NAT64 Deployment Guidelines in Operator and Enterprise Networks

NAT64 () describes a stateful IPv6 to IPv4 translation, which allows IPv6-only hosts to contact IPv4 servers using unicast UDP, TCP or ICMP, by means of a single or a set of IPv4 public addresses assigned to the translator, to be shared by the IPv6-only clients. The translation of the packet headers is done using the IP/ICMP Translation Algorithm defined in and algorithmically translating the IPv4-hosts addresses to IPv6 ones following . To avoid changes in both, the IPv6-only hosts and the IPv4-only server, NAT64 requires also the use of a DNS64 (), in charge for the synthesis of AAAA records from the A records. However, the use of NAT64 and/or DNS64 present three issues: Because DNS64 () modifies DNS answers, and DNSSEC is designed to detect such modifications, DNS64 () can potentially break DNSSEC, depending on a number of factors, such as the location of the DNS64 function (at a DNS server or validator, at the end host, ...), how as been configured, if the end-hosts is validating, etc. Because the need of using DNS64 (), there is a major issue for NAT64 (), as doesn't work when literal addresses or non-IPv6 compliant APIs are being used. NAT64 alone, doesn't provide a solution for IPv4-only hosts or applications located within a network which are connected to a service provider IPv6-only access. The same issues are true if part of an enterprise or similar network, is connected to other parts of the same network or third party networks by means of IPv6-only links. According to that, across this document, the use of "operator network" is interchangeable with equivalent cases of enterprise (or similar) networks. This document looks into different possible NAT64 () deployment scenarios in operators and enterprise networks, and provides guidelines to avoid the above mentioned issues.

As indicated in Section 8 of (DNS64, Security Considerations), because DNS64 modifies DNS answers and DNSSEC is designed to detect such modifications, DNS64 can break DNSSEC. If a device connected to an IPv6-only WAN queries for a domain name in a signed zone, by means of a recursive name server that supports DNS64, and the result is a synthesized AAAA record, and the recursive name server is configured to perform DNSSEC validation and has a valid chain of trust to the zone in question, it will cryptographically validate the negative response from the authoritative name server. So, the recursive name server actually lie to the client device, however in most of the cases, the client will not notice it, because generally they don't perform validation themselves as instead rely on their recursive name servers. If the client device performs DNSSEC validation on the AAAA record, it will fail as it is a synthesized record. Similarly, if the client querying the recursive name server is another name server configured to use it as a forwarder, and is performing DNSSEC validation, it will also fail on any synthesized AAAA record. The obvious solution to avoid DNSSEC issues, will be that all the signed zones also provide IPv6 connectivity, together with the corresponding AAAA records. Previous data seems to indicate, that the figures of DNSSEC broken by using DNS64 will be around 1.7% (information 2 years old).

Section 7 of DNS64 (), provides 3 scenarios, looking at the location of the DNS64. However, since the publication of that document, there are new possible scenarios and NAT64 use cases that need to be considered. The perspective in this document is to broader those scenarios, including a few new ones. However, in order to be able to reduce the number of possible cases, we work under the assumption that the service provider want to make sure that all the customers have a service without failures. This means considering the worst possible case: There are hosts that will be validating DNSSEC. Literal addresses and non-IPv6 compliant APIs are being used. There are IPv4-only hosts or applications beyond the IPv6-only link. We use a common set of possible "participant entities": An IPv6-only access network (IPv6). An IPv4-only remote network (IPv4). The NAT64 of the service provider (NAT64). The DNS64 function (DNS64). An external service provider offering the NAT64 and/or the DNS64 function (extNAT64/extDNS64). 464XLAT customer side translator (CLAT).

In this scenario, the service provider offers a NAT64, however there is no DNS64 function support. As a consequence, an IPv6 host in the IPv6-only access network, will not be able to detect the presence of DNS64 neither learning the IPv6 prefix to be used for the NAT64. This can be sorted out as indicated in . However, despite that, because the lack of the DNS64 function, the IPv6 host will not be able to obtain AAAA synthesised records, so the NAT64 becomes useless. An exception to this "useless" scenario will be manually configure mappings between the A records of each of the IPv4-only remote hosts and the corresponding AAAA records, with the WKP (Well-Known Prefix) or NSP (Network-Specific Prefix) used by the service provider NAT64, as if they were synthesised by a DNS64. This mapping could be done by several means, typically at the authoritative DNS server, or at the service provider resolvers by means of DNS RPZ (Response Policy Zones). The latest, may have implications in DNSSEC, if the zone is signed. Also, if the service provider is using a NSP, having the mapping at the authoritative server, will mean that may create troubles to other parties trying to use different NSP or the WKP, unless multiple DNS "views" are also being used at the authoritative servers. Generally, the mappings alternative, will only make sense if a few set of IPv4-only remote hosts need to be accessed by a single network or reduced set of them, which support IPv6-only in the access, with some kind of mutual agreement for using this procedure, so it doesn't care if they become a trouble for other parties across Internet ("closed services"). In any case, this scenario doesn't solve the issue of literal addresses or non-IPv6 compliant APIs, neither it solves the problem of IPv4-only hosts within that IPv6-only access network.

In this scenario, the service provider offers both, the NAT64 and the DNS64 function. This is probably the most common scenario, however also has the implications related the DNSSEC. This scenario also fails to solve the issue of literal addresses or non-IPv6 compliant APIs, as well as the issue of IPv4-only hosts or applications inside the IPv6-only access network.

A totally equivalent scenario will be if the service provider offers only the DNS64 function, and the NAT64 function is provided by an agreement with an external provider. All the considerations in the previous paragraphs of this section are the same for this sub-case.

As well, is equivalent to the scenario where the agreement with the external provider is to provide both the NAT64 and DNS64 function. Once more, all the considerations in the previous paragraphs of this section are the same for this sub-case.

In this scenario, the service provider offers the NAT64, but not the DNS64. However, the IPv6 hosts have a built-in DNS64 function. This may become common if the DNS64 function is implemented in all the IPv6 hosts/stacks, which is not the actual situation. At this way, the DNSSEC validation is performed on the A record, and then the host can use the DNS64 function so to be able to use the NAT64, without any DNSSEC issues. This scenario fails to solve the issue of literal addresses or non-IPv6 compliant APIs, unless the IPv6 hosts also supports Happy Eyeballs v2 (), which may solve that issue. However, this scenario still fails to solve the problem of IPv4-only hosts or applications inside the IPv6-only access network.

In this scenario, the service provider offers the NAT64 only. The remote IPv4-only network offers the DNS64 function. This is not common, and looks like doesn't make too much sense that a remote network, not deploying IPv6, is providing a DNS64 function and as, in the case of the scenario depicted in , it will only work if both sides are using the WKP or the same NSP, etc., so the same considerations apply. This scenario still fails to solve the issue of literal addresses or non-IPv6 compliant APIs. This scenario also fails to solve the problem of IPv4-only hosts or applications inside the IPv6-only access network.

A totally equivalent scenario will be if the service provider offers the NAT64 only, and the DNS64 function is provided by an external provider without an specific agreement among them. This is an scenario already feasible today, as several "global" service providers provide open DNS64 services and users often configure manually their DNS. All the considerations in the previous paragraphs of this section are the same for this sub-case. If the user instead of configuring a DNS64 uses a regular external DNS, the situation is even much worst, because in that case, NAT64 will not work at all with any IPv4-only remote host. However, if the external DNS64 is agreed with the service provider, then we are in the same case, in terms of considerations of issues, as in .

464XLAT () describes an architecture that provides IPv4 connectivity across a network, or part of it, when it is only natively transporting IPv6. In order to do that, 464XLAT () relies on the combination of existing protocols: The customer-side translator (CLAT) is a stateless IPv4 to IPv6 translator (NAT46) () implemented in the end-user device or CE, located at the "customer" edge of the network. The provider-side translator (PLAT) is a stateful NAT64 (), implemented typically at the opposite edge of the operator network, that provides access to both IPv4 and IPv6 upstreams. Optionally, DNS64 (), implemented as part of the PLAT allows an optimization (a single translation at the NAT64, instead of two translations - NAT46+NAT64), when the application at the end-user device supports IPv6 DNS (uses AAAA RR). In this scenario, using 464XLAT without DNS64, the service provider ensures that DNSSEC is not broken. 464XLAT () is a very simple approach to cope with the major NAT64+DNS64 drawback: Not working with applications or devices that use literal IPv4 addresses or non-IPv6 compliant APIs. 464XLAT () has been used initially in IPv6 cellular networks, so providing an IPv6-only access network, the end-user device applications can access IPv4-only end-networks/applications, despite those applications or devices use literal IPv4 addresses or non-IPv6 compliant APIs. In addition to that, in the same example of the cellular network above, if the User Equipment (UE) provides tethering, other devices behind it will be presented with a traditional NAT44, in addition to the native IPv6 support, so clearly it allows IPv4-only hosts inside the IPv6-only access network. Furthermore, 464XLAT () can be used in non-cellular IPv6 wired (xDSL, DOCSIS, FTTH, Ethernet, ...) and wireless (WiFi) network architectures, by implementing the CLAT functionality at the CE.

In this scenario the service provider deploys 464XLAT with DNS64. As a consequence, the DNSSEC issues remain. However, in this scenario, as in the previous one, there are no issues related to IPv4-only hosts inside the IPv6-only access network, neither to the usage of IPv4 literals or non-IPv6 compliant APIs.

As already mention, the scenarios in the precious section, are in fact somehow simplified, looking at the worst case, because breaking DNSSEC will not happen, if the end-host is not doing validation, and/or some countermeasures are taken, depicted in the next sections.

The ideal solution will be to avoid using DNS64, but as already indicated this is not possible in all the scenarios. However, not having a DNS64, means that is not possible to heuristically discover the NAT64 () and consequently, an IPv6 host in the IPv6-only access network, will not be able to detect the presence of the DNS64, neither to learn the IPv6 prefix to be used for the NAT64. The learning of the IPv6 prefix could be solved by means of adding the relevant AAAA records to the ipv4only.arpa. zone of the service provider recursive servers, i.e., if using the WKP (64:ff9b::/96):

An alternative option to the above, is the use of DNS RPZ (Response Policy Zones). One more alternative, only valid in environments with PCP support (for both the hosts or CEs and for the service provider network), to follow (Discovering NAT64 IPv6 Prefixes using PCP).

In general, DNS servers with DNS64 function, by default, will not synthesize AAAA responses if the DNSSEC OK (DO) flag was set in the query. In this case, as only an A record is available, it means that the CLAT will take the responsibility, as in the case of literal IPv4 addresses, to keep that traffic flow end-to-end as IPv4, so DNSSEC is not broken. However, this will not work if a CLAT is not present as the hosts will not be able to use IPv4 (scenarios without 464XLAT).

If the DO flag is set and the client device performs DNSSEC validation, and the Checking Disabled (CD) flag is set for a query, as the DNS64 recursive server will not synthesize AAAA responses, the client could perform the DNSSEC validation with the A record and then may query the network for a NAT64 prefix () in order to synthesize the AAAA (). This allows the client device to avoid using the CLAT and still use NAT64 even with DNSSEC. If the end-host is IPv4-only, this will not work if a CLAT is not present (scenarios without 464XLAT). Some devices/OSs may implement, instead of CLAT, a simliar function by using Bump-in-the-Host (). In this case, the considerations in the above paragraphs are also applicable.

If a CE includes CLAT support and also a DNS proxy, as indicated in Section 6.4 of , the CE could behave as a stub validator on behalf of the client devices, following the same approach described in the precedent section (Stub validator). So the DNS proxy actually lie to the client devices, which in most of the cases will not notice it unless they perform validation themselves. Again, this allow the clients devices to avoid using the CLAT and still use NAT64 with DNSSEC. Once more, this will not work without a CLAT (scenarios without 464XLAT).

In cases of dual-stack clients, stub resolvers should send the AAAA queries before the A ones. So such clients, if DNS64 is enabled, will never get A records, even for IPv4-only servers, and they may be in the path before the NAT64 and accesible by IPv4. If DNSSEC is being used for all those flows, specific addresses or prefixes can be left-out the DNS64 synthesis by means of ACLs. Once more, this will not work without a CLAT (scenarios without 464XLAT).

If there are well-known specific IPv4 addresses or prefixes using DNSSEC, they can be mapped-out of the DNS64 synthesis. Even if this is not related to DNSSEC, this "mapping-out" feature is actually quite commonly used to ensure that addresses (for example used by LAN servers) are not synthesized to AAAA. Once more, this will not work without a CLAT (scenarios without 464XLAT).

When a client device, using a name server configured to perform DNS64, tries to reverse-map a synthesized IPv6 address, the name server responds with a CNAME record pointing the domain name used to reverse-map the synthesized IPv6 address (the one under ip6.arpa), to the domain name corresponding to the embedded IPv4 address (under in-addr.arpa). This is the expected behaviour, so no issues to be considered regarding DNS reverse mapping.

A hosts or application using literal IPv4 addresses or older APIs, behind a network with IPv6-only access, will not work unless a CLAT is present. A possible alternative approach is described as part of Happy Eyeballs v2 Section 7.1 (), or if not supporting HEv2, directly using Bump-in-the-Host (), and then a DNS64 function. Those alternatives will solve the problem for and end-hosts, however, if that end-hosts is providing "tethering" or an equivalent service to others hosts, that need to be considered as well. In other words, in a case of a cellular network, it resolves the issue for the cellular device itself, but may be not for hosts behind it. Otherwise, 464XLAT is the only valid approach to resolve this issue.

An IPv4-only hosts or application behind a network with IPv6-only access, will not work unless a CLAT is present. 464XLAT is the only valid approach to resolve this issue.

In the case the client device is IPv6-only (either because the stack is IPv6-only, or because it is connected via an IPv6-only LAN) and the server is IPv4-only (either because the stack is IPv4-only, or because it is connected via an IPv4-only LAN), only NAT64 combined with DNS64 will be able to provide access among both. Because DNS64 is then required, DNSSEC validation will be only possible if the recursive name server is validating the negative response from the authoritative name server and the client is not performing validation. However, when the client device is dual-stack and/or connected in a dual-stack LAN by means of a CLAT (or has the built-in CLAT), DNS64 is an option. With DNS64: If DNS64 is used, most of the IPv4 traffic (except if using literal IPv4 addresses or non-IPv6 compliant APIs) will not use the CLAT, so will use the IPv6 path and only one translation will be done at the NAT64. This may break DNSSEC, unless measures as described in the precedent section are taken. Without DNS64: If DNS64 is not used, all the IPv4 traffic will make use of the CLAT, so two translations are required (NAT46 at the CLAT and NAT64 at the PLAT), which adds some overhead in terms of the extra NAT46 translation, however avoids the AAAA synthesis and consequently will never break DNSSEC.

When clients in a service provider network use DNS servers from other networks, for example manually configured by users, they may support or not DNS64, so the considerations in will apply as well. Even in the case that the external DNS supports DNS64 function, we may be in the situation of providing incorrect configurations parameters, as explained in . Having a CLAT and using an external DNS without DNS64, ensures that everything will work. However, it needs to be reinforced, that if there is not a CLAT (scenarios without 464XLAT), an external DNS without DNS64 support, will not only guarantee that DNSSEC is broken, but also disallow any access to IPv4-only networks, so will behave as in the .

As described in Section 6.3 of (IPv6 Prefix Handling), if the CLAT can be configured with a dedicated /64 prefix for the NAT46 translation, then it will be possible to do a more efficient stateless translation. However, if this dedicated prefix is not available, the CLAT will need to do a stateful translation, for example performing stateful NAT44 for all the IPv4 LAN packets, so they appear as coming from a single IPv4 address, and then in turn, stateless translated to a single IPv6 address. The obvious recommended setup, in order to maximize the CLAT performance, is to configure the dedicated translation prefix. This can be easily achieved automatically, if the CE or end-user device is able to obtain a shorter prefix by means of DHCPv6-PD (), so the CE can use a /64 for that. The above recommendation is often not posible for cellular networks, when connecting UEs (some broadband cellular use DHCPv6-PD (), but smartphones, in general, not), as they provide a single /64 for each PDP context and use /64 prefix sharing (). So in this case, the UEs typically have a build-in CLAT client, which is doing a stateful NAT44 before the stateless NAT46.

Service providers willing to deploy NAT64, need to take into account the considerations of this document to avoid the issues depicted in this document. In the case it is a non-cellular network and the operator is providing the CEs to the customers, or suggesting them some specific models, they MUST support the customer-side translator (CLAT), in order to fully support the actual user needs (IPv4-only devices and applications, usage of literals and old APIs). If the operator offers DNS services, in order to increase performance by reducing the double translation for all the IPv4 traffic, and avoid breaking DNSSEC, they MAY support DNS64. In this case, if the DNS service is offering DNSSEC validation, then it MUST be in such way that it is aware of the DNS64. This is considered de simpler and safer approach, and MAY be combined as well with the other possible solutions described in this document: Devices running CLAT SHOULD follow the indications in the "Stub validator" section recommendation. However, most of the time, this is out of the control of the operator. CEs SHOULD include a DNS proxy and validador. This is relevant if the operator is providing the CE or suggesting it to customers. ACL of clients and Mapping-out IPv4 addresses MAY be considered by each operator, depending on their own infrastructure. This "increased performance" approach has the disadvantage of potentially breaking DNSSEC for a small percentage of validating end-hosts. If CE performance is not an issue, then a much safer approach is to not use DNS64 at all, and consequently ensure that all the IPv4 traffic is translated at the CLAT. If DNS64 is not used, one of the alternatives described in , MUST be followed. The ideal configuration for CEs supporting CLAT, is that they support DHCPv6-PD () and internally reserve one /64 for the stateless NAT46 translation. The operator MUST ensure that the customers get allocated prefixes shorter than /64 in order to support this optimization. One way or the other, this is not impacting the performance of the operator network. As indicated in Section 7 of (Deployment Considerations), operators MAY follow those suggestions in order to take advantage of traffic engineering. In the case of cellular networks, the considerations regarding DNSSEC may appear as out-of-scope, because UEs OSs, commonly don't support DNSSEC, however applications running on them may do, or it may be an OS "built-in" support in the future. Moreover, if those devices offer tethering, other client devices may be doing the validation, hence the relevance of a proper DNSSEC support by the operator network. Furthermore, cellular networks supporting 464XLAT () and "Discovery of the IPv6 Prefix Used for IPv6 Address Synthesis" (), allow a progressive IPv6 deployment, with a single APN supporting all types of PDP context (IPv4, IPv6, IPv4v6), in such way that the network is able to automatically serve all the possible combinations of UEs. Finally, if the operator choose to secure the NAT64 prefix, it MUST follow the advise indicated in Section 3.1.1. of (Validation of Discovered Pref64::/n).

The recommendations of this documents can be used as well in enterprise networks, campus and other similar scenarios, when the NAT64 is under the control of that network, and for whatever reasons, there is a need to provide "IPv6-only access" to any part of that network or it is IPv6-only connected to third party networks. An example of that is the IETF meetings network itself, where a NAT64 and DNS64 are provided, presenting in this case the same issues as per . If there is a CLAT in the IETF network, then there is no need to use DNS64 and it falls under the considerations of . Both scenarios have been tested and verified already in the IETF network itself.

This document does not have any new specific security considerations.

This document does not have any new specific IANA considerations. Note: This section is assuming that https://www.rfc-editor.org/errata/eid5152 is resolved, otherwise, this section may include the required text to resolve the issue.

The author would like to acknowledge the inputs of TBD ... Conversations with Marcelo Bagnulo, one of the co-authors of DNS64, as well as several emails in public mailing lists from Mark Andrews, have been very useful for this work. Christian Huitema inspired working in this document by suggesting that DNS64 should never be used, during a discussion regarding the deployment of CLAT in the IETF network.