Network Working Group K. Moore
Request for Comments: 3205 University of Tennessee
BCP: 56 February 2002
Category: Best Current Practice
On the use of HTTP as a Substrate
Status of this Memo
This document specifies an Internet Best Current Practices for the
Internet Community, and requests discussion and suggestions for
improvements. Distribution of this memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (2002). All Rights Reserved.
Abstract
Recently there has been widespread interest in using Hypertext
Transfer Protocol (HTTP) as a substrate for other applications-level
protocols. This document recommends technical particulars of such
use, including use of default ports, URL schemes, and HTTP security
mechanisms.
Recently there has been widespread interest in using Hypertext
Transfer Protocol (HTTP) [1] as a substrate for other applications-
level protocols. Various reasons cited for this interest have
included:
o familiarity and mindshare,
o compatibility with widely deployed browsers,
o ability to reuse existing servers and client libraries,
o ease of prototyping servers using CGI scripts and similar
extension mechanisms,
o ability to use existing security mechanisms such as HTTP digest
authentication [2] and SSL or TLS [3],
o the ability of HTTP to traverse firewalls, and
o cases where a server often needs to support HTTP anyway.
Moore Best Current Practice [Page 1]
RFC 3205 HTTP Layering February 2002
The Internet community has a long tradition of protocol reuse, dating
back to the use of Telnet [4] as a substrate for FTP [5] and SMTP
[6]. However, the recent interest in layering new protocols over
HTTP has raised a number of questions when such use is appropriate,
and the proper way to use HTTP in contexts where it is appropriate.
Specifically, for a given application that is layered on top of HTTP:
o Should the application use a different port than the HTTP default
of 80?
o Should the application use traditional HTTP methods (GET, POST,
etc.) or should it define new methods?
o Should the application use http: URLs or define its own prefix?
o Should the application define its own MIME-types, or use something
that already exists (like registering a new type of MIME-directory
structure)?
This memo recommends certain design decisions in answer to these
questions.
This memo is intended as advice and recommendation for protocol
designers, working groups, implementors, and IESG, rather than as a
strict set of rules which must be adhered to in all cases.
Accordingly, the capitalized key words defined in RFC 2119, which are
intended to indicate conformance to a specification, are not used in
this memo.
Despite the advantages listed above, it's worth asking the question
as to whether HTTP should be used at all, or whether the entire HTTP
protocol should be used.
HTTP started out as a simple protocol, but quickly became much more
complex due to the addition of several features unanticipated by its
original design. These features include persistent connections, byte
ranges, content negotiation, and cache support. All of these are
useful for traditional web applications but may not be useful for the
layered application. The need to support (or circumvent) these
features can add additional complexity to the design and
implementation of a protocol layered on top of HTTP. Even when HTTP
can be "profiled" to minimize implementation overhead, the effort of
specifying such a profile might be more than the effort of specifying
a purpose-built protocol which is better suited to the task at hand.
Moore Best Current Practice [Page 2]
RFC 3205 HTTP Layering February 2002
Even if existing HTTP client and server code can often be re-used,
the additional complexity of layering something over HTTP vs. using a
purpose-built protocol can increase the number of interoperability
problems.
Further, although HTTP can be used as the transport for a "remote
procedure call" paradigm, HTTP's protocol overhead, along with the
connection setup overhead of TCP, can make HTTP a poor choice. A
protocol based on UDP, or with both UDP and TCP variants, should be
considered if the payloads are very likely to be small (less than a
few hundred bytes) for the foreseeable future. This is especially
true if the protocol might be heavily used, or if it might be used
over slow or expensive links.
On the other hand, the connection setup overhead can become
negligible if the layered protocol can utilize HTTP/1.1's persistent
connections, and if the same client and server are likely to perform
several transactions during the time the HTTP connection is open.
Although HTTP appears at first glance to be one of the few "mature"
Internet protocols that can provide good security, there are many
applications for which neither HTTP's digest authentication nor TLS
are sufficient by themselves.
Digest authentication requires a secret (e.g., a password) to be
shared between client and server. This further requires that each
client know the secret to be used with each server, but it does not
provide any means of securely transmitting such secrets between the
parties. Shared secrets can work fine for small groups where
everyone is physically co-located; they don't work as well for large
or dispersed communities of users. Further, if the server is
compromised a large number of secrets may be exposed, which is
especially dangerous if the same secret (or password) is used for
several applications. (Similar concerns exist with TLS based clients
or servers - if a private key is compromised then the attacker can
impersonate the party whose key it has.)
TLS and its predecessor SSL were originally designed to authenticate
web servers to clients, so that a user could be assured (for example)
that his credit card number was not being sent to an imposter.
However, many applications need to authenticate clients to servers,
or to provide mutual authentication of client and server. TLS does
Moore Best Current Practice [Page 3]
RFC 3205 HTTP Layering February 2002
have a capability to provide authentication in each direction, but
such authentication may or may not be suitable for a particular
application.
Web browsers which support TLS or SSL are typically shipped with the
public keys of several certificate authorities (CAs) "wired in" so
that they can verify the identity of any server whose public key was
signed by one of those CAs. For this to work well, every secure web
server's public key has to be signed by one of the CAs whose keys are
wired into popular browsers. This deployment model works when there
are a (relatively) small number of servers whose identities can be
verified, and their public keys signed, by the small number of CAs
whose keys are included in a small number of different browsers.
This scheme does not work as well to authenticate millions of
potential clients to servers. It would take a much larger number of
CAs to do the job, each of which would need to be widely trusted by
servers. Those CAs would also have a more difficult time verifying
the identities of (large numbers of) ordinary users than they do in
verifying the identities of (a smaller number of) commercial and
other enterprises that need to run secure web servers.
Also, in a situation where there were a large number of clients
authenticating with TLS, it seems unlikely that there would be a set
of CAs whose keys were trusted by every server. A client that
potentially needed to authenticate to multiple servers would
therefore need to be configured as to which key to use with which
server when attempting to establish a secure connection to the
server.
For the reasons stated above, client authentication is rarely used
with TLS. A common technique is to use TLS to authenticate the
server to the client and to establish a private channel, and for the
client to authenticate to the server using some other means - for
example, a username and password using HTTP basic or digest
authentication.
For any application that requires privacy, the 40-bit ciphersuites
provided by some SSL implementations (to conform to outdated US
export regulations or to regulations on the use or export of
cryptography in other countries) are unsuitable. Even 56-bit DES
encryption, which is required of conforming TLS implementations, has
been broken in a matter of days with a modest investment in
resources. So if TLS is chosen it may be necessary to discourage use
of small key lengths, or of weak ciphersuites, in order to provide
adequate privacy assurance. If TLS is used to provide privacy for
passwords sent by clients then it is especially important to support
longer keys.
Moore Best Current Practice [Page 4]
RFC 3205 HTTP Layering February 2002
None of the above should be taken to mean that either digest
authentication or TLS are generally inferior to other authentication
systems, or that they are unsuitable for use in other applications
besides HTTP. Many of the limitations of TLS and digest
authentication also apply to other authentication and privacy
systems. The point here is that neither TLS nor digest
authentication is a "magic pixie dust" solution to authentication or
privacy. In every case, an application's designers must carefully
determine the application's users' requirements for authentication
and privacy before choosing an authentication or privacy mechanism.
Note also that TLS can be used with other TCP-based protocols, and
there are SASL [7] mechanisms similar to HTTP's digest
authentication. So it is not necessary to use HTTP in order to
benefit from either TLS or digest-like authentication. However, HTTP
APIs may already support TLS and/or digest.
One oft-cited reason for the use of HTTP is its ability to pass
through proxies, firewalls, or network address translators (NATs).
One unfortunate consequence of firewalls and NATs is that they make
it harder to deploy new Internet applications, by requiring explicit
permission (or even a software upgrade of the firewall or NAT) to
accommodate each new protocol. The existence of firewalls and NATs
creates a strong incentive for protocol designers to layer new
applications on top of existing protocols, including HTTP.
However, if a site's firewall prevents the use of unknown protocols,
this is presumably a conscious policy decision on the part of the
firewall administrator. While it is arguable that such policies are
of limited value in enhancing security, this is beside the point -
well-known port numbers are quite useful for a variety of purposes,
and the overloading of port numbers erodes this utility. Attempting
to circumvent a site's security policy is not an acceptable
justification for doing so.
It would be useful to establish guidelines for "firewall-friendly"
protocols, to make it easier for existing firewalls to be compatible
with new protocols.
o When considering payload size and traffic patterns, is HTTP an
appropriate transport for the anticipated use of this protocol?
Moore Best Current Practice [Page 5]
RFC 3205 HTTP Layering February 2002
(In other words: will the payload size be worth the overhead
associated with TCP and HTTP? Or will the application be able to
make use of HTTP persistent connections to amortize the cost of
that overhead over several requests?)
o Is this new protocol usable by existing web browsers without
modification?
(For example: Is the request transmitted as if it were a filled-in
HTML form? Is the response which is returned viewable from a web
browser, say as HTML?)
o Are the existing HTTP security mechanisms appropriate for the new
application?
o Are HTTP status codes and the HTTP status code paradigm suitable
for this application? (see section 8)
o Does the server for this application need to support HTTP anyway?
IANA has reserved TCP port number 80 for use by HTTP. It would not
be appropriate for a substantially new service, even one which uses
HTTP as a substrate, to usurp port 80 from its traditional use. A
new use of HTTP might be considered a "substantially new service",
thus requiring a new port, if any of the following are true:
o The "new service" and traditional HTTP service are likely to
reference different sets of data, even when they both operate on
the same host.
o There is a good reason for the "new service" to be implemented by
a separate server process, or separate code, than traditional HTTP
service on the same host, at least on some platforms.
o There is a good reason to want to easily distinguish the traffic
of the "new service" from traditional HTTP, e.g., for the purposes
of firewall access control or traffic analysis.
o If none of the above are true, it is arguable that the new use of
HTTP is an "extension" to traditional HTTP, rather than a "new
service". Extensions to HTTP which share data with traditional
HTTP services should probably define new HTTP methods to describe
those extensions, rather than using separate ports. If separate
ports are used, there is no way for a client to know whether they
are separate services or different ways of accessing the same
underlying service.
Moore Best Current Practice [Page 6]
RFC 3205 HTTP Layering February 2002
A number of different URL schemes are in widespread use and many more
are in the process of being standardized. In practice, the URL
scheme not only serves as a "tag" to govern the interpretation of the
remaining portion of the URL, it also provides coarse identification
of the kind of resource or service which is being accessed. For
example, web browsers typically provide a different response when a
user mouse-clicks on an "http" URL, than when the user clicks on a
"mailto" URL.
Some criteria that might be used in making this determination are:
o Whether this URL scheme is likely to become widely used, versus
used only in limited communities or by private agreement.
o Whether a new "default port" is needed. If reuse of port 80 is
not appropriate (see above), a new "default port" is needed. A
new default port in turn requires that a new URL scheme be
registered if that URL scheme is expected to be widely used.
Explicit port numbers in URLs are regarded as an "escape hatch",
not something for use in ordinary circumstances.
o Whether use of the new service is likely to require a
substantially different setup or protocol interaction with the
server, than ordinary HTTP service. This could include the need
to request a different type of service from the network, or to
reserve bandwidth, or to present different TLS authentication
credentials to the server, or different kind of server
provisioning, or any number of other needs.
o Whether user interfaces (such as web browsers) are likely to be
able to exploit the difference in the URL prefix to produce a
significant improvement in usability.
According to the rules in [8] the "http:" URI is part of the "IETF
Tree" for URL scheme names, and IETF is the maintainer of the "IETF
Tree". Since IESG is the decision-making body for IETF, IESG has the
authority to determine whether a resource accessed by a protocol that
is layered on top of HTTP, should use http: or some other URL prefix.
Note that the convention of appending an "s" to the URL scheme to
mean "use TLS or SSL" (as in "http:" vs "https:") is nonstandard and
of limited value. For most applications, a single "use TLS or SSL"
bit is not sufficient to adequately convey the information that a
client needs to authenticate itself to a server, even if it has the
proper credentials. For instance, in order to ensure that adequate
security is provided with TLS an application may need to be
Moore Best Current Practice [Page 7]
RFC 3205 HTTP Layering February 2002
configured with a list of acceptable ciphersuites, or with the client
certificate to be used to authenticate to a particular server. When
it is necessary to specify authentication or other connection setup
information in a URL these should be communicated in URL parameters,
rather than in the URL prefix.
Since HTTP uses the MIME media type system [9] to label its payload,
many applications which layer on HTTP will need to define, or select,
MIME media types for use by that application. Especially when using
a multipart structure, the choice of media types requires careful
consideration. In particular:
o Should some existing framework be used, such as text/directory
[10], or XML [11,12], or should the new content-types be built
from scratch? Just as with HTTP, it's useful if code can be
reused, but protocol designers should not be over-eager to
incorporate a general but complex framework into a new protocol.
Experience with ASN.1, for example, suggests that the advantage of
using a general framework may not be worth the cost.
o Should MIME multipart or message types be allowed? This can be an
advantage if it is desirable to incorporate (for example) the
multipart/alternative construct or the MIME security framework.
On the other hand, these constructs were designed specifically for
use in store-and-forward electronic mail systems, and other
mechanisms may be more appropriate for the application being
considered.
The point here is that a decision to use MIME content-type names
to describe protocol payloads (which is generally desirable if the
same payloads may appear in other applications) does not imply
that the application must accept arbitrary MIME content-types,
including MIME multipart or security mechanisms. Nor does it
imply that the application must use MIME syntax or that it must
recognize or even tolerate existing MIME header fields.
o If the same payload is likely to be sent over electronic mail, the
differences between HTTP encoding of the payload and email
encoding of the payload should be minimized. Ideally, there
should be no differences in the "canonical form" used in the two
environments. Text/* media types can be problematic in this
regard because MIME email requires CRLF for line endings of text/*
body parts, where HTTP traditionally uses LF only.
Moore Best Current Practice [Page 8]
RFC 3205 HTTP Layering February 2002
o A MIME content-type label describes the nature of the object being
labeled. It does not describe, and should not be used to
describe, the semantics which should be applied when the object is
received. For instance, the transmission of an object with a
particular content-type using HTTP POST, should not be taken as a
request for some operation based solely on the type. The request
should be separate from the content-type label and it should be
explicit.
When it is necessary for a protocol layered on HTTP to allow
different operations on the same type of object, this can be
communicated in a number of different ways: HTTP methods, HTTP
request-URI, HTTP request headers, the MIME Content-Disposition
header field, or as part of the payload.
It has been suggested that a new service layered on top of HTTP
should define one or more new HTTP methods, rather than allocating a
new port. The use of new methods may be appropriate, but is not
sufficient in all cases. The definition of one or more new methods
for use in a new protocol, does not by itself alleviate the need for
use of a new port, or a new URL type.
As mentioned earlier, one of the primary reasons for the use of HTTP
as a substrate for new protocols, is to allow reuse of existing HTTP
client, server, or proxy code. However, HTTP was not designed for
such layering. Existing HTTP client and code may have "http"
assumptions wired into them. For instance, client libraries and
proxies may expect "http:" URLs, and clients and servers may send
(and expect) "HTTP/1.1", in requests and responses, as opposed to the
name of the layered protocol and its version number.
Existing client libraries may not understand new URL types. In order
to get a new HTTP-layered application client to work with an existing
client library, it may be necessary for the application to convert
its URLs to an "http equivalent" form. For instance, if service
"xyz" is layered on top of HTTP using port ###, the xyz client may
need, when invoking an HTTP client library, to translate its URLs
from "xyz://host/something" format to "http://host:###/something" for
the purpose of calling that library. This should be done ONLY when
calling the HTTP client library - such URLs should not be used in
other parts of the protocol, nor should they be exposed to users.
Moore Best Current Practice [Page 9]
RFC 3205 HTTP Layering February 2002
Note that when a client is sending requests directly to an origin
server, the URL prefix ("http:") is not normally sent. So
translating xyz: URLs to http: URLs when calling the client library
should not actually cause http: URLs to be sent over the wire. But
when the same client is sending requests to a proxy server, the
client will normally send the entire URL (including the http: prefix)
in those requests. The proxy will remove the http: prefix when the
request is communicated to the origin server.
Existing HTTP client libraries and servers will transmit "HTTP/1.1"
(or a different version) in requests and responses. To facilitate
reuse of such libraries and servers by a new protocol, such a
protocol may therefore need to transmit and accept "HTTP/1.1" rather
than its own protocol name and version number. Designers of
protocols which are layered on top of HTTP should explicitly choose
whether or not to accept "HTTP/1.1" in protocol exchanges.
For certain applications it may be necessary to require or limit use
of certain HTTP features, for example, to defeat caching of responses
by proxies. Each protocol layered on HTTP must therefore specify the
specific way that HTTP will be used, and in particular, how the
client and server should interact with HTTP proxies.
HTTP's three-digit status codes were designed for use with
traditional HTTP applications (e.g., document retrieval, forms-based
queries), and are unlikely to be suitable to communicate the
specifics of errors encountered in dissimilar applications. Even
when it seems like there is a close match between HTTP status codes
and the codes needed by the application, experience with reuse of
other protocols indicates that subtle variations in usage are likely;
and that this is likely to degrade interoperability of both the
original protocol (in this case HTTP) and any layered applications.
HTTP status codes therefore should not be used to indicate subtle
errors of layered applications. At most, the "generic" HTTP codes
200 (for complete success) and 500 (for complete failure) should be
used to indicate errors resulting from the content of the request
message-body. Under certain circumstances, additional detail about
the nature of the error can then be included in the response
message-body. Other status codes than 200 or 500 should only appear
if the error was detected by the HTTP server or by an intermediary.
A layered application should not define new HTTP status codes. The
set of available status codes is small, conflicts in code assignment
between different layered applications are likely, and they may be
needed by future versions of, or extensions to, mainstream HTTP.
Moore Best Current Practice [Page 10]
RFC 3205 HTTP Layering February 2002
Use of HTTP's error codes is problematic when the layered application
does not share same notion of success or failure as HTTP. The
problem exists when the client does not connect directly to the
origin server, but via one or more HTTP caches or proxies. (Since
the ability of HTTP to communicate through intermediaries is often
the primary motivation for reusing HTTP, the ability of the
application to operate in the presence of such intermediaries is
considered very important.) Such caches and proxies will interpret
HTTP's error codes and may take additional action based on those
codes. For instance, on receipt of a 200 error code from an origin
server (and under other appropriate conditions) a proxy may cache the
response and re-issue it in response to a similar request. Or a
proxy may modify the result of a request which returns a 500 error
code in order to add a "helpful" error message. Other response codes
may produce other behaviors.
A few guidelines are therefore in order:
o A layered application should use appropriate HTTP error codes to
report errors resulting from information in the HTTP request-line
and header fields associated with the request. This request
information is part of the HTTP protocol and errors which are
associated with that information should therefore be reported
using HTTP protocol mechanisms.
o A layered application for which all errors resulting from the
message-body can be classified as either "complete success" or
"complete failure" may use 200 and 500 for those conditions,
respectively. However, the specification for such an application
must define the mechanism which ensures that its successful (200)
responses are not cached by intermediaries, or demonstrate that
such caching will do no harm; and it must be able to operate even
if the message-body of an error (500) response is not transmitted
back to the client intact.
o A layered application may return a 200 response code for both
successfully processed requests and errors (or other exceptional
conditions) resulting from the request message-body (but not from
the request headers). Such an application must return its error
code as part of the response message body, and the specification
for that application protocol must define the mechanism by which
the application ensures that its responses are not cached by
intermediaries. In this case a response other than 200 should be
used only to indicate errors with, or the status of, the HTTP
protocol layer (including the request headers), or to indicate the
inability of the HTTP server to communicate with the application
server.
Moore Best Current Practice [Page 11]
RFC 3205 HTTP Layering February 2002
o A layered application which cannot operate in the presence of
intermediaries or proxies that cache and/or alter error responses,
should not use HTTP as a substrate.
1. All protocols should provide adequate security. The security
needs of a particular application will vary widely depending on
the application and its anticipated use environment. Merely using
HTTP and/or TLS as a substrate for a protocol does not
automatically provide adequate security for all environments, nor
does it relieve the protocol developers of the need to analyze
security considerations for their particular application.
2. New protocols - including but not limited to those using HTTP -
should not attempt to circumvent users' firewall policies,
particularly by masquerading as existing protocols.
"Substantially new services" should not reuse existing ports.
3. In general, new protocols or services should not reuse http: or
other URL schemes.
4. Each new protocol specification that uses HTTP as a substrate
should describe the specific way that HTTP is to be used by that
protocol, including how the client and server interact with
proxies.
5. New services should follow the guidelines in section 8 regarding
use of HTTP status codes.
Much of this document is about security. Section 2.3 discusses
whether HTTP security is adequate for the needs of a particular
application, section 2.4 discusses interactions between new HTTP-
based protocols and firewalls, section 3 discusses use of separate
ports so that firewalls are not circumvented, and section 4 discusses
the inadequacy of the "s" suffix of a URL prefix for specifying
security levels.
[1] Fielding, R., Gettys, J., Mogul, J., Frystyk, H., Masinter, L.,
Leach, P. and T. Berners-Lee, "Hypertext Transfer Protocol --
HTTP/1.1", RFC 2616, June 1999.
Moore Best Current Practice [Page 12]
RFC 3205 HTTP Layering February 2002
[2] Franks, J., Hallam-Baker, P., Hostetler, J., Lawrence, S.,
Leach, P., Luotonen, A. and L. Stewart, "HTTP Authentication:
Basic and Digest Access Authentication", RFC 2617, June 1999.
[3] Dierks, T. and C. Allen, "The TLS Protocol Version 1.0", RFC
2246, January 1999.
[4] Postel, J. and J. Reynolds, "Telnet Protocol Specification",
STD 8, RFC 854, May 1983.
[5] Postel, J. and J. Reynolds, "File Transfer Protocol", STD 9,
RFC 959, October 1985.
[6] Klensin, J., "Simple Mail Transfer Protocol", RFC 2821, April
2001.
[7] Myers, J., "Simple Authentication and Security Layer (SASL)",
RFC 2222, October 1997.
[8] Petke, R. and I. King, "Registration Procedures for URL Scheme
Names", BCP 35, RFC 2717, November 1999.
[9] Freed, N. and N. Borenstein, "Multipurpose Internet Mail
Extensions (MIME) Part Two: Media Types", RFC 2046, November
1996.
[10] Howes, T., Smith, M. and F. Dawson, "A MIME Content-Type for
Directory Information", RFC 2425, September 1998.
[11] Bray, T., Paoli, J. and C. Sperberg-McQueen, "Extensible Markup
Language (XML)" World Wide Web Consortium Recommendation REC-
xml-19980210, February 1998. http://www.w3.org/TR/1998/REC-
xml-19980210.
[12] Murata, M., St. Laurent, S. and D. Kohn, "XML Media Types", RFC
3023, January 2001.
Keith Moore
University of Tennessee
Computer Science Department
1122 Volunteer Blvd, Suite 203
Knoxville TN, 37996-3450
USA
EMail: moore@cs.utk.edu
Moore Best Current Practice [Page 13]
RFC 3205 HTTP Layering February 2002
Copyright (C) The Internet Society (2002). All Rights Reserved.
This document and translations of it may be copied and furnished to
others, and derivative works that comment on or otherwise explain it
or assist in its implementation may be prepared, copied, published
and distributed, in whole or in part, without restriction of any
kind, provided that the above copyright notice and this paragraph are
included on all such copies and derivative works. However, this
document itself may not be modified in any way, such as by removing
the copyright notice or references to the Internet Society or other
Internet organizations, except as needed for the purpose of
developing Internet standards in which case the procedures for
copyrights defined in the Internet Standards process must be
followed, or as required to translate it into languages other than
English.
The limited permissions granted above are perpetual and will not be
revoked by the Internet Society or its successors or assigns.
This document and the information contained herein is provided on an
"AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING
TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING
BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION
HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF
MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
Acknowledgement
Funding for the RFC Editor function is currently provided by the
Internet Society.
Moore Best Current Practice [Page 14]