Skip to content

pr-1352/mjcheetham/emu-v4

Following from my original RFC submission [0], this submission is considered
ready for full review. This patch series is now based on top of current
master (9c32cfb49c60fa8173b9666db02efe3b45a8522f) that includes my now
separately submitted patches [1] to fix up the other credential helpers'
behaviour.

In this patch series I update the existing credential helper design in order
to allow for some new scenarios, and future evolution of auth methods that
Git hosts may wish to provide. I outline the background, summary of changes
and some challenges below.

Testing these new additions, I introduce a new test helper test-http-server
that acts as a frontend to git-http-backend; a mini HTTP server based
heavily on git-daemon, with simple authentication configurable by command
line args.

Background
==========

Git uses a variety of protocols [2]: local, Smart HTTP, Dumb HTTP, SSH, and
Git. Here I focus on the Smart HTTP protocol, and attempt to enhance the
authentication capabilities of this protocol to address limitations (see
below).

The Smart HTTP protocol in Git supports a few different types of HTTP
authentication - Basic and Digest (RFC 2617) [3], and Negotiate (RFC 2478)
[4]. Git uses a extensible model where credential helpers can provide
credentials for protocols [5]. Several helpers support alternatives such as
OAuth authentication (RFC 6749) [6], but this is typically done as an
extension. For example, a helper might use basic auth and set the password
to an OAuth Bearer access token. Git uses standard input and output to
communicate with credential helpers.

After a HTTP 401 response, Git would call a credential helper with the
following over standard input:

protocol=https
host=example.com

And then a credential helper would return over standard output:

protocol=https
host=example.com
username=bob@id.example.com
password=<BEARER-TOKEN>

Git then the following request to the remote, including the standard HTTP
Authorization header (RFC 7235 Section 4.2) [7]:

GET /info/refs?service=git-upload-pack HTTP/1.1
Host: git.example
Git-Protocol: version=2
Authorization: Basic base64(bob@id.example.com:<BEARER-TOKEN>)

Credential helpers are encouraged (see gitcredentials.txt) to return the
minimum information necessary.

Limitations
===========

Because this credential model was built mostly for password based
authentication systems, it's somewhat limited. In particular:

 1. To generate valid credentials, additional information about the request
    (or indeed the requestee and their device) may be required. For example,
    OAuth is based around scopes. A scope, like "git.read", might be
    required to read data from the remote. However, the remote cannot tell
    the credential helper what scope is required for this request.

 2. This system is not fully extensible. Each time a new type of
    authentication (like OAuth Bearer) is invented, Git needs updates before
    credential helpers can take advantage of it (or leverage a new
    capability in libcurl).

Goals
=====

 * As a user with multiple federated cloud identities:

   * Reach out to a remote and have my credential helper automatically
     prompt me for the correct identity.
   * Allow credential helpers to differentiate between different authorities
     or authentication/authorization challenge types, even from the same DNS
     hostname (and without needing to use credential.useHttpPath).
   * Leverage existing authentication systems built-in to many operating
     systems and devices to boost security and reduce reliance on passwords.

 * As a Git host and/or cloud identity provider:

   * Enforce security policies (like requiring two-factor authentication)
     dynamically.
   * Allow integration with third party standard based identity providers in
     enterprises allowing customers to have a single plane of control for
     critical identities with access to source code.

Design Principles
=================

 * Use the existing infrastructure. Git credential helpers are an
   already-working model.
 * Follow widely-adopted time-proven open standards, avoid net new ideas in
   the authentication space.
 * Minimize knowledge of authentication in Git; maintain modularity and
   extensibility.

Proposed Changes
================

 1. Teach Git to read HTTP response headers, specifically the standard
    WWW-Authenticate (RFC 7235 Section 4.1) headers.

 2. Teach Git to include extra information about HTTP responses that require
    authentication when calling credential helpers. Specifically the
    WWW-Authenticate header information.

    Because the extra information forms an ordered list, and the existing
    credential helper I/O format only provides for simple key=value pairs,
    we introduce a new convention for transmitting an ordered list of
    values. Key names that are suffixed with a C-style array syntax should
    have values considered to form an order list, i.e. key[]=value, where
    the order of the key=value pairs in the stream specifies the order.

    For the WWW-Authenticate header values we opt to use the key wwwauth[].

Handling the WWW-Authenticate header in detail
==============================================

RFC 6750 [8] envisions that OAuth Bearer resource servers would give
responses that include WWW-Authenticate headers, for example:

HTTP/1.1 401 Unauthorized
WWW-Authenticate: Bearer realm="login.example", scope="git.readwrite"
WWW-Authenticate: Basic realm="login.example"

Specifically, a WWW-Authenticate header consists of a scheme and arbitrary
attributes, depending on the scheme. This pattern enables generic OAuth or
OpenID Connect [9] authorities. Note that it is possible to have several
WWW-Authenticate challenges in a response.

First Git attempts to make a request, unauthenticated, which fails with a
401 response and includes WWW-Authenticate header(s).

Next, Git invokes a credential helper which may prompt the user. If the user
approves, a credential helper can generate a token (or any auth challenge
response) to be used for that request.

For example: with a remote that supports bearer tokens from an OpenID
Connect [9] authority, a credential helper can use OpenID Connect's
Discovery [10] and Dynamic Client Registration [11] to register a client and
make a request with the correct permissions to access the remote. In this
manner, a user can be dynamically sent to the right federated identity
provider for a remote without any up-front configuration or manual
processes.

Following from the principle of keeping authentication knowledge in Git to a
minimum, we modify Git to add all WWW-Authenticate values to the credential
helper call.

Git sends over standard input:

protocol=https
host=example.com
wwwauth[]=Bearer realm="login.example", scope="git.readwrite"
wwwauth[]=Basic realm="login.example"

A credential helper that understands the extra wwwauth[n] property can
decide on the "best" or correct authentication scheme, generate credentials
for the request, and interact with the user.

The credential helper would then return over standard output:

protocol=https
host=example.com
path=foo.git
username=bob@identity.example
password=<BEARER-TOKEN>

Note that WWW-Authenticate supports multiple challenges, either in one
header:

HTTP/1.1 401 Unauthorized
WWW-Authenticate: Bearer realm="login.example", scope="git.readwrite", Basic realm="login.example"

or in multiple headers:

HTTP/1.1 401 Unauthorized
WWW-Authenticate: Bearer realm="login.example", scope="git.readwrite"
WWW-Authenticate: Basic realm="login.example"

These have equivalent meaning (RFC 2616 Section 4.2 [12]). To simplify the
implementation, Git will not merge or split up any of these WWW-Authenticate
headers, and instead pass each header line as one credential helper
property. The credential helper is responsible for splitting, merging, and
otherwise parsing these header values.

An alternative option to sending the header fields individually would be to
merge the header values in to one key=value property, for example:

...
wwwauth=Bearer realm="login.example", scope="git.readwrite", Basic realm="login.example"

Future work
===========

In the future we can further expand the protocol to allow credential helpers
decide the best authentication scheme. Today credential helpers are still
only expected to return a username/password pair to Git, meaning the other
authentication schemes that may be offered still need challenge responses
sent via a Basic Authorization header. The changes outlined above still
permit helpers to select and configure an available authentication mode, but
require the remote for example to unpack a bearer token from a basic
challenge.

More careful consideration is required in the handling of custom
authentication schemes which may not have a username, or may require
arbitrary additional request header values be set.

For example imagine a new "FooBar" authentication scheme that is surfaced in
the following response:

HTTP/1.1 401 Unauthorized
WWW-Authenticate: FooBar realm="login.example", algs="ES256 PS256"

With support for arbitrary authentication schemes, Git would call credential
helpers with the following over standard input:

protocol=https
host=example.com
wwwauth[]=FooBar realm="login.example", algs="ES256 PS256", nonce="abc123"

And then an enlightened credential helper could return over standard output:

protocol=https
host=example.com
authtype=FooBar
username=bob@id.example.com
password=<FooBar credential>
header[]=X-FooBar: 12345
header[]=X-FooBar-Alt: ABCDEF

Git would be expected to attach this authorization header to the next
request:

GET /info/refs?service=git-upload-pack HTTP/1.1
Host: git.example
Git-Protocol: version=2
Authorization: FooBar <FooBar credential>
X-FooBar: 12345
X-FooBar-Alt: ABCDEF

Why not SSH?
============

There's nothing wrong with SSH. However, Git's Smart HTTP transport is
widely used, often with OAuth Bearer tokens. Git's Smart HTTP transport
sometimes requires less client setup than SSH transport, and works in
environments when SSH ports may be blocked. As long as Git supports HTTP
transport, it should support common and popular HTTP authentication methods.

References
==========

 * [0] [PATCH 0/8] [RFC] Enhance credential helper protocol to include auth
   headers
   https://lore.kernel.org/git/pull.1352.git.1663097156.gitgitgadget@gmail.com/

 * [1] [PATCH 0/3] Correct credential helper discrepancies handling input
   https://lore.kernel.org/git/pull.1363.git.1663865974.gitgitgadget@gmail.com/

 * [2] Git on the Server - The Protocols
   https://git-scm.com/book/en/v2/Git-on-the-Server-The-Protocols

 * [3] HTTP Authentication: Basic and Digest Access Authentication
   https://datatracker.ietf.org/doc/html/rfc2617

 * [4] The Simple and Protected GSS-API Negotiation Mechanism
   https://datatracker.ietf.org/doc/html/rfc2478

 * [5] Git Credentials - Custom Helpers
   https://git-scm.com/docs/gitcredentials#_custom_helpers

 * [6] The OAuth 2.0 Authorization Framework
   https://datatracker.ietf.org/doc/html/rfc6749

 * [7] Hypertext Transfer Protocol (HTTP/1.1): Authentication
   https://datatracker.ietf.org/doc/html/rfc7235

 * [8] The OAuth 2.0 Authorization Framework: Bearer Token Usage
   https://datatracker.ietf.org/doc/html/rfc6750

 * [9] OpenID Connect Core 1.0
   https://openid.net/specs/openid-connect-core-1_0.html

 * [10] OpenID Connect Discovery 1.0
   https://openid.net/specs/openid-connect-discovery-1_0.html

 * [11] OpenID Connect Dynamic Client Registration 1.0
   https://openid.net/specs/openid-connect-registration-1_0.html

 * [12] Hypertext Transfer Protocol (HTTP/1.1)
   https://datatracker.ietf.org/doc/html/rfc2616

Updates from RFC
================

 * Submitted first three patches as separate submission:
   https://lore.kernel.org/git/pull.1363.git.1663865974.gitgitgadget@gmail.com/

 * Various style fixes and updates to- and addition of comments.

 * Drop the explicit integer index in new 'array' style credential helper
   attrbiutes ("key[n]=value" becomes just "key[]=value").

 * Added test helper; a mini HTTP server, and several tests.

Updates in v3
=============

 * Split final patch that added the test-http-server in to several, easier
   to review patches.

 * Updated wording in git-credential.txt to clarify which side of the
   credential helper protocol is sending/receiving the new wwwauth and
   authtype attributes.

Updates in v4
=============

 * Drop authentication scheme selection authtype attribute patches to
   greatly simplify the series; auth scheme selection is punted to a future
   series. This series still allows credential helpers to generate
   credentials and intelligently select correct identities for a given auth
   challenge.

Matthew John Cheetham (8):
  http: read HTTP WWW-Authenticate response headers
  credential: add WWW-Authenticate header to cred requests
  test-http-server: add stub HTTP server test helper
  test-http-server: add HTTP error response function
  test-http-server: add HTTP request parsing
  test-http-server: pass Git requests to http-backend
  test-http-server: add simple authentication
  t5556: add HTTP authentication tests

 Documentation/git-credential.txt          |   18 +-
 Makefile                                  |    2 +
 contrib/buildsystems/CMakeLists.txt       |   13 +
 credential.c                              |   13 +
 credential.h                              |   15 +
 http.c                                    |   78 ++
 t/helper/.gitignore                       |    1 +
 t/helper/test-credential-helper-replay.sh |   14 +
 t/helper/test-http-server.c               | 1146 +++++++++++++++++++++
 t/t5556-http-auth.sh                      |  223 ++++
 10 files changed, 1522 insertions(+), 1 deletion(-)
 create mode 100755 t/helper/test-credential-helper-replay.sh
 create mode 100644 t/helper/test-http-server.c
 create mode 100755 t/t5556-http-auth.sh

base-commit: c48035d29b4e524aed3a32f0403676f0d9128863

Submitted-As: https://lore.kernel.org/git/pull.1352.v4.git.1670880984.gitgitgadget@gmail.com
In-Reply-To: https://lore.kernel.org/git/pull.1352.git.1663097156.gitgitgadget@gmail.com
In-Reply-To: https://lore.kernel.org/git/pull.1352.v2.git.1666372083.gitgitgadget@gmail.com
In-Reply-To: https://lore.kernel.org/git/pull.1352.v3.git.1667426969.gitgitgadget@gmail.com
Assets 2