SoundSoftware Code Site

h2. Authentication for Mercurial activity

h3. Requirements

# *Clone/pull from repo for public project*: Any user, no authentication required

# *Clone/pull from repo for private project*: Permitted users only

# *Push to repo for public project*: Permitted users only

# *Push to repo for private project*: Permitted users only

# *Create repo for public project*: User with manager role on project

# *Delete repo or carry out command-line admin tasks*: System admin only

What constitutes a permitted user for limited push or pull activity?

* A user who is a member of the project?

* A user who is identified in the [web] section of the repository?

* A user who is both a member and identified in the [web] section?

* A user who is either a member or identified in the [web] section?

* A user who is identified in the [web] section, if any, or is a member if there is no such section?

h3. Techniques

* Hg repository creation using "reposman.rb":/projects/soundsoftware-site/repository/entry/extra/svn/reposman.rb

* Apache authentication against Redmine user database using mod_auth_mysql (no support for LDAP-authenticated users?)

* Apache authentication against Redmine users using "the mod_perl module Redmine.pm":http://redmine.rubyforge.org/svn/trunk/extra/svn/Redmine.pm ("local copy":/projects/soundsoftware-site/repository/entry/extra/svn/Redmine.pm) or a variant thereof.  Redmine.pm was designed for SVN access via WebDAV, but the code itself handles access and authentication only?

* Hg repository [web]-section authorisation using "hgwebdir.cgi":http://mercurial.selenic.com/wiki/PublishingRepositories#Setting_up_the_hgweb.cgi_script

Other links on this subject:

* "Separation between authentication and authorisation activities":http://markmail.org/message/xmav6qg3is3xptve#query:+page:1+mid:xmav6qg3is3xptve+state:results

* "hgrc [web] section":http://www.selenic.com/mercurial/hgrc.5.html#web

* "Using mod_auth_mysql to authenticate against the Redmine database directly":http://maff.ailoo.net/2009/03/authenticate-apache-against-redmine-with-authmysql/

* "Apache2 logging from mod_perl code":http://perl.apache.org/docs/2.0/api/Apache2/Log.html (for use in debugging Redmine.pm)

* "How to automate repository creation":http://www.redmine.org/wiki/1/HowTo_Automate_repository_creation

h3. Plan

We are currently using proxypass from a front-end server to the actual code server, so this complicates matters somewhat.

It seems reasonable-ish to assume that all hg traffic will use SSL.  Apart from anything else, pushes and any other authenticated traffic really must be over SSL, so this avoids irritations when someone pulls from an http URL and then gets kicked out when attempting to push to the default target.

Therefore,

* Front-end server needs to have the relevant certificates

* Front-end server is currently running too old a version of Apache 2 to support SNI, so it can only support one https domain (question: does Mercurial's own Python client support SNI?) -- fortunately for us it is not currently running any other, but this suggests we really need to replace the front-end arch

* Front-end server redirects all http traffic on the hg domain to https and then proxypasses to the back-end server via http (connection is presumed trusted?)

At the back-end server,

* We will use hgwebdir.cgi instead of serving the hg directories directly as static HTTP

* The server's document root will be just a placeholder

* ScriptAlias is used to point all traffic to hgwebdir.cgi

* We will use our own "SoundSoftware.pm":/projects/soundsoftware-site/repository/entry/extra/svn/SoundSoftware.pm based on Redmine.pm to manage access

* We probably want to prevent people reading the list of repositories (so that per-project authentication is always in force)

The net result is that we should have hg over https with basic authentication based on the Redmine database and LDAP as appropriate, with access controls that we can define based on the project rules set out above.

Because of our front-end server limitations, we will use only a single domain name.  That is, hg will be served through https://code.soundsoftware.ac.uk/hg/ rather than https://hg.soundsoftware.ac.uk/ as we might otherwise do.  We can do this straightforwardly, because a ScriptAlias for /hg in the Apache config will take effect before mod_passenger gets to the request (no rewrite needed).

Thus Apache config:

<pre>

PerlLoadModule Apache::Authn::SoundSoftware

<VirtualHost *:80>

        ServerName code.soundsoftware.ac.uk

        ServerAdmin <me>

        DocumentRoot /var/www/redmine/public

        PassengerRestartDir restart_files

        PassengerHighPerformance on

        PassengerMaxRequests 5000

        PassengerStatThrottleRate 10

        RailsSpawnMethod smart

        ExpiresDefault "access plus 1 minute"

        <DirectoryMatch "^/.*/\.svn/">

                Order allow,deny

                Deny from all

                Satisfy All

        </DirectoryMatch>

        <Directory /var/www/redmine/public>

                Options -MultiViews

        </Directory>

        ScriptAlias /hg "/var/hg/index.cgi"

        <Location /hg>

                AuthName "Mercurial"

                AuthType Basic

                Require valid-user

                PerlAccessHandler Apache::Authn::SoundSoftware::access_handler

                PerlAuthenHandler Apache::Authn::SoundSoftware::authen_handler

                SoundSoftwareDSN "DBI:mysql:database=<db>;host=localhost"

                SoundSoftwareDbUser "<user>"

                SoundSoftwareDbPass "<password>"

                SoundSoftwareRepoPrefix "/var/hg/"

                Options +ExecCGI

                AddHandler cgi-script .cgi

                ExpiresDefault now

        </Location>

        ErrorLog /var/log/apache2/error.log

        LogLevel warn

        ServerSignature Off

</VirtualHost>

</pre>

Note that while SoundSoftware.pm is substantially customised from Redmine.pm, /var/hg/index.cgi is a completely stock hgwebdir.cgi (with UTF-8 and, currently, traceback options uncommented).

*Problem:* hgwebdir.cgi refuses to accept authentication if it is not running under SSL (it just spits out "ssl required").  That's reasonable enough, except that the user connection is under SSL, it's just the connection within the virtual host wrapper that is not.  It can be fixed in @hgweb.config@, see below.  We also want to set the config to permit anyone to push, because we are using the mod_perl authentication module rather than the repository configuration to determine who is permitted.

<pre>

[web]

push_ssl = false

allow_push = *

</pre>

h3. Creating repositories

Reposman appears to be set up to create repositories for all projects listed in Redmine, with each repository having the same name as the project.  If there is a repository (of any name) already configured for a project, it does not create a new one -- but if the repository that is already configured does have the same name as the project, and it exists on the disk, reposman will test its ownership and permissions and attempt to set them to something it things of as sensible.

I suppose the idea is that Reposman is run regularly (by cron?) in order to update and check the repositories for each project.

*What ownership and permissions are sensible?* We want Redmine (user redmine) to be able to read the repository but it doesn't need to be able to modify it.  We need Apache (user www-data) to be able to read and write, so that hgwebdir.cgi can do its work. This suggests we want www-data to own the repository and redmine to be able to read it, so

@drwxr-s--- 17 www-data redmine@

seems sensible.

Example of running reposman with our instance: @./reposman.rb -s /var/hg -r https://code.soundsoftware.ac.uk/ -k <key> -tv --http-user=user --http-pass=pass -o www-data -g redmine -c "hg init"@ -- note the -tv option says to run in test (and verbose) mode so this doesn't actually change anything.

Reposman doesn't seem to have any option to change the permissions (as opposed to ownerships) of a repository.  Actually, we probably don't want it to change the ownerships either -- just to get them right when it first creates a repo.

It would probably be nice for the user who creates or administrates the project to get to choose the repository name.  Rather than just creating a repo for every project, even if it's not needed, it might be better if the user could set the repo path in Redmine, and reposman would then check whether the repo existed and create it if it did not.  Redmine does happily allow you to set a repository path to something that doesn't exist yet -- though the error message from the repository view tab subsequently is not very helpful (and is SCM-specific).

*A big problem with that* and with the current working of reposman is that the user who creates the project doesn't typically know what repository path to use.  For example, a user has just created a project on our test site and typed in "/projects/X/repository" (where X is project name) as the repos path, presumably guessing from the public URL for the browse page.  We should either:

# set a sensible repository path for every project on creation (and create the repository automatically)

# pre-populate the repository path entry field with an appropriate value or prefix

# provide documentation in the repository settings page

# check the repository path in reposman and correct it if it is implausible (i.e. if it is a local path that is not in /var/hg or whatever)

*Note*: Reposman has just started failing for me, with "wrong constant name 0" error (see http://www.redmine.org/boards/1/topics/16274) possibly related to introduction of the Checkout plugin.