| RTGWG | E. Nordmark (ed) |
| Internet-Draft | Arista Networks |
| Intended status: Informational | A. Tian |
| Expires: December 3, 2015 | Ericsson Inc. |
| J. Gross | |
| VMware | |
| J. Hudson | |
| Brocade Communications Systems, Inc. | |
| L. Kreeger | |
| Cisco Systems, Inc. | |
| P. Garg | |
| Microsoft | |
| P. Thaler | |
| Broadcom Corporation | |
| T. Herbert | |
| June 2015 |
Encapsulation Considerations
draft-ietf-rtgwg-dt-encap-00
The IETF Routing Area director has chartered a design team to look at common issues for the different data plane encapsulations being discussed in the NVO3 and SFC working groups and also in the BIER BoF, and also to look at the relationship between such encapsulations in the case that they might be used at the same time. The purpose of this design team is to discover, discuss and document considerations across the different encapsulations in the different WGs/BoFs so that we can reduce the number of wheels that need to be reinvented in the future.
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 http://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 December 3, 2015.
Copyright (c) 2015 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
There have been multiple efforts over the years that have resulted in new or modified data plane behaviors involving encapsulations. That includes IETF efforts like MPLS, LISP, and TRILL but also industry efforts like VXLAN and NVGRE. These collectively can be seen as a source of insight into the properties that data planes need to meet. The IETF is currently working on potentially new encapsulations in NVO3 and SFC and considering working on BIER. In addition there is work on tunneling in the INT area.
This is a short term design team chartered to collect and construct useful advice to parties working on new or modified data plane behaviors that include additional encapsulations. The goal is for the group to document useful advice gathered from interacting with ongoing efforts. An Internet Draft will be produced for IETF92 to capture that advice, which will be discussed in RTGWG.
Data plane encapsulations face a set of common issues such as:
The design team will provide advice on those issues. The intention is that even where we have different encapsulations for different purposes carrying different information, each such encapsulation doesn't have to reinvent the wheel for the above common issues.
The design team will look across the routing area in particular at SFC, NVO3 and BIER. It will not be involved in comparing or analyzing any particular encapsulation formats proposed in those WGs and BoFs but instead focus on common advice.
The references provide background information on NVO3, SFC, and BIER. In particular, NVO3 is introduced in [RFC7364], [RFC7365], and [I-D.ietf-nvo3-arch]. SFC is introduced in [I-D.ietf-sfc-architecture] and [I-D.ietf-sfc-problem-statement]. Finally, the information on BIER is in [I-D.shepherd-bier-problem-statement], [I-D.wijnands-bier-architecture], and [I-D.wijnands-mpls-bier-encapsulation]. We assume the reader has some basic familiarity with those proposed encapsulations. The Related Work section points at some prior work that relates to the encapsulation considerations in this document.
Encapsulation protocols typically have some unique information that they need to carry. In some cases that information might be modified along the path and in other cases it is constant. The in-flight modifications has impacts on what it means to provide security for the encapsulation headers.
Even if information isn't modified in flight there might be devices that wish to inspect that information. For instance, one can envision future NVO3 security devices which filter based on the virtual network identifier.
The need for extensibility is different across the protocols
The extensibility needs and constraints might be different when considering hardware vs. software implementations of the encapsulation headers. NIC hardware might have different constraints than switch hardware.
As the IETF designs these encapsulations the different WGs solve the issues for their own encapsulation. But there are likely to be future cases when the different encapsulations are combined in the same header. For instance, NVO3 might be a "transport" used to carry SFC between the different hops in the service chain.
Most of the issues discussed in this document are not new. The IETF and industry as specified and deployed many different encapsulation or tunneling protocols over time, ranging from simple IP-in-IP and GRE encapsulation, IPsec, pseudo-wires, session-based approached like L2TP, and the use of MPLS control and data planes. IEEE 802 has also defined layered encapsulation for Provider Backbone Bridges (PBB) and IEEE 802.1Qbp (ECMP). This document tries to leverage what we collectively have learned from that experience and summarize what would be relevant for new encapsulations like NVO3, SFC, and BIER.
[This section is mostly a repeat of the charter but with a few modifications and additions.]
Any new encapsulation protocol would need to address a large set of issues that are not central to the new information that this protocol intends to carry. The common issues explored in this document are:
The degree to which these common issues apply to a particular encapsulation can differ based on the intended purpose of the encapsulation. But it is useful to understand all of them before determining which ones apply.
It is important to keep in mind what we are trying to cover and not cover in this document and effort. This is
However, this document and effort is NOT intended to:
While the origin and focus of this document is the routing area and in particular NVO3, SFC, and BIER, the considerations apply to other encapsulations that are being defined in the IETF and elsewhere. There seems to be an increase in the number of encapsulations being defined to run over UDP, where there might already exist an encapsulation over IP or Ethernet. Feedback on how these considerations apply in those contexts is welcome.
The design center for the new encapsulations is a well-managed network. That network can be a datacenter network (plus datacenter interconnect) or a service provider network. Based on the existing and proposed encapsulations in those environment it is reasonable to make these assumptions:
The goal is to implement these encapsulations in hardware and software hence we can't assume that the needs of either implementation approach can trump the needs of the other. In particular, around extensibility the needs and constraints might be quite different.
The capitalized keyword MUST is used as defined in http://en.wikipedia.org/wiki/Julmust
TBD: Refer to existing documents for at least NVO3 and SFC terminology. We use at least the VNI ID in this document.
In many cases the encapsulation format needs to enable ECMP in unmodified routers. Those routers might use different fields in TCP/UDP packets to do ECMP without a risk of reordering a flow. Note that the same entropy might also be used at layer 2 e.g. for Link Aggregation (LAG).
The common way to do ECMP-enabled encapsulation over IP today is to add a UDP header and to use UDP with the UDP source port carrying entropy from the inner/original packet headers as in LISP [RFC6830]. The total entropy consists of 14 bits in the UDP source port (using the ephemeral port range) plus the outer IP addresses which seems to be sufficient for entropy; using outer IPv6 headers would give the option for more entropy should it be needed in the future.
In some environments it might be fine to use all 16 bits of the port range. However, middleboxes might make assumptions about the system ports or user ports. But they should not make any assumptions about the ports in the Dynamic and/or Private Port range, which have the two MSBs set to 11b.
The UDP source port might change over the lifetime of an encapsulated flow, for instance for DoS mitigation or re-balancing load across ECMP. Such changes need to consider reordering if there are packets in flight for the flow.
There is some interaction between entropy and OAM and extensibility mechanism. It is desirable to be able to send OAM packets to follow the same path as network packets. Hence OAM packets should use the same entropy mechanism as data packets. While routers might use information in addition the entropy field and outer IP header, they can not use arbitrary parts of the encapsulation header since that might result in OAM frames taking a different path. Likewise if routers look past the encapsulation header they need to be aware of the extensibility mechanism(s) in the encapsulation format to be able to find the inner headers in the presence of extensions; OAM frames might use some extensions e.g. for timestamps.
Architecturally the entropy and the next header field are really part of enclosing delivery header. UDP with entropy goes hand-in-hand with the outer IP header. Thus the UDP entropy is present for the underlay IP routers the same way that an MPLS entropy label is present for LSRs. The entropy above is all about providing entropy for the outer delivery of the encapsulated packets.
It has been suggested that when IPv6 is used it would not be necessary to add a UDP header for entropy, since the IPv6 flow label can be used for entropy. (This assumes that there is an IP protocol number for the encapsulation in addition to a UDP destination port number since UDP would be used with IPv4 underlay. And any use of UDP checksums would need to be replaced by an encaps-specific checksum or secure hash.) While such an approach would save 8 bytes of headers when the underlay is IPv6, it does assume that the underlay routers use the flow label for ECMP, and it also would make the IPv6 approach different than the IPv4 approach. Currently the leaning is towards recommending using the UDP encapsulation for both IPv4 and IPv6 underlay. The IPv6 flow label can be used for additional entropy if need be. There is more detailed discussion for using the IPv6 flow label for tunnels in [RFC6438].
Note that in the proposed BIER encapsulation [I-D.wijnands-mpls-bier-encapsulation], there is an an 8-bit field which specifies an entropy value that can be used for load balancing purposes. This entropy is for the BIER forwarding decisions, which is independent of any outer delivery ECMP between BIER routers. Thus it is not part of the delivery ECMP discussed in this section.
In summary:
Next-protocol indications appear in three different contexts for encapsulations.
Firstly, the transport delivery mechanism for the encapsulations we discuss in this document need some way to indicate which encapsulation header (or other payload) comes next in the packet. Some encapsulations might be identified by a UDP port; others might be identified by an Ethernet type or IP protocol number. Which approach is used is a function of the preceding header the same way as IPv4 is identified by both an Ethernet type and an IP protocol number (for IP-in-IP). In some cases the header type is implicit in some session (L2TP) or path (MPLS) setup. But this is largely beyond the control of the encapsulation protocol. For instance, if there is a requirement to carry the encapsulation after an Ethernet header, then an Ethernet type is needed. If required to be carried after an IP/UDP header, then a UDP port number is needed. For UDP port numbers there are considerations for port number conservation described in [I-D.ietf-tsvwg-port-use].
It is worth mentioning that in the MPLS case of no implicit protocol type many forwarding devices peek at the first nibble of the payload to determine whether to apply IPv4 or IPv6 L3/L4 hashes for load balancing [RFC7325]. That behavior places some constraints on other payloads carried over MPLS and some protocol define an initial control word in the payload with a value of zero in its first nibble [RFC4385] to avoid confusion with IPv4 and IPv6 payload headers.
Secondly, the encapsulation needs to indicate the type of its payload, which is in scope for the design of the encapsulation. We have existing protocols which use Ethernet types (such as GRE). Here each encapsulation header can potentially makes its own choices between:
Thirdly, if the IETF ends up defining multiple encapsulations at about the same time, and there is some chance that multiple such encapsulations can be combined in the same packet, there is a question whether it makes sense to use a common approach and numbering space for the encapsulation across the different protocols. A common approach might not be beneficial as long as there is only one way to indicate e.g., SFC inside NVO3.
Many Internet protocols use fixed values (typically managed by the IANA function) for their next-protocol field. That facilitates interpretation of packets by middleboxes and e.g., for debugging purposes, but might make the protocol evolution inflexible. Our collective experience with MPLS shows an alternative where the label can be viewed as an index to a table containing processing instructions and the table content can be managed in different ways. Encapsulations might want to consider the tradeoffs between such more flexible versus more fixed approaches.
In summary:
A common approach today is to assume that the underlay have sufficient MTU to carry the encapsulated packets without any fragmentation and reassembly at the tunnel endpoints. That is sufficient when the operator of the ingress and egress have full control of the paths between those endpoints. And it makes for simpler (hardware) implementations if fragmentation and reassembly can be avoided.
However, even under that assumption it would be beneficial to be able to detect when there is some misconfiguration causing packets to be dropped due to MTU issues. One way to do this is to have the encapsulator set the don't-fragment (DF) flag in the outer IPv4 header and receive and log any received ICMP "packet too big" (PTB) errors. Note that no flag needs to be set in an outer IPv6 header [RFC2460].
Encapsulations could also define an optional tunnel fragmentation and reassembly mechanism which would be useful in the case when the operator doesn't have full control of the path, or when the protocol gets deployed outside of its original intended context. Such a mechanism would be required if the underlay might have a path MTU which makes it impossible to carry at least 1518 bytes (if offering Ethernet service), or at least 1280 (if offering IPv6 service). The use of such a protocol mechanism could be triggered by receiving a PTB. But such a mechanism might not be implemented by all encapsulators and decapsulators. [Aerolink is one example of such a protocol.]
Depending on the payload carried by the encapsulation there are some additional possibilities:
In summary:
The OAM area is seeing active development in the IETF with discussions (at least) in NVO3 and SFC working groups, plus the new LIME WG looking at architecture and YANG models.
The design team has take a narrow view of OAM to explore the potential OAM implications on the encapsulation format.
In terms of what we have heard from the various working groups there seem to be needs to:
To ensure that the OAM messages can follow the same path the OAM messages need to get the same ECMP (and LAG hashing) results as a given data flow. An encapsulator can choose between one of:
There can be several ways to prevent OAM packets from accidentally being forwarded to the end station using:
This assumes that the bit or next protocol, respectively, would not affect entropy/ECMP in the underlay. However, the next-protocol field might be used to provide differentiated treatment of packets based on their payload; for instance a TCP vs. IPsec ESP payload might be handled differently. Based on that observation it might be undesirable to overload the next protocol with the OAM drop behavior, resulting in a preference for having a bit to indicate that the packet should be forwarded to the end station after decapsulation.
There has been suggestions that one (or more) marker bits in the encaps header would be useful in order to delineate measurement epochs on the encapsulator and decapsulator and use that to compare counters to determine packet loss.
A result of the above is that OAM is likely to evolve and needs some degree of extensibility from the encapsulation format; a bit or two plus the ability to define additional larger extensions.
An open question is how to handle error messages or other reports relating to OAM. One can think if such reporting as being associated with the encapsulation the same way ICMP is associated with IP. Would it make sense for the IETF to develop a common Encapsulation Error Reporting Protocol as part of OAM, which can be used for different encapsulations? And if so, what are the technical challenges. For instance, how to avoid it being filtered as ICMP often is?
A potential additional consideration for OAM is the possible future existence of gateways that "stitch" together different dataplane encapsulations and might want to carry OAM end-to-end across the different encapsulations.
In summary:
Different encapsulation use cases will have different requirements around security. For instance, when encapsulation is used to build overlay networks for network virtualization, isolation between virtual networks may be paramount. BIER support of multicast may entail different security requirements than encapsulation for unicast.
In real deployment, the security of the underlying network may be considered for determining the level of security needed in the encapsulation layer. However for the purposes of this discussion, we assume that network security is out of scope and that the underlying network does not itself provide adequate or as least uniform security mechanisms for encapsulation.
There are at least three considerations for security:
This section uses a VNI ID in NVO3 as an example. A SFC or BIER encapsulation is likely to have fields with similar security and privacy requirements.
Some of these considerations appear for a new encapsulation, and others are more specific to network virtualization in datacenters.
The first requirement is isolation between virtual networks. Packets sent in one virtual network should never be illegitimately received by a node in another virtual network. Isolation should be protected in the presence of malicious attacks or inadvertent packet corruption.
The second requirement is sender authentication. Sender identity is authenticated to prevent anti-spoofing. Even if an attacker has access to the packets in the network, they cannot send packets into a virtual network. This may have two possibilities:
Possible security solutions:
In any use of a shared key, periodic re-keying should be allowed. This could include use of techniques like generation numbers, key windows, etc. See [I-D.farrelll-mpls-opportunistic-encrypt] for an example application.
We might see firewalls that are aware of the encapsulation and can provide some defense in depth combined with the above example anti-spoofing approaches. An example would be an NVO3-aware firewall being able to check the VNI ID.
Separately and in addition to such filtering, there might be a desire to completely block an encapsulation protocol at certain places in the network, e.g., at the edge of a datacenter. Using a fixed standard UDP destination port number for each encapsulation protocol would facilitate such blocking.
An encapsulated packet may itself be encapsulated in IPsec (e.g. ESP). This should be straightforward and in fact is what would happen today in security gateways. In this case, there is no special consideration for the fact that packet is encapsulated, however since the encapsulation layer headers are included (part of encrypted data for instance) we lose visibility in the network of the encapsulation.
The more interesting case is when security is applied to the encapsulation payload. This will keep the encapsulation headers in the outer header visible to the network (for instance in nvo3 we may way to firewall based on VNI ID even if the payload is encrypted). One possibility is to apply DTLS to the encapsulation payload. In this model the protocol stack may be something like IP|UDP|Encap|DTLS|encrypted_payload. The encapsulation and security should be done together at an encapsulator and resolved at the decapsulator. Since the encapsulation header is outside of the security coverage, this may itself require security (like described above).
In both of the above the security associations (SAs) may be between physical hosts, so for instance in nvo3 we can have packets of different virtual networks using the same SA-- this should not be an issue since it is the VNI ID that ensures isolation (which needs to be secured also).
In the Internet architecture we support QoS using the Differentiated Services Code Points (DSCP) in the formerly named Type-of-Service field in the IPv4 header, and in the Traffic-Class field in the IPv6 header. The ToS and TC fields also contain the two ECN bits, which are discussed in Section 13.
We have existing specifications how to process those bits. See [RFC2983] for diffserv handling, which specifies how the received DSCP value is used to set the DSCP value in an outer IP header when encapsulating. (There are also existing specifications how DSCP can be mapped to layer2 priorities.)
Those specifications apply whether or not there is some intervening headers (e.g., for NVO3 or SFC) between the inner and outer IP headers. Thus the encapsulation considerations in this area are mainly about applying the framework in [RFC2983].
Note that the DSCP and ECN bits are not the only part of an inner packet that might potentially affect the outer packet. For example, [RFC2473] specifies handling of inner IPv6 hop-by-hop options that effectively result in copying some options to the outer header. It is simpler to not have future encapsulations depend on such copying behavior.
There are some other considerations specific to doing OAM for encapsulations. If OAM messages are used to measure latency, it would make sense to treat them the same as data payloads. Thus they need to have the same outer DSCP value as the data packets which they wish to measure.
Due to OAM there are constraints on middleboxes in general. If middleboxes inspect the packet past the outer IP+UDP and encapsulation header and look for inner IP and TCP/UDP headers, that might violate the assumption that OAM packets will be handled the same as regular data packets. That issue is broader than just QoS - applies to firewall filters etc.
In summary:
Additional encapsulation headers does not introduce anything new for Explicit Congestion Notification. It is just like IP-in-IP and IPsec tunnels which is specified in [RFC6040] in terms of how the ECN bits in the inner and outer header are handled when encapsulating and decapsulating packets. Thus new encapsulations can more or less include that by reference.
There are additional considerations around carrying non-congestion controlled traffic. These details have been worked out in [I-D.ietf-mpls-in-udp]. As specified in [RFC5405]: "IP-based traffic is generally assumed to be congestion-controlled, i.e., it is assumed that the transport protocols generating IP-based traffic at the sender already employ mechanisms that are sufficient to address congestion on the path. Consequently, a tunnel carrying IP-based traffic should already interact appropriately with other traffic sharing the path, and specific congestion control mechanisms for the tunnel are not necessary". Those considerations are being captured in [I-D.ietf-tsvwg-rfc5405bis].
For this reason, where an encapsulation method is used to carry IP traffic that is known to be congestion controlled, the UDP tunnels does not create an additional need for congestion control. Internet IP traffic is generally assumed to be congestion-controlled. Similarly, in general Layer 3 VPNs are carrying IP traffic that is similarly assumed to be congestion controlled.
However, some of the encapsulations (at least NVO3) will be able to carry arbitrary Layer 2 packets to provide an L2 service, in which case one can not assume that the traffic is congestion controlled.
One could handle this by adding some congestion control support to the encapsulation header (one instance of which would end up looking like DCCP). However, if the underlay is well-provisioned and managed as opposed to being arbitrary Internet path, it might be sufficient to have a slower reaction to congestion induced by that traffic. There is work underway on a notion of "circuit breakers" for this purpose. See See [I-D.ietf-tsvwg-circuit-breaker]. Encapsulations which carry arbitrary Layer 2 packets want to consider that ongoing work.
If the underlay is provisioned in such a way that it can guarantee sufficient capacity for non-congestion controlled Layer 2 traffic, then such circuit breakers might not be needed.
Two other considerations appear in the context of these encapsulations as applied to overlay networks:
Those issues are really orthogonal to the encapsulation, in that they are present even when no new encapsulation header is in use. However, the application of the new encapsulations are likely to be in environments where those issues are becoming more important. Hence it makes sense to consider them.
One could make the encapsulation header be extensible to that it can carry sufficient information to be able to measure resource usage, delays, and congestion. The suggestions in the OAM section about a single bit for counter synchronization, and optional timestamps and/or sequence numbers, could be part of such an approach. There might also be additional congestion-control extensions to be carried in the encapsulation. Overall this results in a consideration to support sufficient extensibility in the encapsulation to handle potential future developments in this space.
Coarse measurements are likely to suffice, at least for circuit-breaker-like purposes, see [I-D.wei-tsvwg-tunnel-congestion-feedback] and [I-D.briscoe-conex-data-centre] for examples on active work in this area via use of ECN. [RFC6040] Appendix C is also relevant. The outer ECN bits seem sufficient (at least when everything uses ECN) to do this course measurements. Needs some more study for the case when there are also drops; might need to exchange counters between ingress and egress to handle drops.
Circuit breakers are not sufficient to make a network with different congestion control when the goal is to provide a predictable service to different tenants. The fallback would be to rate limit different traffic.
In summary:
Many UDP based encapsulations such as VXLAN [RFC7348] either discourage or explicitly disallow the use of UDP checksums. The reason is that the UDP checksum covers the entire payload of the packet and switching ASICs are typically optimized to look at only a small set of headers as the packet passes through the switch. In these case, computing a checksum over the packet is very expensive. (Software endpoints and the NICs used with them generally do not have the same issue as they need to look at the entire packet anyways.)
The lack a header checksum creates the possibility that bit errors can be introduced into any information carried by the new headers. Specifically, in the case of IPv6, the assumption is that a transport layer checksum - UDP in this case - will protect the IP addresses through the inclusion of a pseudo-header in the calculation. This is different from IPv4 on which many of these encapsulation protocols are initially deployed which contains its own header checksum. In addition to IP addresses, the encapsulation header often contains its own information which is used for addressing packets or other high value network functions. Without a checksum, this information is potentially vulnerable - an issue regardless of whether the packet is carried over IPv4 or IPv6.
Several protocols cite [RFC6935] and [RFC6936] as an exemption to the IPv6 checksum requirements. However, these are intended to be tailored to a fairly narrow set of circumstances - primarily relying on sparseness of the address space to detect invalid values and well managed networks - and are not a one size fits all solution. In these cases, an analysis should be performed of the intended environment, including the probability of errors being introduced and the use of ECC memory in routing equipment.
Conceptually, the ideal solution to this problem is a checksum that covers only the newly added headers of interest. There is little value in the portion of the UDP checksum that covers the encapsulated packet because that would generally be protected by other checksums and this is the expensive portion to compute. In fact, this solution already exists in the form of UDP-Lite and UDP based encapsulations could be easily ported to run on top of it. Unfortunately, the main value in using UDP as part of the encapsulation header is that it is recognized by already deployed equipment for the purposes of ECMP, RSS, and middlebox operations. As UDP-Lite uses a different protocol number than UDP and it is not widely implemented in middleboxes, this value is lost. A possible solution is to incorporate the same partial-checksum concept as UDP-Lite or other header checksum protection into the encapsulation header and continue using UDP as the outer protocol. One potential challenge with this approach is the use of NAT or other form of translation on the outer header will result in an invalid checksum as the translator will not know to update the encapsulation header.
The method chosen to protect headers is often related to the security needs of the encapsulation mechanism. On one hand, the impact of a poorly protected header is not limited to only data corruption but can also introduce a security vulnerability in the form of misdirected packets to an unauthorized recipient. Conversely, high security protocols that already include a secure hash over the valuable portion of the header (such as by encrypting the entire IP packet using IPsec, or some secure hash of the encap header) do not require additional checksum protection as the hash provides stronger assurance than a simple checksum.
If the sender has included a checksum, then the receiver should verify that checksum or, if incapable, drop the packet. The assumption is that configuration and/or control-plane capability exchanges can be used when different receiver have different checksum validation capabilities.
In summary:
Protocol extensibility is the concept that a networking protocol may be extended to include new use cases or functionality that were not part of the original protocol specification. Extensibility may be used to add security, control, management, or performance features to a protocol. A solution may allow private extensions for customization or experimentation.
Extending a protocol often implies that a protocol header must carry new information. There are two usual methods to accomplish this:
It is also possible to create a new protocol version, but this is more associated with defining a protocol than extending it (IPv6 being a successor to IPv4 is an example of protocol versioning).
In some cases it might be more appropriate to define a new inner protocol which can carry the new functionality instead of extending the outer protocol. Examples where this works well is in the IP/transport split, where the earlier architecture had a single NCP [RFC0033] protocol which carried both the hop-by-hop semantics which are now in IP, and the end-to-end semantics which are now in TCP. Such a split is effective when different nodes need to act upon the different information. Applying this for general protocol extensibility through nesting is not well understood, and does result in longer header chains. Furthermore, our experience with IPv6 extension headers [RFC2460] in middleboxes indicates that the header chaining approach does not help with middlebox traversal.
Many protocol definitions include some number of reserved fields or bits which can be used for future extension. VXLAN is an example of a protocol that includes reserved bits which are subsequently being allocated for new purposes. Another technique employed is to re-purpose existing header fields with new meanings. A classic example of this is the definition of DSCP code point which redefines the ToS field originally specified in IPv4. When a field is redefined, some mechanism may be needed to ensure that all interested parties agree on the meaning of the field. The techniques of defining meaning for reserved bits or redefining existing fields have the advantage that a protocol header can be kept a fixed length. The disadvantage is that the extensibility is limited. For instance, the number reserved bits in a fixed protocol header is limited. For standard protocols the decision to commit to a definition for a field can be wrenching since it is difficult to retract later. Also, it is difficult to predict a priori how many reserved fields or bits to put into a protocol header to satisfy the extensions create over the lifetime of the protocol.
Extending a protocol header with new fields can be done in several ways.
The requirements for receiving unknown or unimplemented extensible elements in an encapsulation protocol (flags, TLVs, optional fields) need to be specified. There are two parties to consider, middle boxes and terminal endpoints of encapsulation (at the decapsulator).
A protocol may allow or expect nodes in a path to modify fields in an encapsulation (example use of this is BIER). In this case, the middleboxes should follow the same requirements as nodes terminating the encapsulation. In the case that middle boxes do not modify the encapsulation, we can assume that they may still inspect any fields of the encapsulation. Missing or unknown fields should be accepted per protocol specification, however it is permissible for a site to implement a local policy otherwise (e.g. a firewall may drop packets with unknown options).
For handling unknown options at terminal nodes, there are two possibilities: drop packet or accept while ignoring the unknown options. Many Internet protocols specify that reserved flags must be set to zero on transmission and ignored on reception. L2TP is example data protocol that has such flags. GRE is a notable exception to this rule, reserved flag bits 1-5 cannot be ignored [RFC2890]. For TCP and IPv4, implementations must ignore optional TLVs with unknown type; however in IPv6 if a packet contains an unknown extension header (unrecognized next header type) the packet must be dropped with an ICMP error message returned. The IPv6 options themselves (encoded inside the destinations options or hop-by-hop options extension header) have more flexibility. There are bits in the option code are used to instruct the receiver whether to ignore, silently drop, or drop and send error if the option is unknown. Some protocols define a "mandatory bit" that can is set with TLVs to indicate that an option must not be ignored. Conceptually, optional data elements can only be ignored if they are idempotent and do not alter how the rest of the packet is parsed or processed.
Depending on what type of protocol evolution one can predict, it might make sense to have a way for a sender to express that the packet should be dropped by a terminal node which does not understand the new information. In other cases it would make sense to have the receiver silently ignore the new info. The former can be expressed by having a version field in the encapsulation, or a notion of "mandatory bit" as discussed above.
A security mechanism which use some form secure hash over the encapsulation header would need to be able to know which extensions can be changed in flight.
In summary:
One can envision that SFC might use NVO3 as a delivery/transport mechanism. With more imagination that in turn might be delivered using BIER. Thus it is useful to think about what things look like when we have BIER+NVO3+SFC+payload. Also, if NVO3 is widely deployed there might be cases of NVO3 nesting where a customer uses NVO3 to provide network virtualization e.g., across departments. That customer uses a service provider which happens to use NVO3 to provide transport for their customers.Thus NVO3 in NVO3 might happen.
A key question we set out to answer is what the packets might look like in such a case, and in particular whether we would end up with multiple UDP headers for entropy.
Based on the discussion in the Entropy section, the entropy is associated with the outer delivery IP header. Thus if there are multiple IP headers there would be a UDP header for each one of the IP headers. But SFC does not require its own IP header. So a case of NVO3+SFC would be IP+UDP+NVO3+SFC. A nested NVO3 encapsulation would have independent IP+UDP headers.
The layering also has some implications for middleboxes.
The design team asked itself some additional questions:
The IP service is lossy and subject to reordering. In order to avoid a performance impact on transports like TCP the handling of packets is designed to avoid reordering packets that are in the same transport flow (which is typically identified by the 5-tuple). But across such flows the receiver can see different ordering for a given sender. That is the case for a unicast vs. a multicast flow from the same sender.
There is a general tussle between the desire for high capacity utilization across a multipath network and the impact on packet ordering within the same flow (which results in lower transport protocol performance). That isn't affected by the introduction of an encapsulation. However, the encapsulation comes with some entropy, and there might be cases where folks want to change that in response to overload or failures. For instance, one might want to change UDP source port to try different ECMP route. Such changes can result in packet reordering within a flow, hence would need to be done infrequently and with care e.g., by identifying packet trains.
There might be some applications/services which are not able to handle reordering across flows. The IETF has defined pseudo-wires [RFC3985] which provides the ability to ensure ordering (implemented using sequence numbers and/or timestamps).
Architectural such services would make sense, but as a separate layer on top of an encapsulation protocol. They could be deployed between ingress and egress of a tunnel which uses some encaps. Potentially the tunnel control points at the ingress and egress could become a platform for fixing suboptimal behavior elsewhere in the network. That would clearly be undesirable in the general case. However, handling encapsulation of non-IP traffic hence non-congestion-controlled traffic is likely to be required, which implies some fairness and/or QoS policing on the ingress and egress devices.
But the tunnels could potentially do more like increase reliability (retransmissions, FEC) or load spreading using e.g. MP-TCP between ingress and egress.
Hosts, switches and routers often leverage capabilities in the hardware to accelerate packet encapsulation, decapsulation and forwarding.
Some design considerations in encapsulation that leverage these hardware capabilities may result in more efficiently packet processing and higher overall protocol throughput.
While "hardware friendliness" can be viewed as unnecessary considerations for a design, part of the motivation for considering this is ease of deployment; being able to leverage existing NIC and switch chips for at least a useful subset of the functionality that the new encapsulation provides. The other part is the ease of implementing new NICs and switch/router chips that support the encapsulation at ever increasing line rates.
[disclaimer] There are many different types of hardware in any given network, each maybe better at some tasks while worse at others. We would still recommend protocol designers to examine the specific hardware that are likely to be used in their networks and make decisions on a case by case basis.
Some considerations are:
This section provides guidelines to provide support of common offloads for encapsulation in Network Interface Cards (NICs). Offload mechanisms are techniques that are implemented separately from the normal protocol implementation of a host networking stack and are intended to optimize or speed up protocol processing. Hardware offload is performed within a NIC device on behalf of a host.
There are three basic offload techniques of interest:
Contemporary NICs support multiple receive descriptor queues (multi-queue). Multi-queue enables load balancing of network processing for a NIC across multiple CPUs. On packet reception, a NIC must select the appropriate queue for host processing. Receive Side Scaling (RSS) is a common method which uses the flow hash for a packet to index an indirection table where each entry stores a queue number.
UDP encapsulation, where the source port is used for entropy, should be compatible with multi-queue NICs that support five-tuple hash calculation for UDP/IP packets as input to RSS. The source port ensures classification of the encapsulated flow even in the case that the outer source and destination addresses are the same for all flows (e.g. all flows are going over a single tunnel).
Many NICs provide capabilities to calculate standard ones complement payload checksum for packets in transmit or receive. When using encapsulation over UDP there are at least two checksums that may be of interest: the encapsulated packet's transport checksum, and the UDP checksum in the outer header.
NICs may provide a protocol agnostic method to offload transmit checksum (NETIF_F_HW_CSUM in Linux parlance) that can be used with UDP encapsulation. In this method the host provides checksum related parameters in a transmit descriptor for a packet. These parameters include the starting offset of data to checksum, the length of data to checksum, and the offset in the packet where the computed checksum is to be written. The host initializes the checksum field to pseudo header checksum. In the case of encapsulated packet, the checksum for an encapsulated transport layer packet, a TCP packet for instance, can be offloaded by setting the appropriate checksum parameters.
NICs typically can offload only one transmit checksum per packet, so simultaneously offloading both an inner transport packet's checksum and the outer UDP checksum is likely not possible. In this case setting UDP checksum to zero (per above discussion) and offloading the inner transport packet checksum might be acceptable.
There is a proposal in [I-D.herbert-remotecsumoffload] to leverage NIC checksum offload when an encapsulator is co-resident with a host.
Protocol encapsulation is compatible with NICs that perform a protocol agnostic receive checksum (CHECKSUM_COMPLETE in Linux parlance). In this technique, a NIC computes a ones complement checksum over all (or some predefined portion) of a packet. The computed value is provided to the host stack in the packet's receive descriptor. The host driver can use this checksum to "patch up" and validate any inner packet transport checksum, as well as the outer UDP checksum if it is non-zero.
Many legacy NICs don't provide checksum-complete but instead provide an indication that a checksum has been verified (CHECKSUM_UNNECESSARY in Linux). Usually, such validation is only done for simple TCP/IP or UDP/IP packets. If a NIC indicates that a UDP checksum is valid, the checksum-complete value for the UDP packet is the "not" of the pseudo header checksum. In this way, checksum-unnecessary can be converted to checksum-complete. So if the NIC provides checksum-unnecessary for the outer UDP header in an encapsulation, checksum conversion can be done so that the checksum-complete value is derived and can be used by the stack to validate an checksums in the encapsulated packet.
Segmentation offload refers to techniques that attempt to reduce CPU utilization on hosts by having the transport layers of the stack operate on large packets. In transmit segmentation offload, a transport layer creates large packets greater than MTU size (Maximum Transmission Unit). It is only at much lower point in the stack, or possibly the NIC, that these large packets are broken up into MTU sized packet for transmission on the wire. Similarly, in receive segmentation offload, small packets are coalesced into large, greater than MTU size packets at a point low in the stack receive path or possibly in a device. The effect of segmentation offload is that the number of packets that need to be processed in various layers of the stack is reduced, and hence CPU utilization is reduced.
Transmit Segmentation Offload (TSO) is a NIC feature where a host provides a large (larger than MTU size) TCP packet to the NIC, which in turn splits the packet into separate segments and transmits each one. This is useful to reduce CPU load on the host.
The process of TSO can be generalized as:
Following this general process, TSO can be extended to support TCP encapsulation UDP. For each segment the Ethernet, outer IP, UDP header, encapsulation header, inner IP header if tunneling, and TCP headers are replicated. Any packet length header fields need to be set properly (including the length in the outer UDP header), and checksums need to be set correctly (including the outer UDP checksum if being used).
To facilitate TSO with encapsulation it is recommended that optional fields should not contain values that must be updated on a per segment basis-- for example an encapsulation header should not include checksums, lengths, or sequence numbers that refer to the payload. If the encapsulation header does not contain such fields then the TSO engine only needs to copy the bits in the encapsulation header when creating each segment and does not need to parse the encapsulation header.
Large Receive Offload (LRO) is a NIC feature where packets of a TCP connection are reassembled, or coalesced, in the NIC and delivered to the host as one large packet. This feature can reduce CPU utilization in the host.
LRO requires significant protocol awareness to be implemented correctly and is difficult to generalize. Packets in the same flow need to be unambiguously identified. In the presence of tunnels or network virtualization, this may require more than a five-tuple match (for instance packets for flows in two different virtual networks may have identical five-tuples). Additionally, a NIC needs to perform validation over packets that are being coalesced, and needs to fabricate a single meaningful header from all the coalesced packets.
The conservative approach to supporting LRO for encapsulation would be to assign packets to the same flow only if they have identical five-tuple and were encapsulated the same way. That is the outer IP addresses, the outer UDP ports, encapsulated protocol, encapsulation headers, and inner five tuple are all identical.
In summary, for NIC offload:
This document has touched upon middleboxes in different section. The reason for this is as encapsulations get widely deployed one would expect different forms of middleboxes might become aware of the encapsulation protocol just as middleboxes have been made aware of other protocols where there are business and deployment opportunities. Such middleboxes are likely to do more than just drop packets based on the UDP port number used by an encapsulation protocol.
We note that various forms of encapsulation gateways that stitch one encapsulation protocol together with another form of protocol could have similar effects.
An example of a middlebox that could see some use would be an NVO3-aware firewall that would filter on the VNI IDs to provide some defense in depth inside or across NVO3 datacenters.
A question for the IETF is whether we should document what to do or what not to do in such middleboxes. This document touches on areas of OAM and ECMP as it relates to middleboxes and it might make sense to document how encapsulation-aware middleboxes should avoid unintended consequences in those (and perhaps other) areas.
In summary:
The IETF and industry has defined encapsulations for a long time, with examples like GRE [RFC2890], VXLAN [RFC7348], and NVGRE [I-D.sridharan-virtualization-nvgre] being able to carry arbitrary Ethernet payloads, and various forms of IP-in-IP and IPsec encapsulations that can carry IP packets. As part of NVO3 there has been additional proposals like Geneve [I-D.gross-geneve] and GUE [I-D.herbert-gue] which look at more extensibility. NSH [I-D.quinn-sfc-nsh] is an example of an encapsulation that tries to provide extensibility mechanisms which target both hardware and software implementations.
There is also a large body of work around MPLS encapsulations [RFC3032]. The MPLS-in-UDP work [I-D.ietf-mpls-in-udp] and GRE over UDP [I-D.ietf-tsvwg-gre-in-udp-encap] have worked on some of the common issues around checksum and congestion control. MPLS also introduced a entropy label [RFC6790]. There is also a proposal for MPLS encryption [I-D.farrelll-mpls-opportunistic-encrypt].
The idea to use a UDP encapsulation with a UDP source port for entropy for the underlay routers' ECMP dates back to LISP [RFC6830].
The pseudo-wire work [RFC3985] is interesting in the notion of layering additional services/characteristics such as ordered delivery or timely deliver on top of an encapsulation. That layering approach might be useful for the new encapsulations as well. For instance, the control word [RFC4385]. There is also material on congestion control for pseudo-wires in [I-D.ietf-pwe3-congcons].
Both MPLS and L2TP [RFC3931] rely on some control or signaling to establish state (for the path/labels in the case of MPLS, and for the session in the case of L2TP). The NVO3, SFC, and BIER encapsulations will also have some separation between the data plane and control plane, but the type of separation appears to be different.
IEEE 802.1 has defined encapsulations for L2 over L2, in the form of Provider backbone Bridging (PBB) [IEEE802.1Q-2014] and Equal Cost Multipath (ECMP) [IEEE802.1Q-2014]. The latter includes something very similar to the way the UDP source port is used as entropy: "The flow hash, carried in an F-TAG, serves to distinguish frames belonging to different flows and can be used in the forwarding process to distribute frames over equal cost paths"
TRILL, which is also a L2 over L2 encapsulation, took a different approach to entropy but preserved the ability for OAM frames [RFC7174] to use the same entropy hence ECMP path as data frames. In [I-D.ietf-trill-oam-fm] there 96 bytes of headers for entropy in the OAM frames, followed by the actual OAM content. This ensures that any headers, which fit in those 96 bytes except the OAM bit in the TRILL header, can be used for ECMP hashing.
As encapsulations evolve there might be a desire to fit multiple inner packets into one outer packet. The work in [I-D.saldana-tsvwg-simplemux] might be interesting for that purpose.
The authors acknowledge the comments from Alia Atlas, Fred Baker, David Black, Bob Briscoe, Stewart Bryant, Mike Cox, Andy Malis, Radia Perlman, Carlos Pignataro, Jamal Hadi Salim, Michael Smith, and Lucy Yong.
The changes from draft-rtg-dt-encap-01 based on feedback at the Dallas IETF meeting:
| [I-D.briscoe-conex-data-centre] | Briscoe, B. and M. Sridharan, "Network Performance Isolation in Data Centres using Congestion Policing", Internet-Draft draft-briscoe-conex-data-centre-02, February 2014. |
| [I-D.farrelll-mpls-opportunistic-encrypt] | Farrel, A. and S. Farrell, "Opportunistic Security in MPLS Networks", Internet-Draft draft-farrelll-mpls-opportunistic-encrypt-05, June 2015. |
| [I-D.gross-geneve] | Gross, J., Sridhar, T., Garg, P., Wright, C., Ganga, I., Agarwal, P., Duda, K., Dutt, D. and J. Hudson, "Geneve: Generic Network Virtualization Encapsulation", Internet-Draft draft-gross-geneve-02, October 2014. |
| [I-D.herbert-gue] | Herbert, T., Yong, L. and O. Zia, "Generic UDP Encapsulation", Internet-Draft draft-herbert-gue-03, March 2015. |
| [I-D.herbert-remotecsumoffload] | Herbert, T., "Remote checksum offload for encapsulation", Internet-Draft draft-herbert-remotecsumoffload-01, November 2014. |
| [I-D.ietf-mpls-in-udp] | Xu, X., Sheth, N., Yong, L., Callon, R. and D. Black, "Encapsulating MPLS in UDP", Internet-Draft draft-ietf-mpls-in-udp-11, January 2015. |
| [I-D.ietf-nvo3-arch] | Black, D., Hudson, J., Kreeger, L., Lasserre, M. and T. Narten, "An Architecture for Overlay Networks (NVO3)", Internet-Draft draft-ietf-nvo3-arch-03, March 2015. |
| [I-D.ietf-pwe3-congcons] | Stein, Y., Black, D. and B. Briscoe, "Pseudowire Congestion Considerations", Internet-Draft draft-ietf-pwe3-congcons-02, July 2014. |
| [I-D.ietf-sfc-architecture] | Halpern, J. and C. Pignataro, "Service Function Chaining (SFC) Architecture", Internet-Draft draft-ietf-sfc-architecture-09, June 2015. |
| [I-D.ietf-sfc-problem-statement] | Quinn, P. and T. Nadeau, "Service Function Chaining Problem Statement", Internet-Draft draft-ietf-sfc-problem-statement-13, February 2015. |
| [I-D.ietf-trill-oam-fm] | Senevirathne, T., Finn, N., Salam, S., Kumar, D., Eastlake, D., Aldrin, S. and L. Yizhou, "TRILL Fault Management", Internet-Draft draft-ietf-trill-oam-fm-11, October 2014. |
| [I-D.ietf-tsvwg-circuit-breaker] | Fairhurst, G., "Network Transport Circuit Breakers", Internet-Draft draft-ietf-tsvwg-circuit-breaker-01, March 2015. |
| [I-D.ietf-tsvwg-gre-in-udp-encap] | Crabbe, E., Yong, L., Xu, X. and T. Herbert, "GRE-in-UDP Encapsulation", Internet-Draft draft-ietf-tsvwg-gre-in-udp-encap-07, July 2015. |
| [I-D.ietf-tsvwg-port-use] | Touch, J., "Recommendations on Using Assigned Transport Port Numbers", Internet-Draft draft-ietf-tsvwg-port-use-11, April 2015. |
| [I-D.quinn-sfc-nsh] | Quinn, P., Guichard, J., Surendra, S., Smith, M., Henderickx, W., Nadeau, T., Agarwal, P., Manur, R., Chauhan, A., Halpern, J., Majee, S., Elzur, U., Melman, D., Garg, P., McConnell, B., Wright, C. and K. Kevin, "Network Service Header", Internet-Draft draft-quinn-sfc-nsh-07, February 2015. |
| [I-D.saldana-tsvwg-simplemux] | Saldana, J., "Simplemux. A generic multiplexing protocol", Internet-Draft draft-saldana-tsvwg-simplemux-02, January 2015. |
| [I-D.shepherd-bier-problem-statement] | Shepherd, G., Dolganow, A. and a. arkadiy.gulko@thomsonreuters.com, "Bit Indexed Explicit Replication (BIER) Problem Statement", Internet-Draft draft-shepherd-bier-problem-statement-02, February 2015. |
| [I-D.sridharan-virtualization-nvgre] | Garg, P. and Y. Wang, "NVGRE: Network Virtualization using Generic Routing Encapsulation", Internet-Draft draft-sridharan-virtualization-nvgre-08, April 2015. |
| [I-D.wei-tsvwg-tunnel-congestion-feedback] | Wei, X., Zhu, L., Deng, L. and B. Briscoe, "Tunnel Congestion Feedback", Internet-Draft draft-wei-tsvwg-tunnel-congestion-feedback-04, June 2015. |
| [I-D.wijnands-bier-architecture] | Wijnands, I., Rosen, E., Dolganow, A., Przygienda, T. and S. Aldrin, "Multicast using Bit Index Explicit Replication", Internet-Draft draft-wijnands-bier-architecture-05, March 2015. |
| [I-D.wijnands-mpls-bier-encapsulation] | Wijnands, I., Rosen, E., Dolganow, A., Tantsura, J. and S. Aldrin, "Encapsulation for Bit Index Explicit Replication in MPLS Networks", Internet-Draft draft-wijnands-mpls-bier-encapsulation-02, December 2014. |
| [I-D.xu-bier-encapsulation] | Xu, X., Somasundaram, S., Jacquenet, C. and R. Raszuk, "BIER Encapsulation", Internet-Draft draft-xu-bier-encapsulation-02, February 2015. |
| [IEEE802.1Q-2014] | IEEE, "IEEE Standard for Local and metropolitan area networks—Bridges and Bridged Networks", IEEE Std 802.1Q-2014, 2014. (Access Controlled link within page) |
| [RFC0033] | Crocker, S., "New Host-Host Protocol", RFC 33, February 1970. |