Authentication ring =================== When accessing a remote branch (specified as an URL), it may occur that the server requests an authentication. This authentication can be provided in different ways: 1. Embedding the user and password in the URL:: bzr branch ://:@host:port/path * ``scheme``: Any transport protocol requiring authentication. * ``user``: The login used to authenticate. * ``password``: The associated password. * ``host``: The address of the server. * ``port``: The port the server is listening to. * ``path``: The path on the server. 2. Embedding the user in the URL and let bzr find the right password or prompt for one:: bzr branch ://@host/path 3. Embedding nothing in the URL and let bzr find user and password or prompt for user and/or password:: bzr branch ://host/path This specification proposes a mechanism that will allow users to just use ``bzr branch ://host/path`` or ``bzr branch ://@host/path`` and leaves bzr find the ``user`` and ``password`` in its configuration files. When no user is specified for ``FTP``, ``SFTP`` or ``SSH``, the actual behavior of ``bzr`` is to default to ``getpass.get_user()``. Any implementation of this specification should respect that behaviour. This specification also proposes a way to describe credentials so that several remote branches can use the same definition. This is particularily important for users handling a lot of passwords who need to update them on a regular basis. Rationale --------- Embedding user and passwords in the command line is a security hazard (see `bug #34685 `_). Storing passwords in ``~/.bazaar/bazaar.conf`` or ``~/.bazaar/locations.conf`` is also a security risk. Typing user and passwords is error-prone and boring. Yet, a safe way to store passwords, while allowing bzr to retrieve them, when needed, could improve the bzr user experience. This specification describes a way to provide user and passwords to bzr while storing them in a relatively safe way. Note that ssh servers can be configured to use keys instead of (``user``, ``password``) and, when used with appropriate agents, provide the same kind of comfort this specification aims to provide for all other schemes. These specification do not try to cover these configurations by providing pass-phrases, but the mechanisms presented *can* be used to provide users. Authentication definitions -------------------------- There are two kinds of authentication used by the various schemes: 1. user and password ``FTP`` and ``SFTP`` needs a (``user``, ``password``) to authenticate against a ``host`` (SFTP can use ssh keys too, but we don't talk about that in this specification as ssh agents provide a better solution). 2. user, realm and password ``HTTP`` and ``HTTPS`` needs a (``user, realm, password``) to authenticate against a host. But, by using ``.htaccess`` files, for example, it is possible to define several (``user, realm, password``) for a given ``host``. So what is really needed is (``user``, ``password``, ``host``, ``path``). The ``realm`` can be ignored [#ignored_realm]_ as long as it is still presented to the user when prompting for the password (unless someone found a way to declare two different realms for the same path). ``HTTP proxy`` can be handled as ``HTTP`` (or ``HTTPS``) by explicitely specifying the appropriate port. .. [#ignored_realm] The true purpose of realms is to allow the same credentials to be reused for disjoint hierarchies. Ignoring them in this specification aims to simplify the user experience while still allowing to share the same credentials for a whole hierarchy. To take all schemes into account, the password will be deduced from a set of authentication definitions (``scheme``, ``host``, ``port``, ``path``, ``user``, ``password``). * ``scheme``: can be empty (meaning the rest of the definition can be used for any scheme), * ``host``: can be empty (to act as a default for any host), * ``port`` can be empty (useful when an host provides several servers for the same scheme), only numerical values are allowed, * ``path``: can be empty (FTP or SFTP will never user it), * ``user``: can be empty (and will use the same default as ``bzr``, python's ``getpass.get_user()``), * ``password``: can be empty (for security reasons, a user may use the definitions without storing the passwords but want to be prompted). * ``password_encoding``: can be empty (default is ``plaintext``). Also note that an optional ``verify_certificates=no`` field will allow the connection to ``HTTPS`` hosts that provides a self certified certificate (the default should be to refuse the connection and inform the user). Multiple definitions can be provided and, for a given URL, bzr will select a (``user`` [, ``password``]) based on the following rules : 1. the first match wins, 2. empty fields match everything, 3. ``scheme`` matches even if decorators are used in the requested URL, 4. ``host`` matches if included in the requested URL. ``foo.net`` will match a requested ``bzr.foo.net``. 5. ``port`` matches if included in the requested URL (exact matches only) 6. ``path`` matches if included in the requested URL (and by rule #2 above, empty paths will match any provided path). An optional ``password_encoding`` field may specify how the password is encoded but has no impact on the definition selection. Possible values are ``plaintext`` (no encoding at all) and ``base64``. When the field is absent, ``plaintext`` is assumed. Additional encodings may be added in future versions. Encoding passwords in ``base64``, while weak, provides protection against accidental reading (if an administrator have to look into the file, he will not see the passwords in clear). This specification intend to ease the authentication providing, not to secure it in the best possible way. Future versions of this specification may provide additional encodings [#password_encoding]_. .. [#password_encoding] Additional password encoding methods may be defined that will rely on external means to store the password which, in these cases, will not appear anymore in the definition but be provided. An encoding named ``netrc`` for example will provide passwords by using the ``.netrc`` file. File format ----------- Even if ``~/.bazaar/bazaar.conf`` and ``~/.bazaar/locations.conf`` seems to provide most of the needed infrastructure, we choose to use a dedicated file for the authentication info ``~/.bazaar/authentication.conf`` for the following reasons: * allow the user to protect the content of one file only, relaxing security constraints on the others, * while ``locations.conf`` is organized around *local* branches, ``authentication.conf`` is organized around *remote* branches or more generally servers. The same authentification definition can even be used for several schemes for servers providing those schemes. ``~/.bazaar/authentication.conf`` will use the same file format than ``~/.bazaar/bazaar.conf``. Each section will define an authentication definition. The section name is an arbitrary string, only the ``DEFAULT`` value is reserved and should appear as the *last* section. Each section should define: * ``user``: the login to be used, Each section could define: * ``host``: the remote server, * ``port``: the port the server is listening, * ``verify_certificates``: to control certificate verification (useful for self certified hosts). This applies to HTTP[S] only. Accepted values are yes and no, default to yes. * ``path``: the branch location, * ``password``: the password, * ``password_encoding``: the method used to encode the password if any, The default content of the file will be:: [DEFAULT] This section could define: * ``user``: default user to be used. * ``password_encoding``: default password encoding. Use Cases --------- The use cases described below use the file format defined above. * all FTP connections to the foo.net domain are done with the same (``user``, ``password``):: # Identity on foo.net [foo.net] scheme=ftp host=foo.net user=joe password=secret-pass will provide ('joe', 'secret-pass') for:: bzr branch ftp://foo.net/bzr/branch bzr pull ftp://bzr.foo.net/bzr/product/branch/trunk * all connections are done with the same ``user`` (the local one) and the password is always prompted with some exceptions:: # Pet projects on hobby.net [hobby] scheme=https host=r.hobby.net verify_certificates=no user=jim password=obvious1234 # Home server [home] host=home.net user=joe password='c2VjcmV0LXBhc3M=' password_encoding=base64 [DEFAULT] user=foobar * an HTTP server that also acts as a proxy (weird):: # development branches on dev server [dev] scheme=https host=dev.company.com path=/dev user=user1 password=pass1 # toy branches [localhost] scheme=http host=dev.company.com path=/ user=user2 password=pass2 # Pesky proxy [proxy] scheme=http host=dev.company.com user=proxyuser1 password=proxypass1 Note that the proxy should be specified last because it uses no path. An alternative is to specify the port used by the proxy. UI Changes ---------- Depending on the info provided in the URL, bzr will interact with the user in different ways: 1. ``user`` and ``password`` given in the URL. Nothing to do. 2. ``user`` given in the URL. Get a password from ``~/.bazaar/authentication.conf`` or prompt for one if none is found. 3. No ``user`` given in the URL (and no ``password``). Get a user from ``~/.bazaar/authentication.conf`` or prompt for one if none is found. Continue as 2. Note: A user will be queried only if the server requires it for ``HTTP``, other protocols always require a user. In any case, if the server refuses the authentication, bzr reports to the user and terminates. Implementation constraints -------------------------- * bzr should be able to prompt for a ``user`` for a given (``scheme``, ``host`` [, ``realm``]). Note that ``realm`` may be available only after a first connection attempt to the server. * No assumptions should be made about the clients of this service (i.e. Transport is the primary target but plugins must be able to use it as well, the definitions used: (``scheme, host, [port,] path``) are general enough to described credentials for ``svn`` servers or LaunchPad xmlrpc calls). * Policies regarding default users may be taken into account by the implementations, there is no good way to represent that in this specification and stays flexible enough to accommodate various needs (default user policies may differ for different schemes and that may be easier to handle in the code than in the authentication file itself). * If no user can be found by the mechanism described above, bzr should still default to ``getpass.get_user()`` and may attempt a second matching to obtain a password. * As this specification proposes a matching between some credentials definitions and real urls, the implementation should provide an optional UI feedback about which credential definition is used. That will allow the user to validate his definitions. Questions and Answers --------------------- * What if a ``.authinfo`` file exists ? * It will be ignored, * Automatic (one-time) conversions may be proposed if sufficient demand exists, * What if a ``.netrc`` file exists ? * It will be honored if the definition specifies ``password_encoding=netrc`` once the appropriate plugin have been written. * What mode should the authentication file use ? * 600 read/write for owner only by default, if another mode (more permissive) is used, a warning will be issued to inform the users of the potential risks. * What about using ``seahorse`` on Ubuntu or ``KeyChain Access`` on Mac OS X ? * plugins can be written and registered to handle the associated ``password_encoding``. * Could it be possible to encode the whole authentication file with a ssh key ? * yes and if the user configure a ssh-agent it will not be queried for pass-phrase every time we want to query the file for a password. But that seems a bit extreme for a first version. * Why can't bzr update the authentication file when it queried the user for a password ? * a future version may address that but: 1. The user may want to decide which passwords are stored in the file and which aren't. 2. The user should decide if the passwords are encoded (and how) or not (but we may default to base64). 3. The right definition may be hard to get right, but reducing it to (``scheme, host, [port,] user, password``) may be a good start. I.e. no path so that all paths on the host will match. The user will have to modify it for more complex configurations anyway.