RFC 1781 Network Working Group S. Kille Request for Comments: 1781 ISODE Consortium Obsoletes: RFC 1484 March 1995 Category: Standards Track Using the OSI Directory to Achieve User Friendly Naming Status of this Memo This document specifies an Internet standards track protocol for the Internet community, and requests discussion and suggestions for improvements. Please refer to the current edition of the "Internet Official Protocol Standards" (STD 1) for the standardization state and status of this protocol. Distribution of this memo is unlimited. Abstract The OSI Directory has user friendly naming as a goal. A simple minded usage of the directory does not achieve this. Two aspects not achieved are: o A user oriented notation o Guessability This proposal sets out some conventions for representing names in a friendly manner, and shows how this can be used to achieve really friendly naming. This then leads to a specification of a standard format for representing names, and to procedures to resolve them. This leads to a specification which allows directory names to be communicated between humans. The format in this specification is identical to that defined in [5], and it is intended that these specifications are compatible. Table of Contents 1. Why a notation is needed ................................... 2 2. The Notation ............................................... 3 3. Communicating Directory Names .............................. 7 4. Matching a purported name .................................. 9 4.1 Environment .......................................... 9 4.2 Matching ............................................. 10 4.3 Top Level ............................................ 12 4.4 Intermediate Level ................................... 13 4.5 Bottom Level ......................................... 14 5. Examples ................................................... 14 6. Support required from the standard ......................... 15 Kille [Page 1] RFC 1781 User Friendly Naming March 1995 7. Support of OSI Services .................................... 15 8. Experience ................................................. 16 9. Relationship to other work ................................. 17 10. Issues ..................................................... 19 11. References ................................................. 20 12. Security Considerations .................................... 21 13. Author's Address ........................................... 21 A. Pseudo-code for the matching algorithm ..................... 22 List of Figures 1. Example usage of User Friendly Naming ................ 18 2. Matching Algorithm ................................... 22 List of Tables 1. Local environment for private DUA .................... 10 2. Local environment for US Public DUA .................. 11 1. Why a notation is needed Many OSI Applications make use of Distinguished Names (DN) as defined in the OSI Directory [1]. The main reason for having a notation for name format is to interact with a user interface. This specification is coming dangerously close to the sin of standardising interfaces. However, there are aspects of presentation which it is desirable to standardise. It is important to have a common format to be able to conveniently refer to names. This might be done to represent a directory name on a business card or in an email message. There is a need for a format to support human to human communication, which must be string based (not ASN.1) and user oriented. In very many cases, a user will be required to input a name. This notation is designed to allow this to happen in a uniform manner across many user interfaces. The intention is that the name can just be typed in. There should not be any need to engage in form filling or complex dialogue. It should be possible to take the "human" description given at the meeting, and use it directly. The means in which this happens will become clear later. This approach uses the syntax defined in [5] for representing distinguished names. By relaxing some of the constraints on this specification, it is argued that a more user oriented specification is produced. However, this syntax cannot be mapped algorithmically onto a distinguished name without the use of a directory. This notation is targeted towards a general user oriented system, and in particular to represent the names of humans. Other syntaxes may be more appropriate for other uses of the directory. For example, the OSF Syntax may be more appropriate for some system oriented uses. Kille [Page 2] RFC 1781 User Friendly Naming March 1995 (The OSF Syntax uses "/" as a separator, and forms names in a manner intended to resemble UNIX filenames). This notation is targeted towards names which follow a particular DIT structure: organisationally oriented. This may make it inappropriate for some types of application. There may be a requirement to extend this notation to deal more cleanly with fully geographical names. This approach effectively defines a definition of descriptive names on top of the primitive names defined by the OSI Directory. 2. The Notation The notation used in this specification is defined in [5]. This notation defines an unambiguous representation of distinguished name, and this specification is designed to be used in conjunction with this format. Both specifications arise from the same piece of research work [4]. Some examples of the specification are given here. The author's User Friendly Name (UFN) might be written: Steve Kille, Computer Science, University College London, GB or S. Kille, Computer Science, University College London, GB This may be folded, perhaps to display in multi-column format. For example: Steve Kille, Computer Science, University College London, GB Another UFN might be: Christian Huitema, INRIA, FR or James Hacker, Basingstoke, Widget Inc, GB The final example shows quoting of a comma in an Organisation name: L. Eagle, "Sue, Grabbit and Runn", GB Kille [Page 3] RFC 1781 User Friendly Naming March 1995 A purported name is what a user supplies to an interface for resolution into one or more distinguished names. A system should almost always store a name as a distinguished name. This will be more efficient, and avoid problems with purported names which become ambiguous when a new name appears. A user interface may display a distinguished name, using the distinguished name notation. However, it may display a purported name in cases where this will be more pleasing to the user. Examples of this might be: o Omission of the higher components of the distinguished name are not displayed (abbreviation). o Omission of attribute types, where the type is unlikely to be needed to resolve ambiguity. The ways in which a purported name may vary from a distinguished name are now described: Type Omission There are two cases of this. o Schema defaulting. In this case, although the type is not present, a schema defaulting is used to deduce the type. The first two types of schema defaulting may be used to deduce a distinguished name without the use of the directory. The use of schema defaulting may be useful to improve the performance of UFN resolution. The types of schema defaulting are: -- Default Schema -- Context Dependent Default Schema -- Data Dependent Default Schema o Omission of the type to be resolved by searching. Default Schema The attribute type of an attribute may always be present. This may be done to emphasise the type structure of a name. In some cases, the typing may be omitted. This is done in a way so that in many common cases, no attribute types are needed. The following type hierarchy (schema) is assumed: Kille [Page 4] RFC 1781 User Friendly Naming March 1995 Common Name, (((Organisational Unit)*, Organisation,) Country). Explicitly typed RDNs may be inserted into this hierarchy at any point. The least significant component is always of type Common Name. Other types follow the defined organisational hierarchy. The following are equivalent: Filestore Access, Bells, Computer Science, University College London, GB and CN=Filestore Access, OU=Bells, OU=Computer Science, O=University College London, C=GB To interpet a distinguished name presented in this format, with some or all of the attributes with the type not specified, the types are derived according to the type hierarchy by the following algorithm: 1. If the first attribute type is not specified, it is CommonName. 2. If the last attribute type is not specified, it is Country. 3. If there is no organisation explicitly specified, the last attribute with type not specified is of type Organisation. 4. Any remaining attribute with type unspecified must be before an Organisation or OrganisationalUnit attribute, and is of type OrganisationalUnit. To take a distinguished name, and generate a name of this format with attribute types omitted, the following steps are followed. 1. If the first attribute is of type CommonName, the type may be omitted. 2. If the last attribute is of type Country, the type may be omitted. 3. If the last attribute is of type Country, the last Organisation attribute may have the type omitted. 4. All attributes of type OrganisationalUnit may have the type omitted, unless they are after an Organisation attribute or the first attribute is of type OrganisationalUnit. Kille [Page 5] RFC 1781 User Friendly Naming March 1995 Context Dependent Default Schema The distinguished name notation defines a fixed schema for type defaulting. It may be useful to have different defaults in different contexts. For example, the defaulting convention may be applied in a modified fashion to objects which are known not to be common name objects. This will always be followed if the least significant component is explicitly typed. In this case, the following hierarchy is followed: ((Organisational Unit)*, Organisation,) Country Data Dependent Defaulting There are cases where it would be optimal to default according to the data. For example, in: Einar Stefferud, Network Management Associates, CA, US It would be useful to default "CA" to type State. This might be done by defaulting all two letter attributes under C=US to type State. General Defaulting A type may be omitted in cases where it does not follow a default schema hierarchy, and then type variants can be explored by searching. Thus a distinguished name could be represented by a uniquely matching purported name. For example, James Hacker, Basingstoke, Widget Inc, GB Would match the distinguished name: CN=James Hacker, L=Basingstoke, O=Widget Inc, C=GB Abbreviation Some of the more significant components of the DN will be omitted, and then defaulted in some way (e.g., relative to a local context). For example: Steve Kille Kille [Page 6] RFC 1781 User Friendly Naming March 1995 Could be interpreted in the context of an organisational default. Local Type Keywords Local values can be used to identify types, in addition to the keywords defined in [5]. For example, "Organisation" may be recognised as an alternative to "O". Component Omission An intermediate component of the name may be omitted. Typically this will be an organisational unit. For example: Steve Kille, University College London, GB In some cases, this can be combined with abbreviation. For example: Steve Kille, University College London Approximation Approximate renditions or alternate values of one or more of the components will be supplied. For example: Stephen Kille, CS, UCL, GB or Steve Keill, Comp Sci, Univarstiy College London, GB Friendly Country A "friendly country name" can be used instead of the ISO 3166 two letter code. For example: UK; USA; France; Deutchland. 3. Communicating Directory Names A goal of this standard is to provide a means of communicating directory names. Two approaches are given, one defined in [5], and the other here. A future version of these specifications may contain only one of these approaches, or recommend use of one approach. The approach can usually be distinguished implicitly, as types are normally omitted in the UFN approach, and are always present in the Distinguished Name approach. No recommendation is made here, but the merits of each approach is given. Kille [Page 7] RFC 1781 User Friendly Naming March 1995 1. Distinguished Name or DN. A representation of the distinguished name, according to the specification of [5]. 2. User Friendly Name or UFN. A purported name, which is expected to unambiguously resolve onto the distinguished name. When a UFN is communicated, a form which should efficiently and unambiguously resolve onto a distinguished name should be chosen. Thus it is reasonable to omit types, or to use alternate values which will unambiguously identify the entry in question (e.g., by use of an alternate value of the RDN attribute type). It is not reasonable to use keys which are (or are likely to become) ambiguous. The approach used should be implicit from the context, rather than wired into the syntax. The terms "Directory Name" and "X.500 Name" should be used to refer to a name which might be either a DN or UFN. An example of appropriate usage of both forms is given in the Section which defines the Author's location in Section 12. Advantages of communicating the DN are: o The Distinguished Name is an unambiguous and stable reference to the user. o The DN will be used efficiently by the directory to obtain information. Advantages of communicating the UFN are: o Redundant type information can be omitted (e.g., "California", rather than "State=California", where there is known to be no ambiguity. o Alternate values can be used to identify a component. This might be used to select a value which is meaningful to the recipient, or to use a shorter form of the name. Often the uniqueness requirements of registration will lead to long names, which users will wish to avoid. o Levels of the hierarchy may be omitted. For example in a very small organisation, where a level of hierarchy has been used to represent company structure, and the person has a unique name within the organisation. Where UFN form is used, it is important to specify an unambiguous form. In some ways, this is analogous to writing a postal address. There are many legal ways to write it. Care needs to be taken to make the address unambiguous. Kille [Page 8] RFC 1781 User Friendly Naming March 1995 4. Matching a purported name The following approach specifies a default algorithm to be used with the User Friendly Naming approach. It is appropriate to modify this algorithm, and future specifications may propose alternative algorithms. Two simple algorithms are noted in passing, which may be useful in some contexts: 1. Use type omission only, but otherwise require the value of the RDN attribute to be present. 2. Require each RDN to be identified as in 1), or by an exact match on an alternate value of the RDN attribute. These algorithms do not offer the flexibility of the default algorithm proposed, but give many of the benefits of the approach in a very simple manner. The major utility of the purported name is to provide the important "user friendly" characteristic of guessability. A user will supply a purported name to a user interface, and this will be resolved onto a distinguished name. When a user supplies a purported name there is a need to derive the DN. In most cases, it should be possible to derive a single name from the purported name. In some cases, ambiguities will arise and the user will be prompted to select from a multiple matches. This should also be the case where a component of the name did not "match very well". There is an assumption that the user will simply enter the name correctly. The purported name variants are designed to make this happen! There is no need for fancy window based interfaces or form filling for many applications of the directory. Note that the fancy interfaces still have a role for browsing, and for more complex matching. This type of naming is to deal with cases where information on a known user is desired and keyed on the user's name. 4.1 Environment All matches occur in the context of a local environment. The local environment defines a sequence of names of a non-leaf objects in the DIT. This environment effectively defines a list of acceptable name abbreviations where the DUA is employed. The environment should be controllable by the individual user. It also defines an order in which to operate. This list is defined in the context of the number of name components supplied. This allows varying heuristics, depending on the environment, to make the approach have the "right" behaviour. In Kille [Page 9] RFC 1781 User Friendly Naming March 1995 most cases, the environment will start at a local point in the DIT, and move upwards. Examples are given in Tables 1 and 2. Table 1 shows an example for a typical local DUA, which has the following characteristics: One component Assumed first to be a user in the department, then a user or department within the university, then a national organisation, and finally a country. Two components Most significant component is first assumed to be a national organisation, then a department (this might be reversed in some organisations), and finally a country. Three or more components The most significant component is first assumed to be a country, then a national organisation, and finally a department. 4.2 Matching A purported name will be supplied, usually with a small number of components. This will be matched in the context of an environment. Where there are multiple components to be matched, these should be matched sequentially. If an unambiguous DN is determined, the match continues as if the full DN had been supplied. For example, if +-----------+--------------------------------------+ |Number of |Environment | |Components | | +-----------+--------------------------------------+ |1 |Physics, University College London, GB| | |University College London, GB | | |GB | +-----------+--------------------------------------+ |2 |GB | | |University College London, GB | | |-- | +-----------+--------------------------------------+ |3+ |-- | | |GB | | |University College London, GB | +-----------+--------------------------------------+ Table 1: Local environment for private DUA Kille [Page 10] RFC 1781 User Friendly Naming March 1995 +------------+-----------+ | Number of |Environment| | Components | | +------------+-----------+ | 1,2 | US | | | CA | | | -- | +------------+-----------+ | 3+ | -- | | | US | | | CA | +------------+-----------+ Table 2: Local environment for US Public DUA Stephen Kille, UCL is being matched in the context of environment GB, first UCL is resolved to the distinguished name: University College London, GB Then the next component of the purported name is taken to determine the final name. If there is an ambiguity (e.g., if UCL had made two matches, both paths are explored to see if the ambiguity can be resolved. Eventually a set of names will be passed back to the user. Each component of the environment is taken in turn. If the purported name has more components than the maximum depth, the environment element is skipped. The advantage of this will be seen in the example given later. A match of a name is considered to have three levels: Exact A DN is specified exactly Good Initially, a match should be considered good if it is unambiguous, and exactly matches an attribute value in the entry. For human names, a looser metric is probably desirable (e.g., S Kille should be a good match of S. Kille, S.E. Kille or Steve Kille even if these are not explicit alternate values). Poor Any other substring or approximate match Following a match, the reference can be followed, or the user prompted. If there are multiple matches, more than one path may be followed. There is also a shift/reduce type of choice: should any partial matches be followed or should the next element of the Kille [Page 11] RFC 1781 User Friendly Naming March 1995 environment be tried. The following heuristics are suggested, which may be modified in the light of experience. The overall aim is to resolve cleanly specified names with a minimum of fuss, but give sufficient user control to prevent undue searching and delay. 1. Always follow an exact match. 2. Follow all good matches if there are no exact matches. 3. If there are only poor matches, prompt the user. If the user accepts one or more matches, they can be considered as good. If all are rejected, this can be treated as no matches. 4. Automatically move to the next element of the environment if no matches are found. When the final component is matched, a set of names will be identified. If none are identified, proceed to the next environment element. If the user rejects all of the names, processing of the next environment element should be confirmed. The exact approach to matching will depend on the level of the tree at which matching is being done. We can now consider how attributes are matched at various levels of the DIT. There is an issue of approximate matching. Sometimes it helps, and sometimes just returns many spurious matches. When a search is requested, all relevant attributes should be returned, so that distinguished and non-distinguished values can be looked at. This will allow a distinction to be made between good and poor matches. It is important that where, for example, an acronym exactly matches an organisation, that the user is not prompted about other organisations where it matches as a substring. 4.3 Top Level In this case, a match is being done at the root of the DIT. Three approaches are suggested, dependent on the length of supplied name. All lead to a single level search of the top level of the DIT. Exactly 2 This is assumed to be a 3166 two letter country code, or an exact match on a friendly country or organisation (e.g., UK or UN). Do exact match on country and friendly country. Kille [Page 12] RFC 1781 User Friendly Naming March 1995 Greater than 2 Make an approximate and substring match on friendly country and organisation. 4.4 Intermediate Level Once the root level has been dealt with, intermediate levels will be looking for organisational components (Organisation, Locality, Org Unit). In some cases, private schema control will allow the system to determine which is at the next level. In general this will not be possible. In each case, make a substring and approximate match search of one level. The choice depends on the base object used in the search. 1. If DN has no Organisation or Locality, filter on Organisation and Locality. 2. If DN has Org Unit, filter on Org Unit. 3. If DN has Organisation, filter on Locality and Org Unit. 4. If DN has Locality, filter on Organisation. These allow some optimisation, based on legal choices of schema. Keeping filters short is usually desirable to improve performance. A few examples of this, where a base object has been determined (either by being the environment or by partial resolution of a purported name), and the next element of a purported name is being considered. This will generate a single level search. What varies is the types being filtered against. If the DN is: University College London, GB The search should be for Org Unit or Locality. If the DN is: Organisation=UN the search should be for Org Unit or Locality. There may be some improvements with respect to very short keys. Not making approximate or substring matches in these cases seems sensible (It might be desirable to allow "*" as a part of the purported name notation.) Kille [Page 13] RFC 1781 User Friendly Naming March 1995 4.5 Bottom Level The "Bottom Level" is to deal with leaf entries in the DIT. This will often be a person, but may also be a role, an application entity or something else. The last component of a purported name may either reference a leaf or non-leaf. For this reason, both should be tested for. As a heuristic, if the base object for the search has two or more components it should be tested first as a bottom level name and then intermediate. Reverse this for shorter names. This optimises for the (normal) case of non-leaves high up the tree and leaves low down the tree. For bottom level names, make an approximate and substring match against Common Name, Surname, and User ID. Where common name is looked for, a full subtree search will be used when at the second level of the DIT or lower, otherwise a single level search. For example, if I have resolved a purported name to the distinguished name University College London, GB and have a single component Bloggs, this will generate a subtree search. 5. Examples This is all somewhat confusing, and a few examples are given. These are all in the context of the environment shown in Table 1 on Page 13. If "Joe Bloggs" is supplied, a subtree search of Physics, University College London, GB will be made, and the user prompted for "Joseph Z. Bloggs" as the only possible match. If "Computer Science" is supplied, first Physics, University College London, GB will be searched, and the user will reject the approximate match of "Colin Skin". Then a subtree search of University College London, GB Kille [Page 14] RFC 1781 User Friendly Naming March 1995 will be made, looking for a person. Then a single level search will be made looking for Org Unit, and Computer Science, University College London, GB will be returned without prompting (exact match). Supplying "Steve Kille" will lead to a failed subtree search of Physics, University College London, GB and lead straight to a subtree search of University College London, GB This will lead to an exact value match, and so a single entry returned without prompting. If "Andrew Findlay, Brunel" is supplied, the first element of the environment will be skipped, single level search of "Brunel" under "GB" will find: Brunel University, GB and a subtree search for "Andrew Findlay" initiated. This will yield Andrew Findlay, Computing and Media Services, Brunel University, GB Dr A J Findlay, Manufacturing and Engineering Systems, Brunel University, GB and the user will be prompted with a choice. This approach shows how a simple format of this nature will "do the right thing" in many cases. 6. Support required from the standard Fortunately, all that is needed is there! It would be useful to have "friendly country name" as a standard attribute. 7. Support of OSI Services The major focus of this work has been to provide a mechanism for identifying Organisations and Users. A related function is to identify applications. Where the Application is identified by an AET (Application Entity Title) with an RDN of Common Name, this specification leads to a natural usage. For example, if a filestore is named "gannet", then this could easily be identified by the name: Kille [Page 15] RFC 1781 User Friendly Naming March 1995 Gannet, Computer Laboratory, Cambridge University, GB In normal usage, this might lead to access (using a purported name) of: FTAM gannet,cambridge A second type of access is where the user identifies an Organisation (Organisational Unit), and expects to obtain a default service. The service is implied by the application, and should not require any additional naming as far as the user is concerned. It is proposed that this is supported by User Friendly Naming in the following way. 1. Determine that the purported name identifies a non-leaf object, which is of object class Organisation or Organisational Unit or Locality. 2. Perform a single level search for Application Entities which support the required application contexts. This assumes that all services which are supporting default access for the organisation are registered at one level below (possibly by the use of aliases), and that other services (specific machines or parts of the organisation) are represented further down the tree. This seems to be a reasonable layout, and its utility can be evaluated by experiment. 8. Experience An experimental implementation of this has been written by Colin Robbins. The example in Figure 1 shows that it can be very effective at locating known individuals with a minimum of effort. This code has been deployed within the "FRED" interface of the PSI Pilot [9], and within an prototype interface for managing distribution lists. The user reaction has been favourable: Some issues have arisen from this experience: o Where there is more than one level of Organisational Unit, and the user guesses one which is not immediately below the organisation, the algorithm works badly. There does not appear to be an easy fix for this. It is not clear if this is a serious deficiency. o Substring searching is currently done with leading and trailing wildcards. As many implementations will not implement leading wildcards efficiently, it may be preferable to only use trailing wildcards. The effect of this on the algorithm needs to be investigated. Kille [Page 16] RFC 1781 User Friendly Naming March 1995 Implementors of this specification are encouraged to investigate variants of the basic algorithm. A final specification should depend on experience with such variants. 9. Relationship to other work Colin Robbin's work on the interface "Tom" and implementation of a distribution list interface strongly influenced this specification [6]. Some of the ideas used here originally came from a UK Proposal to the ISO/CCITT Directory Group on "New Name Forms" [2]. This defined, and showed how to implement, four different types of names: Typed and Ordered The current Distinguished Name is a restricted example of this type of name. Kille [Page 17] RFC 1781 User Friendly Naming March 1995 -> t hales, csiro, australia Found good match(es) for 'australia' Found exact match(es) for 'csiro' Please select from the following: Trevor Hales, OC, HPCC, DIT, IICT, CSIRO, AU [y/n] ? y The following were matched... Trevor Hales, OC, HPCC, DIT, IICT, CSIRO, AU -> g michaelson, queensland, au Found exact match(es) for 'au' Please select from the following: University of Queensland, AU [y/n] ? y Axolotl, AU [y/n] ? n Please select from the following: George Michaelson, Prentice Computer Centre, University of Queensland, AU [y/n] ? y Manager, University of Queensland, AU [y/n] ? n The following were matched... George Michaelson, Prentice Computer Centre, University of Queensland, AU -> r needham, cambridge Found good match(es) for 'cambridge' Please select from the following: Roger Needham, Computer Lab, Cambridge University [y/n] ? y The following were matched... Roger Needham, Computer Lab, Cambridge University -> kirstein Found good match(es) for 'kirstein' The following were matched... Peter Kirstein Figure 1: Example usage of User Friendly Naming Untyped and Ordered This is the type of name proposed here (with some extensions to allow optional typing). It is seen as meeting the key user requirement of disliking typed names, and is efficient to implement. Typed and Unordered This sort of name is proposed by others as the key basis for user friendly naming. Neufeld shows how X.500 can be used to provide this [7], and Peterson proposes the Profile system to provide this [8]. Kille [Page 18] RFC 1781 User Friendly Naming March 1995 The author contends that whilst typed naming is interesting for some types of searching (e.g., yellow page searching), it is less desirable for naming objects. This is borne out by operational experience with OSI Directories [3]. Untyped and Unordered Surprisingly this form of name can be supported quite easily. However, a considerable gain in efficiency can be achieved by requiring ordering. In practice, users can supply this easily. Therefore, this type of name is not proposed. 10. Issues The following issues are noted, which would need to be resolved before this document is progressed as an Internet Standard. Potential Ambiguity Whilst the intention of the notation is to allow for specification of alternate values, it inherently allows for ambiguous names to be specified. It needs to be demonstrated that problems of this characteristic are outweighed by other benefits of the notation. Utility Determine that the specification is being implemented and used. Performance Measurements on the performance implications of using this approach should be made. Alogrithm The utility of the algorithm, and possible variants, should be investigated. This format, and the procedures for resolving purported names, should be evolved to an Internet Standard. The syntax can be expected to be stable. In light of experience, the algorithm for resolving purported names may be changed. Kille [Page 19] RFC 1781 User Friendly Naming March 1995 11. References [1] The Directory --- overview of concepts, models and services, 1993. CCITT X.500 Series Recommendations. [2] S.E. Kille. New name forms, May 1989. ISO/IEC/JTC 21/ WG4/N797 UK National Body Contribution to the Oslo Directory Meeting. [3] S.E. Kille. The THORN large scale pilot exercise. Computer Networks and ISDN Systems, 16(1):143--145, January 1989. [4] S.E. Kille. Using the OSI directory to achieve user friendly naming. Research Note RN/20/29, Department of Computer Science, University College London, February 1990. [5] Kille, S., "A String Representation of Distinguished Names", RFC 1779, ISODE Consortium, March 1995. [6] S.E. Kille and C.J. Robbins. The ISO development environment: User's manual (version 7.0), July 1991. Volume 5: QUIPU. [7] G.W. Neufeld. Descriptive names in X.500. In SIGCOMM 89 Symposiun Communications Architectures and Protocols, pages 64-- 71, September 1989. [8] L.L. Petersen. The profile naming service. ACM Transactions on Computing Systems, 6(4):341--364, November 1988. [9] M.T. Rose. Realizing the White Pages using the OSI Directory Service. Technical Report 90--05--10--1, Performance Systems International, Inc., May 1990. Kille [Page 20] RFC 1781 User Friendly Naming March 1995 12. Security Considerations Security issues are not discussed in this memo. 13. Author's Address Steve Kille ISODE Consortium The Dome The Square Richmond, Surrey TW9 1DT England Phone:+44-181-332-9091 EMail: S.Kille@ISODE.COM DN: CN=Steve Kille, O=ISODE Consortium, C=GB UFN: S. Kille, ISODE Consortium, GB Kille [Page 21] RFC 1781 User Friendly Naming March 1995 A. Pseudo-code for the matching algorithm The following pseudo-code is intended to clarify the matching algorithm. The language uses ASN.1 data types, with flow control "C"-like, but with keywords upper-cased. PurportedName ::= SEQUENCE OF String -- simplication, as attribute types can optionally be -- specified -- Each element of the Purported Name is a string -- which has been parsed from the BNF Attribute ::= SEQUENCE { type OBJECT IDENTIFIER, value ANY } RDN ::= Attribute -- simplification, as can be multi-value DN ::= SEQUENCE OF RDN Environment ::= SEQUENCE OF DN EnvironmentList ::= SEQUENCE OF SEQUENCE { lower-bound INTEGER, upper-bound INTEGER, environment Environment } friendlyMatch(p: PurportedName; el: EnvironmentList): SET OF DN { -- Find correct environment IF length(el) == 0 THEN return(NULL); IF length(p) <= head(el).upper-bound && length(p) >= head(el).lower-bound THEN return envMatch (p, head(el).environment); ELSE return(friendlyMatch(p, tail(el)); } envMatch(p: PurportedName; e: Environment): SET OF DN { -- Check elements of environment -- in the defined order matches: SET OF DN; Kille [Page 22] RFC 1781 User Friendly Naming March 1995 IF length(e) == 0 THEN return(NULL); matches = purportedMatch(head(e).DN, p) IF matches != NULL THEN return(matches); ELSE return(envMatch(p, tail(e)); } purportedMatch(base: DN; p: PurportedName): SET OF DN { s: String = head(p); matches: SET OF DN = NULL; IF length(p) == 1 THEN IF length(base) == 0 THEN IF (matches = rootSearch(s)) != NULL THEN return(matches); ELSE return(leafSearch(base, s, one-level); ELSE IF length(base) == 1 THEN IF (matches = intSearch(base, s)) != NULL THEN return(matches); ELSE return(leafSearch(base, s, one-level); ELSE IF (matches = leafSearch(base, s, subtree)) != NULL THEN return(matches); ELSE return(intsearch(base, s); IF length(base) == 0 THEN FOR x IN rootSearch(s) DO matches += (purportedMatch(x, tail(p)); ELSE FOR x IN intSearch(base, s) DO matches += (purportedMatch(x, tail(p)); return(matches); } Kille [Page 23] RFC 1781 User Friendly Naming March 1995 -- General. Might need to tighten the filter for short strings, -- in order to stop being flooded. Alternatively, this could be -- done if the loose search hits a size limit rootSearch(s: String): SET OF DN { IF length(s) == 2 THEN return(search(NULL, one-level, s, {CountryName, FriendlyCountryName, OrganizationName}, {exact}, {Country, Organisation})); -- test exact match only -- probably a country code ELSE return(search(NULL, one-level, s, {OrganizationName, FriendlyCountryName}, {substring, approx}, {Country, Organisation})); } intSearch( base: DN; s: String) { IF present(base, OrgUnitName) THEN return(search(base, one-level, s, {OrgUnitName}, {substring, approx}, {OrgUnit})); ELSE IF present(base, OrganisationName) THEN return(search(base, one-level, s, {OrgUnitName, LocalityName}, {substring, approx}, {Organization, OrgUnit, Locality})); ELSE IF present(base, LocalityName) THEN return(search(base, one-level, s, {OrganisationName}, {substring, approx}, {Locality}); ELSE return(search(base, one-level, s, {OrganisationName, LocalityName}, {substring, approx}, {Organisation, Locality})); } present(d: DN; t: AttributeType): BOOLEAN { FOR x IN d DO IF x.type == t THEN return(TRUE); return(FALSE); } SearchScope := ENUMERATED (base-object, one-level, subtree) leafSearch(base: DN; s: String; search-scope: SearchScope) Kille [Page 24] RFC 1781 User Friendly Naming March 1995 { return(search(base, search-scope, s, {CommonName, Surname, UserId}, {substring, approx})); } search(base: DN; search-scope: SearchScope; s: string; alist SET OF AttributeType; matchtypes SET OF MatchType objectClasses SET OF ObjectClass OPTIONAL): SET OF DN { -- mapped onto Directory Search, with OR conjunction -- of filter items return dNSelect (s, search-results, alist); } read(base: DN; alist SET OF AttributeType): SET OF Attribute; { -- mapped onto Directory Read -- Types repeated to deal with multiple values -- This would be implemented by returning selected info -- with the search operation } dNSelect(s: String; dlist SET OF DN; alist: SET OF AttributeType):16SET0OF DN { exact, good: SET OF DN; FOR x IN dlist DO IF last(DN).Value == s THEN exact += x; ELSE IF FOR y IN read(x, alist) DO IF y.value == s THEN good += x; IF exact != NULL THEN return(exact); IF good != NULL THEN return(good); return(userQuery(dlist)); } userQuery(dlist SET OF DN): SET OF DN { -- pass back up for manual checking -- user can strip all matches to force progres.... } head() -- return first element of list tail() -- return list with first element removed Kille [Page 25] RFC 1781 User Friendly Naming March 1995 length() -- return size of list last() -- return last element of list Figure 2: Matching Algorithm Kille [Page 26]