NetBSD/usr.sbin/sendmail/doc/rfc819.lpr

1045 lines
34 KiB
Plaintext
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

Network Working Group Zaw-Sing Su (SRI)
Request for Comments: 819 Jon Postel (ISI)
August 1982
The Domain Naming Convention for Internet User Applications
1. Introduction
For many years, the naming convention "<user>@<host>" has served the
ARPANET user community for its mail system, and the substring
"<host>" has been used for other applications such as file transfer
(FTP) and terminal access (Telnet). With the advent of network
interconnection, this naming convention needs to be generalized to
accommodate internetworking. A decision has recently been reached to
replace the simple name field, "<host>", by a composite name field,
"<domain>" [2]. This note is an attempt to clarify this generalized
naming convention, the Internet Naming Convention, and to explore the
implications of its adoption for Internet name service and user
applications.
The following example illustrates the changes in naming convention:
ARPANET Convention: Fred@ISIF
Internet Convention: Fred@F.ISI.ARPA
The intent is that the Internet names be used to form a
tree-structured administrative dependent, rather than a strictly
topology dependent, hierarchy. The left-to-right string of name
components proceeds from the most specific to the most general, that
is, the root of the tree, the administrative universe, is on the
right.
The name service for realizing the Internet naming convention is
assumed to be application independent. It is not a part of any
particular application, but rather an independent name service serves
different user applications.
2. The Structural Model
The Internet naming convention is based on the domain concept. The
name of a domain consists of a concatenation of one or more <simple
names>. A domain can be considered as a region of jurisdiction for
name assignment and of responsibility for name-to-address
translation. The set of domains forms a hierarchy.
Using a graph theory representation, this hierarchy may be modeled as
a directed graph. A directed graph consists of a set of nodes and a
Su & Postel [Page 1]
RFC 819 August 1982;
collection of arcs, where arcs are identified by ordered pairs of
distinct nodes [1]. Each node of the graph represents a domain. An
ordered pair (B, A), an arc from B to A, indicates that B is a
subdomain of domain A, and B is a simple name unique within A. We
will refer to B as a child of A, and A a parent of B. The directed
graph that best describes the naming hierarchy is called an
"in-tree", which is a rooted tree with all arcs directed towards the
root (Figure 1). The root of the tree represents the naming universe,
ancestor of all domains. Endpoints (or leaves) of the tree are the
lowest-level domains.
U
/ | \
/ | \ U -- Naming Universe
^ ^ ^ I -- Intermediate Domain
| | | E -- Endpoint Domain
I E I
/ \ |
^ ^ ^
| | |
E E I
/ | \
^ ^ ^
| | |
E E E
Figure 1
The In-Tree Model for Domain Hierarchy
The simple name of a child in this model is necessarily unique within
its parent domain. Since the simple name of the child's parent is
unique within the child's grandparent domain, the child can be
uniquely named in its grandparent domain by the concatenation of its
simple name followed by its parent's simple name.
For example, if the simple name of a child is "C1" then no other
child of the same parent may be named "C1". Further, if the
parent of this child is named "P1", then "P1" is a unique simple
name in the child's grandparent domain. Thus, the concatenation
C1.P1 is unique in C1's grandparent domain.
Similarly, each element of the hierarchy is uniquely named in the
universe by its complete name, the concatenation of its simple name
and those for the domains along the trail leading to the naming
universe.
The hierarchical structure of the Internet naming convention supports
decentralization of naming authority and distribution of name service
capability. We assume a naming authority and a name server
Su & Postel [Page 2]
RFC 819 August 1982;
associated with each domain. In Sections 5 and 6 respectively the
name service and the naming authority are discussed.
Within an endpoint domain, unique names are assigned to <user>
representing user mailboxes. User mailboxes may be viewed as
children of their respective domains.
In reality, anomalies may exist violating the in-tree model of naming
hierarchy. Overlapping domains imply multiple parentage, i.e., an
entity of the naming hierarchy being a child of more than one domain.
It is conceivable that ISI can be a member of the ARPA domain as well
as a member of the USC domain (Figure 2). Such a relation
constitutes an anomaly to the rule of one-connectivity between any
two points of a tree. The common child and the sub-tree below it
become descendants of both parent domains.
U
/ | \
/ . \
. . ARPA
. . | \
USC | \
\ | .
\ | .
ISI
Figure 2
Anomaly in the In-Tree Model
Some issues resulting from multiple parentage are addressed in
Appendix B. The general implications of multiple parentage are a
subject for further investigation.
3. Advantage of Absolute Naming
Absolute naming implies that the (complete) names are assigned with
respect to a universal reference point. The advantage of absolute
naming is that a name thus assigned can be universally interpreted
with respect to the universal reference point. The Internet naming
convention provides absolute naming with the naming universe as its
universal reference point.
For relative naming, an entity is named depending upon the position
of the naming entity relative to that of the named entity. A set of
hosts running the "unix" operating system exchange mail using a
method called "uucp". The naming convention employed by uucp is an
example of relative naming. The mail recipient is typically named by
a source route identifying a chain of locally known hosts linking the
Su & Postel [Page 3]
RFC 819 August 1982;
sender's host to the recipient's. A destination name can be, for
example,
"alpha!beta!gamma!john",
where "alpha" is presumably known to the originating host, "beta" is
known to "alpha", and so on.
The uucp mail system has demonstrated many of the problems inherent
to relative naming. When the host names are only locally
interpretable, routing optimization becomes impossible. A reply
message may have to traverse the reverse route to the original sender
in order to be forwarded to other parties.
Furthermore, if a message is forwarded by one of the original
recipients or passed on as the text of another message, the frame of
reference of the relative source route can be completely lost. Such
relative naming schemes have severe problems for many of the uses
that we depend upon in the ARPA Internet community.
4. Interoperability
To allow interoperation with a different naming convention, the names
assigned by a foreign naming convention need to be accommodated.
Given the autonomous nature of domains, a foreign naming environment
may be incorporated as a domain anywhere in the hierarchy. Within
the naming universe, the name service for a domain is provided within
that domain. Thus, a foreign naming convention can be independent of
the Internet naming convention. What is implied here is that no
standard convention for naming needs to be imposed to allow
interoperations among heterogeneous naming environments.
For example:
There might be a naming convention, say, in the FOO world,
something like "<user>%<host>%<area>". Communications with an
entity in that environment can be achieved from the Internet
community by simply appending ".FOO" on the end of the name in
that foreign convention.
John%ISI-Tops20-7%California.FOO
Another example:
One way of accommodating the "uucp world" described in the last
section is to declare it as a foreign system. Thus, a uucp
name
"alpha!beta!gamma!john"
Su & Postel [Page 4]
RFC 819 August 1982;
might be known in the Internet community as
"alpha!beta!gamma!john.UUCP".
Communicating with a complex subdomain is another case which can
be treated as interoperation. A complex subdomain is a domain
with complex internal naming structure presumably unknown to the
outside world (or the outside world does not care to be concerned
with its complexity).
For the mail system application, the names embedded in the message
text are often used by the destination for such purpose as to reply
to the original message. Thus, the embedded names may need to be
converted for the benefit of the name server in the destination
environment.
Conversion of names on the boundary between heterogeneous naming
environments is a complex subject. The following example illustrates
some of the involved issues.
For example:
A message is sent from the Internet community to the FOO
environment. It may bear the "From" and "To" fields as:
From: Fred@F.ISI.ARPA
To: John%ISI-Tops20-7%California.FOO
where "FOO" is a domain independent of the Internet naming
environment. The interface on the boundary of the two
environments may be represented by a software module. We may
assume this interface to be an entity of the Internet community
as well as an entity of the FOO community. For the benefit of
the FOO environment, the "From" and "To" fields need to be
modified upon the message's arrival at the boundary. One may
view naming as a separate layer of protocol, and treat
conversion as a protocol translation. The matter is
complicated when the message is sent to more than one
destination within different naming environments; or the
message is destined within an environment not sharing boundary
with the originating naming environment.
While the general subject concerning conversion is beyond the scope
of this note, a few questions are raised in Appendix D.
Su & Postel [Page 5]
RFC 819 August 1982;
5. Name Service
Name service is a network service providing name-to-address
translation. Such service may be achieved in a number of ways. For
a simple networking environment, it can be accomplished with a single
central database containing name-to-address correspondence for all
the pertinent network entities, such as hosts.
In the case of the old ARPANET host names, a central database is
duplicated in each individual host. The originating module of an
application process would query the local name service (e.g., make a
system call) to obtain network address for the destination host. With
the proliferation of networks and an accelerating increase in the
number of hosts participating in networking, the ever growing size,
update frequency, and the dissemination of the central database makes
this approach unmanageable.
The hierarchical structure of the Internet naming convention supports
decentralization of naming authority and distribution of name service
capability. It readily accommodates growth of the naming universe.
It allows an arbitrary number of hierarchical layers. The addition
of a new domain adds little complexity to an existing Internet
system.
The name service at each domain is assumed to be provided by one or
more name servers. There are two models for how a name server
completes its work, these might be called "iterative" and
"recursive".
For an iterative name server there may be two kinds of responses.
The first kind of response is a destination address. The second
kind of response is the address of another name server. If the
response is a destination address, then the query is satisfied. If
the response is the address of another name server, then the query
must be repeated using that name server, and so on until a
destination address is obtained.
For a recursive name server there is only one kind of response --
a destination address. This puts an obligation on the name server
to actually make the call on another name server if it can't
answer the query itself.
It is noted that looping can be avoided since the names presented for
translation can only be of finite concatenation. However, care
should be taken in employing mechanisms such as a pointer to the next
simple name for resolution.
We believe that some name servers will be recursive, but we don't
believe that all will be. This means that the caller must be
Su & Postel [Page 6]
RFC 819 August 1982;
prepared for either type of server. Further discussion and examples
of name service is given in Appendix C.
The basic name service at each domain is the translation of simple
names to addresses for all of its children. However, if only this
basic name service is provided, the use of complete (or fully
qualified) names would be required. Such requirement can be
unreasonable in practice. Thus, we propose the use of partial names
in the context in which their uniqueness is preserved. By
construction, naming uniqueness is preserved within the domain of a
common ancestry. Thus, a partially qualified name is constructed by
omitting from the complete name ancestors common to the communicating
parties. When a partially qualified name leaves its context of
uniqueness it must be additionally qualified.
The use of partially qualified names places a requirement on the
Internet name service. To satisfy this requirement, the name service
at each domain must be capable of, in addition to the basic service,
resolving simple names for all of its ancestors (including itself)
and their children. In Appendix B, the required distinction among
simple names for such resolution is addressed.
6. Naming Authority
Associated with each domain there must be a naming authority to
assign simple names and ensure proper distinction among simple names.
Note that if the use of partially qualified names is allowed in a
sub-domain, the uniqueness of simple names inside that sub-domain is
insufficient to avoid ambiguity with names outside the subdomain.
Appendix B discusses simple name assignment in a sub-domain that
would allow the use of partially qualified names without ambiguity.
Administratively, associated with each domain there is a single
person (or office) called the registrar. The registrar of the naming
universe specifies the top-level set of domains and designates a
registrar for each of these domains. The registrar for any given
domain maintains the naming authority for that domain.
7. Network-Oriented Applications
For user applications such as file transfer and terminal access, the
remote host needs to be named. To be compatible with ARPANET naming
convention, a host can be treated as an endpoint domain.
Many operating systems or programming language run-time environments
provide functions or calls (JSYSs, SVCs, UUOs, SYSs, etc.) for
standard services (e.g., time-of-day, account-of-logged-in-user,
convert-number-to-string). It is likely to be very helpful if such a
Su & Postel [Page 7]
RFC 819 August 1982;
function or call is developed for translating a host name to an
address. Indeed, several systems on the ARPANET already have such
facilities for translating an ARPANET host name into an ARPANET
address based on internal tables.
We recommend that this provision of a standard function or call for
translating names to addresses be extended to accept names of
Internet convention. This will promote a consistent interface to the
users of programs involving internetwork activities. The standard
facility for translating Internet names to Internet addresses should
include all the mechanisms available on the host, such as checking a
local table or cache of recently checked names, or consulting a name
server via the Internet.
8. Mail Relaying
Relaying is a feature adopted by more and more mail systems.
Relaying facilitates, among other things, interoperations between
heterogeneous mail systems. The term "relay" is used to describe the
situation where a message is routed via one or more intermediate
points between the sender and the recipient. The mail relays are
normally specified explicitly as relay points in the instructions for
message delivery. Usually, each of the intermediate relays assume
responsibility for the relayed message [3].
A point should be made on the basic difference between mail
relaying and the uucp naming system. The difference is that
although mail relaying with absolute naming can also be considered
as a form of source routing, the names of each intermediate points
and that of the destination are universally interpretable, while
the host names along a source route of the uucp convention is
relative and thus only locally interpretable.
The Internet naming convention explicitly allows interoperations
among heterogeneous systems. This implies that the originator of a
communication may name a destination which resides in a foreign
system. The probability is that the destination network address may
not be comprehensible to the transport system of the originator.
Thus, an implicit relaying mechanism is called for at the boundary
between the domains. The function of this implicit relay is the same
as the explicit relay.
Su & Postel [Page 8]
RFC 819 August 1982;
9. Implementation
The Actual Domains
The initial set of top-level names include:
ARPA
This represents the set of organizations involved in the
Internet system through the authority of the U.S. Defense
Advanced Research Projects Agency. This includes all the
research and development hosts on the ARPANET and hosts on
many other nets as well. But note very carefully that the
top-level domain "ARPA" does not map one-to-one with the
ARPANET -- domains are administrative, not topological.
Transition
In the transition from the ARPANET naming convention to the
Internet naming convention, a host name may be used as a simple
name for an endpoint domain. Thus, if "USC-ISIF" is an ARPANET
host name, then "USC-ISIF.ARPA" is the name of an Internet domain.
10. Summary
A hierarchical naming convention based on the domain concept has been
adopted by the Internet community. It is an absolute naming
convention defined along administrative rather than topological
boundaries. This naming convention is adaptive for interoperations
with other naming conventions. Thus, no standard convention needs to
be imposed for interoperations among heterogeneous naming
environments.
This Internet naming convention allows distributed name service and
naming authority functions at each domain. We have specified these
functions required at each domain. Also discussed are implications
on network-oriented applications, mail systems, and administrative
aspects of this convention.
Su & Postel [Page 9]
RFC 819 August 1982;
APPENDIX A
The BNF Specification
We present here a rather detailed "BNF" definition of the allowed
form for a computer mail "mailbox" composed of a "local-part" and a
"domain" (separated by an at sign). Clearly, the domain can be used
separately in other network-oriented applications.
<mailbox> ::= <local-part> "@" <domain>
<local-part> ::= <string> | <quoted-string>
<string> ::= <char> | <char> <string>
<quoted-string> ::= """ <qtext> """
<qtext> ::= "\" <x> | "\" <x> <qtext> | <q> | <q> <qtext>
<char> ::= <c> | "\" <x>
<domain> ::= <naming-domain> | <naming-domain> "." <domain>
<naming-domain> ::= <simple-name> | <address>
<simple-name> ::= <a> <ldh-str> <let-dig>
<ldh-str> ::= <let-dig-hyp> | <let-dig-hyp> <ldh-str>
<let-dig> ::= <a> | <d>
<let-dig-hyp> ::= <a> | <d> | "-"
<address> :: = "#" <number> | "[" <dotnum> "]"
<number> ::= <d> | <d> <number>
<dotnum> ::= <snum> "." <snum> "." <snum> "." <snum>
<snum> ::= one, two, or three digits representing a decimal integer
value in the range 0 through 255
<a> ::= any one of the 52 alphabetic characters A through Z in upper
case and a through z in lower case
<c> ::= any one of the 128 ASCII characters except <s> or <SP>
<d> ::= any one of the ten digits 0 through 9
Su & Postel [Page 10]
RFC 819 August 1982;
<q> ::= any one of the 128 ASCII characters except CR, LF, quote ("),
or backslash (\)
<x> ::= any one of the 128 ASCII characters (no exceptions)
<s> ::= "<", ">", "(", ")", "[", "]", "\", ".", ",", ";", ":", "@",
""", and the control characters (ASCII codes 0 through 31 inclusive
and 127)
Note that the backslash, "\", is a quote character, which is used to
indicate that the next character is to be used literally (instead of
its normal interpretation). For example, "Joe\,Smith" could be used
to indicate a single nine character user field with comma being the
fourth character of the field.
The simple names that make up a domain may contain both upper and
lower case letters (as well as digits and hyphen), but these names
are not case sensitive.
Hosts are generally known by names. Sometimes a host is not known to
the translation function and communication is blocked. To bypass
this barrier two forms of addresses are also allowed for host
"names". One form is a decimal integer prefixed by a pound sign, "#".
Another form, called "dotted decimal", is four small decimal integers
separated by dots and enclosed by brackets, e.g., "[123.255.37.2]",
which indicates a 32-bit ARPA Internet Address in four 8-bit fields.
(Of course, these numeric address forms are specific to the Internet,
other forms may have to be provided if this problem arises in other
transport systems.)
Su & Postel [Page 11]
RFC 819 August 1982;
APPENDIX B
An Aside on the Assignment of Simple Names
In the following example, there are two naming hierarchies joining at
the naming universe 'U'. One consists of domains (S, R, N, J, P, Q,
B, A); and the other (L, E, F, G, H, D, C, K, B, A). Domain B is
assumed to have multiple parentage as shown.
U
/ \
/ \
J L
/ \
N E
/ \ / \
R P D F
/ \ | \ \
S Q C (X) G
\ / \ \
B K H
/
A
Figure 3
Illustration of Requirements for the Distinction of Simple Names
Suppose someone at A tries to initiate communication with destination
H. The fully qualified destination name would be
H.G.F.E.L.U
Omitting common ancestors, the partially qualified name for the
destination would be
H.G.F
To permit the case of partially qualified names, name server at A
needs to resolve the simple name F, i.e., F needs to be distinct from
all the other simple names in its database.
To enable the name server of a domain to resolve simple names, a
simple name for a child needs to be assigned distinct from those of
all of its ancestors and their immediate children. However, such
distinction would not be sufficient to allow simple name resolution
at lower-level domains because lower-level domains could have
multiple parentage below the level of this domain.
In the example above, let us assume that a name is to be assigned
Su & Postel [Page 12]
RFC 819 August 1982;
to a new domain X by D. To allow name server at D to resolve
simple names, the name for X must be distinct from L, E, D, C, F,
and J. However, allowing A to resolve simple names, X needs to be
also distinct from A, B, K, as well as from Q, P, N, and R.
The following observations can be made.
Simple names along parallel trails (distinct trails leading from
one domain to the naming universe) must be distinct, e.g., N must
be distinct from E for B or A to properly resolve simple names.
No universal uniqueness of simple names is called for, e.g., the
simple name S does not have to be distinct from that of E, F, G,
H, D, C, K, Q, B, or A.
The lower the level at which a domain occurs, the more immune it
is to the requirement of naming uniqueness.
To satisfy the required distinction of simple names for proper
resolution at all levels, a naming authority needs to ensure the
simple name to be assigned distinct from those in the name server
databases at the endpoint naming domains within its domain. As an
example, for D to assign a simple name for X, it would need to
consult databases at A and K. It is, however, acceptable to have
simple names under domain A identical with those under K. Failure of
such distinct assignment of simple names by naming authority of one
domain would jeopardize the capability of simple name resolution for
entities within the subtree under that domain.
Su & Postel [Page 13]
RFC 819 August 1982;
APPENDIX C
Further Discussion of Name Service and Name Servers
The name service on a system should appear to the programmer of an
application program simply as a system call or library subroutine.
Within that call or subroutine there may be several types of methods
for resolving the name string into an address.
First, a local table may be consulted. This table may be a
complete table and may be updated frequently, or it may simply be
a cache of the few latest name to address mappings recently
determined.
Second, a call may be made to a name server to resolve the string
into a destination address.
Third, a call may be made to a name server to resolve the string
into a relay address.
Whenever a name server is called it may be a recursive server or an
interactive server.
If the server is recursive, the caller won't be able to tell if
the server itself had the information to resolve the query or
called another server recursively (except perhaps for the time it
takes).
If the server is iterative, the caller must be prepared for either
the answer to its query, or a response indicating that it should
call on a different server.
It should be noted that the main name service discussed in this memo
is simply a name string to address service. For some applications
there may be other services needed.
For example, even within the Internet there are several procedures
or protocols for actually transferring mail. One need is to
determine which mail procedures a destination host can use.
Another need is to determine the name of a relay host if the
source and destination hosts do not have a common mail procedure.
These more specialized services must be specific to each
application since the answers may be application dependent, but
the basic name to address translation is application independent.
Su & Postel [Page 14]
RFC 819 August 1982;
APPENDIX D
Further Discussion of Interoperability and Protocol Translations
The translation of protocols from one system to another is often
quite difficult. Following are some questions that stem from
considering the translations of addresses between mail systems:
What is the impact of different addressing environments (i.e.,
environments of different address formats)?
It is noted that the boundary of naming environment may or may not
coincide with the boundary of different mail systems. Should the
conversion of naming be independent of the application system?
The boundary between different addressing environments may or may
not coincide with that of different naming environments or
application systems. Some generic approach appears to be
necessary.
If the conversion of naming is to be independent of the
application system, some form of interaction appears necessary
between the interface module of naming conversion with some
application level functions, such as the parsing and modification
of message text.
To accommodate encryption, conversion may not be desirable at all.
What then can be an alternative to conversion?
Su & Postel [Page 15]
RFC 819 August 1982;
GLOSSARY
address
An address is a numerical identifier for the topological location
of the named entity.
name
A name is an alphanumeric identifier associated with the named
entity. For unique identification, a name needs to be unique in
the context in which the name is used. A name can be mapped to an
address.
complete (fully qualified) name
A complete name is a concatenation of simple names representing
the hierarchical relation of the named with respect to the naming
universe, that is it is the concatenation of the simple names of
the domain structure tree nodes starting with its own name and
ending with the top level node name. It is a unique name in the
naming universe.
partially qualified name
A partially qualified name is an abbreviation of the complete name
omitting simple names of the common ancestors of the communicating
parties.
simple name
A simple name is an alphanumeric identifier unique only within its
parent domain.
domain
A domain defines a region of jurisdiction for name assignment and
of responsibility for name-to-address translation.
naming universe
Naming universe is the ancestor of all network entities.
naming environment
A networking environment employing a specific naming convention.
Su & Postel [Page 16]
RFC 819 August 1982;
name service
Name service is a network service for name-to-address mapping.
name server
A name server is a network mechanism (e.g., a process) realizing
the function of name service.
naming authority
Naming authority is an administrative entity having the authority
for assigning simple names and responsibility for resolving naming
conflict.
parallel relations
A network entity may have one or more hierarchical relations with
respect to the naming universe. Such multiple relations are
parallel relations to each other.
multiple parentage
A network entity has multiple parentage when it is assigned a
simple name by more than one naming domain.
Su & Postel [Page 17]
RFC 819 August 1982;
REFERENCES
[1] F. Harary, "Graph Theory", Addison-Wesley, Reading,
Massachusetts, 1969.
[2] J. Postel, "Computer Mail Meeting Notes", RFC-805,
USC/Information Sciences Institute, 8 February 1982.
[3] J. Postel, "Simple Mail Transfer Protocol", RFC-821,
USC/Information Sciences Institute, August 1982.
[4] D. Crocker, "Standard for the Format of ARPA Internet Text
Messages", RFC-822, Department of Electrical Engineering, University
of Delaware, August 1982.
Su & Postel [Page 18]