Re: [http-auth] WGLC on the MutualAuth drafts

[Julian Reschke <[e-mail]>]

Hi there,

find below some unstructured (sorry for that) feedback on the extension 
draft.

I believe this contains lots of interesting ideas, but it'll need more 
work on the details; in particular how to deploy it in practice (such 
as: do we need a way for clients to signal that they implement this spec?).



Abstract

    This document specifies extensions of HTTP authentication framework

s/of/for the/

    for interactive clients.  Recently, fundamental features of HTTP-

Maybe: "currently"?

    level authentication are insufficient for complex requirements of
    various Web-based applications.  This forces these applications to
    implement their own authentication frameworks using HTML Forms and
    other means, which becomes one of the hurdles against introducing

"by means like HTML forms"

    secure authentication mechanisms handled jointly by servers and user-
    agent.  The extended framework fills gaps between Web application
    requirements and HTTP authentication provisions to solve the above
    problems, while maintaining compatibility against existing Web and
    non-Web uses of HTTP authentications.

s/against/with/

1.  Introduction

    The document proposes several extensions to the current HTTP

"This document defines..."

    authentication framework, to provide functionality comparable with
    current widely-used form-based Web authentication.  A majority of the
    recent websites on the Internet use custom application-layer
    authentication implementations using Web forms.  The reasons for
    these may vary, but many people believe that the current HTTP Basic
    and Digest authentication methods do not have enough functionality
    (including good user interfaces) to support most realistic Web-based
    applications.  However, this method is very weak against phishing and

clarify "this method"...

    other attacks, because all behavior of the authentication is

"...against attacks like phishing...."

1.1.  Terminology

    The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
    "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
    "OPTIONAL" in this document are to be interpreted as described in
    [RFC2119].

    The terms "encouraged" and "advised" are used for suggestions that do
    not constitute "SHOULD"-level requirements.  People MAY freely choose
    not to include the suggested items regarding [RFC2119], but complying
    with those suggestions would be a best practice; it will improve the
    security, interoperability, and/or operational performance.

Not sure what the "regarding [RFC2119]" is about...


2.  Definitions

2.1.  Terms for describing authentication protocol flow

    HTTP Authentication defined in [RFC7235] may involve several pairs of
    HTTP requests/responses.  Throughout this document, the following
    terms are used to categorize those messages: for requests,

    1) A non-authenticating request is a request not attempting any
       authentication: a request without any Authorization header.

s/header/header field/ (throughout)


    3) A successfully-authenticated response is a response for an
       authenticating request meaning that the authentication attempt was
       granted.  (Note: if the authentication scheme used does not use an
       Authentication-Info header, it may be indistinguishable from a
       non-authenticated response.)

...avoid lowercase BCP14 keywords (throughout). Here maybe s/may/can/.


    5) A negatively-authenticated response is a response for an
       authenticating request which means that the authentication attempt
       was declined and can not continue without a different set of
       authentication credentials.  Clients typically erase memory of the
       active credentials and ask the user for other ones.

       Usually the format of these responses are as same as the one for
       authentication-initializing responses.  Client can distinguish

s/Client/Clients/

       negatively-authenticated responses from authentication-
       initializing responses by comparing the protection spaces
       contained in the request and in the response.

    Figure 1 shows a state diagram of generic HTTP authentication with
    the above message categorization.  Note that many authentication
    schemes use only a subset of the transitions described on the
    diagram.  Labels in the figure show the abbreviated names of response
    types.


2.2.  Syntax Notation

    This specification uses an extended BNF syntax defined in [RFC7230].

s/BNF/ABNF/

Also, 7230 only defines extensions to 5234, so that should be referenced 
here as well.

    Additionally, this specification uses the following syntax
    definitions as a refinement for token and the right-hand-side of
    auth-param in [RFC7235].  (Note: these definitions are consistent
    with those in [I-D.ietf-httpauth-mutual].)

     bare-token        = 1*(%x30-39 / %x41-5A / %x61-7A / "-" / "_")
     extension-token   = "-" bare-token 1*("." bare-token)
     extensive-token   = bare-token / extension-token
     integer           = "0" / (%x31-39 *%x30-39)      ; no leading zeros

                Figure 2: the BNF syntax for common notations

    Extensive-tokens are used in this protocol where the set of
    acceptable tokens may include private extensions.  Any private
    extensions of this protocol MUST use extension-tokens with the format
    "-<token>.<domain-name>", where <domain-name> is a valid (sub-)domain
    name on the Internet owned by the party who defines the extension.

a) it would be good to state how bare-token is different from token, and 
why it is needed.

b) this introduces a syntax for private extensions that is awfully 
similar to the "x-" convention that we just deprecated a few years ago...


3.  Optional Authentication

    The Optional-WWW-Authenticate header enables a non-mandatory
    authentication, which is not possible under the current HTTP
    authentication mechanism.  In several Web applications, users can

Insert paragraph break here.

    access the same contents as both a guest user and an authenticated
    user.  In most Web applications, This functionality is implemented

s/THis/this/

    using HTTP cookies [RFC6265] and custom form-based authentication.
    The new authentication method using this message will provide a
    replacement for these authentication systems.

    Servers MAY send HTTP non-interim responses containing the
    Optional-WWW-Authenticate header as a replacement of a 401 response
    when it the response is authentication-initializing.  The
    Optional-WWW-Authenticate header MUST NOT sent on 401 responses (i.e.
    a usual WWW-Authenticate header MUST be used on 401 responses.)

       HTTP/1.1 200 OK
       Optional-WWW-Authenticate: Basic realm="xxxx"


     Optional-WWW-Authenticate = 1#challenge

          Figure 3: BNF syntax for Optional-WWW-Authenticate header

I'm not totally convinced that a new header field is needed here? Is 
there a reason why sending "WWW-Authenticate" with a 2xx response 
wouldn't work?

(And yes, we probably discussed this several times already, but it would 
be good to summarize the outcome over here)

    The challenges contained in the Optional-WWW-Authenticate header are
    the same as those for a 401 responses corresponding to the same
    request.  For authentication-related matters, an optional
    authentication request will have the same meaning as a 401 message
    with a corresponding WWW-Authenticate header (as an authentication-
    initializing response).  (The behavior for other matters, such as
    caching, MAY be different between the optional authentication and 401
    messages.)

This "MAY" is really vague; the behavior needs to be spelled out 
completely, I think.

    Support of this header is OPTIONAL; Clients MAY also choose any set
    of authentication schemes for which optional authentication is
    supported (in other words, its support MAY be scheme-dependent).

Hard to understand. Does this mean: "clients are free to implement this 
only for selected schemes"?

    However, some authentication schemes MAY require mandatory/
    recommended support for this header, so that server-side applications
    MAY assume that clients supporting such schemes are likely to support
    the extension as well.

"New authentication schemes though can make support of optional auth 
mandatory" ?


4.  Authentication-Control header

     Authentication-Control = 1#Auth-Control-Entry
     Auth-Control-Entry     = auth-scheme 1*SP 1#auth-control-param
     auth-control-param     = extensive-token BWS "=" BWS token
                            / extensive-token "*" BWS "=" BWS ext-value
     ext-value              = <see RFC 5987, Section 3.2>

       Figure 4: the BNF syntax for the Authentication-Control header

FYI: I'm in the process of revising RFC 5987, and that ABNF production 
is going to be removed. Seems we need to coordinate here.

    Among the multiple entries in the header, the meaningful entries in
    the header are those corresponding to an auth-scheme and a realm (if
    any), for which "the authentication process is being performed, or
    going to be performed".  In more detail,

    (1)  If the response is either an authentication-initializing
         response or a negatively-authenticated response, there may be
         multiple challenges in the WWW-Authenticate header (or the
         Optional-WWW-Authenticate header defined in this extension),
         each of which corresponds to a different scheme and realm.  In
         this case, the client has a choice on the scheme and realm they
         will use to authenticate.  Only the entry in the
         &Authentication-Control; header corresponding to that scheme and
         realm are meaningful.

&Authentication-Control;?

Also, maybe use "relevant" instead of "meaningful"?

    Each entry contains one or more parameters, each of which is a name-
    value pair.  The name of each parameter MUST be an extensive-token.
    Clients MUST ignore any unknown parameters contained in this header.
    The entries for the same auth-scheme and the realm MUST NOT contain
    duplicated parameters for the same name.

...otherwise? Do we want to state what to do with invalid values?

    The type of parameter value depends on the parameter name as defined
    in the following subsections.  Regardless of the type, however, the
    recipients SHOULD accept both quoted and unquoted representations of

MUST...

    values as defined in HTTP.  If the parameter is defined to have a
    string value, implementations SHOULD send the parameter in a quoted
    form or an ext-value form (see Section 4.1).  If the parameter is

Why, it it doesn't require quoting?

    defined as a token (or similar) or an integer, the value SHOULD
    follow the corresponding ABNF syntax after possible unquoting of the
    quoted-string value (as defined in HTTP), and SHOULD be sent in a
    unquoted form.

Again, why? ...don't make requirements that do not help on the wire, and 
which make it harder to use generic libraries.

    Server-side applications SHOULD be aware that any parameters
    contained in this header MAY be ignored by clients.  Also, even when

What does it mean to "be aware" of it? Don't need a SHOULD here...

    a client accepts this header, users may always be able to circumvent

s/may always be/are/

    the semantics of this header.  Therefore, if this header is used for
    security purposes, its use MUST be limited to providing some non-
    fundamental additional security measures valuable for end-users (such
    as client-side log-out for protecting against console takeover).
    Server-side applications MUST NOT rely on the use of this header for
    protecting server-side resources.

+1.

    Note: The header syntax allows servers to specify Authentication-
    Control for multiple authentication schemes, either as multiple
    occurrences of this header or as a combined single header (see
    Section 3.2.2 of [RFC7230] for rationale).  The same care as for
    parsing multiple authentication challenges SHALL be taken.

s/SHALL/needs to/

4.1.  Non-ASCII extended header parameters

    Parameters contained in the Authentication-Control header MAY be
    extended to ISO 10646-1 values using the framework described in

s/ISO 16046-1/non-ASCII/

    [RFC5987].  All servers and clients MUST be capable of receiving and
    sending values encoded in [RFC5987] syntax.

    If a value to be sent contains only ASCII characters, the field MUST
    be sent using plain RFC 7235 syntax.  The syntax as extended by RFC
    5987 MUST NOT be used in this case.

IMHO it would be better to use the ABNF terms instead of the RFC numbers 
here.

    If a value (except the "realm" header) contains one or more non-ASCII
    characters, the parameter SHOULD be sent using the ext-value syntax
    defined in Section 3.2 of [RFC5987].  Such a parameter MUST have a
    charset value of "UTF-8", and the language value MUST always be
    omitted (have an empty value).  The same parameter MUST NOT be sent
    more than once, regardless of the used syntax.

The "realm" parameter exception needs a rational.

    For example, a parameter "username" with value "Renee of France"
    SHOULD be sent as < username="Renee of France" >.  If the value is
    "Ren<e acute>e of France", it SHOULD be sent as < username*=UTF-
    8''Ren%C3%89e%20of%20France > instead.

Unfortunate line breaking;  I'd move the examples into <artwork> elements.

4.2.  Auth-style parameter

    Authentication-Control: Digest realm="protected space",
            auth-style=modal

(The above is an example, right? Maybe say so...).

    The default behavior for clients is implementation-dependent, and
    clients MAY choose different defaults for different authentication
    schemes.  The proposed default behavior is "modal" for all
    authentication schemes, but specifications for authentication schemes
    MAY propose a different default.

Do we really need to allow all this variation?

    The above two different methods of authentication may introduce a
    observable difference of semantics when the response contains state-
    changing side effects; for example, it may change whether Cookie

s/may change whether/can affect how/

    headers [RFC6265] in 401 responses are processed or not.  However,
    the server applications SHOULD NOT depend on both existence and non-
    existence of such side effects.

remove "and non-existence"

4.3.  Location-when-unauthenticated parameter

    Authentication-Control: Mutual realm="auth-space-1",
        location-when-unauthenticated="http://www.example.com/login.html"

(see above)

    The parameter "location-when-unauthenticated" specifies a location
    where any unauthenticated clients should be redirected to.  This
    header may be used, for example, when there is a central login page
    for the entire Web application.  The value of this parameter is a
    string that contains an absolute URL location.  Senders MUST always
    send an absolute URL location.  If a received URL is not absolute,
    the clients SHOULD either ignore it or consider it a relative URL
    from the current location.

Too much variation. There's no reason to require an absolute URL here, 
in particular if we have a SHOULD-level requirement to accept relative 
references as well.

    This parameter MAY be used with a 401 response for an authentication-
    initializing response.  It can also be contained, although this is
    NOT RECOMMENDED, in a positive response with an
    Optional-WWW-Authenticate header.  The clients MUST ignore this
    parameter when a response is either successfully-authenticated or
    intermediately-authenticated.  The clients SHOULD ignore this
    parameter when a response is a negatively-authenticated one (the case
    is unlikely to happen, though).

Is this really a SHOULD?

    When a client receives an authentication-initiating response with
    this parameter, if the client has to ask users for authentication

s/if/and if/

    credentials, the client will treat the entire response as if it were
    a 303 "See Other" response with a Location header that contains the
    value of this parameter (i.e., client will be redirected to the
    specified location with a GET request).  Unlike a normal 303
    response, if the client can process authentication without the user's
    interaction, this parameter MUST be ignored.

4.4.  No-auth parameter

    Authentication-Control: Basic realm="entrance", no-auth=true

(see above)

    The parameter "no-auth" is a variant of the
    location-when-unauthenticated parameter; it specifies that new
    authentication attempts are not to be performed on this location in
    order to improve the user experience, without specifying the
    redirection on the HTTP level.  This header may be used, for example,

s/may/can/

    when there is a central login page for the entire Web application,
    and when an explicit user interaction with the Web content is desired
    before authentications.  The value of this parameter MUST be a token
    "true".  If the value is incorrect, client MAY ignore this parameter.

    This parameter MAY be used with authentication-initiating responses.
    It can also be contained, although this is NOT RECOMMENDED, in a
    positive response with an Optional-WWW-Authenticate header.  The
    clients MUST ignore this parameter when a response is either
    successfully-authenticated or intermediately-authenticated.  The
    clients SHOULD ignore this parameter when a response is a negatively-
    authenticated one (the case is unlikely to happen, though).

(see above)

    When a client receives an authentication-initiating response with
    this parameter, if the client has to ask users for authentication
    credentials, the client will ignore the WWW-Authenticate header
    contained in the response and treat the whole response as a normal
    negative 4xx-class response instead of giving the user an opportunity
    to start authentication.  If the client can process authentication
    without the user's interaction, this parameter MUST be ignored.

    This parameter SHOULD NOT be used along with the
    location-when-unauthenticated parameter.  If both were supplied,
    clients MAY choose which one is to be honored.

    This parameter SHOULD NOT be used as a security measure to prevent
    authentication attempts, as it is easily circumvented by users.  This
    parameter SHOULD be used solely for improving user experience of Web
    applications.

4.5.  Location-when-logout parameter

    Authentication-Control: Digest realm="protected space",
        location-when-logout="http://www.example.com/byebye.html"

(see above)

    The parameter "location-when-logout" specifies a location where the
    client is to be redirected when the user explicitly requests a
    logout.  The value of this parameter MUST be a string that contains
    an absolute URL location.  If a given URL is not absolute, the
    clients MAY consider it a relative URL from the current location.

(see above)

    This parameter MAY be used with successfully-authenticated responses.
    If this parameter is contained in other kinds of responses, the
    clients MUST ignore this parameter.

    When the user requests termination of an authentication period, and
    if the client currently displays a page supplied by a response with
    this parameter, the client will be redirected to the specified
    location by a new GET request (as if it received a 303 response).
    The log-out operation (e.g. erasing memories of user name,
    authentication credential and all related one-time credentials such
    as nonce or keys) SHOULD occur before processing a redirection.

    When the user requests termination of an authentication period, if

s/if/and if/

    the client supports this parameter but the server response does not
    contain this parameter, the client's RECOMMENDED behavior is as
    follows: if the request corresponding to the current content was safe
    (e.g.  GET), reload the page without the authentication credential.
    If the request was non-idempotent (e.g.  POST), keep the current
    content as-is and simply forget the authentication status.  The

Is it intentional that the alternatives are "safe" and "non-idempotent"? 
What about, for instance, about DELETE?

    client SHOULD NOT replay a non-idempotent request without the user's
    explicit approval.

    Web applications are encouraged to send this parameter with an
    appropriate value for any responses (except those with redirection
    (3XX) statuses) for non-GET requests.

General comment: it makes me a bit uneasy that we define response header 
fields that essentially change the semantics t a redirect, without the 
status code saying so. There's a long history of redirect-related issues 
in browsers (such as what to change the method to, how to repeat them, 
what header fields to include etc), so adding a new way to redirect that 
doesn't look like a redirect might be a concern.

4.6.  Logout-timeout parameter

    Authentication-Control: Basic realm="entrance", logout-timeout=300

(see above)

4.7.  Username parameter

    Authentication-Control: Basic realm="configuration", username="admin"

(see above)

    If the used authentication scheme requires a specific style of text
    preparation for the user name (e.g., PRECIS string preparation or
    Unicode normalization), the specified user name SHOULD follow such
    requirements.

Needs references. Also: if the auth scheme has a "MUST" requirement wrt 
names, why would we only say "SHOULD" here?

5.  Usage examples (informative)

Drop "(informative)". That's implied by "examples".

    This section shows some examples for applying this extension to
    typical websites which are using Forms and cookies for managing
    authentication and authorization.  The content of this section is not
    normative and for illustrative purposes only.

    We assume that all features described in the previous sections are
    implemented in clients (Web browsers).  We also assume that browsers

That's a bold assumption :-)

    will have a user interface which allows users to deactivate (log-out
    from) current authentication sessions.  If this assumption is not
    value, the text below provides another approach with de-

/not value/not valid/ I assume?

    authentication pages used instead of such a UI.


5.4.  Parallel deployment with Form/Cookie authentications

    In some transition periods, sites may need to support both HTTP-layer
    and form-based authentication.  The following example shows one way
    to achieve that.

    o  If Cookies are used even for HTTP-authenticated users, each
       session determined by Cookies should identify which authentication
       has been used for the session.

SHOULD?

    o  First, set up any of the above settings for enabling HTTP-layer
       authentication.

    o  For unauthenticated users, add the following things to the Web
       pages, unless the client supports this extension and HTTP-level
       authentication.

       *  For non-mandatory authenticated pages, put a link to Form-based
          authenticated pages.

       *  For mandatory authenticated pages, either put a link to Form-
          based authenticated pages, or put a HTML-level redirection
          (using META element) to such pages.

http-equiv=refresh? If so, please say so.

    o  In Form-based authenticated pages, if users are not authenticated,
       the page may have a diversion for HTTP-level authentication by
       "location-when-unauthenticated" setting.

Diversion?

    o  Users are identified to authorization and content customization by
       the following logic.

       *  First, check the result of the HTTP-level authentication.  If
          there is a Cookie session tied to a specific user, both should
          match.

       *  If the user is not authenticated on the HTTP-level, use the
          conventional Form-based method to determine the user.

       *  If there is a Cookie tied to HTTP authentication, but there is
          no corresponding HTTP authentication result, that session will
          be discarded (because it means that authentication is
          deactivated by the corresponding user).


7.  IANA Considerations

    This document defines two new entries for the "Permanent Message
    Header Field Names" registry.

    +---------------------------+----------+----------------------------+
    | Header Field Name         | Protocol | Specification              |
    +---------------------------+----------+----------------------------+
    | Optional-WWW-Authenticate | http     | Section 3 of this document |
    | Authentication-Control    | http     | Section 4 of this document |
    +---------------------------+----------+----------------------------+

    This document also establishes a registry for HTTP authentication
    control parameters.  The registry manages a case-insensitive ASCII
    strings.  The string MUST follow the extensive-token syntax defined
    in Section 2.2.

/manages a/manages/

    To acquire registered tokens, a specification for the use of such
    tokens MUST be available as a publicly-accessible documents, as
    outlined as "Specification Required" level in [RFC5226].

    Registrations for authentication algorithms are required to include a
    description of the control extension.  New registrations are advised
    to provide the following information:

Hmmm... Why are these code "algorithms"? That's very confusing.

    o  Token: a token used in HTTP headers for identifying the algorithm.

    o  Specification: A reference for a specification defining the
       algorithm.


8.  Security Considerations

    All parameters in Authentication-Control header SHOULD NOT be used

s/All parameters in/Parameters in the/

    for any security-enforcement purposes.  Server-side applications MUST
    always consider that the header may be either ignored by clients or
    even bypassed by users.

Then it' a MUST NOT, no?


Authors' Addresses

  ...

  Nit: an author should supply at least sufficient information so he/she 
can be contacted. Why no email addresses here?


Best regards, Julian