 3168 are documented in an Experimental RFC in the IETF document stream."¶
Section 4.3 of RFC 8311 amends every statement in RFC 3168 that precludes the use of ECT on control packets and retransmissions to add "unless otherwise specified by an Experimental RFC in the IETF document stream". The present specification is such an Experimental RFC. Therefore, In order for the present RFC 8311 experiment to be useful, TCP receivers will need to satisfy the following requirements:¶
The following sections give further requirements specific to each type of control packet.¶
These measures are derived from the robustness principle of "... be liberal in what you accept from others", not only to ensure compatibility with the present experimental specification, but also any future protocol changes that allow ECT on any TCP packet.¶
RFC 3168 negotiates the use of ECN for the connection end-to-end using the ECN flags in the TCP header. RFC 3168 originally said that "A host MUST NOT set ECT on SYN ... packets." but it was silent as to what a TCP server ought to do if it receives a SYN packet with a non-zero IP/ECN field anyway.¶
For the avoidance of doubt, the normative statements for all TCP control packets in Section 3.3.1 are interpreted for the specific case when a SYN is received as follows:¶
Rationale: At the time of the writing, some implementations of TCP servers (see Section 4.2.2.2) assume that, if a host receives a SYN with a non-zero IP/ECN field, it must be due to network mangling, and they disable ECN for the rest of the connection. Section 4.2.2.2 cites a measurement study run in 2017 that found no occurrence of this type of network mangling. However, a year earlier, when ECN was enabled on connections from Apple clients, there was a case of a whole network that re-marked the ECN field of every packet to CE (it was rapidly fixed).¶
When ECN was not allowed on SYNs, it made sense to look for a non-zero ECN field on the SYN to detect this type of network mangling. But now that ECN is being allowed on a SYN, detection needs to be more nuanced. A server needs to disable the test on the SYN alone for AccECN SYNs (which was done for Linux RFC 3168 servers in 2019 [relax-strict-ecn]) and for RFC 3168 SYNs it needs to watch for three or four packets all set to CE at the start of a flow. If such mangling is indeed now so rare, it would also be preferable to log each case detected and manually report it to the responsible network, so that the problem will eventually be eliminated.¶
For the avoidance of doubt, the normative statements for all TCP control packets in Section 3.3.1 are interpreted for the specific case when a Pure ACK is received as follows:¶
The question of whether and how the receiver of pure ACKs is required to feed back any CE marks on them is outside the scope of the present specification because it is a matter for the relevant feedback specification ([RFC3168] or [I-D.ietf-tcpm-accurate-ecn]). AccECN feedback is required to count CE marking of any control packet including pure ACKs. Whereas RFC 3168 is silent on this point, so feedback of CE-markings might be implementation specific (see Section 4.4.1.1).¶
The TCP data receiver MUST ignore the CE codepoint on incoming FINs that fail any validity check. The validity check in section 5.2 of [RFC5961] is RECOMMENDED.¶
The "challenge ACK" approach to checking the validity of RSTs (section 3.2 of [RFC5961] is RECOMMENDED at the data receiver.¶
The TCP data receiver MUST ignore the CE codepoint on incoming segments that fail any validity check. The validity check in section 5.2 of [RFC5961] is RECOMMENDED. This will effectively mitigate an attack that uses spoofed data packets to fool the receiver into feeding back spoofed congestion indications to the sender, which in turn would be fooled into continually reducing its congestion window.¶
This section is informative, not normative. It presents counter-arguments against the justifications in the RFC series for disabling ECN on TCP control segments and retransmissions. It also gives rationale for why ECT is safe on control segments that have not, so far, been mentioned in the RFC series. First it addresses over-arching arguments used for most packet types, then it addresses the specific arguments for each packet type in turn.¶
Section 5.2 of RFC 3168 states:¶
We believe this argument is misplaced. TCP does not deliver most control packets reliably. So it is more important to allow control packets to be ECN-capable, which greatly improves reliable delivery of the control packets themselves (see motivation in Section 1.1). ECN also improves the reliability and latency of delivery of any congestion notification on control packets, particularly because TCP does not detect the loss of most types of control packet anyway. Both these points outweigh by far the concern that a CE marking applied to a control packet by one node might subsequently be dropped by another node.¶
The principle to determine whether a packet can be ECN-capable ought to be "do no extra harm", meaning that the reliability of a congestion signal's delivery ought to be no worse with ECN than without. In particular, setting the CE codepoint on the very same packet that would otherwise have been dropped fulfills this criterion, since either the packet is delivered and the CE signal is delivered to the endpoint, or the packet is dropped and the original congestion signal (packet loss) is delivered to the endpoint.¶
The concern about a CE marking being dropped at a subsequent node might be motivated by the idea that ECN-marking a packet at the first node does not remove the packet, so it could go on to worsen congestion at a subsequent node. However, it is not useful to reason about congestion by considering single packets. The departure rate from the first node will generally be the same (fully utilized) with or without ECN, so this argument does not apply.¶
RFC 5562 presents two arguments against ECT marking of SYN packets (quoted verbatim):¶
The first point actually describes two subtly different issues. So below three arguments are countered in turn.¶
This argument certainly applied at the time RFC 5562 was written, when no ECN responder mechanism had any logic to recognize a CE marking on a SYN and, even if logic were added, there was no field in the SYN-ACK to feed it back. The problem was that, during the 3WHS, the flag in the TCP header for ECN feedback (called Echo Congestion Experienced) had been overloaded to negotiate the use of ECN itself.¶
The accurate ECN (AccECN) protocol [I-D.ietf-tcpm-accurate-ecn] has since been designed to solve this problem. Two features are important here:¶
This still does not seem to allow a client to set ECT on a SYN, it only finds out whether the server would have supported it afterwards. The trick the client uses for ECN++ is to set ECT on the SYN optimistically then, if the SYN-ACK reveals that the server wouldn't have understood CE on the SYN, the client responds conservatively as if the SYN was marked with CE.¶
The recommended conservative congestion response is to reduce the initial window, which does not affect the performance of very popular protocols such as HTTP, since it is extremely rare for an HTTP client to send more than one packet as its initial request anyway (for data on HTTP/1 & HTTP/2 request sizes see Fig 3 in [Manzoor17]). Any clients that do frequently use a larger initial window for their first message to the server can cache which servers will not understand ECT on a SYN (see Section 4.2.3 below). If caching is not practical, such clients could reduce the initial window to say IW2 or IW3.¶
Given, until now, ECT-marked SYN packets have been prohibited, it cannot be assumed they will be accepted, by TCP middleboxes or servers.¶
According to a study using 2014 data [ecn-pam] from a limited range of fixed vantage points, for the top 1M Alexa web sites, adding the ECN capability to SYNs was increasing connection establishment failures by about 0.4%.¶
From a wider range of fixed and mobile vantage points, a more recent study in Jan-May 2017 [Mandalari18] found no occurrences of blocking of ECT on SYNs. However, in more than half the mobile networks tested it found wiping of the ECN codepoint at the first hop.¶
Silent failures introduce a retransmission timeout delay (default 1 second) at the initiator before it attempts any fall back strategy (whereas explicit RSTs can be dealt with immediately). Ironically, making SYNs ECN-capable is intended to avoid the timeout when a SYN is lost due to congestion. Fortunately, if there is any discard of ECN-capable SYNs due to policy, it will occur predictably, not randomly like congestion. So the initiator should be able to avoid it by caching those sites that do not support ECN-capable SYNs (see the last paragraph of Section 3.2.1.2).¶
A study conducted in Nov 2017 [Kuehlewind18] found that, of the 82% of the Alexa top 50k web servers that supported ECN, 84% disabled ECN if the IP/ECN field on the SYN was ECT0, CE or either. Given most web servers use Linux, this behaviour can most likely be traced to a patch contributed in May 2012 that was first distributed in v3.5 of the Linux kernel [strict-ecn]. The comment says "RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set. If we receive a SYN packet with these bits set, it means a network is playing bad games with TOS bits. In order to avoid possible false congestion notifications, we disable TCP ECN negociation." Of course, some of the 84% might be due to similar code in other OSs.¶
For brevity we shall call this the "over-strict" ECN test, because it is over-conservative with what it accepts, contrary to Postel's robustness principle. A robust protocol will not usually assume network mangling without comparing with the value originally sent, and one packet is not sufficient to make an assumption with such irreversible consequences anyway.¶
Ironically, networks rarely seem to alter the IP/ECN field on a SYN from zero to non-zero anyway. In a study conducted in Jan-May 2017 over millions of paths from vantage points in a few dozen mobile and fixed networks [Mandalari18], no such transition was observed. With such a small or non-existent incidence of this sort of network mangling, it would be preferable to report any residual problem paths so that they can be fixed.¶
Whatever, the widespread presence of this 'over-strict' test proves that RFC 5562 was correct to expect that ECT would be considered invalid on SYNs. Nonetheless, it is not an insurmountable problem - the over-strict test in Linux was patched in Apr 2019 [relax-strict-ecn] and caching can work round it where previous versions of Linux are running. The prevalence of these "over-strict" ECN servers makes it challenging to cache them all. However, Section 4.2.3 below explains how a cache of limited size can alleviate this problem for a client's most popular sites.¶
For the future, [RFC8311] updates RFC 3168 to clarify that the IP/ECN field does not have to be zero on a SYN if documented in an experimental RFC such as the present ECN++ specification.¶
Given the server handling of ECN on SYNs outlined in Section 4.2.2.2 above, an initiator might combine AccECN with three candidate caching strategies for setting ECT on a SYN:¶
Pessimistic ECT and cache successes: The initiator always requests AccECN, but by default without ECT on the SYN. Then it caches those servers that confirm that they support AccECN as 'ECT SYN OK'. On a subsequent connection to any server that supports AccECN, the initiator can then set ECT on the SYN. When connecting to other servers (non-ECN or classic ECN) it will not set ECT on the SYN, so it will not fail the 'over-strict' ECN test.¶
Longer term, as servers upgrade to AccECN, the initiator is still requesting AccECN, so it will add them to the cache and use ECT on subsequent SYNs to those servers. However, assuming it has to cap the size of the cache, the client will not have the benefit of ECT SYNs to those less frequently used AccECN servers expelled from its cache.¶
Optimistic ECT: The initiator always requests AccECN and by default sets ECT on the SYN. Then, if the server response shows it has no AccECN logic (so it cannot feed back a CE mark), the initiator conservatively behaves as if the SYN was CE-marked, by reducing its initial window.¶
Cache failures: The optimistic ECT strategy can be improved by caching solely those servers that do not support AccECN as 'ECT SYN NOK'. This would include non-ECN servers and all Classic ECN servers whether 'over-strict' or not. On subsequent connections to these non-AccECN servers, the initiator will still request AccECN but not set ECT on the SYN. Then, the connection can still fall back to Classic ECN, if the server supports it, and the initiator can use its full initial window (if it has enough request data to need it).¶
Longer term, as servers upgrade to AccECN, the initiator will remove them from the cache and use ECT on subsequent SYNs to that server.¶
Where an access network operator mediates Internet access via a proxy that does not support AccECN, the optimistic ECT strategy will always fail. This scenario is more likely in mobile networks. Therefore, a mobile host could cache lack of AccECN support per attached access network operator. Whenever it attached to a new operator, it could check a well-known AccECN test server and, if it found no AccECN support, it would add a cache entry for the attached operator. It would only use ECT when neither network nor server were cached. It would only populate its per server cache when not attached to a non-AccECN proxy.¶
For unmanaged environments like the public Internet, pragmatically the choice is between strategies (S1), (S2A) and (S2B). The normative specification for ECT on a SYN in Section 3.2.1 recommends the "optimistic ECT and cache failures" strategy (S2B) but the choice depends on the implementer's motivation for using ECN++, and the deployment prevalence of different technologies and bug-fixes.¶
Note that, if AccECN deployment grows, caching successes (S1) starts off small then grows, while caching failures (S2B) becomes large at first, then shrinks. At half-way, the size of the cache has to be capped with either approach, so the default behaviour for all the servers that do not fit in the cache is as important as the behaviour for the popular servers that do fit.¶
Another strategy would be to send a not-ECT SYN a short delay (below the typical lowest RTT) after an ECT SYN and only accept the non-ECT connection if it returned first. This would reduce the performance penalty for those deploying ECT SYN support. However, this 'happy eyeballs' approach becomes complex when multiple optional features are all tried on the first SYN (or on multiple SYNs), so it is not recommended.¶
[RFC5562] says that ECT SYN packets could be misused by malicious clients to augment "the well-known TCP SYN attack". It goes on to say "a malicious host might be able to inject a large number of TCP SYN packets through a potentially congested ECN-enabled router, congesting it even further."¶
We assume this is a reference to the TCP SYN flood attack (see https://en.wikipedia.org/wiki/SYN_flood), which is an attack against a responder end point. We assume the idea of this attack is to use ECT to get more packets through an ECN-enabled router in preference to other non-ECN traffic so that they can go on to use the SYN flooding attack to inflict more damage on the responder end point. This argument could apply to flooding with any type of packet, but we assume SYNs are singled out because their source address is easier to spoof, whereas floods of other types of packets are easier to block.¶
Mandating Not-ECT in an RFC does not stop attackers using ECT for flooding. Nonetheless, if a standard says SYNs are not meant to be ECT it would make it legitimate for firewalls to discard them. However this would negate the considerable benefit of ECT SYNs for compliant transports and seems unnecessary because RFC 3168 already provides the means to address this concern. In section 7, RFC 3168 says "During periods where ... the potential packet marking rate would be high, our recommendation is that routers drop packets rather then set the CE codepoint..." and this advice is repeated in [RFC7567] (section 4.2.1). This makes it harder for flooding packets to gain from ECT.¶
[ecn-overload] showed that ECT can only slightly augment flooding attacks relative to a non-ECT attack. It was hard to overload the link without causing the queue to grow, which in turn caused the AQM to disable ECN and switch to drop, thus negating any advantage of using ECT. This was true even with the switch-over point set to 25% drop probability (i.e. the arrival rate was 133% of the link rate).¶
The proposed approach in Section 3.2.2 for experimenting with ECN-capable SYN-ACKs is effectively identical to the scheme called ECN+ [ECN-PLUS]. In 2005, the ECN+ paper demonstrated that it could reduce the average Web response time by an order of magnitude. It also argued that adding ECT to SYN-ACKs did not raise any new security vulnerabilities.¶
The feedback behaviour by the initiator in response to a CE-marked SYN-ACK from the responder depends on whether classic ECN feedback [RFC3168] or AccECN feedback [I-D.ietf-tcpm-accurate-ecn] has been negotiated. In either case no change is required to RFC 3168 or the AccECN specification.¶
Some classic ECN client implementations might ignore a CE-mark on a SYN-ACK, or even ignore a SYN-ACK packet entirely if it is set to ECT or CE. This is a possibility because an RFC 3168 implementation would not necessarily expect a SYN-ACK to be ECN-capable. This issue already came up when the IETF first decided to experiment with ECN on SYN-ACKs [RFC5562] and it was decided to go ahead without any extra precautionary measures. This was because the probability of encountering the problem was believed to be low and the harm if the problem arose was also low (see Appendix B of RFC 5562).¶
The IETF has already specified an experiment with ECN-capable SYN-ACK packets [RFC5562]. It was inspired by the ECN+ paper, but it specified a much more conservative congestion response to a CE-marked SYN-ACK, called ECN+/TryOnce. This required the server to reduce its initial window to 1 segment (like ECN+), but then the server had to send a second SYN-ACK and wait for its ACK before it could continue with its initial window of 1 SMSS. The second SYN-ACK of this 5-way handshake had to carry no data, and had to disable ECN, but no justification was given for these last two aspects.¶
The present ECN++ experimental specification obsoletes RFC 5562 because it uses the ECN+ congestion response, not ECN+/TryOnce. First we argue against the rationale for ECN+/TryOnce given in sections 4.4 and 6.2 of [RFC5562]. It starts with a rather too literal interpretation of the requirement in RFC 3168 that says TCP's response to a single CE mark has to be "essentially the same as the congestion control response to a *single* dropped packet." TCP's response to a dropped initial (SYN or SYN-ACK) packet is to wait for the retransmission timer to expire (currently 1s). However, this long delay assumes the worst case between two possible causes of the loss: a) heavy overload; or b) the normal capacity-seeking behaviour of other TCP flows. When the network is still delivering CE-marked packets, it implies that there is an AQM at the bottleneck and that it is not overloaded. This is because an AQM under overload will disable ECN (as recommended in section 7 of RFC 3168 and repeated in section 4.2.1 of RFC 7567). So scenario (a) can be ruled out. Therefore, TCP's response to a CE-marked SYN-ACK can be similar to its response to the loss of any packet, rather than backing off as if the special initial packet of a flow has been lost.¶
How TCP responds to the loss of any single packet depends what it has just been doing. But there is not really a precedent for TCP's response when it experiences a CE mark having sent only one (small) packet. If TCP had been adding one segment per RTT, it would have halved its congestion window, but it hasn't established a congestion window yet. If it had been exponentially increasing it would have exited slow start, but it hasn't started exponentially increasing yet so it hasn't established a slow-start threshold.¶
Therefore, we have to work out a reasoned argument for what to do. If an AQM is CE-marking packets, it implies there is already a queue and it is probably already somewhere around the AQM's operating point - it is unlikely to be well below and it might be well above. So, the more data packets that the client sends in its IW, the more likely at least one will be CE marked, leading it to exit slow-start early. On the other hand, it is highly unlikely that the SYN-ACK itself pushed the AQM into congestion, so it will be safe to introduce another single segment immediately (1 RTT after the SYN-ACK). Therefore, starting to probe for capacity with a slow start from an initial window of 1 segment seems appropriate to the circumstances. This is the approach adopted in Section 3.2.2.¶
An alternative to the server caching failed connection attempts would be for the server to rely on the client caching failed attempts (on the basis that the client would cache a failure whether ECT was blocked on the SYN or the SYN-ACK). This strategy cannot be used if the SYN does not request AccECN support. It works as follows: if the server receives a SYN that requests AccECN support but is set to not-ECT, it replies with a SYN-ACK also set to not-ECT. If a middlebox only blocks ECT on SYNs, not SYN-ACKs, this strategy might disable ECN on a SYN-ACK when it did not need to, but at least it saves the server from maintaining a cache.¶
Section 5.2 of RFC 3168 gives the following arguments for not allowing the ECT marking of pure ACKs (ACKs not piggy-backed on data):¶
Later on, in section 6.1.4 it reads:¶
We next address each of the arguments presented above.¶
The first argument is a specific instance of the reliability argument for the case of pure ACKs. This has already been addressed by countering the general reliability argument in Section 4.1.¶
The second argument says that ECN ought not to be enabled unless there is a mechanism to respond to it. This argument actually comprises three sub-arguments:¶
The last two arguments involve value judgements, but they both depend on the concrete technical question of mechanism feasibility, which will therefore be addressed first in Section 4.4.1 below. Then Section 4.4.2 draws conclusions by addressing the value judgements in the other two questions.¶
The question of whether the receiver of pure ACKs is required to detect and feed back any CE-marking is outside the scope of the present specification - it is a matter for the relevant feedback specification (classic ECN [RFC3168] and AccECN [I-D.ietf-tcpm-accurate-ecn]). The response to congestion feedback is also out of scope, because it would be defined in the base TCP congestion control specification [RFC5681] or its variants.¶
Nonetheless, in order to decide whether the present ECN++ experimental specification should require a host to set ECT on pure ACKs, we only need to know whether a response mechanism would be feasible - we do not have to standardize it. So the bullets below assess, for each type of feedback, whether the three stages of the congestion response mechanism could all work.¶
Can the receiver of a pure ACK detect a CE marking on it?:¶
As a general rule, TCP does not ACK a pure ACK. However, even if the receiver of a CE-mark on a pure ACK does not feed it back immediately, it could still include it within subsequent feedback, for instance when it later sends a data segment (if it ever does):¶
Therefore a congestion response mechanism is clearly feasible if AccECN has been negotiated, but the position is unknown for the installed base of classic ECN feedback.¶
This subsection explores issues that congestion control designers will need to consider when defining a cwnd response to CE-marked Pure ACKs.¶
A CE-mark on a Pure ACK does not mean that only Pure ACKs are causing congestion. It only means that the marked Pure ACK is part of an aggregate that is collectively causing a bottleneck queue to randomly CE-mark a fraction of the packets. A CE-mark on a Pure ACK might be due to data packets in other flows through the same bottleneck, due to data packets interspersed between Pure ACKs in the same half-connection, or just due to the rate of Pure ACKs alone. (RFC 3168 only considered the last possibility, which led to the argument that ECN-enabled Pure ACKs had to be deferred, because ACK congestion control was a research issue.)¶
If a host has been sending a mix of Pure ACKs and data, it doesn't need to work out whether a particular CE mark was on a Pure ACK or not; it just needs to respond to congestion feedback as a whole by reducing its congestion window (cwnd), which limits the data it can launch into flight through the congested bottleneck. If it is purely receiving data and sending only Pure ACKs, reducing cwnd will have caused it no harm, having no effect on its ACK rate (the next subsection addresses that).¶
However, when a host is sending data as well as Pure ACKs, it would not be right for CE-marks on Pure ACKs and on data packets to induce the same reduction in cwnd. A possible way to address this issue would be to weight the response by the size of the marked packets (assuming the congestion control supports a weighted response, e.g. [RFC8257]). For instance, one could calculate the fraction of CE-marked bytes (headers and data) over each round trip (say) as follows:¶
Header bytes can be calculated by multiplying a packet count by a nominal header size, which is possible with AccECN feedback, because it gives a count of CE-marked packets (as well as CE-marked bytes). The above simple aggregate calculation caters for the full range of scenarios; from all Pure ACKs to just a few interspersed with data packets.¶
Note that any mechanism that reduces cwnd due to CE-marked Pure ACKs would need to be integrated with the congestion window validation mechanism [RFC7661], which already conservatively reduces cwnd over time because cwnd becomes stale if it is not used to fill the pipe.¶
Reducing the congestion window will have no effect on the rate of pure ACKs. The worst case here is if the bottleneck is congested solely with pure ACKs, but it could also be problematic if a large fraction of the load was from unresponsive ACKs, leaving little or no capacity for the load from responsive data.¶
Since RFC 3168 was published, experimental Acknowledgement Congestion Control (AckCC) techniques have been documented in [RFC5690] (informational). So any pair of TCP end-points can choose to agree to regulate the delayed ACK ratio in response to lost or CE-marked pure ACKs. However, the protocol has a number of open issues concerning deployment (e.g. it requires support from both ends, it relies on two new TCP options, one of which is required on the SYN where option space is at a premium and, if either option is blocked by a middlebox, no fall-back behaviour is specified).¶
The new TCP options address two problems, namely that TCP had: i) no mechanism to allow ECT to be set on pure ACKs; and ii) no mechanism to feed back loss or CE-marking of pure ACKs. A combination of the present specification and AccECN addresses both these problems, at least for CE-marking. So it might now be possible to design an ECN-specific ACK congestion control scheme without the extra TCP options proposed in RFC 5690. However, such a mechanism is out of scope of the present document.¶
Setting aside the practicality of RFC 5690, the need for AckCC has not been conclusively demonstrated. It has been argued that the Internet has survived so far with no mechanism to even detect loss of pure ACKs. However, it has also been argued that ECN is not the same as loss. Packet discard can naturally thin the ACK load to whatever the bottleneck can support, whereas ECN marking does not (it queues the ACKs instead). Nonetheless, RFC 3168 (section 7) recommends that an AQM switches over from ECN marking to discard when the marking probability becomes high. Therefore discard can still be relied on to thin out ECN-enabled pure ACKs as a last resort.¶
In the case when AccECN has been negotiated, it provides a feasible congestion response mechanism, so the arguments for ECT on pure ACKs heavily outweigh those against. ECN is always more and never less reliable for delivery of congestion notification. A cwnd reduction needs to be considered by congestion control designers as a response to congestion on pure ACKs. Separately, AckCC (or an improved variant exploiting AccECN) could optionally be used to regulate the spacing between pure ACKs. However, it is not clear whether AckCC is justified. If it is not, packet discard will still act as the "congestion response of last resort" by thinning out the traffic. In contrast, not setting ECT on pure ACKs is certainly detrimental to performance, because when a pure ACK is lost it can prevent the release of new data.¶
In the case when Classic ECN has been negotiated, the argument for ECT on pure ACKs is less clear-cut. Some of the installed base of RFC 3168 implementations might happen to (unintentionally) provide a feedback mechanism to support a cwnd response. For those that did not, setting ECT on pure ACKs would be better for the flow's own performance than not setting it. However, where there was no feedback mechanism, setting ECT could do slightly more harm than not setting it. AckCC could provide a complementary response mechanism, because it is designed to work with RFC 3168 ECN, but it has deployment challenges. In summary, a congestion response mechanism is unlikely to be feasible with the installed base of classic ECN.¶
This specification uses a safe approach. Allowing hosts to set ECT on Pure ACKs without a feasible response mechanism could result in risk. It would certainly improve the flow's own performance, but it would slightly increase potential harm to others. Morevoer, if would set an undesirable precedent for setting ECT on packets with no mechanism to respond to any resulting congestion signals. Therefore, Section 3.2.3 allows ECT on Pure ACKs if AccECN feedback has been negotiated, but not with classic RFC 3168 ECN feedback.¶
Section 6.1.6 of RFC 3168 presents only the reliability argument for prohibiting ECT on Window probes:¶
The reliability argument has already been addressed in Section 4.1.¶
Allowing ECT on window probes could considerably improve performance because, once the receive window has reopened, if a window probe is lost the sender will stall until the next window probe reaches the receiver, which might be after the maximum retransmission timeout (at least 1 minute [RFC6928]).¶
On the bright side, RFC 3168 at least specifies the receiver behaviour if a CE-marked window probe arrives, so changing the behaviour ought to be less painful than for other packet types.¶
RFC 3168 is silent on whether a TCP sender can set ECT on a FIN. A FIN is considered as part of the sequence of data, and the rate of pure ACKs sent after a FIN could be controlled by a CE marking on the FIN. Therefore there is no reason not to set ECT on a FIN.¶
RFC 3168 is silent on whether a TCP sender can set ECT on a RST. The host generating the RST message does not have an open connection after sending it (either because there was no such connection when the packet that triggered the RST message was received or because the packet that triggered the RST message also triggered the closure of the connection).¶
Moreover, the receiver of a CE-marked RST message can either: i) accept the RST message and close the connection; ii) emit a so-called challenge ACK in response (with suitable throttling) [RFC5961] and otherwise ignore the RST (e.g. because the sequence number is in-window but not the precise number expected next); or iii) discard the RST message (e.g. because the sequence number is out-of-window). In the first two cases there is no point in echoing any CE mark received because the sender closed its connection when it sent the RST. In the third case it makes sense to discard the CE signal as well as the RST.¶
Although a congestion response following a CE-marking on a RST does not appear to make sense, the following factors have been considered before deciding whether the sender ought to set ECT on a RST message:¶
RST messages are used to both mount and mitigate attacks:¶
AQMs are advised to disable ECN marking during persistent overload, so:¶
If ECT were prohibited on RSTs¶
So, on balance, it has been decided that it is worth experimenting with ECT on RSTs. During experiments, if the ECN capability on RSTs is found to open a vulnerability that is hard to close, this decision can be reversed, before it is specified for the standards track.¶
RFC 3168 says the sender "MUST NOT" set ECT on retransmitted packets. The rationale for this consumes nearly 2 pages of RFC 3168, so the reader is referred to section 6.1.5 of RFC 3168, rather than quoting it all here. There are essentially three arguments, namely: reliability; DoS attacks; and over-reaction to congestion. We address them in order below.¶
The reliability argument has already been addressed in Section 4.1.¶
Protection against DoS attacks is not afforded by prohibiting ECT on retransmitted packets. An attacker can set CE on spoofed retransmissions whether or not it is prohibited by an RFC. Protection against the DoS attack described in section 6.1.5 of RFC 3168 is solely afforded by the requirement that "the TCP data receiver SHOULD ignore the CE codepoint on out-of-window packets". Therefore in Section 3.2.7 the sender is allowed to set ECT on retransmitted packets, in order to reduce the chance of them being dropped. We also strengthen the receiver's requirement from "SHOULD ignore" to "MUST ignore". And we generalize the receiver's requirement to include failure of any validity check, not just out-of-window checks, in order to include the more stringent validity checks in RFC 5961 that have been developed since RFC 3168.¶
A consequence is that, for those retransmitted packets that arrive at the receiver after the original packet has been properly received (so-called spurious retransmissions), any CE marking will be ignored. There is no problem with that because the fact that the original packet has been delivered implies that the sender's original congestion response (when it deemed the packet lost and retransmitted it) was unnecessary.¶
Finally, the third argument is about over-reacting to congestion. The argument goes that, if a retransmitted packet is dropped, the sender will not detect it, so it will not react again to congestion (it would have reduced its congestion window already when it retransmitted the packet). Whereas, if retransmitted packets can be CE tagged instead of dropped, senders could potentially react more than once to congestion. However, we argue that it is legitimate to respond again to congestion if it still persists in subsequent round trip(s).¶
Therefore, in all three cases, it is not incorrect to set ECT on retransmissions.¶
Extensive experiments have found no evidence of any traversal problems with ECT on any TCP control packet [Mandalari18]. Nonetheless, Sections 3.2.1.4 and 3.2.2.3 specify fall-back measures if ECT on the first packet of each half-connection (SYN or SYN-ACK) appears to be blocking progress. Here, the question of fall-back measures for ECT on other control packets is explored. It supports the advice given in Section 3.2.8; until there's evidence that something's broken, don't fix it.¶
If an implementation has had to disable ECT to ensure the first packet of a flow (SYN or SYN-ACK) gets through, the question arises whether it ought to disable ECT on all subsequent control packets within the same TCP connection. Without evidence of any such problems, this seems unnecessarily cautious. Particularly given it would be hard to detect loss of most other types of TCP control packets that are not ACK'd. And particularly given that unnecessarily removing ECT from other control packets could lead to performance problems, e.g. by directing them into another queue [I-D.ietf-tsvwg-ecn-l4s-id] or over a different path, because some broken multipath equipment (erroneously) routes based on all 8 bits of the Diffserv field.¶
In the case where a connection starts without ECT on the SYN (perhaps because problems with previous connections had been cached), there will have been no test for ECT traversal in the client-server direction until the pure ACK that completes the handshake. It is possible that some middlebox might block ECT on this pure ACK or on later retransmissions of lost packets. Similarly, after a route change, the new path might include some middlebox that blocks ECT on some or all TCP control packets. However, without evidence of such problems, the complexity of a fix does not seem worthwhile.¶
The following subsections discuss any interactions between setting ECT on all packets and using the following popular variants of TCP: IW10 and TFO. It also briefly notes the possibility that the principles applied here should translate to protocols derived from TCP. This section is informative not normative, because no interactions have been identified that require any change to specifications. The subsection on IW10 discusses potential changes to specifications but recommends that no changes are needed.¶
The designs of the following TCP variants have also been assessed and found not to interact adversely with ECT on TCP control packets: SYN cookies (see Appendix A of [RFC4987] and section 3.1 of [RFC5562]), TCP Fast Open (TFO [RFC7413]) and L4S [I-D.ietf-tsvwg-l4s-arch].¶
IW10 is an experiment to determine whether it is safe for TCP to use an initial window of 10 SMSS [RFC6928].¶
This subsection does not recommend any additions to the present specification in order to interwork with IW10. The specifications as they stand are safe, and there is only a corner-case with ECT on the SYN where performance could be occasionally improved, as explained below.¶
As specified in Section 3.2.1.1, a TCP initiator will typically only set ECT on the SYN if it requests AccECN support. If, however, the SYN-ACK tells the initiator that the responder does not support AccECN, Section 3.2.1.1 advises the initiator to conservatively reduce its initial window, preferably to 1 SMSS because, if the SYN was CE-marked, the SYN-ACK has no way to feed that back.¶
If the initiator implements IW10, it seems rather over-conservative to reduce IW from 10 to 1 just in case a congestion marking was missed. Nonetheless, a reduction to 1 SMSS will rarely harm performance, because:¶
If a responder receives feedback that the SYN-ACK was CE-marked, Section 3.2.2.2 recommends that it reduces its initial window, preferably to 1 SMSS. When the responder also implements IW10, it might again seem rather over-conservative to reduce IW from 10 to 1. But in this case the rationale is somewhat different:¶
Experimentation will be needed to determine the best strategy. It should be noted that experience from recent congestion avoidance experiments where the window is reduced by less than half is not necessarily applicable to a flow start scenario. Reducing cwnd by less is one thing. Reducing an increase in cwnd by less is another.¶
TCP Fast Open (TFO [RFC7413]) is an experiment to remove the round trip delay of TCP's 3-way hand-shake (3WHS). A TFO initiator caches a cookie from a previous connection with a TFO-enabled server. Then, for subsequent connections to the same server, any data included on the SYN can be passed directly to the server application, which can then return up to an initial window of response data on the SYN-ACK and on data segments straight after it, without waiting for the ACK that completes the 3WHS.¶
The TFO experiment and the present experiment to add ECN-support for TCP control packets can be combined without altering either specification, which is justified as follows:¶
A Low Latency Low Loss Scalable throughput (L4S) variant of TCP such as TCP Prague [PragueLinux] is mandated to negotiate AccECN feedback, and strongly recommended to use ECN++ [I-D.ietf-tsvwg-ecn-l4s-id].¶
The L4S experiment and the present ECN++ experiment can be combined without altering any of the specifications. The only difference would be in the recommendation of the best SYN cache strategy.¶
The normative specification for ECT on a SYN in Section 3.2.1 recommends the "optimistic ECT and cache failures" strategy (S2B defined in Section 4.2.3) for the general Internet. However, if a user's Internet access bottleneck supported L4S ECN but not Classic ECN, the "optimistic ECT without a cache" strategy (S2A) would make most sense, because there would be little point trying to avoid the 'over-strict' test and negotiate Classic ECN, if L4S ECN but not Classic ECN was available on that user's access link (as is the case with Low Latency DOCSIS [DOCSIS3.1]).¶
Strategy (S2A) is the simplest, because it requires no cache. It would satisfy the goal of an implementer who is solely interested in ultra-low latency using AccECN and ECN++ (e.g. accessing L4S servers) and is not concerned about fall-back to Classic ECN (e.g. when accessing other servers).¶
Experience from experiments on adding ECN support to all TCP packets ought to be directly transferable between TCP and other transport protocols, like SCTP or QUIC.¶
Stream Control Transmission Protocol (SCTP [RFC4960]) is a standards track transport protocol derived from TCP. SCTP currently does not include ECN support, but Appendix A of RFC 4960 broadly describes how it would be supported and a (long-expired) draft on the addition of ECN to SCTP has been produced [I-D.stewart-tsvwg-sctpecn]. This draft avoided setting ECT on control packets and retransmissions, closely following the arguments in RFC 3168.¶
QUIC [RFC9000] is another standards track transport protocol offering similar services to TCP but intended to exploit some of the benefits of running over UDP. Building on the arguments in the current draft, a QUIC sender sets ECT(0) on all packets.¶
Section 3.2.6 considers the question of whether ECT on RSTs will allow RST attacks to be intensified. There are several security arguments presented in RFC 3168 for preventing the ECN marking of TCP control packets and retransmitted segments. We believe all of them have been properly addressed in Section 4, particularly Section 4.2.4 and Section 4.8 on DoS attacks using spoofed ECT-marked SYNs and spoofed CE-marked retransmissions.¶
Section 3.2.6 on sending TCP RSTs points out that implementers need to take care to ensure that the ECN field on a RST does not depend on TCP's state machine. Otherwise the internal information revealed could be of use to potential attackers. This point applies more generally to all control packets, not just RSTs.¶
There are no IANA considerations in this memo.¶
Thanks to Mirja Kuehlewind, David Black, Padma Bhooma, Gorry Fairhurst, Michael Scharf, Yuchung Cheng and Christophe Paasch for their useful reviews. Richard Scheffenegger provided useful advice gained from implementing ECN++ for FreeBSD.¶
The work of Marcelo Bagnulo has been performed in the framework of the H2020-ICT-2014-2 project 5G NORMA. His contribution reflects the consortium's view, but the consortium is not liable for any use that may be made of any of the information contained therein.¶
Bob Briscoe's contribution was partly funded by the Research Council of Norway through the TimeIn project, partly by CableLabs and partly by the Comcast Innovation Fund. The views expressed here are solely those of the authors.¶