Internet Engineering Task Force (IETF) J. Klensin Request for Comments: 5894 August 2010 Category: Informational ISSN: 2070-1721
Internationalized Domain Names for Applications (IDNA): Background, Explanation, and Rationale
Abstract
Several years have passed since the original protocol for Internationalized Domain Names (IDNs) was completed and deployed. During that time, a number of issues have arisen, including the need to update the system to deal with newer versions of Unicode. Some of these issues require tuning of the existing protocols and the tables on which they depend. This document provides an overview of a revised system and provides explanatory material for its components.
Status of This Memo
This document is not an Internet Standards Track specification; it is published for informational purposes.
This document is a product of the Internet Engineering Task Force (IETF). It represents the consensus of the IETF community. It has received public review and has been approved for publication by the Internet Engineering Steering Group (IESG). Not all documents approved by the IESG are a candidate for any level of Internet Standard; see Section 2 of RFC 5741.
Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at http://www.rfc-editor.org/info/rfc5894.
Klensin Informational [Page 1]
RFC 5894 IDNA Rationale August 2010
Copyright Notice
Copyright (c) 2010 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.
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.
Internationalized Domain Names in Applications (IDNA) is a collection of standards that allow client applications to convert some mnemonic strings expressed in Unicode to an ASCII-compatible encoding form ("ACE") that is a valid DNS label containing only LDH syntax (see the Definitions document [RFC5890]). The specific form of ACE label used by IDNA is called an "A-label". A client can look up an exact A-label in the existing DNS, so A-labels do not require any extensions to DNS, upgrades of DNS servers, or updates to low-level client libraries. An A-label is recognizable from the prefix "xn--" before the characters produced by the Punycode algorithm [RFC3492]; thus, a user application can identify an A-label and convert it into Unicode (or some local coded character set) for display.
On the registry side, IDNA allows a registry to offer Internationalized Domain Names (IDNs) for registration as A-labels. A registry may offer any subset of valid IDNs, and may apply any restrictions or bundling (grouping of similar labels together in one registration) appropriate for the context of that registry. Registration of labels is sometimes discussed separately from lookup, and it is subject to a few specific requirements that do not apply to lookup.
DNS clients and registries are subject to some differences in requirements for handling IDNs. In particular, registries are urged to register only exact, valid A-labels, while clients might do some mapping to get from otherwise-invalid user input to a valid A-label.
The first version of IDNA was published in 2003 and is referred to here as IDNA2003 to contrast it with the current version, which is known as IDNA2008 (after the year in which IETF work started on it). IDNA2003 consists of four documents: the IDNA base specification [RFC3490], Nameprep [RFC3491], Punycode [RFC3492], and Stringprep [RFC3454]. The current set of documents, IDNA2008, is not dependent on any of the IDNA2003 specifications other than the one for Punycode encoding. References to "IDNA2008", "these specifications", or "these documents" are to the entire IDNA2008 set listed in a separate Definitions document [RFC5890]. The characters that are valid in A-labels are identified from rules listed in the Tables document [RFC5892], but validity can be derived from the Unicode properties of those characters with a very few exceptions.
Traditionally, DNS labels are matched case-insensitively (as described in the DNS specifications [RFC1034][RFC1035]). That convention was preserved in IDNA2003 by a case-folding operation that
Klensin Informational [Page 4]
RFC 5894 IDNA Rationale August 2010
generally maps capital letters into lowercase ones. However, if case rules are enforced from one language, another language sometimes loses the ability to treat two characters separately. Case- insensitivity is treated slightly differently in IDNA2008.
IDNA2003 used Unicode version 3.2 only. In order to keep up with new characters added in new versions of Unicode, IDNA2008 decouples its rules from any particular version of Unicode. Instead, the attributes of new characters in Unicode, supplemented by a small number of exception cases, determine how and whether the characters can be used in IDNA labels.
This document provides informational context for IDNA2008, including terminology, background, and policy discussions. It contains no normative material; specifications for conformance to the IDNA2008 protocols appears entirely in the other documents in the series.
Terminology for IDNA2008 appears in the Definitions document [RFC5890]. That document also contains a road map to the IDNA2008 document collection. No attempt should be made to understand this document without the definitions and concepts that appear there.
In the context of IDNs, the DNS term "name" has introduced some confusion as people speak of DNS labels in terms of the words or phrases of various natural languages. Historically, many of the "names" in the DNS have been mnemonics to identify some particular concept, object, or organization. They are typically rooted in some language because most people think in language-based ways. But, because they are mnemonics, they need not obey the orthographic conventions of any language: it is not a requirement that it be possible for them to be "words".
This distinction is important because the reasonable goal of an IDN effort is not to be able to write the great Klingon (or language of one's choice) novel in DNS labels but to be able to form a usefully broad range of mnemonics in ways that are as natural as possible in a very broad range of scripts.
IDNA2008 introduces new terminology. Precise definitions are provided in the Definitions document for the terms U-label, A-Label, LDH label (to which all valid pre-IDNA hostnames conformed), Reserved LDH label (R-LDH label), XN-label, Fake A-label, and Non-Reserved LDH label (NR-LDH label).
In addition, the term "putative label" has been adopted to refer to a label that may appear to meet certain definitional constraints but has not yet been sufficiently tested for validity.
These definitions are also illustrated in Figure 1 of the Definitions document. R-LDH labels contain "--" in the third and fourth character positions from the beginning of the label. In IDNA-aware applications, only a subset of these reserved labels is permitted to be used, namely the A-label subset. A-labels are a subset of the R-LDH labels that begin with the case-insensitive string "xn--". Labels that bear this prefix but that are not otherwise valid fall into the "Fake A-label" category. The Non-Reserved labels (NR-LDH labels) are implicitly valid since they do not bear any resemblance to the labels specified by IDNA.
The creation of the Reserved-LDH category is required for three reasons:
o to prevent confusion with pre-IDNA coding forms;
o to permit future extensions that would require changing the prefix, no matter how unlikely those might be (see Section 7.4); and
o to reduce the opportunities for attacks via the Punycode encoding algorithm itself.
As with other documents in the IDNA2008 set, this document uses the term "registry" to describe any zone in the DNS. That term, and the terms "zone" or "zone administration", are interchangeable.
o Use a more recent version of Unicode and allow IDNA to be independent of Unicode versions, so that IDNA2008 need not be updated for implementations to adopt code points from new Unicode versions.
Klensin Informational [Page 6]
RFC 5894 IDNA Rationale August 2010
o Fix a very small number of code point categorizations that have turned out to cause problems in the communities that use those code points.
o Reduce the dependency on mapping, in favor of valid A-labels. This will result in pre-mapped forms that are not valid IDNA labels appearing less often in various contexts.
o Fix some details in the bidirectional code point handling algorithms.
The IDNA specification solves the problem of extending the repertoire of characters that can be used in domain names to include a large subset of the Unicode repertoire.
IDNA does not extend DNS. Instead, the applications (and, by implication, the users) continue to see an exact-match lookup service. Either there is a single name that matches exactly (subject to the base DNS requirement of case-insensitive ASCII matching) or there is no match. This model has served the existing applications well, but it requires, with or without internationalized domain names, that users know the exact spelling of the domain names that are to be typed into applications such as web browsers and mail user agents. The introduction of the larger repertoire of characters potentially makes the set of misspellings larger, especially given that in some cases the same appearance, for example on a business card, might visually match several Unicode code points or several sequences of code points.
The IDNA standard does not require any applications to conform to it, nor does it retroactively change those applications. An application can elect to use IDNA in order to support IDNs while maintaining interoperability with existing infrastructure. For applications that want to use non-ASCII characters in public DNS domain names, IDNA is the only option that is defined at the time this specification is published. Adding IDNA support to an existing application entails changes to the application only, and leaves room for flexibility in front-end processing and more specifically in the user interface (see Section 6).
A great deal of the discussion of IDN solutions has focused on transition issues and how IDNs will work in a world where not all of the components have been updated. Proposals that were not chosen by the original IDN Working Group would have depended on updating user applications, DNS resolvers, and DNS servers in order for a user to apply an internationalized domain name in any form or coding
Klensin Informational [Page 7]
RFC 5894 IDNA Rationale August 2010
acceptable under that method. While processing must be performed prior to or after access to the DNS, IDNA requires no changes to the DNS protocol, any DNS servers, or the resolvers on users' computers.
IDNA allows the graceful introduction of IDNs not only by avoiding upgrades to existing infrastructure (such as DNS servers and mail transport agents), but also by allowing some limited use of IDNs in applications by using the ASCII-encoded representation of the labels containing non-ASCII characters. While such names are user- unfriendly to read and type, and hence not optimal for user input, they can be used as a last resort to allow rudimentary IDN usage. For example, they might be the best choice for display if it were known that relevant fonts were not available on the user's computer. In order to allow user-friendly input and output of the IDNs and acceptance of some characters as equivalent to those to be processed according to the protocol, the applications need to be modified to conform to this specification.
This version of IDNA uses the Unicode character repertoire for continuity with the original version of IDNA.
1.5. Comprehensibility of IDNA Mechanisms and Processing
One goal of IDNA2008, which is aided by the main goal of reducing the dependency on mapping, is to improve the general understanding of how IDNA works and what characters are permitted and what happens to them. Comprehensibility and predictability to users and registrants are important design goals for this effort. End-user applications have an important role to play in increasing this comprehensibility.
Any system that tries to handle international characters encounters some common problems. For example, a User Interface (UI) cannot display a character if no font containing that character is available. In some cases, internationalization enables effective localization while maintaining some global uniformity but losing some universality.
It is difficult to even make suggestions as to how end-user applications should cope when characters and fonts are not available. Because display functions are rarely controlled by the types of applications that would call upon IDNA, such suggestions will rarely be very effective.
Conversion between local character sets and normalized Unicode, if needed, is part of this set of user interface issues. Those conversions introduce complexity in a system that does not use Unicode as its primary (or only) internal character coding system. If a label is converted to a local character set that does not have
Klensin Informational [Page 8]
RFC 5894 IDNA Rationale August 2010
all the needed characters, or that uses different character-coding principles, the user interface program may have to add special logic to avoid or reduce loss of information.
The major difficulty may lie in accurately identifying the incoming character set and applying the correct conversion routine. Even more difficult, the local character coding system could be based on conceptually different assumptions than those used by Unicode (e.g., choice of font encodings used for publications in some Indic scripts). Those differences may not easily yield unambiguous conversions or interpretations even if each coding system is internally consistent and adequate to represent the local language and script.
IDNA2008 shifts responsibility for character mapping and other adjustments from the protocol (where it was located in IDNA2003) to pre-processing before invoking IDNA itself. The intent is that this change will lead to greater usage of fully-valid A-Labels or U-labels in display, transit, and storage, which should aid comprehensibility and predictability. A careful look at pre-processing raises issues about what that pre-processing should do and at what point pre-processing becomes harmful; how universally consistent pre-processing algorithms can be; and how to be compatible with labels prepared in an IDNA2003 context. Those issues are discussed in Section 6 and in the Mapping document [IDNA2008-Mapping].
IDNA2008 separates Domain Name Registration and Lookup in the protocol specification (RFC 5891, Sections 4 and 5 [RFC5891]). Although most steps in the two processes are similar, the separation reflects current practice in which per-registry (DNS zone) restrictions and special processing are applied at registration time but not during lookup. Another significant benefit is that separation facilitates incremental addition of permitted character groups to avoid freezing on one particular version of Unicode.
The actual registration and lookup protocols for IDNA2008 are specified in the Protocol document.
IDNA2008 adopts the inclusion model. A code point is assumed to be invalid for IDN use unless it is included as part of a Unicode property-based rule or, in rare cases, included individually by an exception. When an implementation moves to a new version of Unicode, the rules may indicate new valid code points.
Klensin Informational [Page 9]
RFC 5894 IDNA Rationale August 2010
This section provides an overview of the model used to establish the algorithm and character lists of the Tables document [RFC5892] and describes the names and applicability of the categories used there. Note that the inclusion of a character in the PROTOCOL-VALID category group (Section 3.1.1) does not imply that it can be used indiscriminately; some characters are associated with contextual rules that must be applied as well.
The information given in this section is provided to make the rules, tables, and protocol easier to understand. The normative generating rules that correspond to this informal discussion appear in the Tables document, and the rules that actually determine what labels can be registered or looked up are in the Protocol document.
3.1. A Tiered Model of Permitted Characters and Labels
Moving to an inclusion model involves a new specification for the list of characters that are permitted in IDNs. In IDNA2003, character validity is independent of context and fixed forever (or until the standard is replaced). However, globally context- independent rules have proved to be impractical because some characters, especially those that are called "Join_Controls" in Unicode, are needed to make reasonable use of some scripts but have no visible effect in others. IDNA2003 prohibited those types of characters entirely by discarding them. We now have a consensus that under some conditions, these "joiner" characters are legitimately needed to allow useful mnemonics for some languages and scripts. In general, context-dependent rules help deal with characters (generally characters that would otherwise be prohibited entirely) that are used differently or perceived differently across different scripts, and allow the standard to be applied more appropriately in cases where a string is not universally handled the same way.
IDNA2008 divides all possible Unicode code points into four categories: PROTOCOL-VALID, CONTEXTUAL RULE REQUIRED, DISALLOWED, and UNASSIGNED.
Characters identified as PROTOCOL-VALID (often abbreviated PVALID) are permitted in IDNs. Their use may be restricted by rules about the context in which they appear or by other rules that apply to the entire label in which they are to be embedded. For example, any label that contains a character in this category that has a "right-to-left" property must be used in context with the Bidi rules [RFC5893]. The term PROTOCOL-VALID is used to stress the fact that the presence of a character in this category does not imply that a given registry need accept registrations containing any of the
Klensin Informational [Page 10]
RFC 5894 IDNA Rationale August 2010
characters in the category. Registries are still expected to apply judgment about labels they will accept and to maintain rules consistent with those judgments (see the Protocol document [RFC5891] and Section 3.3).
Characters that are placed in the PROTOCOL-VALID category are expected to never be removed from it or reclassified. While theoretically characters could be removed from Unicode, such removal would be inconsistent with the Unicode stability principles (see UTR 39: Unicode Security Mechanisms [Unicode52], Appendix F) and hence should never occur.
Some characters may be unsuitable for general use in IDNs but necessary for the plausible support of some scripts. The two most commonly cited examples are the ZERO WIDTH JOINER and ZERO WIDTH NON-JOINER characters (ZWJ, U+200D and ZWNJ, U+200C), but other characters may require special treatment because they would otherwise be DISALLOWED (typically because Unicode considers them punctuation or special symbols) but need to be permitted in limited contexts. Other characters are given this special treatment because they pose exceptional danger of being used to produce misleading labels or to cause unacceptable ambiguity in label matching and interpretation.
Characters with contextual restrictions are identified as CONTEXTUAL RULE REQUIRED and are associated with a rule. The rule defines whether the character is valid in a particular string, and also whether the rule itself is to be applied on lookup as well as registration.
A distinction is made between characters that indicate or prohibit joining and ones similar to them (known as CONTEXT-JOINER or CONTEXTJ) and other characters requiring contextual treatment (CONTEXT-OTHER or CONTEXTO). Only the former require full testing at lookup time.
It is important to note that these contextual rules cannot prevent all uses of the relevant characters that might be confusing or problematic. What they are expected to do is to confine applicability of the characters to scripts (and narrower contexts) where zone administrators are knowledgeable enough about the use of those characters to be prepared to deal with them appropriately.
Klensin Informational [Page 11]
RFC 5894 IDNA Rationale August 2010
For example, a registry dealing with an Indic script that requires ZWJ and/or ZWNJ as part of the writing system is expected to understand where the characters have visible effect and where they do not and to make registration rules accordingly. By contrast, a registry dealing primarily with Latin or Cyrillic script might not be actively aware that the characters exist, much less about the consequences of embedding them in labels drawn from those scripts and therefore should avoid accepting registrations containing those characters, at least in labels using characters from the Latin or Cyrillic scripts.
Rules have descriptions such as "Must follow a character from Script XYZ", "Must occur only if the entire label is in Script ABC", or "Must occur only if the previous and subsequent characters have the DFG property". The actual rules may be DEFINED or NULL. If present, they may have values of "True" (character may be used in any position in any label), "False" (character may not be used in any label), or may be a set of procedural rules that specify the context in which the character is permitted.
Because it is easier to identify these characters than to know that they are actually needed in IDNs or how to establish exactly the right rules for each one, a rule may have a null value in a given version of the tables. Characters associated with null rules are not permitted to appear in putative labels for either registration or lookup. Of course, a later version of the tables might contain a non-null rule.
The actual rules and their descriptions are in Sections 2 and 3 of the Tables document [RFC5892]. That document also specifies the creation of a registry for future rules.
Some characters are inappropriate for use in IDNs and are thus excluded for both registration and lookup (i.e., IDNA-conforming applications performing name lookup should verify that these characters are absent; if they are present, the label strings should be rejected rather than converted to A-labels and looked up. Some of these characters are problematic for use in IDNs (such as the FRACTION SLASH character, U+2044), while some of them (such as the various HEART symbols, e.g., U+2665, U+2661, and U+2765, see Section 7.6) simply fall outside the conventions for typical identifiers (basically letters and numbers).
Klensin Informational [Page 12]
RFC 5894 IDNA Rationale August 2010
Of course, this category would include code points that had been removed entirely from Unicode should such removals ever occur.
Characters that are placed in the DISALLOWED category are expected to never be removed from it or reclassified. If a character is classified as DISALLOWED in error and the error is sufficiently problematic, the only recourse would be either to introduce a new code point into Unicode and classify it as PROTOCOL-VALID or for the IETF to accept the considerable costs of an incompatible change and replace the relevant RFC with one containing appropriate exceptions.
There is provision for exception cases but, in general, characters are placed into DISALLOWED if they fall into one or more of the following groups:
o The character is a compatibility equivalent for another character. In slightly more precise Unicode terms, application of Normalization Form KC (NFKC) to the character yields some other character.
o The character is an uppercase form or some other form that is mapped to another character by Unicode case folding.
o The character is a symbol or punctuation form or, more generally, something that is not a letter, digit, or a mark that is used to form a letter or digit.
For convenience in processing and table-building, code points that do not have assigned values in a given version of Unicode are treated as belonging to a special UNASSIGNED category. Such code points are prohibited in labels to be registered or looked up. The category differs from DISALLOWED in that code points are moved out of it by the simple expedient of being assigned in a later version of Unicode (at which point, they are classified into one of the other categories as appropriate).
The rationale for restricting the processing of UNASSIGNED characters is simply that the properties of such code points cannot be completely known until actual characters are assigned to them. For example, assume that an UNASSIGNED code point were included in a label to be looked up. Assume that the code point was later assigned to a character that required some set of contextual rules. With that combination, un-updated instances of IDNA-aware software might permit lookup of labels containing the previously unassigned characters while updated versions of the software might restrict use of the same
Klensin Informational [Page 13]
RFC 5894 IDNA Rationale August 2010
label in lookup, depending on the contextual rules. It should be clear that under no circumstance should an UNASSIGNED character be permitted in a label to be registered as part of a domain name.
While these recommendations cannot and should not define registry policies, registries should develop and apply additional restrictions as needed to reduce confusion and other problems. For example, it is generally believed that labels containing characters from more than one script are a bad practice although there may be some important exceptions to that principle. Some registries may choose to restrict registrations to characters drawn from a very small number of scripts. For many scripts, the use of variant techniques such as those as described in the JET specification for the CJK script [RFC3743] and its generalization [RFC4290], and illustrated for Chinese by the tables provided by the Chinese Domain Name Consortium [RFC4713] may be helpful in reducing problems that might be perceived by users.
In general, users will benefit if registries only permit characters from scripts that are well-understood by the registry or its advisers. If a registry decides to reduce opportunities for confusion by constructing policies that disallow characters used in historic writing systems or characters whose use is restricted to specialized, highly technical contexts, some relevant information may be found in Section 2.4 (Specific Character Adjustments) of Unicode Identifier and Pattern Syntax [Unicode-UAX31], especially Table 4 (Candidate Characters for Exclusion from Identifiers), and Section 3.1 (General Security Profile for Identifiers) in Unicode Security Mechanisms [Unicode-UTS39].
The requirement (in Section 4.1 of the Protocol document [RFC5891]) that registration procedures use only U-labels and/or A-labels is intended to ensure that registrants are fully aware of exactly what is being registered as well as encouraging use of those canonical forms. That provision should not be interpreted as requiring that registrants need to provide characters in a particular code sequence. Registrant input conventions and management are part of registrant- registrar interactions and relationships between registries and registrars and are outside the scope of these standards.
It is worth stressing that these principles of policy development and application apply at all levels of the DNS, not only, e.g., top level domain (TLD) or second level domain (SLD) registrations. Even a trivial, "anything is permitted that is valid under the protocol" policy is helpful in that it helps users and application developers know what to expect.
Klensin Informational [Page 14]
RFC 5894 IDNA Rationale August 2010
3.3. Layered Restrictions: Tables, Context, Registration, and Applications
The character rules in IDNA2008 are based on the realization that there is no single magic bullet for any of the security, confusability, or other issues associated with IDNs. Instead, the specifications define a variety of approaches. The character tables are the first mechanism, protocol rules about how those characters are applied or restricted in context are the second, and those two in combination constitute the limits of what can be done in the protocol. As discussed in the previous section (Section 3.2), registries are expected to restrict what they permit to be registered, devising and using rules that are designed to optimize the balance between confusion and risk on the one hand and maximum expressiveness in mnemonics on the other.
In addition, there is an important role for user interface programs in warning against label forms that appear problematic given their knowledge of local contexts and conventions. Of course, no approach based on naming or identifiers alone can protect against all threats.
Domain names are always transmitted in network order (the order in which the code points are sent in protocols), but they may have a different display order (the order in which the code points are displayed on a screen or paper). When a domain name contains characters that are normally written right to left, display order may be affected although network order is not. It gets even more complicated if left-to-right and right-to-left labels are adjacent to each other within a domain name. The decision about the display order is ultimately under the control of user agents -- including Web browsers, mail clients, hosted Web applications and many more -- which may be highly localized. Should a domain name abc.def, in which both labels are represented in scripts that are written right to left, be displayed as fed.cba or cba.fed? Applications that are in deployment today are already diverse, and one can find examples of either choice.
The picture changes once again when an IDN appears in an Internationalized Resource Identifier (IRI) [RFC3987]. An IRI or internationalized email address contains elements other than the domain name. For example, IRIs contain protocol identifiers and field delimiter syntax such as "http://" or "mailto:" while email addresses contain the "@" to separate local parts from domain names.
Klensin Informational [Page 15]
RFC 5894 IDNA Rationale August 2010
An IRI in network order begins with "http://" followed by domain labels in network order, thus "http://abc.def".
User interface programs are not required to display and allow input of IRIs directly but often do so. Implementers have to choose whether the overall direction of these strings will always be left to right (or right to left) for an IRI or email address. The natural order for a user typing a domain name on a right-to-left system is fed.cba. Should the right-to-left (RTL) user interface reverse the entire domain name each time a domain name is typed? Does this change if the user types "http://" right before typing a domain name, thus implying that the user is beginning at the beginning of the network-order IRI? Experience in the 1980s and 1990s with mixing systems in which domain name labels were read in network order (left to right) and those in which those labels were read right to left would predict a great deal of confusion.
If each implementation of each application makes its own decisions on these issues, users will develop heuristics that will sometimes fail when switching applications. However, while some display order conventions, voluntarily adopted, would be desirable to reduce confusion, such suggestions are beyond the scope of these specifications.
Applications can accept and display domain names using any character set or character coding system. The IDNA protocol does not necessarily affect the interface between users and applications. An IDNA-aware application can accept and display internationalized domain names in two formats: as the internationalized character set(s) supported by the application (i.e., an appropriate local representation of a U-label) and as an A-label. Applications may allow the display of A-labels, but are encouraged not to do so except as an interface for special purposes, possibly for debugging, or to cope with display limitations. In general, they should allow, but not encourage, user input of A-labels. A-labels are opaque and ugly, and malicious variations on them are not easily detected by users. Where possible, they should thus only be exposed when they are absolutely needed. Because IDN labels can be rendered either as A-labels or U-labels, the application may reasonably have an option for the user to select the preferred method of display. Rendering the U-label should normally be the default.
Domain names are often stored and transported in many places. For example, they are part of documents such as mail messages and web pages. They are transported in many parts of many protocols, such as both the control commands of SMTP and associated message body parts,
Klensin Informational [Page 16]
RFC 5894 IDNA Rationale August 2010
and in the headers and the body content in HTTP. It is important to remember that domain names appear both in domain name slots and in the content that is passed over protocols, and it would be helpful if protocols explicitly define what their domain name slots are.
In protocols and document formats that define how to handle specification or negotiation of charsets, labels can be encoded in any charset allowed by the protocol or document format. If a protocol or document format only allows one charset, the labels must be given in that charset. Of course, not all charsets can properly represent all labels. If a U-label cannot be displayed in its entirety, the only choice (without loss of information) may be to display the A-label.
Where a protocol or document format allows IDNs, labels should be in whatever character encoding and escape mechanism the protocol or document format uses in the local environment. This provision is intended to prevent situations in which, e.g., UTF-8 domain names appear embedded in text that is otherwise in some other character coding.
All protocols that use domain name slots (see Section 2.3.2.6 in the Definitions document [RFC5890]) already have the capacity for handling domain names in the ASCII charset. Thus, A-labels can inherently be handled by those protocols.
IDNA2008 does not specify required mappings between one character or code point and others. An extended discussion of mapping issues appears in Section 6 and specific recommendations appear in the Mapping document [IDNA2008-Mapping]. In general, IDNA2008 prohibits characters that would be mapped to others by normalization or other rules. As examples, while mathematical characters based on Latin ones are accepted as input to IDNA2003, they are prohibited in IDNA2008. Similarly, uppercase characters, double-width characters, and other variations are prohibited as IDNA input although mapping them as needed in user interfaces is strongly encouraged.
Since the rules in the Tables document [RFC5892] have the effect that only strings that are not transformed by NFKC are valid, if an application chooses to perform NFKC normalization before lookup, that operation is safe since this will never make the application unable to look up any valid string. However, as discussed above, the application cannot guarantee that any other application will perform that mapping, so it should be used only with caution and for informed users.
Klensin Informational [Page 17]
RFC 5894 IDNA Rationale August 2010
In many cases, these prohibitions should have no effect on what the user can type as input to the lookup process. It is perfectly reasonable for systems that support user interfaces to perform some character mapping that is appropriate to the local environment. This would normally be done prior to actual invocation of IDNA. At least conceptually, the mapping would be part of the Unicode conversions discussed above and in the Protocol document [RFC5891]. However, those changes will be local ones only -- local to environments in which users will clearly understand that the character forms are equivalent. For use in interchanges among systems, it appears to be much more important that U-labels and A-labels can be mapped back and forth without loss of information.
One specific, and very important, instance of this strategy arises with case folding. In the ASCII-only DNS, names are looked up and matched in a case-independent way, but no actual case folding occurs. Names can be placed in the DNS in either uppercase or lowercase form (or any mixture of them) and that form is preserved, returned in queries, and so on. IDNA2003 approximated that behavior for non-ASCII strings by performing case folding at registration time (resulting in only lowercase IDNs in the DNS) and when names were looked up.
As suggested earlier in this section, it appears to be desirable to do as little character mapping as possible as long as Unicode works correctly (e.g., Normalization Form C (NFC) mapping to resolve different codings for the same character is still necessary although the specifications require that it be performed prior to invoking the protocol) in order to make the mapping between A-labels and U-labels idempotent. Case mapping is not an exception to this principle. If only lowercase characters can be registered in the DNS (i.e., be present in a U-label), then IDNA2008 should prohibit uppercase characters as input even though user interfaces to applications should probably map those characters. Some other considerations reinforce this conclusion. For example, in ASCII case mapping for individual characters, uppercase(character) is always equal to uppercase(lowercase(character)). That may not be true with IDNs. In some scripts that use case distinctions, there are a few characters that do not have counterparts in one case or the other. The relationship between uppercase and lowercase may even be language- dependent, with different languages (or even the same language in different areas) expecting different mappings. User interface programs can meet the expectations of users who are accustomed to the case-insensitive DNS environment by performing case folding prior to IDNA processing, but the IDNA procedures themselves should neither require such mapping nor expect them when they are not natural to the localized environment.
Klensin Informational [Page 18]
RFC 5894 IDNA Rationale August 2010
4.3. Linguistic Expectations: Ligatures, Digraphs, and Alternate Character Forms
Users have expectations about character matching or equivalence that are based on their own languages and the orthography of those languages. These expectations may not always be met in a global system, especially if multiple languages are written using the same script but using different conventions. Some examples:
o A Norwegian user might expect a label with the ae-ligature to be treated as the same label as one using the Swedish spelling with a-diaeresis even though applying that mapping to English would be astonishing to users.
o A German user might expect a label with an o-umlaut and a label that had "oe" substituted, but was otherwise the same, to be treated as equivalent even though that substitution would be a clear error in Swedish.
o A Chinese user might expect automatic matching of Simplified and Traditional Chinese characters, but applying that matching for Korean or Japanese text would create considerable confusion.
o An English user might expect "theater" and "theatre" to match.
A number of languages use alphabetic scripts in which single phonemes are written using two characters, termed a "digraph", for example, the "ph" in "pharmacy" and "telephone". (Such characters can also appear consecutively without forming a digraph, as in "tophat".) Certain digraphs may be indicated typographically by setting the two characters closer together than they would be if used consecutively to represent different phonemes. Some digraphs are fully joined as ligatures. For example, the word "encyclopaedia" is sometimes set with a U+00E6 LATIN SMALL LIGATURE AE. When ligature and digraph forms have the same interpretation across all languages that use a given script, application of Unicode normalization generally resolves the differences and causes them to match. When they have different interpretations, matching must utilize other methods, presumably chosen at the registry level, or users must be educated to understand that matching will not occur.
The nature of the problem can be illustrated by many words in the Norwegian language, where the "ae" ligature is the 27th letter of a 29-letter extended Latin alphabet. It is equivalent to the 28th letter of the Swedish alphabet (also containing 29 letters), U+00E4 LATIN SMALL LETTER A WITH DIAERESIS, for which an "ae" cannot be substituted according to current orthographic standards. That character (U+00E4) is also part of the German alphabet where, unlike
Klensin Informational [Page 19]
RFC 5894 IDNA Rationale August 2010
in the Nordic languages, the two-character sequence "ae" is usually treated as a fully acceptable alternate orthography for the "umlauted a" character. The inverse is however not true, and those two characters cannot necessarily be combined into an "umlauted a". This also applies to another German character, the "umlauted o" (U+00F6 LATIN SMALL LETTER O WITH DIAERESIS) which, for example, cannot be used for writing the name of the author "Goethe". It is also a letter in the Swedish alphabet where, like the "a with diaeresis", it cannot be correctly represented as "oe" and in the Norwegian alphabet, where it is represented, not as "o with diaeresis", but as "slashed o", U+00F8.
Some of the ligatures that have explicit code points in Unicode were given special handling in IDNA2003 and now pose additional problems in transition. See Section 7.2.
Additional cases with alphabets written right to left are described in Section 4.5.
Matching and comparison algorithm selection often requires information about the language being used, context, or both -- information that is not available to IDNA or the DNS. Consequently, IDNA2008 makes no attempt to treat combined characters in any special way. A registry that is aware of the language context in which labels are to be registered, and where that language sometimes (or always) treats the two-character sequences as equivalent to the combined form, should give serious consideration to applying a "variant" model [RFC3743][RFC4290] or to prohibiting registration of one of the forms entirely, to reduce the opportunities for user confusion and fraud that would result from the related strings being registered to different parties.
In the DNS, ASCII letters are stored with their case preserved. Matching during the query process is case-independent, but none of the information that might be represented by choices of case has been lost. That model has been accidentally helpful because, as people have created DNS labels by catenating words (or parts of words) to form labels, case has often been used to distinguish among components and make the labels more memorable.
Since DNS servers do not get involved in parsing IDNs, they cannot do case-independent matching. Thus, keeping the cases separate in lookup or registration, and doing matching at the server, is not feasible with IDNA or any similar approach. Matching of characters that are considered to differ only by case must be done, if desired, by programs invoking IDNA lookup even though it wasn't done by ASCII-
Klensin Informational [Page 20]
RFC 5894 IDNA Rationale August 2010
only DNS clients. That situation was recognized in IDNA2003 and nothing in IDNA2008 fundamentally changes it or could do so. In IDNA2003, all characters are case folded and mapped by clients in a standardized step.
Even in scripts that generally support case distinctions, some characters do not have uppercase forms. For example, the Unicode case-folding operation maps Greek Final Form Sigma (U+03C2) to the medial form (U+03C3) and maps Eszett (German Sharp S, U+00DF) to "ss". Neither of these mappings is reversible because the uppercase of U+03C3 is the uppercase Sigma (U+03A3) and "ss" is an ASCII string. IDNA2008 permits, at the risk of some incompatibility, slightly more flexibility in this area by avoiding case folding and treating these characters as themselves. Approaches to handling one- way mappings are discussed in Section 7.2.
Because IDNA2003 maps Final Sigma and Eszett to other characters, and the reverse mapping is never possible, neither Final Sigma nor Eszett can be represented in the ACE form of IDNA2003 IDN nor in the native character (U-label) form derived from it. With IDNA2008, both characters can be used in an IDN and so the A-label used for lookup for any U-label containing those characters is now different. See Section 7.1 for a discussion of what kinds of changes might require the IDNA prefix to change; after extended discussions, the IDNABIS Working Group came to consensus that the change for these characters did not justify a prefix change.
In order to be sure that the directionality of right-to-left text is unambiguous, IDNA2003 required that any label in which right-to-left characters appear both starts and ends with them and that it does not include any characters with strong left-to-right properties (that excludes other alphabetic characters but permits European digits). Any other string that contains a right-to-left character and does not meet those requirements is rejected. This is one of the few places where the IDNA algorithms (both in IDNA2003 and in IDNA2008) examine an entire label, not just individual characters. The algorithmic model used in IDNA2003 rejects the label when the final character in a right-to-left string requires a combining mark in order to be correctly represented.
That prohibition is not acceptable for writing systems for languages written with consonantal alphabets to which diacritical vocalic systems are applied, and for languages with orthographies derived from them where the combining marks may have different functionality. In both cases, the combining marks can be essential components of the orthography. Examples of this are Yiddish, written with an extended
Klensin Informational [Page 21]
RFC 5894 IDNA Rationale August 2010
Hebrew script, and Dhivehi (the official language of Maldives), which is written in the Thaana script (which is, in turn, derived from the Arabic script). IDNA2008 removes the restriction on final combining characters with a new set of rules for right-to-left scripts and their characters. Those new rules are specified in the Bidi document [RFC5893].
The "Robustness Principle" is often stated as "Be conservative about what you send and liberal in what you accept" (see, e.g., Section 1.2.2 of the applications-layer Host Requirements specification [RFC1123]). This principle applies to IDNA. In applying the principle to registries as the source ("sender") of all registered and useful IDNs, registries are responsible for being conservative about what they register and put out in the Internet. For IDNs to work well, zone administrators (registries) must have and require sensible policies about what is registered -- conservative policies -- and implement and enforce them.
Conversely, lookup applications are expected to reject labels that clearly violate global (protocol) rules (no one has ever seriously claimed that being liberal in what is accepted requires being stupid). However, once one gets past such global rules and deals with anything sensitive to script or locale, it is necessary to assume that garbage has not been placed into the DNS, i.e., one must be liberal about what one is willing to look up in the DNS rather than guessing about whether it should have been permitted to be registered.
If a string cannot be successfully found in the DNS after the lookup processing described here, it makes no difference whether it simply wasn't registered or was prohibited by some rule at the registry. Application implementers should be aware that where DNS wildcards are used, the ability to successfully resolve a name does not guarantee that it was actually registered.
6. Front-end and User Interface Processing for Lookup
Domain names may be identified and processed in many contexts. They may be typed in by users themselves or embedded in an identifier such as an email address, URI, or IRI. They may occur in running text or be processed by one system after being provided in another. Systems may try to normalize URLs to determine (or guess) whether a reference is valid or if two references point to the same object without actually looking the objects up (comparison without lookup is necessary for URI types that are not intended to be resolved). Some of these goals may be more easily and reliably satisfied than others.
Klensin Informational [Page 22]
RFC 5894 IDNA Rationale August 2010
While there are strong arguments for any domain name that is placed "on the wire" -- transmitted between systems -- to be in the zero- ambiguity forms of A-labels, it is inevitable that programs that process domain names will encounter U-labels or variant forms.
An application that implements the IDNA protocol [RFC5891] will always take any user input and convert it to a set of Unicode code points. That user input may be acquired by any of several different input methods, all with differing conversion processes to be taken into consideration (e.g., typed on a keyboard, written by hand onto some sort of digitizer, spoken into a microphone and interpreted by a speech-to-text engine, etc.). The process of taking any particular user input and mapping it into a Unicode code point may be a simple one: if a user strikes the "A" key on a US English keyboard, without any modifiers such as the "Shift" key held down, in order to draw a Latin small letter A ("a"), many (perhaps most) modern operating system input methods will produce to the calling application the code point U+0061, encoded in a single octet.
Sometimes the process is somewhat more complicated: a user might strike a particular set of keys to represent a combining macron followed by striking the "A" key in order to draw a Latin small letter A with a macron above it. Depending on the operating system, the input method chosen by the user, and even the parameters with which the application communicates with the input method, the result might be the code point U+0101 (encoded as two octets in UTF-8 or UTF-16, four octets in UTF-32, etc.), the code point U+0061 followed by the code point U+0304 (again, encoded in three or more octets, depending upon the encoding used) or even the code point U+FF41 followed by the code point U+0304 (and encoded in some form). These examples leave aside the issue of operating systems and input methods that do not use Unicode code points for their character set.
In every case, applications (with the help of the operating systems on which they run and the input methods used) need to perform a mapping from user input into Unicode code points.
IDNA2003 used a model whereby input was taken from the user, mapped (via whatever input method mechanisms were used) to a set of Unicode code points, and then further mapped to a set of Unicode code points using the Nameprep profile [RFC3491]. In this procedure, there are two separate mapping steps: first, a mapping done by the input method (which might be controlled by the operating system, the application, or some combination) and then a second mapping performed by the Nameprep portion of the IDNA protocol. The mapping done in Nameprep includes a particular mapping table to re-map some characters to other characters, a particular normalization, and a set of prohibited characters.
Klensin Informational [Page 23]
RFC 5894 IDNA Rationale August 2010
Note that the result of the two-step mapping process means that the mapping chosen by the operating system or application in the first step might differ significantly from the mapping supplied by the Nameprep profile in the second step. This has advantages and disadvantages. Of course, the second mapping regularizes what gets looked up in the DNS, making for better interoperability between implementations that use the Nameprep mapping. However, the application or operating system may choose mappings in their input methods, which when passed through the second (Nameprep) mapping result in characters that are "surprising" to the end user.
The other important feature of IDNA2003 is that, with very few exceptions, it assumes that any set of Unicode code points provided to the Nameprep mapping can be mapped into a string of Unicode code points that are "sensible", even if that means mapping some code points to nothing (that is, removing the code points from the string). This allowed maximum flexibility in input strings.
The present version of IDNA (IDNA2008) differs significantly in approach from the original version. First and foremost, it does not provide explicit mapping instructions. Instead, it assumes that the application (perhaps via an operating system input method) will do whatever mapping it requires to convert input into Unicode code points. This has the advantage of giving flexibility to the application to choose a mapping that is suitable for its user given specific user requirements, and avoids the two-step mapping of the original protocol. Instead of a mapping, IDNA2008 provides a set of categories that can be used to specify the valid code points allowed in a domain name.
In principle, an application ought to take user input of a domain name and convert it to the set of Unicode code points that represent the domain name the user intends. As a practical matter, of course, determining user intent is a tricky business, so an application needs to choose a reasonable mapping from user input. That may differ based on the particular circumstances of a user, depending on locale, language, type of input method, etc. It is up to the application to make a reasonable choice.
Klensin Informational [Page 24]
RFC 5894 IDNA Rationale August 2010
7. Migration from IDNA2003 and Unicode Version Synchronization
As mentioned above and in the IAB review and recommendations for IDNs [RFC4690], two key goals of the IDNA2008 design are:
o to enable applications to be agnostic about whether they are being run in environments supporting any Unicode version from 3.2 onward.
o to permit incrementally adding new characters, character groups, scripts, and other character collections as they are incorporated into Unicode, doing so without disruption and, in the long term, without "heavy" processes (an IETF consensus process is required by the IDNA2008 specifications and is expected to be required and used until significant experience accumulates with IDNA operations and new versions of Unicode).
7.1.1. Summary and Discussion of IDNA Validity Criteria
The general criteria for a label to be considered valid under IDNA are (the actual rules are rigorously defined in the Protocol [RFC5891] and Tables [RFC5892] documents):
o The characters are "letters", marks needed to form letters, numerals, or other code points used to write words in some language. Symbols, drawing characters, and various notational characters are intended to be permanently excluded. There is no evidence that they are important enough to Internet operations or internationalization to justify expansion of domain names beyond the general principle of "letters, digits, and hyphen". (Additional discussion and rationale for the symbol decision appears in Section 7.6.)
o Other than in very exceptional cases, e.g., where they are needed to write substantially any word of a given language, punctuation characters are excluded. The fact that a word exists is not proof that it should be usable in a DNS label, and DNS labels are not expected to be usable for multiple-word phrases (although they are certainly not prohibited if the conventions and orthography of a particular language cause that to be possible).
o Characters that are unassigned (have no character assignment at all) in the version of Unicode being used by the registry or application are not permitted, even on lookup. The issues involved in this decision are discussed in Section 7.7.
Klensin Informational [Page 25]
RFC 5894 IDNA Rationale August 2010
o Any character that is mapped to another character by a current version of NFKC is prohibited as input to IDNA (for either registration or lookup). With a few exceptions, this principle excludes any character mapped to another by Nameprep [RFC3491].
The principles above drive the design of rules that are specified exactly in the Tables document. Those rules identify the characters that are valid under IDNA. The rules themselves are normative, and the tables are derived from them, rather than vice versa.
Any label registered in a DNS zone must be validated -- i.e., the criteria for that label must be met -- in order for applications to work as intended. This principle is not new. For example, since the DNS was first deployed, zone administrators have been expected to verify that names meet "hostname" requirements [RFC0952] where those requirements are imposed by the expected applications. Other applications contexts, such as the later addition of special service location formats [RFC2782] imposed new requirements on zone administrators. For zones that will contain IDNs, support for Unicode version-independence requires restrictions on all strings placed in the zone. In particular, for such zones (the exact rules appear in Section 4 of the Protocol document [RFC5891]):
o Any label that appears to be an A-label, i.e., any label that starts in "xn--", must be valid under IDNA, i.e., they must be valid A-labels, as discussed in Section 2 above.
o The Unicode tables (i.e., tables of code points, character classes, and properties) and IDNA tables (i.e., tables of contextual rules such as those that appear in the Tables document), must be consistent on the systems performing or validating labels to be registered. Note that this does not require that tables reflect the latest version of Unicode, only that all tables used on a given system are consistent with each other.
Under this model, registry tables will need to be updated (both the Unicode-associated tables and the tables of permitted IDN characters) to enable a new script or other set of new characters. The registry will not be affected by newer versions of Unicode, or newly authorized characters, until and unless it wishes to support them. The zone administrator is responsible for verifying validity for IDNA as well as its local policies -- a more extensive set of checks than are required for looking up the labels. Systems looking up or
Klensin Informational [Page 26]
RFC 5894 IDNA Rationale August 2010
resolving DNS labels, especially IDN DNS labels, must be able to assume that applicable registration rules were followed for names entered into the DNS.
Any application processing a label through IDNA so it can be looked up in a DNS zone is required to (the exact rules appear in Section 5 of the Protocol document [RFC5891]):
o Maintain IDNA and Unicode tables that are consistent with regard to versions, i.e., unless the application actually executes the classification rules in the Tables document [RFC5892], its IDNA tables must be derived from the version of Unicode that is supported more generally on the system. As with registration, the tables need not reflect the latest version of Unicode, but they must be consistent.
o Validate the characters in labels to be looked up only to the extent of determining that the U-label does not contain "DISALLOWED" code points or code points that are unassigned in its version of Unicode.
o Validate the label itself for conformance with a small number of whole-label rules. In particular, it must verify that:
* there are no leading combining marks,
* the Bidi conditions are met if right-to-left characters appear,
* any required contextual rules are available, and
* any contextual rules that are associated with joiner characters (and CONTEXTJ characters more generally) are tested.
o Do not reject labels based on other contextual rules about characters, including mixed-script label prohibitions. Such rules may be used to influence presentation decisions in the user interface, but not to avoid looking up domain names.
To further clarify the rules about handling characters that require contextual rules, note that one can have a context-required character (i.e., one that requires a rule), but no rule. In that case, the character is treated the same way DISALLOWED characters are treated, until and unless a rule is supplied. That state is more or less equivalent to "the idea of permitting this character is accepted in principle, but it won't be permitted in practice until consensus is reached on a safe way to use it".
Klensin Informational [Page 27]
RFC 5894 IDNA Rationale August 2010
The ability to add a rule more or less exempts these characters from the prohibition against reclassifying characters from DISALLOWED to PVALID.
And, obviously, "no rule" is different from "have a rule, but the test either succeeds or fails".
Lookup applications that follow these rules, rather than having their own criteria for rejecting lookup attempts, are not sensitive to version incompatibilities with the particular zone registry associated with the domain name except for labels containing characters recently added to Unicode.
An application or client that processes names according to this protocol and then resolves them in the DNS will be able to locate any name that is registered, as long as those registrations are valid under IDNA and its version of the IDNA tables is sufficiently up to date to interpret all of the characters in the label. Messages to users should distinguish between "label contains an unallocated code point" and other types of lookup failures. A failure on the basis of an old version of Unicode may lead the user to a desire to upgrade to a newer version, but will have no other ill effects (this is consistent with behavior in the transition to the DNS when some hosts could not yet handle some forms of names or record types).
As a consequence of the elimination of mapping, the current version of IDNA changes the interpretation of a few characters relative to its predecessors. This subsection outlines the issues and discusses possible transition strategies.
In those scripts that make case distinctions, there are a few characters for which an obvious and unique uppercase character has not historically been available to match a lowercase one, or vice versa. For those characters, the mappings used in constructing the Stringprep tables for IDNA2003, performed using the Unicode toCaseFold operation (see Section 5.18 of the Unicode Standard [Unicode52]), generate different characters or sets of characters. Those operations are not reversible and lose even more information than traditional uppercase or lowercase transformations, but are more useful than those transformations for comparison purposes. Two notable characters of this type are the German character Eszett (Sharp S, U+00DF) and the Greek Final Form Sigma (U+03C2). The former is case folded to the ASCII string "ss", the latter to a medial (lowercase) Sigma (U+03C3).
Klensin Informational [Page 28]
RFC 5894 IDNA Rationale August 2010
7.2.2. Character Changes: Zero Width Joiner and Zero Width Non-Joiner
IDNA2003 mapped both ZERO WIDTH JOINER (ZWJ, U+200D) and ZERO WIDTH NON-JOINER (ZWNJ, U+200C) to nothing, effectively dropping these characters from any label in which they appeared and treating strings containing them as identical to strings that did not. As discussed in Section 3.1.2 above, those characters are essential for writing many reasonable mnemonics for certain scripts. However, treating them as valid in IDNA2008, even with contextual restrictions, raises approximately the same problem as exists with Eszett and Final Sigma: strings that were valid under IDNA2003 have different interpretations as labels, and different A-labels, than the same strings under this newer version.
7.2.3. Character Changes and the Need for Transition
The decision to eliminate mandatory and standardized mappings, including case folding, from the IDNA2008 protocol in order to make A-labels and U-labels idempotent made these characters problematic. If they were to be disallowed, important words and mnemonics could not be written in orthographically reasonable ways. If they were to be permitted as distinct characters, there would be no information loss and registries would have more flexibility, but IDNA2003 and IDNA2008 lookups might result in different A-labels.
With the understanding that there would be incompatibility either way but a judgment that the incompatibility was not significant enough to justify a prefix change, the Working Group concluded that Eszett and Final Form Sigma should be treated as distinct and Protocol-Valid characters.
Since these characters are interpreted in different ways under the older and newer versions of IDNA, transition strategies and policies will be necessary. Some actions can reasonably be taken by applications' client programs (those that perform lookup operations or cause them to be performed), but because of the diversity of situations and uses of the DNS, much of the responsibility will need to fall on registries.
Registries, especially those maintaining zones for third parties, must decide how to introduce a new service in a way that does not create confusion or significantly weaken or invalidate existing identifiers. This is not a new problem; registries were faced with similar issues when IDNs were introduced (potentially, and especially for Latin-based scripts, in conflict with existing labels that had been rendered in ASCII characters by applying more or less standardized conventions) and when other new forms of strings have been permitted as labels.
There are several approaches to the introduction of new characters or changes in interpretation of existing characters from their mapped forms in the earlier version of IDNA. The transition issue is complicated because the forms of these labels after the ToUnicode(ToASCII()) translation in IDNA2003 not only remain valid but do not provide strong indications of what the registrant intended: a string containing "ss" could have simply been intended to be that string or could have been intended to contain an Eszett; a string containing lowercase Sigma could have been intended to contain Final Sigma (one might make heuristic guesses based on position in a string, but the long tradition of forming labels by concatenating words makes such heuristics unreliable), and strings that do not contain ZWJ or ZWNJ might have been intended to contain them. Without any preference or claim to completeness, some of these, all of which have been used by registries in the past for similar transitions, are:
1. Do not permit use of the newly available character at the registry level. This might cause lookup failures if a domain name were to be written with the expectation of the IDNA2003 mapping behavior, but would eliminate any possibility of false matches.
2. Hold a "sunrise"-like arrangement in which holders of labels containing "ss" in the Eszett case, lowercase Sigma in that case, or that might have contained ZWJ or ZWNJ in context, are given priority (and perhaps other benefits) for registering the corresponding string containing Eszett, Final Sigma, or the appropriate zero-width character respectively.
3. Adopt some sort of "variant" approach in which registrants obtain labels with both character forms.
4. Adopt a different form of "variant" approach in which registration of additional strings that would produce the same A-label if interpreted according to IDNA2003 is either not permitted at all or permitted only by the registrant who already has one of the names.
5. Ignore the issue and assume that the marketplace or other mechanisms will sort things out.
In any event, a registry (at any level of the DNS tree) that chooses to permit labels to be registered that contains these characters, or considers doing so, will have to address the relationship with existing, possibly conflicting, labels in some way, just as
Klensin Informational [Page 30]
RFC 5894 IDNA Rationale August 2010
registries that already had a considerable number of labels did when IDNs were first introduced.
As discussed at length in Section 6, IDNA2003, via Nameprep (see Section 7.5), mapped many characters into related ones. Those mappings no longer exist as requirements in IDNA2008. These specifications strongly prefer that only A-labels or U-labels be used in protocol contexts and as much as practical more generally. IDNA2008 does anticipate situations in which some mapping at the time of user input into lookup applications is appropriate and desirable. The issues are discussed in Section 6 and specific recommendations are made in the Mapping document [IDNA2008-Mapping].
The conditions that would have required a change in the IDNA ACE prefix ("xn--", used in IDNA2003) were of great concern to the community. A prefix change would have clearly been necessary if the algorithms were modified in a manner that would have created serious ambiguities during subsequent transition in registrations. This section summarizes the working group's conclusions about the conditions under which a change in the prefix would have been necessary and the implications of such a change.
An IDN prefix change would have been needed if a given string would be looked up or otherwise interpreted differently depending on the version of the protocol or tables being used. This IDNA upgrade would have required a prefix change if, and only if, one of the following four conditions were met:
1. The conversion of an A-label to Unicode (i.e., a U-label) would have yielded one string under IDNA2003 and a different string under IDNA2008.
2. In a significant number of cases, an input string that was valid under IDNA2003 and also valid under IDNA2008 would have yielded two different A-labels with the different versions. This condition is believed to be essentially equivalent to the one above except for a very small number of edge cases that were not found to justify a prefix change (see Section 7.2).
Note that if the input string was valid under one version and not valid under the other, this condition would not apply. See the first item in Section 7.4.2, below.
Klensin Informational [Page 31]
RFC 5894 IDNA Rationale August 2010
3. A fundamental change was made to the semantics of the string that would be inserted in the DNS, e.g., if a decision were made to try to include language or script information in the encoding in addition to the string itself.
4. A sufficiently large number of characters were added to Unicode so that the Punycode mechanism for block offsets would no longer reference the higher-numbered planes and blocks. This condition is unlikely even in the long term and certain not to arise in the next several years.
As a result of the principles described above, none of the following changes required a new prefix:
1. Prohibition of some characters as input to IDNA. Such a prohibition might make names that were previously registered inaccessible, but did not change those names.
2. Adjustments in IDNA tables or actions, including normalization definitions, that affected characters that were already invalid under IDNA2003.
3. Changes in the style of the IDNA definition that did not alter the actions performed by IDNA.
While it might have been possible to make a prefix change, the costs of such a change are considerable. Registries could not have converted all IDNA2003 ("xn--") registrations to a new form at the same time and synchronize that change with applications supporting lookup. Unless all existing registrations were simply to be declared invalid (and perhaps even then), systems that needed to support both labels with old prefixes and labels with new ones would be required to first process a putative label under the IDNA2008 rules and try to look it up and then, if it were not found, would be required to process the label under IDNA2003 rules and look it up again. That process would probably have significantly slowed down all processing that involved IDNs in the DNS, especially since a fully-qualified name might contain a mixture of labels that were registered with the old and new prefixes. That would have made DNS caching very difficult. In addition, looking up the same input string as two separate A-labels would have created some potential for confusion and attacks, since the labels could map to different targets and then resolve to different entries in the DNS.
Klensin Informational [Page 32]
RFC 5894 IDNA Rationale August 2010
Consequently, a prefix change should have been, and was, avoided if at all possible, even if it means accepting some IDNA2003 decisions about character distinctions as irreversible and/or giving special treatment to edge cases.
The Nameprep specification [RFC3491], a key part of IDNA2003, is a profile of Stringprep [RFC3454]. While Nameprep is a Stringprep profile specific to IDNA, Stringprep is used by a number of other protocols. Were Stringprep to have been modified by IDNA2008, those changes to improve the handling of IDNs could cause problems for non-DNS uses, most notably if they affected identification and authentication protocols. Several elements of IDNA2008 give interpretations to strings prohibited under IDNA2003 or prohibit strings that IDNA2003 permitted. Those elements include the new inclusion information in the Tables document [RFC5892], the reduction in the number of characters permitted as input for registration or lookup (Section 3), and even the changes in handling of right-to-left strings as described in the Bidi document [RFC5893]. IDNA2008 does not use Nameprep or Stringprep at all, so there are no side-effect changes to other protocols.
It is particularly important to keep IDNA processing separate from processing for various security protocols because some of the constraints that are necessary for smooth and comprehensible use of IDNs may be unwanted or undesirable in other contexts. For example, the criteria for good passwords or passphrases are very different from those for desirable IDNs: passwords should be hard to guess, while domain names should normally be easily memorable. Similarly, internationalized Small Computer System Interface (SCSI) identifiers and other protocol components are likely to have different requirements than IDNs.
One of the major differences between this specification and the original version of IDNA is that IDNA2003 permitted non-letter symbols of various sorts, including punctuation and line-drawing symbols, in the protocol. They were always discouraged in practice. In particular, both the "IESG Statement" about IDNA and all versions of the ICANN Guidelines specify that only language characters be used in labels. This specification disallows symbols entirely. There are several reasons for this, which include:
1. As discussed elsewhere, the original IDNA specification assumed that as many Unicode characters as possible should be permitted, directly or via mapping to other characters, in IDNs. This
Klensin Informational [Page 33]
RFC 5894 IDNA Rationale August 2010
specification operates on an inclusion model, extrapolating from the original "hostname" rules (LDH, see the Definitions document [RFC5890]) -- which have served the Internet very well -- to a Unicode base rather than an ASCII base.
2. Symbol names are more problematic than letters because there may be no general agreement on whether a particular glyph matches a symbol; there are no uniform conventions for naming; variations such as outline, solid, and shaded forms may or may not exist; and so on. As just one example, consider a "heart" symbol as it might appear in a logo that might be read as "I love...". While the user might read such a logo as "I love..." or "I heart...", considerable knowledge of the coding distinctions made in Unicode is needed to know that there is more than one "heart" character (e.g., U+2665, U+2661, and U+2765) and how to describe it. These issues are of particular importance if strings are expected to be understood or transcribed by the listener after being read out loud.
3. Design of a screen reader used by blind Internet users who must listen to renderings of IDN domain names and possibly reproduce them on the keyboard becomes considerably more complicated when the names of characters are not obvious and intuitive to anyone familiar with the language in question.
4. As a simplified example of this, assume one wanted to use a "heart" or "star" symbol in a label. This is problematic because those names are ambiguous in the Unicode system of naming (the actual Unicode names require far more qualification). A user or would-be registrant has no way to know -- absent careful study of the code tables -- whether it is ambiguous (e.g., where there are multiple "heart" characters) or not. Conversely, the user seeing the hypothetical label doesn't know whether to read it -- try to transmit it to a colleague by voice -- as "heart", as "love", as "black heart", or as any of the other examples below.
5. The actual situation is even worse than this. There is no possible way for a normal, casual, user to tell the difference between the hearts of U+2665 and U+2765 and the stars of U+2606 and U+2729 without somehow knowing to look for a distinction. We have a white heart (U+2661) and few black hearts. Consequently, describing a label as containing a heart is hopelessly ambiguous: we can only know that it contains one of several characters that look like hearts or have "heart" in their names. In cities where "Square" is a popular part of a location name, one might well want to use a square symbol in a label as well and there are far more squares of various flavors in Unicode than there are hearts or stars.
Klensin Informational [Page 34]
RFC 5894 IDNA Rationale August 2010
The consequence of these ambiguities is that symbols are a very poor basis for reliable communication. Consistent with this conclusion, the Unicode standard recommends that strings used in identifiers not contain symbols or punctuation [Unicode-UAX31]. Of course, these difficulties with symbols do not arise with actual pictographic languages and scripts which would be treated like any other language characters; the two should not be confused.
7.7. Migration between Unicode Versions: Unassigned Code Points
In IDNA2003, labels containing unassigned code points are looked up on the assumption that, if they appear in labels and can be mapped and then resolved, the relevant standards must have changed and the registry has properly allocated only assigned values.
In the IDNA2008 protocol, strings containing unassigned code points must not be either looked up or registered. In summary, the status of an unassigned character with regard to the DISALLOWED, PROTOCOL-VALID, and CONTEXTUAL RULE REQUIRED categories cannot be evaluated until a character is actually assigned and known. There are several reasons for this, with the most important ones being:
o Tests involving the context of characters (e.g., some characters being permitted only adjacent to others of specific types) and integrity tests on complete labels are needed. Unassigned code points cannot be permitted because one cannot determine whether particular code points will require contextual rules (and what those rules should be) before characters are assigned to them and the properties of those characters fully understood.
o It cannot be known in advance, and with sufficient reliability, whether a newly assigned code point will be associated with a character that would be disallowed by the rules in the Tables document [RFC5892] (such as a compatibility character). In IDNA2003, since there is no direct dependency on NFKC (many of the entries in Stringprep's tables are based on NFKC, but IDNA2003 depends only on Stringprep), allocation of a compatibility character might produce some odd situations, but it would not be a problem. In IDNA2008, where compatibility characters are DISALLOWED unless character-specific exceptions are made, permitting strings containing unassigned characters to be looked up would violate the principle that characters in DISALLOWED are not looked up.
o The Unicode Standard specifies that an unassigned code point normalizes (and, where relevant, case folds) to itself. If the code point is later assigned to a character, and particularly if the newly assigned code point has a combining class that
Klensin Informational [Page 35]
RFC 5894 IDNA Rationale August 2010
determines its placement relative to other combining characters, it could normalize to some other code point or sequence.
It is possible to argue that the issues above are not important and that, as a consequence, it is better to retain the principle of looking up labels even if they contain unassigned characters because all of the important scripts and characters have been coded as of Unicode 5.2 (or even earlier), and hence unassigned code points will be assigned only to obscure characters or archaic scripts. Unfortunately, that does not appear to be a safe assumption for at least two reasons. First, much the same claim of completeness has been made for earlier versions of Unicode. The reality is that a script that is obscure to much of the world may still be very important to those who use it. Cultural and linguistic preservation principles make it inappropriate to declare the script of no importance in IDNs. Second, we already have counterexamples, e.g., in the relationships associated with new Han characters being added (whether in the BMP or in Unicode Plane 2).
Independent of the technical transition issues identified above, it can be observed that any addition of characters to an existing script to make it easier to use or to better accommodate particular languages may lead to transition issues. Such additions may change the preferred form for writing a particular string, changes that may be reflected, e.g., in keyboard transition modules that would necessarily be different from those for earlier versions of Unicode where the newer characters may not exist. This creates an inherent transition problem because attempts to access labels may use either the old or the new conventions, requiring registry action whether or not the older conventions were used in labels. The need to consider transition mechanisms is inherent to evolution of Unicode to better accommodate writing systems and is independent of how IDNs are represented in the DNS or how transitions among versions of those mechanisms occur. The requirement for transitions of this type is illustrated by the addition of Malayalam Chillu in Unicode 5.1.0.
The 2003 IDNA model includes several odd artifacts of the context in which it was developed. Many, if not all, of these are potential avenues for exploits, especially if the registration process permits "source" names (names that have not been processed through IDNA and Nameprep) to be registered. As one example, since the character Eszett, used in German, is mapped by IDNA2003 into the sequence "ss" rather than being retained as itself or prohibited, a string containing that character, but that is otherwise in ASCII, is not really an IDN (in the U-label sense defined above). After Nameprep maps out the Eszett, the result is an ASCII string and so it does not
Klensin Informational [Page 36]
RFC 5894 IDNA Rationale August 2010
get an xn-- prefix, but the string that can be displayed to a user appears to be an IDN. IDNA2008 eliminates this artifact. A character is either permitted as itself or it is prohibited; special cases that make sense only in a particular linguistic or cultural context can be dealt with as localization matters where appropriate.
Existing DNS servers do not know the IDNA rules for handling non-ASCII forms of IDNs, and therefore need to be shielded from them. All existing channels through which names can enter a DNS server database (for example, master files (as described in RFC 1034) and DNS update messages [RFC2136]) could not be IDNA-aware because they predate IDNA. Other sections of this document provide the needed shielding by ensuring that internationalized domain names entering DNS server databases through such channels have already been converted to their equivalent ASCII A-label forms.
Because of the distinction made between the algorithms for Registration and Lookup in Sections 4 and 5 (respectively) of the Protocol document [RFC5891] (a domain name containing only ASCII code points cannot be converted to an A-label), there cannot be more than one A-label form for any given U-label.
As specified in clarifications to the DNS specification [RFC2181], the DNS protocol explicitly allows domain labels to contain octets beyond the ASCII range (0000..007F), and this document does not change that. However, although the interpretation of octets 0080..00FF is well-defined in the DNS, many application protocols support only ASCII labels and there is no defined interpretation of these non-ASCII octets as characters and, in particular, no interpretation of case-independent matching for them (e.g., see the clarification on DNS case insensitivity [RFC4343]). If labels containing these octets are returned to applications, unpredictable behavior could result. The A-label form, which cannot contain those characters, is the only standard representation for internationalized labels in the DNS protocol.
IDNs in A-label form will generally be somewhat longer than current domain names, so the bandwidth needed by the root servers is likely to go up by a small amount. Also, queries and responses for IDNs will probably be somewhat longer than typical queries historically,
Klensin Informational [Page 37]
RFC 5894 IDNA Rationale August 2010
so Extension Mechanisms for DNS (EDNS0) [RFC2671] support may be more important (otherwise, queries and responses may be forced to go to TCP instead of UDP).
DNS labels and fully-qualified domain names provide mnemonics that assist in identifying and referring to resources on the Internet. IDNs expand the range of those mnemonics to include those based on languages and character sets other than Western European and Roman- derived ones. But domain "names" are not, in general, words in any language. The recommendations of the IETF policy on character sets and languages (BCP 18 [RFC2277]) are applicable to situations in which language identification is used to provide language-specific contexts. The DNS is, by contrast, global and international and ultimately has nothing to do with languages. Adding languages (or similar context) to IDNs generally, or to DNS matching in particular, would imply context-dependent matching in DNS, which would be a very significant change to the DNS protocol itself. It would also imply that users would need to identify the language associated with a particular label in order to look that label up. That knowledge is generally not available because many labels are not words in any language and some may be words in more than one.
This section gives an overview of IANA registries required for IDNA. The actual definitions of, and specifications for, the first two, which have been newly created for IDNA2008, appear in the Tables document [RFC5892]. This document describes the registries, but it does not specify any IANA actions.
The distinction among the major categories "UNASSIGNED", "DISALLOWED", "PROTOCOL-VALID", and "CONTEXTUAL RULE REQUIRED" is made by special categories and rules that are integral elements of the Tables document. While not normative, an IANA registry of characters and scripts and their categories, updated for each new version of Unicode and the characters it contains, are convenient for programming and validation purposes. The details of this registry are specified in the Tables document.
IANA has created and now maintains a list of approved contextual rules for characters that are defined in the IDNA Character Registry list as requiring a Contextual Rule (i.e., the types of rules described in Section 3.1.2). The details for those rules appear in the Tables document.
This registry, historically described as the "IANA Language Character Set Registry" or "IANA Script Registry" (both somewhat misleading terms), is maintained by IANA at the request of ICANN. It is used to provide a central documentation repository of the IDN policies used by top level domain (TLD) registries who volunteer to contribute to it and is used in conjunction with ICANN Guidelines for IDN use.
It is not an IETF-managed registry and, while the protocol changes specified here may call for some revisions to the tables, IDNA2008 has no direct effect on that registry and no IANA action is required as a result.
This document is purely explanatory and informational and consequently introduces no new security issues. It would, of course, be a poor idea for someone to try to implement from it; such an attempt would almost certainly lead to interoperability problems and might lead to security ones. A discussion of security issues with IDNA, including some relevant history, appears in the Definitions document [RFC5890].
The editor and contributors would like to express their thanks to those who contributed significant early (pre-working group) review comments, sometimes accompanied by text, Paul Hoffman, Simon Josefsson, and Sam Weiler. In addition, some specific ideas were incorporated from suggestions, text, or comments about sections that were unclear supplied by Vint Cerf, Frank Ellerman, Michael Everson, Asmus Freytag, Erik van der Poel, Michel Suignard, and Ken Whistler. Thanks are also due to Vint Cerf, Lisa Dusseault, Debbie Garside, and Jefsey Morfin for conversations that led to considerable improvements in the content of this document and to several others, including Ben
Klensin Informational [Page 39]
RFC 5894 IDNA Rationale August 2010
Campbell, Martin Duerst, Subramanian Moonesamy, Peter Saint-Andre, and Dan Winship, for catching specific errors and recommending corrections.
A meeting was held on 30 January 2008 to attempt to reconcile differences in perspective and terminology about this set of specifications between the design team and members of the Unicode Technical Consortium. The discussions at and subsequent to that meeting were very helpful in focusing the issues and in refining the specifications. The active participants at that meeting were (in alphabetic order, as usual) Harald Alvestrand, Vint Cerf, Tina Dam, Mark Davis, Lisa Dusseault, Patrik Faltstrom (by telephone), Cary Karp, John Klensin, Warren Kumari, Lisa Moore, Erik van der Poel, Michel Suignard, and Ken Whistler. We express our thanks to Google for support of that meeting and to the participants for their contributions.
Useful comments and text on the working group versions of the working draft were received from many participants in the IETF "IDNABIS" working group and a number of document changes resulted from mailing list discussions made by that group. Marcos Sanz provided specific analysis and suggestions that were exceptionally helpful in refining the text, as did Vint Cerf, Martin Duerst, Andrew Sullivan, and Ken Whistler. Lisa Dusseault provided extensive editorial suggestions during the spring of 2009, most of which were incorporated.
While the listed editor held the pen, the core of this document and the initial working group version represents the joint work and conclusions of an ad hoc design team consisting of the editor and, in alphabetic order, Harald Alvestrand, Tina Dam, Patrik Faltstrom, and Cary Karp. Considerable material describing mapping principles has been incorporated from a draft of the Mapping document [IDNA2008-Mapping] by Pete Resnick and Paul Hoffman. In addition, there were many specific contributions and helpful comments from those listed in the Acknowledgments section and others who have contributed to the development and use of the IDNA protocols.
[RFC3490] Faltstrom, P., Hoffman, P., and A. Costello, "Internationalizing Domain Names in Applications (IDNA)", RFC 3490, March 2003.
Klensin Informational [Page 40]
RFC 5894 IDNA Rationale August 2010
[RFC3492] Costello, A., "Punycode: A Bootstring encoding of Unicode for Internationalized Domain Names in Applications (IDNA)", RFC 3492, March 2003.
[RFC5890] Klensin, J., "Internationalized Domain Names for Applications (IDNA): Definitions and Document Framework", RFC 5890, August 2010.
[RFC5891] Klensin, J., "Internationalized Domain Names in Applications (IDNA): Protocol", RFC 5891, August 2010.
[RFC5892] Faltstrom, P., "The Unicode Code Points and Internationalized Domain Names for Applications (IDNA)", RFC 5892, August 2010.
[RFC5893] Alvestrand, H. and C. Karp, "Right-to-Left Scripts for Internationalized Domain Names for Applications (IDNA)", RFC 5893, August 2010.
[Unicode52] The Unicode Consortium. The Unicode Standard, Version 5.2.0, defined by: "The Unicode Standard, Version 5.2.0", (Mountain View, CA: The Unicode Consortium, 2009. ISBN 978-1-936213-00-9). <http://www.unicode.org/versions/Unicode5.2.0/>.
[IDNA2008-Mapping] Resnick, P. and P. Hoffman, "Mapping Characters in Internationalized Domain Names for Applications (IDNA)", Work in Progress, April 2010.
[RFC0952] Harrenstien, K., Stahl, M., and E. Feinler, "DoD Internet host table specification", RFC 952, October 1985.
[RFC1034] Mockapetris, P., "Domain names - concepts and facilities", STD 13, RFC 1034, November 1987.
[RFC1035] Mockapetris, P., "Domain names - implementation and specification", STD 13, RFC 1035, November 1987.
[RFC1123] Braden, R., "Requirements for Internet Hosts - Application and Support", STD 3, RFC 1123, October 1989.
[RFC2136] Vixie, P., Thomson, S., Rekhter, Y., and J. Bound, "Dynamic Updates in the Domain Name System (DNS UPDATE)", RFC 2136, April 1997.
Klensin Informational [Page 41]
RFC 5894 IDNA Rationale August 2010
[RFC2181] Elz, R. and R. Bush, "Clarifications to the DNS Specification", RFC 2181, July 1997.
[RFC2277] Alvestrand, H., "IETF Policy on Character Sets and Languages", BCP 18, RFC 2277, January 1998.
[RFC2671] Vixie, P., "Extension Mechanisms for DNS (EDNS0)", RFC 2671, August 1999.
[RFC2782] Gulbrandsen, A., Vixie, P., and L. Esibov, "A DNS RR for specifying the location of services (DNS SRV)", RFC 2782, February 2000.
[RFC3454] Hoffman, P. and M. Blanchet, "Preparation of Internationalized Strings ("stringprep")", RFC 3454, December 2002.
[RFC3491] Hoffman, P. and M. Blanchet, "Nameprep: A Stringprep Profile for Internationalized Domain Names (IDN)", RFC 3491, March 2003.
[RFC3743] Konishi, K., Huang, K., Qian, H., and Y. Ko, "Joint Engineering Team (JET) Guidelines for Internationalized Domain Names (IDN) Registration and Administration for Chinese, Japanese, and Korean", RFC 3743, April 2004.
[RFC3987] Duerst, M. and M. Suignard, "Internationalized Resource Identifiers (IRIs)", RFC 3987, January 2005.
[RFC4290] Klensin, J., "Suggested Practices for Registration of Internationalized Domain Names (IDN)", RFC 4290, December 2005.
[RFC4343] Eastlake, D., "Domain Name System (DNS) Case Insensitivity Clarification", RFC 4343, January 2006.
[RFC4690] Klensin, J., Faltstrom, P., Karp, C., and IAB, "Review and Recommendations for Internationalized Domain Names (IDNs)", RFC 4690, September 2006.
[RFC4713] Lee, X., Mao, W., Chen, E., Hsu, N., and J. Klensin, "Registration and Administration Recommendations for Chinese Domain Names", RFC 4713, October 2006.