Network Working Group J. Klensin
Internet-Draft
Updates: 5892, 5894 (if approved) P. Faltstrom
Intended status: Standards Track Netnod
Expires: April 11, 2018 October 8, 2017
IDNA Update for Unicode 7.0 and Later Versions
draft-klensin-idna-5892upd-unicode70-05
Abstract
The current version of the IDNA specifications anticipated that each
new version of Unicode would be reviewed to verify that no changes
had been introduced that required adjustments to the set of rules
and, in particular, whether new exceptions or backward compatibility
adjustments were needed. The review for Unicode 7.0.0 first
identified a potentially problematic new code point and then a much
more general and difficult issue with Unicode normalization. This
specification discusses those issues and proposes updates to IDNA
and, potentially, the way the IETF handles comparison of identifiers
more generally, especially when there is no associated language or
language identification. It also applies an editorial clarification
to RFC 5892 that was the subject of an earlier erratum and updates
RFC 5894 to point to the issues involved.
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on April 11, 2018.
Copyright Notice
Copyright (c) 2017 IETF Trust and the persons identified as the
document authors. All rights reserved.
Klensin & Faltstrom Expires April 11, 2018 [Page 1]
Internet-Draft IDNA Unicode Update October 2017
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(https://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include 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.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1. Origins and Discovery of the Issue . . . . . . . . . . . 4
1.2. IDNA2008 and Special or Exceptional Cases . . . . . . . . 5
1.3. Terminology . . . . . . . . . . . . . . . . . . . . . . . 7
2. Document Aspirations . . . . . . . . . . . . . . . . . . . . 8
3. Problem Description . . . . . . . . . . . . . . . . . . . . . 8
3.1. IDNA assumptions about Unicode normalization . . . . . . 8
3.2. The discovery and the Arabic script cases . . . . . . . . 10
3.2.1. New code point U+08A1, decomposition, and language
dependency . . . . . . . . . . . . . . . . . . . . . 10
3.2.2. Other examples of the same behavior within the Arabic
Script . . . . . . . . . . . . . . . . . . . . . . . 11
3.2.3. Hamza and Combining Sequences . . . . . . . . . . . . 11
3.3. Precomposed characters without decompositions more
generally . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3.1. Description of the general problem . . . . . . . . . 12
3.3.2. Latin Examples and Cases . . . . . . . . . . . . . . 14
3.3.2.1. The font exclusion and compatability
relationships . . . . . . . . . . . . . . . . . . 14
3.3.2.2. The phonetic notation characters and extensions . 14
3.3.2.3. The stroke (solidus) ambiguity . . . . . . . . . 14
3.3.2.3.1. Combining dots and other shapes combine...
unless... . . . . . . . . . . . . . . . . . . 15
3.3.2.3.2. "Legacy" characters and new additions . . . . 16
3.3.3. Unexpected Combining Sequances . . . . . . . . . . . 16
3.3.4. Examples and Cases from Other Scripts . . . . . . . . 17
3.3.4.1. Scripts with precomposed preferences and ones
with combining preferences . . . . . . . . . . . 17
3.3.4.2. The Han and Kangxu Cases . . . . . . . . . . . . 17
3.4. Confusion and the Casual User . . . . . . . . . . . . . . 17
4. Implementation options and issues: Unicode properties,
exceptions, and the nature of stability . . . . . . . . . . . 18
4.1. Unicode Stability compared to IETF (and ICANN) Stability 18
4.2. New Unicode Properties . . . . . . . . . . . . . . . . . 19
4.3. The need for exception lists . . . . . . . . . . . . . . 20
5. Proposed/ Alternative Changes to RFC 5892 for the issues
Klensin & Faltstrom Expires April 11, 2018 [Page 2]
Internet-Draft IDNA Unicode Update October 2017
first exposed by new code point U+08A1 . . . . . . . . . . . 20
5.1. Disallow This New Code Point . . . . . . . . . . . . . . 20
5.2. Disallow This New Code Point and All Future Precomposed
Additions that Do Not Decompose . . . . . . . . . . . . . 22
5.3. Disallow the combining sequences for these characters . . 22
5.4. Use Combinnig Classes to Develop Additional Contextual
Rules . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.5. Disallow all Combining Characters for Specific Scripts . 23
5.6. Do Nothing Other Than Warn . . . . . . . . . . . . . . . 24
5.7. Normalization Form IETF (NFI)) . . . . . . . . . . . . . 25
6. Editorial clarification to RFC 5892 . . . . . . . . . . . . . 26
7. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 26
8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 26
9. Security Considerations . . . . . . . . . . . . . . . . . . . 27
10. References . . . . . . . . . . . . . . . . . . . . . . . . . 28
10.1. Normative References . . . . . . . . . . . . . . . . . . 28
10.2. Informative References . . . . . . . . . . . . . . . . . 30
Appendix A. Change Log . . . . . . . . . . . . . . . . . . . . . 33
A.1. Changes from version -00 (2014-07-21)to -01 . . . . . . . 33
A.2. Changes from version -01 (2014-12-07) to -02 . . . . . . 33
A.3. Changes from version -02 (2014-12-07) to -03 . . . . . . 33
A.4. Changes from version -03 (2015-01-06) to -04 . . . . . . 33
A.5. Changes from version -04 (2015-03-11) to -05 . . . . . . 34
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 34
1. Introduction
Note in/about -04 and -05 Drafts: These two versions of the
document contains a very large amount of new material as compared
to the -03 version. The new material reflects an evolution of
community understanding in the first quarter of 2015 and further
evolution between then and mid-2017 from an assumption that the
problem involved only a few code points and one combining
character in a single script (Hamza Above and Arabic) to an
understanding that the problem we have come to call "non-
decomposing code points" and several closely related ones are
quite pervasive and may represent fundamental misunderstandings or
omissions from IDNA2008 (and, by extension, the basics of PRECIS
[RFC8264]) that must be corrected if those protocols are going to
be used in a way that supports internationalized identifiers on
the Internet predictably (as seen by the end user) and securely.
This version is still necessarily incomplete: not only is our
understanding probably still not comprehensive, but there are a
number of placeholders for text and references. Nonetheless, the
document in its current form should be useful as both the
beginning of a comprehensive overview is the issues and a source
of references to other relevant materials.
Klensin & Faltstrom Expires April 11, 2018 [Page 3]
Internet-Draft IDNA Unicode Update October 2017
This draft could almost certainly be better organized to improve
its readability: specific suggestions would be welcome.
1.1. Origins and Discovery of the Issue
The current version of the IDNA specifications, known as "IDNA2008"
[RFC5890], anticipated that each new version of Unicode would be
reviewed to verify that no changes had been introduced that required
adjustments to IDNA's rules and, in particular, whether new
exceptions or backward compatibility adjustments were needed. When
that review was carefully conducted for Unicode 7.0.0 [Unicode7],
comparing it to prior versions including the text in Unicode 6.2
[Unicode62], it identified a problematic new code point (U+08A1,
ARABIC LETTER BEH WITH HAMZA ABOVE). The code point was added for
Arabic Script use with the Fula (also known as Fulfulde, Pulaar, amd
Pular'Fulaare) language. That language is apparently most often
written in Latin characters today [Omniglot-Fula] [Dalby] [Daniels].
The specific problem is discussed in detail in Section 3. In very
broad terms, IDNA (and other IETF work) assume that, if one can
represent "the same character" either as a combining sequence or as a
single code point, strings that are identical except for those
alternate forms will compare equal after normalization. Part of the
difficulty that has characterized this discussion is that "the same"
differs depending on the criteria that are chosen. It may be further
complicated in practice by differences in preferred type styles or
rendering, but Unicode code point choices are not supposed to depend
on type style (font) variations and, again, IDNA has no mechanism for
specifying language choices that might affect rendering.
The behavior of the newly-added code point, while non-optimal for
IDNA, follows that of a few code points that predate Unicode 7.x and
even the IDNA 2008 specifications and Unicode 6.0. Those existing
code points, which may not be easy to accurately characterize as a
group, make the question of what, if anything, to do about this new
exceedingly problematic one and, perhaps separately, what to do about
existing sets of code points with the same behavior, because
different reasonable criteria yield different decisions,
specifically:
o To disallow it (and future, but not existing, characters with
similar characteristics) as an IDNA exception case creates
inconsistencies with how those earlier code points were handled.
o To disallow it and the similar code points as well would
necessitate invalidating some potential labels that would have
been valid under IDNA2008 until this time. Depending on how the
Klensin & Faltstrom Expires April 11, 2018 [Page 4]
Internet-Draft IDNA Unicode Update October 2017
collection of similar code points is characterized, a few of them
are almost certainly used in reasonable labels.
o To permit the new code point to be treated as PVALID creates a
situation in which it is possible, within the same script, to
compose the same character symbol (glyph or grapheme) in two
different ways that do not compare equal even after normalization.
That condition would then apply to it and the earlier code points
with the same behavior. That situation contradicts a fundamental
assumption of IDNA that is discussed in more detail below.
NOTE IN DRAFT:
This working draft discusses six alternatives, including an idea
(an IETF-specific normalization form) that seemed too drastic to
be considered when IDNA2008 was designed or even when the review
of Unicode 7.0 for IDAN purposes began. In retrospect, it not
only would have been appropriate to discuss when the IDNA2008
specifications were being developed but is appearing more
attractive now. The authors suggest that the community discuss
the relevant tradeoffs and make a decision and that the document
then be revised to reflect that decision, with the other
alternatives discussed as options not chosen. Because there is no
ideal choice, the discussion of the issues in Section 3 is
probably as or more important than the particular choice of how to
handle this code point. In addition to providing information for
this document, that section should be considered as an updating
addendum to RFC 5894 [RFC5894] and should be incorporated into any
future revision of that document.
As the result of this version of the document containing several
alternate proposals, some of the text is also a little bit
redundant. That will be corrected in future versions.
1.2. IDNA2008 and Special or Exceptional Cases
IDNA2008 contains several type of explicit provisions for characters
(code points) that require special treatment when the requirements of
the DNS cannot easily be met by calculations based on stable Unicode
properties. Those provisions are
[[CREF1: ... to be supplied]]
As anticipated when IDNA2008, and RFC 5892 in particular, were
written, exceptions and explicit updates are likely to be needed only
if there is disagreement between the Unicode Consortium's view about
what is best for the Standard and its very diverse user community and
the IETF's view of what is best for IDNs, the DNS, and IDNA. It was
hoped that a situation would never arise in which the the two
Klensin & Faltstrom Expires April 11, 2018 [Page 5]
Internet-Draft IDNA Unicode Update October 2017
perspectives would disagree, but the possibility was anticipated and
considerable mechanism added to RFC 5890 and 5982 as a result. It is
probably important to note that a disagreement in this context does
not imply that anyone is "wrong", only that the two different groups
have different needs and therefore criteria about what is acceptable.
In particular, it appears that the Unicode Consortium has made
assumptions about the availability (by explicit designation or
context) of information about applicable languages or other context
for a give string that are not possible for IDNA. For that reason,
the IETF has, in the past, allowed some characters for IDNA that
active Unicode Technical Committee members suggested be disallowed to
avoid a change in derived tables [RFC6452]. This document describes
a set of cases for which the IETF must consider disallowing sets of
characters that the various properties would otherwise treat as
PVALID.
This document provides the "flagging for the IESG" specified by
Section 5.1 of RFC 5892. As specified there, the change itself
requires IETF review because it alters the rules of Section 2 of that
document.
[[RFC Editor: please remove the following comment and note if they
get to you.]]
[[IESG: It might not be a bad idea to incorporate some version of
the following into the Last Call announcement.]]
NOTE IN DRAFT to IETF Reviewers: The issues in this document, and
particularly the choices among options for either adding exception
cases to RFC 5892 or ignoring the issue, warning people, and
hoping the results do not include or enable serious problems, are
fairly esoteric. Understanding them requires that one have at
least some understanding of how scripts in which precomposed
characters are preferred over combining sequences as a Unicode
design and extension principle work. Those scripts include Arabic
but, unlike the assumption when the issues were first discovered,
are by no means limited to it. Readers should also understand the
reasons the Unicode Standard gives various Arabic Script
characters a fairly extended discussion [Unicode70-Arabic] but
should treat that only as an example and note that most other
cases are much less well documented. It also requires
understanding of a number of Unicode principles, including the
Normalization Stability rules [UAX15-Versioning] as applied to new
precomposed characters and guidelines for adding new characters.
There is considerable discussion of the issues in Section 3 and
references are provided for those who want to pursue them, but
potential reviewers should assume that the background needed to
understand the reasons for this change is no less deep in the
Klensin & Faltstrom Expires April 11, 2018 [Page 6]
Internet-Draft IDNA Unicode Update October 2017
subject matter than would be expected of someone reviewing a
proposed change in, e.g., the fundamentals of BGP, TCP congestion
control, or some cryptographic algorithm. Put more bluntly, one's
ability to read or speak languages other than English, or even one
or more languages that use the Arabic script or other scripts
similarly affected, does not make one an expert in these matters.
1.3. Terminology
This document assumes that the reader is reasonably familiar with the
terminology of IDNA [RFC5890] and Unicode [Unicode7] and with the
IETF conventions for representing Unicode code points [RFC5137].
Some terms used here may not be used in the same way in those two
sets of documents. From one point of view, those differences may
have been the results of, or led to, misunderstandings that may, in
turn, be part of the root cause of the problems explored in this
document. In particular, this document uses the term "precomposed
character" to describe characters that could reasonably be composed
by a combining sequence using code points with appropriate appearance
in common type styles but for which a single code point that does not
require combining sequences is available. That definition is
strictly about mechanical composition and does not involve any
considerations about how the character is used. It is closely
related to this document's definition of "identical". When a
precomposed character exists and either applying NFC to the combining
sequence does not yield that character or applying NFD to that
character's code point does not yield the combining sequence, it is
referred to in this document as "non-decomposable".
The document also uses some terms that are familiar to those who have
been involved with IDNs and IDNA for a long time, but uses them more
precisely than may be common in other quarters. For example, the
term "Punycode" is not used at all in the rest of this document
because it is the name of a very specific encoding algorithm
[RFC3492] that does not incorporate the rules and algorithms for
domain name labels that are produced by that encoding. Instead, the
generic terms "ACE" or "ACE string" for "ASCII-compatible encoding"
is used to refer to strings that abstractly contain characters
outside the ASCII repertoire [RFC0020] but are encoded so that only
ASCII characters appear in the string that would be encountered by a
user or protocol and the terms "A-label" and "U-label", as defined in
RFC 5890, to refer to the ACE and more conventional (or "native")
character forms in which those non-ASCII characters appear in
conventional Unicode encodings (typically UTF-8).
Klensin & Faltstrom Expires April 11, 2018 [Page 7]
Internet-Draft IDNA Unicode Update October 2017
2. Document Aspirations
This document, in its present form, is not a proposal for a solution.
Instead, it is intended to be (or evolve into) a comprehensive
description of the issues and problems and to outline some possible
approaches to a solution. A perfect solution -- one that would
resolve all of the issues identified in this document -- would
involve a relatively small set of relatively simple rules and hence
would be comprehensible and predictable for and by non-expert end
users, would not require code point by code point or even block by
block exception lists, and would not leave uses of any script or
language feeling that their particular writing system have been
treated less fairly than others.
Part of the reality we need to accept is that IDNA, in its present
form, represents compromises that does not completely satisfy those
criteria and whatever is done about these issues will probably make
it (or the job of administering zones containing IDNs) more complex.
Similarly, as the Unicode Standard suggests when it identifies ten
Design Principles and the text then says "Not all of these principles
can be satisfied simultaneously..." [Unicode70-Design], while there
are guidelines and principles, a certain amount of subjective
judgment is involved in making determinations about normalization,
decomposition, and some property values. For Unicode itself, those
issues are resolved by multiple statements (at least one cited below)
that one needs to rely on per-code point information in the Unicode
Character Database rather than on rules or principles. The design of
IDNA and the effort to keep it largely independent of Unicode
versions requires rules, categories, and principles that can be
relied upon and applied algorithmically. There is obviously some
tension between the two approaches.
3. Problem Description
3.1. IDNA assumptions about Unicode normalization
IDNA makes several assumptions about Unicode, Unicode "characters",
and the effects of normalization. Those assumptions were based on
careful reading of the Unicode Standard at the time [Unicode5],
guided by advice and commitments by members of the Unicode Technical
Committee. Those assumptions, and the associated requirements, are
necessitated by three properties of DNS labels that typically do not
apply to blocks of running text:
1. There is no language context for a label. While particular DNS
zones may impose restrictions, including language or script
restrictions, on what labels can be registered, neither the DNS
nor IDNA impose either type of restriction or give the user of a
Klensin & Faltstrom Expires April 11, 2018 [Page 8]
Internet-Draft IDNA Unicode Update October 2017
label any indication about the registration or other restrictions
that may have been imposed.
2. Labels are often mnemonics rather than words in any language.
They may be abbreviations or acronyms or contain embedded digits
and have other characteristics that are not typical of words.
3. Labels are, in practice, usually short. Even when they are the
maximum length allowed by the DNS and IDNA, they are typically
too short to provide significant context. Statements that
suggest that languages can almost always be determined from
relatively short paragraphs or equivalent bodies of text do not
apply to DNS labels because of their typical short length and
because, as noted above, they are not required to be formed
according to language-based rules.
At the same time, because the DNS is an exact-match system, there
must be no ambiguity about whether two labels are equal. Although
there have been extensive discussions about "confusingly similar"
characters, labels, and strings, such tests between scripts are
always somewhat subjective: they are affected by choices of type
styles and by what the user expects to see. In spite of the fact
that the glyphs that represent many characters in different scripts
are identical in appearance (e.g., basic Latin "a" (U+0061) and the
identical-appearing Cyrillic character (U+0430), the most important
test is that, if two glyphs are the same within a given script, they
must represent the same character no matter how they are formed.
Unicode normalization, as explained in [UAX15], is expected to
resolve those "same script, same glyph, different formation methods"
issues. Within the Latin script, the code point sequence for lower
case "o" (U+006F) and combining diaeresis (U+0308) will, when
normalized using the "NFC" method required by IDNA, produce the
precomposed small letter o with diaeresis (U+00F6) and hence the two
ways of forming the character will compare equal (and the combining
sequence is effectively prohibited from U-labels).
NFC was preferred over other normalization methods for IDNA because
it is more compact, more likely to be produced on keyboards on which
the relevant characters actually appeared, and because it does not
lose substantive information (e.g., some types of compatibility
equivalence involves judgment calls as to whether two characters are
actually the same -- they may be "the same" in some contexts but not
others -- while canonical equivalence is about different ways to
produce the glyph for the same abstract character).
IDNA also assumed that the extensive Unicode stability rules would be
applied and work as specified when new code points were added. Those
Klensin & Faltstrom Expires April 11, 2018 [Page 9]
Internet-Draft IDNA Unicode Update October 2017
rules, as described in The Unicode Standard and the normative annexes
identified below, provide that:
1. New code points representing precomposed characters that can be
formed from combining sequences will not be added to Unicode
unless neither the relevant base character nor required combining
character(s) are part of the Standard within the relevant script
[UAX15-Versioning].
2. If circumstances require that principle be violated,
normalization stability requires that the newly-added character
decompose (even under NFC) to the previously-available combining
sequence [UAX15-Exclusion].
At least at the time IDNA2008 was being developed, there was no
explicit provision in the Standard's discussion of conditions for
adding new code points, nor of normalization stability, for an
exception based on different languages using the same script or
ambiguities about the shape or positioning of combining characters.
3.2. The discovery and the Arabic script cases
While the set of problems with normalization discussed above were
discovered with a newly-added code point for the Arabic Script and
some characteristics of Unicode handling of that script seem to make
the problem more complex going forward, these are not issues specific
to Arabic. This section describes the Arabic-specific problems;
subsequent ones (starting with Section 3.3) discuss the problem more
generally and include illustrations from other scripts.
3.2.1. New code point U+08A1, decomposition, and language dependency
Unicode 7.0.0 introduces the new code point U+08A1, ARABIC LETTER BEH
WITH HAMZA ABOVE. As can be deduced from the name, it is visually
identical to the glyph that can be formed from a combining sequence
consisting of the code point for ARABIC LETTER BEH (U+0628) and the
code point for Combining Hamza Above (U+0654). The two rules
summarized above (see the last part of Section 3.1) suggest that
either the new code point should not be allocated at all or that it
should have a decomposition to \u'0628'\u'0654'.
Had the issues outlined in this document been better understood at
the time, it probably would have been wise for RFC 5892 to disallow
either the precomposed character or the combining sequence of each
pair in those cases in which Unicode normalization rules do not cause
the right thing to happen, i.e., the combining sequence and
precomposed character to be treated as equivalent. Failure to do so
at the time places an extra burden on registries to be sure that
Klensin & Faltstrom Expires April 11, 2018 [Page 10]
Internet-Draft IDNA Unicode Update October 2017
conflicts (and the potential for confusion and attacks) do not exist.
Oddly, had the exclusion been made part of the specification at that
time, the preference for precomposed forms noted above would probably
have dictated excluding the combining sequence, something not
otherwise done in IDNA2008 because the NFC requirement serves the
same purpose. Today, the only thing that can be excluded without the
potential disruption of disallowing a previously-PVALID combining
sequence is the to exclude the newly-added code point so whatever is
done, or might have been contemplated with hindsight, will be
somewhat inconsistent.
3.2.2. Other examples of the same behavior within the Arabic Script
One of the things that complicates the issue with the new U+08A1 code
point is that there are several other Arabic-script code points that
behave in the same way for similar language-specific reasons.
In particular, at least three other grapheme clusters that have been
present for many version of Unicode can be seen as involving issues
similar to those for the newly-added ARABIC LETTER BEH WITH HAMZA
ABOVE. ARABIC LETTER HAH WITH HAMZA ABOVE (U+0681) and ARABIC LETTER
REH WITH HAMZA ABOVE (U+076C) do not have decomposition forms and are
preferred over combining sequences using HAMZA ABOVE (U+0654)
[Unicode70-Hamza]. By contrast, ARABIC LETTER ALEF WITH HAMZA ABOVE
(U+0623) decomposes into \u'0627'\u'0654', ARABIC LETTER WAW WITH
HAMZA ABOVE (U+0624) decomposes into \u'0648'\u'0654', and ARABIC
LETTER YEH WITH HAMZA ABOVE (U+0626) decomposes into \u'064A'\u'0654'
so the precomposed character and combining sequences compare equal
when both are normalized, as this specification prefers.
There are other variations in which a precomposed character involving
HAMZA ABOVE has a decomposition to a combining sequence that can form
it. For example, ARABIC LETTER U WITH HAMZA ABOVE (U+0677) has a
compatibility decomposition. but not a canonical one, into the
combining sequence \u'06C7'\u'0674'.
3.2.3. Hamza and Combining Sequences
As the Unicode Standard points out at some length [Unicode70-Arabic],
Hamza is a problematic abstract character and the "Hamza Above"
construction even more so [Unicode70-Hamza]. Those sections explain
a distinction made by Unicode between the use of a Hamza mark to
denote a glottal stop and one used as a diacritic mark to denote a
separate letter. In the first case, the combining sequence is used.
In the second, a precomposed character is assigned.
Unlike Unicode generally and because of concerns about identifier
spoofing and attacks based on similarities, character distinctions in
Klensin & Faltstrom Expires April 11, 2018 [Page 11]
Internet-Draft IDNA Unicode Update October 2017
IDNA are based much more strictly on the appearance of characters;
language and pronunciation distinctions within a script are not
considered. So, for IDNA, BEH WITH HAMZA ABOVE is not-quite-
tautologically the same as BEH WITH HAMZA ABOVE, even if one of them
is written as U+08A1 (new to Unicode 7.0.0) and the other as the
sequence \u'0628'\u'0654' (feasible with Unicode 7.0.0 but also
available in versions of Unicode going back at least to the version
[Unicode32] used in the original version of IDNA [RFC3490]. Because
the precomposed form and combining sequence are, for IDNA purposes,
the same, IDNA expects that normalization (specifically the
requirement that all U-labels be in NFC form) will cause them to
compare equal.
If Unicode also considered them the same, then the principle would
apply that new precomposed ("composition") forms are not added unless
one of the code points that could be used to construct it did not
exist in an earlier version (and even then is discouraged)
[UAX15-Versioning]. When exceptions are made, they are expected to
conform to the rules and classes in the "Composition Exclusion
Table", with class 2 being relevant to this case [UAX15-Exclusion].
That rule essentially requires that the normalization for the old
combining sequence to itself be retained (for stability) but that the
newly-added character be treated as canonically decomposable and
decompose back to the older sequence even under NFC. That was not
done for this particular case, presumably because of the distinction
about pronunciation modifiers versus separate letters noted above.
Because, for IDNA and the DNS, there is a possibility that the
composing sequence \u'0628'\u'0654' already appears in labels, the
only choice other than allowing an otherwise-identical, and
identically-appearing, label with U+08A1 substituted to identify a
different DNS entry is to DISALLOW the new character.
3.3. Precomposed characters without decompositions more generally
3.3.1. Description of the general problem
As mentioned above, IDNA made a strong assumption that, if there were
two ways to form the same abstract character in the same script,
normalization would result in them comparing equal. Work on IDNA2008
recognized that early version of Unicode might also contain some
inconsistencies; see Section 3.3.2.3.2 below.
Having precomposed code points exist that don't have decompositions,
or having code points of that nature allocated in the future, is
problematic for those IDNA assumptions about character comparison.
It seems to call for either excluding some set of code points that
IDNA's rules do not now identify, development and use of a
normalization procedure that behaves as expected (those two options
Klensin & Faltstrom Expires April 11, 2018 [Page 12]
Internet-Draft IDNA Unicode Update October 2017
may be nearly equivalent for many purposes), or deciding to accept a
risk that, apparently, will only increase over time.
It is not clear whether the reasons the IDNABIS WG did not understand
and allow for these cases are important except insofar as they inform
considerations about what to do in the future. It seemed (and still
seems to some people) that the Unicode Standard is very clear on the
matter (or at least was when IDNA2008 was being developed). In
addition to the normalization stability rules cited in the last part
of Section 3.1. the discussion in the Core Standard seems quite
clear. For example, "Where characters are used in different ways in
different languages, the relevant properties are normally defined
outside the Unicode Standard" in Section 2.2, subsection titled
"Semantics" [Unicode7] did not suggest to most readers that sometimes
separate code points would be allocated within a script based on
language considerations. Similarly, the same section of the Standard
says, in a subsection titled "Unification", "The Unicode Standard
avoids duplicate encoding of characters by unifying them within
scripts across language" and does not list exceptions to that rule or
limit it to a single script although it goes on to list "CJK" as an
example. Another subsection, "Equivalent Sequences" indicates
"Common precomposed forms ... are included for compatibility with
current standards. For static precomposed forms, the standard
provides a mapping to an equivalent dynamically composed sequence of
characters". The latter appears to be precisely the "all precomposed
characters decompose into the relevant combining sequences if the
relevant base and combining characters exist in the Standard" rule
that IDNA needs and assumed and, again, there is no mention of
exceptions, language-dependent of otherwise. The summary of
stability policies cited in the Standard [Unicode70-Stability] does
not appear to shed any additional light on these issues.
The Standard now contains a subsection titled "Non-decomposition of
Overlaid Diacritics" [Unicode70-Overlay] that identifies a list of
diacritics that do not normally form characters that have
decompositions. The rule given has its own exceptions and the text
clearly states that there is actually no way to know whether a code
point has a decomposition other than consulting the Unicode Character
Database entry for that code point. The subsequent section notes
that this can be a security problem. While the issues with IDNA go
well beyond what is normally considered security, that comment now
seems clear. While that subsection is helpful in explaining the
problem, especially for European scripts, it does not appear in the
Unicode versions that were current when IDNA2008 was being developed.
Klensin & Faltstrom Expires April 11, 2018 [Page 13]
Internet-Draft IDNA Unicode Update October 2017
3.3.2. Latin Examples and Cases
While this set of problems was discovered because of a code point
added to the Arabic script in precombined form to support a
particular language, there are actually far more examples for, e.g.,
Latin script than there are for Arabic script. Many of them are
associated with the "non-decomposition of combining diacriticals"
issues mentioned above, but the next subsections describe other cases
that are not directly bound to decomposition.
3.3.2.1. The font exclusion and compatability relationships
Unicode contains a large collection of characters that are identified
as "Mathematical Symbols". A large subset of them are basic or
decorated Latin characters, differing from the ordinary ones only by
their usage and, in appearance, by font or type styling (despite the
general principle that font distinctions are not used as the basis
for assigning separate code points. Most of these have canonical
mappings to the base form, which eliminates them from IDNA, but
others do not and, because the same marks that are used as phonetic
diacritical markings in conventional alphabetical use have special
mathematical meanings, applications that permit the use of these
characters have their own issues with normalization and equality.
3.3.2.2. The phonetic notation characters and extensions
Another example involves various Phonetic Alphabet and Extension
characters. many of which, unlike the Mathematical ones, do not have
normalizations that would make them compare equal to the basic
characters with essentially identical representations. This would
not be a problem for IDNA if they were identified with a specialized
script or as symbols rather than letters, but neither is the case:
they are generally identified as lower case Latin Script letters even
when they are visually upper-case, another issue for IDNA.
3.3.2.3. The stroke (solidus) ambiguity
Some combining characters have two or more forms. for example, in
the case of the character popularly known as "slash", "stroke", or
"solidus" (sometime prefixed by "forward"), there are "short" and
"long" combining forms, U+0337 (COMBINING SHORT SOLIDUS OVERLAY) and
U+0338 (COMBINING LONG SOLIDUS OVERLAY). It is not clear how long a
short one needs to be to make it "long" or how short a long one needs
to be to make it "short". Perhaps for that reason, U+00F8 has no
decomposition and neither U+006F U+0337 nor U+006F U+0338 combine to
it with NFC.
Klensin & Faltstrom Expires April 11, 2018 [Page 14]
Internet-Draft IDNA Unicode Update October 2017
Adding to the confusion, at least when one attempts to use Unicode
character names to identify places to look for problems, U+00F8 is
formally called LATIN SMALL LETTER O WITH STROKE but, in combining
character terminology, the term "stroke" refers to a horizontal bar,
not an angled one, as in U+0335 and U+0336 (also short and long
versions). However, when one overlays one of those on an "o"
(U+006F), one gets U+0275, LATIN SMALL LETTER BARRED O, not "...o
with stroke". That character, by the way, does not decompose either.
This does illustrate the principle that it is not feasible to rely on
Unicode code point names to identify confusable character sequences,
even ones that produce the same, more or less font-independent,
grapheme clusters.
3.3.2.3.1. Combining dots and other shapes combine... unless...
The discussion of "Non-decomposition of Overlaid Diacritics"
[Unicode70-Overlay] indirectly exhibits at least one reason why it
has been difficult to characterize the problem. If one combines that
subsection with others, one gets a set of rules that might be
described as:
1. If the precomposed character and the code points that make up the
combining sequence exist, then canonical composition and
decomposition work as expected, except...
2. If the precomposed character was added to Unicode after the code
points that make up the combining sequence, normalization
stability for the combining sequences requires that NFC applied
to the precomposed character decomposes rather than having the
combining sequence compose to the new character, however...
3. If the combining sequence involves a diacritic or other mark that
actually touches the base character when composed, the
precomposed character does not have a decomposition, unless...
4. The combining diacritic involved is Cedilla (U+0327), Ogonek
(U+0328), or Horn (U+031B), in which case the precomposed
characters that contain them "regularly" (but presumably not
always) decomposes, and...
5. There are further exceptions for Hamza which does not overlay the
associated base character in the same way the Latin-derived
combining diacritics and other marks do. Those decisions to
decompose a precomposed character (or not) are based on language
or phonetic considerations, not the combining mechanism or
appearance, or perhaps,...
Klensin & Faltstrom Expires April 11, 2018 [Page 15]
Internet-Draft IDNA Unicode Update October 2017
6. Some characters have compatibility decompositions rather than
canonical ones [Unicode70-CompatDecomp]. Because compatibility
relationships are treated differently by IDNA, PRECIS [RFC8264],
and, potentially, other protocols involving identifiers for
Internet use, the existence of compatibility relationship may or
may not be helpful. Finally,...
7. There is no reason to believe the above list is complete. In
particular, if whether a precomposed character decomposes or not
is determined by language or phonetic distinctions or by a
decision that all new characters for some scripts will be
precomposed while new ones for others will be added (if needed)
as combining sequences, one may need additional rules on a per-
script and/or per-character basis.
The above list only covers the cases involving combining sequences.
It does not cover cases such as those in Section 3.3.2.1 and
Section 3.3.2.2 and there may be additional groups of cases not yet
identified.
3.3.2.3.2. "Legacy" characters and new additions
The development of categories and rules for IDNA recognized that
early version of Unicode might contain some inconsistencies if
evaluated using more contemporary rules about code point assignments
and stability. In particular, there might be some exceptions from
different practices in early version of Unicode or anomalies caused
by copying existing single- or dual-script standards into Unicode as
block rather than individual character additions to the repertoire.
The possibility of such "legacy" exceptions was one reason why the
IDNA category rules include explicit provisions for exception lists
(even though no such code points were identified prior to 2014).
3.3.3. Unexpected Combining Sequances
Most combining characters have the script property "Inherited" or
"Common", i.e., are not members of any particular script and will not
cause rules against mixed-script labels to be triggered.
Normalization rules are generally structured around the base
character, so unexpected combinations of base characters with
combining ones may lead to cases where normalization might normally
be expected to produce a precombined character but does not do so (in
the most common situation because no such precombined character
exists. For example, the Latin script characters "a" and "a with
acute accent" are both coded (as U+0061 and U+00E1). If the latter
is coded as the combining sequence U+0061 U+0301, NFC will turn that
sequence into U+00E1 and everything will work as users expect.
However, the Cyrillic "a" character (U+0430) is notoriously similar
Klensin & Faltstrom Expires April 11, 2018 [Page 16]
Internet-Draft IDNA Unicode Update October 2017
in appearance in most type styles to U+0061 and the U+0439 U+0301 and
that sequence does not normalize to anything else. Because thre is
no code point assigned for Cyrillic small letter a with acute accent
and unlike many of the other examples in this document, that is
Unicode working exactly as would be expected. Whether it is an issue
or not depends on the questions that are being asked and what rules
are being applied.
3.3.4. Examples and Cases from Other Scripts
Research into these issues has not yet turned up a comprehensive list
of affected scripts and code points. As discussed elsewhere in this
document, it is clear that Arabic and Latin Scripts are significantly
affected, that some Han and Kangxu radicals and ideographs are
affected, and that other examples do exist -- it is just not known
how many of those examples there are and what patterns, if any,
characterize them.
3.3.4.1. Scripts with precomposed preferences and ones with combining
preferences
While the authors have been unable to find an explanation for the
differentiation in the Unicode Standard, we have been told that there
are differences among scripts as to whether the action preference is
to add new combining sequences only (and resist adding precomposed
characters) as suggested in Section 3.3.2.3.1 or to add precomposed
characters, often ones that do not have decompositions. If those
difference in preference do exist, it is probably important to have
them documented so that they can be reflected in IDNA review
procedures and elsewhere. It will also require IETF discussion of
whether combining sequences should be deprecated when the
corresponding precomposed characters are added or to disallow
combining sequences entirely for those scripts (as has been
implicitly suggested for Arabic language use [RFC5564]).
[[CREF2: The above isn't quite right and probably needs additional
discussion and text.]]
3.3.4.2. The Han and Kangxu Cases
[[CREF3: .. to be supplied .. ]]
3.4. Confusion and the Casual User
To the extent to which predictability for relatively casual users is
a desired and important feather of relevant application or
application support protocols, it is probably worth observing that
the complex of rules and cases suggested or implied above is almost
Klensin & Faltstrom Expires April 11, 2018 [Page 17]
Internet-Draft IDNA Unicode Update October 2017
certainly too involved for the typical such user to develop a good
intuitive understanding of how things behave and what relationships
exist. Conversely, the nature of writing systems for natural
languages, especially those that have evolved and diverged over
centuries, implies that no set of rules about allowable characters
will guarantee complete safety (however that is defined).
4. Implementation options and issues: Unicode properties, exceptions,
and the nature of stability
4.1. Unicode Stability compared to IETF (and ICANN) Stability
The various stability rules in Unicode [Unicode70-Stability] all
appear to be based on the model that once a value is assigned, it can
never be changed. That is probably appropriate for a character
coding system with multiple uses and applications. It is probably
the only option when normative relationships are expressed in tables
of values rather than by rules. One consequence of such a model is
that it is difficult or impossible to fix mistakes (for some
stability rules, the Unicode Standard does provide for exceptions)
and even harder to make adjustments that would normally be dictated
by evolution.
"No changes" provides a very strong and predictable type of
stability. There are many reasons to take that path. As in some of
the cases that motivated this document, the difficulty is that simply
adding new code points (in Unicode) or features (in a protocol or
application) may be destabilizing. One then has complete stability
for systems that never use or allow the new code points or features,
but rough edges for newer systems that see the discrepancies and
rough edges. IDNA2003 (inadvertently) took that approach by freezing
on Unicode 3.2 -- if no code points added after Unicode 3.2 had ever
been allowed, we would have had complete stability even as Unicode
libraries changed. Unicode has been quite ingenious about working
around those difficulties with such provisions as having code points
for newly-added precomposed characters decompose rather than altering
the normalization for the combining sequences. Other cases, such as
newly-added precomposed characters that do not decompose for, e.g.,
language or phonetic reasons, are more problematic.
The IETF (and ICANN and standards development bodies such as ISO and
ISO/IEC JTC1) have generally adopted a different type of stability
model, one which considers experience in use and the ill effects of
not making changes as well as the disruptive effects of doing so. In
the IETF model, if an earlier decision is causing sufficient harm and
there is consensus in the communities that are most affected that a
change is desirable enough to make transition costs acceptable, then
the change is made.
Klensin & Faltstrom Expires April 11, 2018 [Page 18]
Internet-Draft IDNA Unicode Update October 2017
The difference and its implications are perhaps best illustrated by a
disagreement when IDNA2008 was being approved. IDNA2003 had
effectively prevented some characters, notably (measured by intensity
of the protests) the Sharp S character (U+00DF) from being used in
DNS labels by mapping them to other characters before conversion to
ACE form. It has also prohibited some other code points, notably ZWJ
(U+200D) and ZWNJ (U+200C), by discarding them. In both cases, there
were strong voices from the relevant language communities, supported
by the registry communities, that the characters were important
enough that it was more desirable to undergo the short-term pain of a
transition and some uncertainty than to continue to exclude those
characters and the IDNA2008 rules and repertoire are consistent with
that preference. The Unicode Consortium apparently believed that
stability --elimination of any possibility of label invalidation or
different interpretations of the same string-- was more important
than those writing system requirements and community preferences.
That view was expressed through what was effectively a fork in (or
attempt to nullify) the IETF Standard [UTS46] a result that has
probably been worse for the overall Internet than either of the
possible decision choices.
4.2. New Unicode Properties
One suggestion about the way out of these problems would be to create
one or more new Unicode properties, maintained along with the rest of
Unicode, and then incorporated into new or modified rules or
categories in IDNA. Given the analysis in this document, it appears
that that property (or properties) would need to provide:
1. Identification of combining characters that, when used in
combining sequences, do not produce decomposable characters.
[[CREF4: Wording on the above is not quite right but, for the
present, maybe the intent is clear.]]
2. Identification of precomposed characters that might reasonably be
expected to decompose, but that do not.
3. Identification of character forms that are distinct only because
of language or phonetic distinctions within a script.
4. Identification of scripts for which precomposed forms are
strongly preferred and combining sequences should either be
viewed as temporary mechanisms until precomposed characters are
assigned or banned entirely.
5. Identification of code points that represent symbols for
specific, non-language, purposes even if identified as letters or
numerals by their General Property. This would include all
Klensin & Faltstrom Expires April 11, 2018 [Page 19]
Internet-Draft IDNA Unicode Update October 2017
characters given separate code points because of specialized
"mathematical" and "phonetic" characters (see Section 3.3.2.2 and
Section 3.3.2.1), but there are probably additional cases.
Some of these properties (or characteristics or values of a single
property) would be suitable for disallowing characters, code points,
or contextual sequences that otherwise might be allowed by IDNA.
Others would be more suitable for making equality comparisons come
out as needed by IDNA, particularly to eliminate distinctions based
on language context.
While it would appear that appropriate rules and categories could be
developed for IDNA (and, presumably, for PRECIS, etc.) if the problem
areas are those identified in this document, it is not yet known
whether the list is complete (and, hence, whether additional
properties or information would be needed).
Even with such properties, IDNA would still almost certainly need
exception lists. In addition, it is likely that stability rules for
those properties would need to reflect IETF norms with arrangements
for bringing the IETF and other communities into the discussion when
tradeoffs are reviewed.
4.3. The need for exception lists
[[CREF5: Note in draft: this section is a partial placeholder and may
need more elaboration.]]
Issues with exception lists and the requirements for them are
discussed in Section 2 above and in RFC 5894 [RFC5894].
5. Proposed/ Alternative Changes to RFC 5892 for the issues first
exposed by new code point U+08A1
NOTE IN DRAFT: See the comments in the Introduction, Section 1 and
the first paragraph of each Subsection below for the status of the
Subsections that follow. Each one, in combination with the material
in Section 3 above, also provides information about the reasons why
that particular strategy might or might not be appropriate.
When the term "Category" followed by an upper-case letter appears
below, it is s reference to a rule in RFC 5892.
5.1. Disallow This New Code Point
This option is almost certainly too Arabic-specific and does not
solve, or even address, the underlying problem. It also does not
inherently generalize to non-decomposing precomposed code points that
might be added in the future (whether to Arabic or other scripts)
Klensin & Faltstrom Expires April 11, 2018 [Page 20]
Internet-Draft IDNA Unicode Update October 2017
even though one could add more code points to Category F in the same
way.
If chosen by the community, this subsection would update the portion
of the IDNA2008 specification that identifies rules for what
characters are permitted [RFC5892] to disallow that code point.
With the publication of this document, Section 2.6 ("Exceptions (F)")
of RFC 5892 [RFC5892] is updated by adding 08A1 to the rule in
Category F so that the rule itself reads:
F: cp is in {00B7, 00DF, 0375, 03C2, 05F3, 05F4, 0640, 0660,
0661, 0662, 0663, 0664, 0665, 0666, 0667, 0668,
0669, 06F0, 06F1, 06F2, 06F3, 06F4, 06F5, 06F6,
06F7, 06F8, 06F9, 06FD, 06FE, 07FA, 08A1, 0F0B,
3007, 302E, 302F, 3031, 3032, 3033, 3034, 3035,
303B, 30FB}
and then add to the subtable designated
"DISALLOWED -- Would otherwise have been PVALID"
after the line that begins "07FA", the additional line:
08A1; DISALLOWED # ARABIC LETTER BEH WITH HAMZA ABOVE
This has the effect of making the cited code point DISALLOWED
independent of application of the rest of the IDNA rule set to the
current version of Unicode. Those wishing to create domain name
labels containing Beh with Hamza Above may continue to use the
sequence
U+0628, ARABIC LETTER BEH
followed by
U+0654, ARABIC HAMZA ABOVE
which was valid for IDNA purposes in Unicode 5.0 and earlier and
which continues to be valid.
In principle, much the same thing could be accomplished by using the
IDNA "BackwardCompatible" category (IDNA Category G, RFC 5892
Section 5.3). However, that category is described as applying only
when "property values in versions of Unicode after 5.2 have changed
in such a way that the derived property value would no longer be
PVALID or DISALLOWED". Because U+08A1 is a newly-added code point in
Unicode 7.0.0 and no property values of code points in prior versions
have changed, category G does not apply. If that section of RFC 5892
were to be replaced in the future, perhaps consideration should be
Klensin & Faltstrom Expires April 11, 2018 [Page 21]
Internet-Draft IDNA Unicode Update October 2017
given to adding Normalization Stability and other issues to that
description but, at present, it is not relevant.
5.2. Disallow This New Code Point and All Future Precomposed Additions
that Do Not Decompose
At least in principle, the approach suggested above (Section 5.1)
could be expanded to disallow all future allocations of non-
decomposing precomposed characters. This would probably require
either a new Unicode property to identify such characters and/or more
emphasis on the manual, individual code point, checking of the new
Unicode version review proces (i.e,. not just application of the
existing rules and algorithm). It might require either a new rule in
IDNA or a modification to the structure of Category F to make
additions less tedious. It would do nothing for different ways to
form identical characters within the same script that were not
associated with decomposition and so would have to be used in
conjunction with other appropaches. Finally, for scripts (such as
Arabic) where there is a very strong preference to avoid combining
sequences, this approach would exclude exactly the wrong set of
characters.
5.3. Disallow the combining sequences for these characters
As in the approach discussed in Section 5.1, this approach is too
Arabic-specific to address the more general problem. However, it
illustrates a single-script approach and a possible mechanism for
excluding combining sequences whose handling is connected to language
information (information that, as discussed above, is not relevant to
the DNS).
If chosen by the community, this subsection would update the portion
of the IDNA2008 specification that identifies contextual rules
[RFC5892] to prohibit (combining) Hamza Above (U+0654) in conjunction
with Arabic BEH (U+0628), HAH (U+062D), and REH (U+0631). Note that
the choice of this option is consistent with the general preference
for precomposed characters discussed above but would ban some labels
that are valid today and that might, in principle, be in use.
The required prohibition could be imposed by creating a new
contextual rule in RFC 5892 to constrain combining sequences
containing Hamza Above.
As the Unicode Standard points out at some length [Unicode70-Arabic],
Hamza is a problematic abstract character and the "Hamza Above"
construction even more so. IDNA has historically associated
characters whose use is reasonable in some contexts but not others
with the special derived property "CONTEXTO" and then specified
Klensin & Faltstrom Expires April 11, 2018 [Page 22]
Internet-Draft IDNA Unicode Update October 2017
specific, context-dependent, rules about where they may be used.
Because Hamza Above is problematic (and spawns edge cases, as
discussed in the Unicode Standard section cited above), it was
suggested that a contextual rule might be appropriate. There are at
least two reasons why a contextual rule would not be suitable for the
present situation.
1. As discussed above, the present situation is a normalization
stability and predictability problem, not a contextual one. Had
the same issues arisen with a newly-added precomposed character
that could previously be constructed from non-problematic base
and combining characters, it would be even more clearly a
normalization issue and, following the principles discussed there
and particularly in UAX 15 [UAX15-Exclusion], might not have been
assigned at all.
2. The contextual rule sets are designed around restricting the use
of code points to a particular script or adjacent to particular
characters within that script. Neither of these cases applies to
the newly-added character even if one could imagine rules for the
use of Hamza Above (U+0654) that would reflect the considerations
of Chapter 8 of Unicode 6.2. Even had the latter been desired,
it would be somewhat late now -- Hamza Above has been present as
a combining character (U+0654) in many versions of Unicode.
While that section of the Unicode Standard describes the issues,
it does not provide actionable guidance about what to do about it
for cases going forward or when visual identity is important.
5.4. Use Combinnig Classes to Develop Additional Contextual Rules
This option may not be of any practical use, but Unicode supports a
property called "Combining_Class". That property has been used in
IDNA only to construct a contextual rule for Zero-Width Non-Joiner
[RFC5892, Appendix A.1] but speculation has arisen during discussions
of work on Arabic combining characters and rendering [UTR53] as to
whether Combining Classes could be used to build additional
contextual rules that would restrict problematic cases. Unless such
rules were applied only to new code points, they would also not be
backward compatable.
The question of whether Combining Classes could be used to reduce the
number of problematic labels is at least worth examination.
5.5. Disallow all Combining Characters for Specific Scripts
[[CREF6: This subsection needs to be turned into prose, but the
follow bullet points are probably sufficient to identify the
issues.]]
Klensin & Faltstrom Expires April 11, 2018 [Page 23]
Internet-Draft IDNA Unicode Update October 2017
o Might work for Arabic and other "precomposed preference" scripts
if those can be identified in an orderly and stable way (see
Section 3.3.4.1; recommended by the Arabic language community for
IDNs [RFC5564]).
o Unworkable for Latin because many characters that do not decompose
are, at least in part, historical accidents resulting from
combining prior national standards (this probably may exist for
other scripts as well).
o No effect at all on special-use representations of identical
characters within a script (see Section 3.3.2.1 and
Section 3.3.2.2).
o Not backwards compatible.
5.6. Do Nothing Other Than Warn
A recommendation from UTC and others has been to simply warn
registries, at all levels of the tree, to be careful with this set of
characters. Doing that well would probably require making language
distinctions within zones, which would violate the important IDNA
principles that labels are not necessarily "words", do not carry
language information, and may, at the protocol level, even
deliberately mix languages and scripts. It is also problematic
because the relevant set of characters is not easily defined in a
precise way. This suggestion is problematic because the DNS and IDNA
cannot make or enforce language distinctions, but it would avoid
having the IETF either invalidate label strings that are potentially
now in use or creating inconsistencies among the characters that
combine with selected base characters but that also have precomposed
forms that do not have decompositions. The potential would still
exist for registries to respect the warning and deprecate such labels
if they existed.
More generally, while there are already requirements in IDNA for
registries to be knowledgeable and responsible about the labels they
register (a separate document discusses that requirement
[Klensin-rfc5891bis]), experience indicates that those requirements
are often ignored. At least as important, warning registries about
what should or should not be registered and even calling out specific
code points as dangerous and in need of extra attention
[Freytag-dangerous] does nothing to address the many cases in which
lookup-time checking for IDNA conformance and deliberately misleading
label constructions is important.
Klensin & Faltstrom Expires April 11, 2018 [Page 24]
Internet-Draft IDNA Unicode Update October 2017
5.7. Normalization Form IETF (NFI))
The most radical possibility for the comparison issue would be to
decide that none of the Unicode Normalization Forms specified in UAX
15 [UAX15] are adequate for use with the DNS because, contrary to
their apparent descriptions, normalization tables are actually
determined using language information. However, use of language
information is unacceptable for IDNA for reasons described elsewhere
in this document. The remedy would be to define an IETF-specific (or
DNS-specific) normalization form (sometimes called "NFI" in
discussions), building on NFC but adhering strictly to the rule that
normalization causes two different forms of the same character (glyph
image) within the same script to be treated as equal. In practice
such a form could be implemented for IDNA purposes as an additional
rule within RFC 5892 (and its successors) that constituted an
exception list for the NFC tables. For this set of characters, the
special IETF normalization form would be equivalent to the exclusion
discussed in Section 5.3 above.
An Internet-identifier-specific normalization form, especially if
specified somewhat separately from the IDNA core, would have a small
marginal advantage over the other strategies in this section (or in
combination with some of them), even though most of the end result
and much of the implementation would be the same in practice. While
the design of IDNA requires that strings be normalized as part of the
process of determining label validity (and hence before either
storage of values in the DNS or name resolution), there is an ongoing
debate about whether normalization should be performed before storing
a string or putting it on the wire or only when the string is
actually compared or otherwise used.
If a normalization procedure with the right properties for the IETF
was defined, that argument could be bypassed and the best decisions
made for different circumstances. The separation would also allow
better comparison of strings that lack language context in
applications environments in which the additional processing and
character classifications of IDNA and/or PRECIS were not applicable.
Having such a normalization procedure defined outside IDNA would also
minimize changes to IDNA itself, which is probably an advantage.
If the new normalizstion form were, in practice, simply an overlay on
NFC with modifications dictated by exception and/or property lists,
keeping its definition separate from IDNA would also avoid
interweaving those exceptions and property lists with the rules and
categories of IDNA itself, avoiding some unnecessary complexity.
Klensin & Faltstrom Expires April 11, 2018 [Page 25]
Internet-Draft IDNA Unicode Update October 2017
6. Editorial clarification to RFC 5892
Verified RFC Editor Erratum 3312 [RFC5892Erratum] provides a
clarification to Appendix A and Section A.1 of RFC 5892. This
section of this document updates the RFC to apply that clarification.
1. In Appendix A, add a new paragraph after the paragraph that
begins "The code point...". The new paragraph should read:
"For the rule to be evaluated to True for the label, it MUST be
evaluated separately for every occurrence of the Code point in
the label; each of those evaluations must result in True."
2. In Appendix A, Section A.1, replace the "Rule Set" by
Rule Set:
False;
If Canonical_Combining_Class(Before(cp)) .eq. Virama Then True;
If cp .eq. \u200C And
RegExpMatch((Joining_Type:{L,D})(Joining_Type:T)*cp
(Joining_Type:T)*(Joining_Type:{R,D})) Then True;
7. Acknowledgements
The Unicode 7.0.0 changes were extensively discussed within the IAB's
Internationalization Program. The authors are grateful for the
discussions and feedback there, especially from Andrew Sullivan and
David Thaler. Additional information was requested and received from
Mark Davis and Ken Whistler and while they probably do not agree with
the necessity of excluding this code point or taking even more
drastic action as their responsibility is to look at the Unicode
Consortium requirements for stability, the decision would not have
been possible without their input. Thanks to Bill McQuillan and Ted
Hardie for reading versions of the document carefully enough to
identify and report some confusing typographical errors. Several
experts and reviewers who prefer to remain anonymous also provided
helpful input and comments on preliminary versions of this document.
8. IANA Considerations
When the IANA registry and tables are updated to reflect Unicode
7.0.0, changes should be made according to the decisions the IETF
makes about Section 5.
Klensin & Faltstrom Expires April 11, 2018 [Page 26]
Internet-Draft IDNA Unicode Update October 2017
9. Security Considerations
From at least one point of view, this document is entirely a
discussion of a security issue or set of such issues. While the
"similar-looking characters" issue that has been a concern since the
earliest days of IDNs [HomographAttack] and that has driven assorted
"character confusion" projects [ICANN-VIP], if a user types in a
string on one device and can get different results that do not
compare equal when it is typed on a different device (with both
behaving correctly and both keyboards appearing to be the same and
for the same script) then all security mechanism that depend on the
underlying identifiers, including the practical applications of DNS
response integrity checks via DNSSEC [RFC4033] and DNS-embedded
public key mechanisms [RFC6698], are at risk if different parties, at
least one of them malicious, obtain or register some of the
identical-appearing and identically-typed strings and get them into
appropriate zones.
Mechanisms that depend on trusting registration systems (e.g.,
registries and registrars in the DNS IDN case, see Section 5.6 above)
are likely to be of only limited utility because fully-qualified
domains that may be perfectly reasonable at the first level or two of
the DNS may have differences of this type deep in the tree, into
levels where name management, and often accountability, are weak.
Similar issues obviously apply when names are user-selected or
unmanaged.
When the issue is not a deliberate attack but simple accidental
confusion among similar strings, most of our strategies depend on the
acceptability of false negatives on matching if there is low risk of
false positives (see, for example, the discussion of false negatives
in identifier comparison in Section 2.1 of RFC 6943 [RFC6943]).
Aspects of that issue appear in, for example, RFC 3986 [RFC3986] and
the PRECIS effort [RFC8264]. However, because the cases covered here
are connected, not just to what the user sees but to what is typed
and where, there is an increased risk of false positives (accidental
as well as deliberate).
[[CREF7: Note in Draft: The paragraph that follows was written for a
much earlier version of this document. It is obsolete, but is being
retained as a placeholder for future developments.]]
This specification excludes a code point for which the Unicode-
specified normalization behavior could result in two ways to form a
visually-identical character within the same script not comparing
equal. That behavior could create a dream case for someone intending
to confuse the user by use of a domain name that looked identical to
Klensin & Faltstrom Expires April 11, 2018 [Page 27]
Internet-Draft IDNA Unicode Update October 2017
another one, was entirely in the same script, but was still
considered different.
Internet Security in areas that involve internationalized identifiers
that might contain the relevant characters is therefore significantly
dependent on some effective resolution for the issues identified in
this document, not just hand waving, devout wishes, or appointment of
study committees about it.
10. References
10.1. Normative References
[RFC5137] Klensin, J., "ASCII Escaping of Unicode Characters",
BCP 137, RFC 5137, DOI 10.17487/RFC5137, February 2008,
<https://www.rfc-editor.org/info/rfc5137>.
[RFC5890] Klensin, J., "Internationalized Domain Names for
Applications (IDNA): Definitions and Document Framework",
RFC 5890, DOI 10.17487/RFC5890, August 2010,
<https://www.rfc-editor.org/info/rfc5890>.
[RFC5892] Faltstrom, P., Ed., "The Unicode Code Points and
Internationalized Domain Names for Applications (IDNA)",
RFC 5892, DOI 10.17487/RFC5892, August 2010,
<https://www.rfc-editor.org/info/rfc5892>.
[RFC5892Erratum]
"RFC5892, "The Unicode Code Points and Internationalized
Domain Names for Applications (IDNA)", August 2010, Errata
ID: 3312", Errata ID 3312, August 2012,
<http://www.rfc-editor.org/errata_search.php/doc/html/rfc5892>.
[RFC5894] Klensin, J., "Internationalized Domain Names for
Applications (IDNA): Background, Explanation, and
Rationale", RFC 5894, DOI 10.17487/RFC5894, August 2010,
<https://www.rfc-editor.org/info/rfc5894>.
[RFC6943] Thaler, D., Ed., "Issues in Identifier Comparison for
Security Purposes", RFC 6943, DOI 10.17487/RFC6943, May
2013, <https://www.rfc-editor.org/info/rfc6943>.
[RFC8264] Saint-Andre, P. and M. Blanchet, "PRECIS Framework:
Preparation, Enforcement, and Comparison of
Internationalized Strings in Application Protocols",
RFC 8264, DOI 10.17487/RFC8264, October 2017,
<https://www.rfc-editor.org/info/rfc8264>.
Klensin & Faltstrom Expires April 11, 2018 [Page 28]
Internet-Draft IDNA Unicode Update October 2017
[UAX15] Davis, M., Ed., "Unicode Standard Annex #15: Unicode
Normalization Forms", June 2014,
<http://www.unicode.org/reports/tr15/>.
[UAX15-Exclusion]
"Unicode Standard Annex #15: ob. cit., Section 5",
<http://www.unicode.org/reports/
tr15/#Primary_Exclusion_List_Table>.
[UAX15-Versioning]
"Unicode Standard Annex #15, ob. cit., Section 3",
<http://www.unicode.org/reports/tr15/#Versioning>.
[Unicode5]
The Unicode Consortium, "The Unicode Standard, Version
5.0", ISBN 0-321-48091-0, 2007.
Boston, MA, USA: Addison-Wesley. ISBN 0-321-48091-0.
This printed reference has now been updated online to
reflect additional code points. For code points, the
reference at the time RFC 5890-5894 were published is to
Unicode 5.2.
[Unicode62]
The Unicode Consortium, "The Unicode Standard, Version
6.2.0", ISBN 978-1-936213-07-8, 2012,
<http://www.unicode.org/versions/Unicode6.2.0/>.
Preferred citation: The Unicode Consortium. The Unicode
Standard, Version 6.2.0, (Mountain View, CA: The Unicode
Consortium, 2012. ISBN 978-1-936213-07-8)
[Unicode7]
The Unicode Consortium, "The Unicode Standard, Version
7.0.0", ISBN 978-1-936213-09-2, 2014,
<http://www.unicode.org/versions/Unicode7.0.0/>.
Preferred Citation: The Unicode Consortium. The Unicode
Standard, Version 7.0.0, (Mountain View, CA: The Unicode
Consortium, 2014. ISBN 978-1-936213-09-2)
[Unicode70-Arabic]
"The Unicode Standard, Version 7.0.0, ob.cit., Chapter
9.2: Arabic", Chapter 9, 2014,
<http://www.unicode.org/versions/Unicode7.0.0/ch09.pdf>.
Subsection titled "Encoding Principles", paragraph
numbered 4, starting on page 362.
Klensin & Faltstrom Expires April 11, 2018 [Page 29]
Internet-Draft IDNA Unicode Update October 2017
[Unicode70-CompatDecomp]
"The Unicode Standard, Version 7.0.0, ob.cit., Chapter
2.3: Compatibility Characters", Chapter 2, 2014,
<http://www.unicode.org/versions/Unicode7.0.0/ch02.pdf>.
Subsection titled "Compatibility Decomposable Characters"
starting on page 26.
[Unicode70-Design]
"The Unicode Standard, Version 7.0.0, ob.cit., Chapter
2.2: Unicode Design Principles", Chapter 2, 2014,
<http://www.unicode.org/versions/Unicode7.0.0/ch02.pdf>.
[Unicode70-Hamza]
"The Unicode Standard, Version 7.0.0, ob.cit., Chapter
9.2: Arabic", Chapter 9, 2014,
<http://www.unicode.org/versions/Unicode7.0.0/ch09.pdf>.
Subsection titled "Combining Hamza Above" starting on page
378.
[Unicode70-Overlay]
"The Unicode Standard, Version 7.0.0, ob.cit., Chapter
2.2: Unicode Design Principles", Chapter 2, 2014,
<http://www.unicode.org/versions/Unicode7.0.0/ch02.pdf>.
Subsection titled "Non-decomposition of Overlaid
Diacritics" starting on page 64.
[Unicode70-Stability]
"The Unicode Standard, Version 7.0.0, ob.cit., Chapter
2.2: Unicode Design Principles", Chapter 2, 2014,
<http://www.unicode.org/versions/Unicode7.0.0/ch02.pdf>.
Subsection titled "Stability" starting on page 23 and
containing a link to http://www.unicode.org/policies/
stability_policy.html..
[UTS46] Davis, M. and M. Suignard, "Unicode Technical Standard
#46: Unicode IDNA Compatibility Processing",
Version 7.0.0, June 2014,
<http://unicode.org/reports/tr46/>.
10.2. Informative References
[Dalby] Dalby, A., "Dictionary of Languages: The definitive
reference to more than 400 languages", Columbia Univeristy
Press , 2004.
Klensin & Faltstrom Expires April 11, 2018 [Page 30]
Internet-Draft IDNA Unicode Update October 2017
pages 206-207
[Daniels] Daniels, P. and W. Bright, "The World's Writing Systems",
Oxford University Press , 1986.
page 744
[Freytag-dangerous]
Freytag, A., Klensin, J., and A. Sullivan, "Those
Troublesome Characters: A Registry of Unicode Code Points
Needing Special Consideration When Used in Network
Identifiers", June 2017,
<https://datatracker.ietf.org/doc/
draft-freytag-troublesome-characters/>.
[HomographAttack]
Gabrilovich, E. and A. Gontmakher, "The Homograph Attack",
Communications of the ACM 45(2):128, February 2002,
<http://www.cs.technion.ac.il/~gabr/papers/
homograph_full.pdf>.
[ICANN-VIP]
ICANN, "The IDN Variant Issues Project: A Study of Issues
Related to the Management of IDN Variant TLDs (Integrated
Issues Report)", February 2012,
<https://www.icann.org/en/system/files/files/
idn-vip-integrated-issues-final-clean-20feb12-en.pdf>.
[Klensin-rfc5891bis]
Klensin, J., "Internationalized Domain Names in
Applications (IDNA): Registry Restrictions and
Recommendations", September 2017,
<https://datatracker.ietf.org/doc/
draft-klensin-idna-rfc5891bis/>.
[Omniglot-Fula]
Ager, S., "Omniglot: Fula (Fulfulde, Pulaar,
Pular'Fulaare)",
<http://www.omniglot.com/writing/fula.htm>.
Captured 2015-01-07
[RFC0020] Cerf, V., "ASCII format for network interchange", STD 80,
RFC 20, DOI 10.17487/RFC0020, October 1969,
<https://www.rfc-editor.org/info/rfc20>.
Klensin & Faltstrom Expires April 11, 2018 [Page 31]
Internet-Draft IDNA Unicode Update October 2017
[RFC3490] Faltstrom, P., Hoffman, P., and A. Costello,
"Internationalizing Domain Names in Applications (IDNA)",
RFC 3490, DOI 10.17487/RFC3490, March 2003,
<https://www.rfc-editor.org/info/rfc3490>.
[RFC3492] Costello, A., "Punycode: A Bootstring encoding of Unicode
for Internationalized Domain Names in Applications
(IDNA)", RFC 3492, DOI 10.17487/RFC3492, March 2003,
<https://www.rfc-editor.org/info/rfc3492>.
[RFC3986] Berners-Lee, T., Fielding, R., and L. Masinter, "Uniform
Resource Identifier (URI): Generic Syntax", STD 66,
RFC 3986, DOI 10.17487/RFC3986, January 2005,
<https://www.rfc-editor.org/info/rfc3986>.
[RFC4033] Arends, R., Austein, R., Larson, M., Massey, D., and S.
Rose, "DNS Security Introduction and Requirements",
RFC 4033, DOI 10.17487/RFC4033, March 2005,
<https://www.rfc-editor.org/info/rfc4033>.
[RFC5564] El-Sherbiny, A., Farah, M., Oueichek, I., and A. Al-Zoman,
"Linguistic Guidelines for the Use of the Arabic Language
in Internet Domains", RFC 5564, DOI 10.17487/RFC5564,
February 2010, <https://www.rfc-editor.org/info/rfc5564>.
[RFC6452] Faltstrom, P., Ed. and P. Hoffman, Ed., "The Unicode Code
Points and Internationalized Domain Names for Applications
(IDNA) - Unicode 6.0", RFC 6452, DOI 10.17487/RFC6452,
November 2011, <https://www.rfc-editor.org/info/rfc6452>.
[RFC6698] Hoffman, P. and J. Schlyter, "The DNS-Based Authentication
of Named Entities (DANE) Transport Layer Security (TLS)
Protocol: TLSA", RFC 6698, DOI 10.17487/RFC6698, August
2012, <https://www.rfc-editor.org/info/rfc6698>.
[Unicode32]
The Unicode Consortium, "The Unicode Standard, Version
3.2.0".
The Unicode Standard, Version 3.2.0 is defined by The
Unicode Standard, Version 3.0 (Reading, MA, Addison-
Wesley, 2000. ISBN 0-201-61633-5), as amended by the
Unicode Standard Annex #27: Unicode 3.1
(http://www.unicode.org/reports/tr27/) and by the Unicode
Standard Annex #28: Unicode 3.2
(http://www.unicode.org/reports/tr28/).
Klensin & Faltstrom Expires April 11, 2018 [Page 32]
Internet-Draft IDNA Unicode Update October 2017
[UTR53] Unicode Consortium, "Proposed Draft: Unicode Technical
Report #53: Unicode Arabic Mark Ordering Algorithm",
August 2017, <http://www.unicode.org/reports/tr53/>.
Note: this is a Proposed Draft, out for public review when
this version of the current I-D is posted, and should not
be considered either an approved/ final document or a
stable reference.
Appendix A. Change Log
RFC Editor: Please remove this appendix before publication.
A.1. Changes from version -00 (2014-07-21)to -01
o Version 01 of this document is an extensive rewrite and
reorganization, reflecting discussions with UTC members and adding
three more options for discussion to the original proposal to
simply disallow the new code point.
A.2. Changes from version -01 (2014-12-07) to -02
Corrected a typographical error in which Hamza Above was incorrectly
listed with the wrong code point.
A.3. Changes from version -02 (2014-12-07) to -03
Corrected a typographical error in the Abstract in which RFC 5892 was
incorrectly shown as 5982.
A.4. Changes from version -03 (2015-01-06) to -04
o Explicitly identified the applicability of U+08A1 with Fula and
added references that discuss that language and how it is written.
o Updated several Unicode 6.2 references to point to Unicode 7.0
since the latter is now available in stable form (it was done when
work on this I-D started).
o Extensively revised to discuss the non-Arabic cases, non-
decomposing diacritics, other types of characters that don't
compare equal after normalization, and more general problem and
approaches.
Klensin & Faltstrom Expires April 11, 2018 [Page 33]
Internet-Draft IDNA Unicode Update October 2017
A.5. Changes from version -04 (2015-03-11) to -05
o Modified a few citation labels to make them more obvious.
o Restructured Section 1 and added additional terminology comments.
o Added discussion about non-decomposable character cases, including
the "slash" example, and associated references for which -04
contained only placeholders.
o The examples and discussion of Latin script issues has been
expanded considerably. It is unfortunate that many readers in the
IETF community apparently cannot understand examples well enough
to believe a problem is significant unless they is a discussion of
Latin script examples, but, at least for this working draft, that
is the way it is.
o Rewrote the discussion of several of the alternatives and added
the discussion of combining classes.
o Rewrote and extended the discussion of the "warn only"
alternative.
o Several other sections modified to improve technical or editorial
clarity.
o Note that, while some references have been updated, others have
not. In particular, Unicode references are still tied to versions
6 or 7. In some cases, those non-historical references are and
will remain appropriate; others will best be replaced with
information about current versions of documents.
Authors' Addresses
John C Klensin
1770 Massachusetts Ave, Ste 322
Cambridge, MA 02140
USA
Phone: +1 617 245 1457
Email: john-ietf@jck.com
Klensin & Faltstrom Expires April 11, 2018 [Page 34]
Internet-Draft IDNA Unicode Update October 2017
Patrik Faltstrom
Netnod
Franzengatan 5
Stockholm 112 51
Sweden
Phone: +46 70 6059051
Email: paf@netnod.se
Klensin & Faltstrom Expires April 11, 2018 [Page 35]