We have been using GitLab to provide access control and repository management for Git at work. Originally I looked for something to manage the user access and found gitolite looked like best suiting our needs with one exception. Gitolite access is managed by making commits and pushing the gitolite-admin repository which is not especially user friendly. It is easy to automate however and so in searching for a web front end to gitolite I found GitLab.
GitLab is a Rails web application that can handle user account management and repository access for git using gitolite underneath. It has a number of plugins to allow for a selection of authentication mechanism and in a Windows culture we can set it up for LDAP authentication against the local Active Directory servers so that our users can use their normal Windows login details and do not have to remember yet another password. They do however have to create an SSH key and upload it to GitLab but GitExtensions provides a reasonable user interface for dealing with this. In practice we find Windows users can create an account and create repositories with minimal assistance.
However, I find that at least on the server we are using that the git repository browsing from GitLab is quite slow and all repository access requires an account. I have some automated processes that don't really need an SSH key as they are read-only and these are using the git protocol to update their clones. I also like using gitweb for repository browsing.
gitolite has a solution to this that involves using two pseudo-user accounts but GitLab erases these whenever it updates the gitolite configuration file. Further, if we simply go and create the git-daemon-export-ok file that git-daemon uses to test for exported repositories, either GitLab or GitExtensions will delete these on the next configuration update. So in searching for some way around this I found the git-daemon access-hook parameter. This allows git-daemon to call an external program to check for access. If we combine this with the gitweb project file which can be used to git gitweb a specfic set of visible directories then we don't have to hack GitLab itself.
This is achieved by creating a text file that just lists the repository names we wish to be publicly readable via git-daemona and gitweb. Then edit /etc/gitweb.conf and set the $projects_list variable to point to this file.
$projects_list = "/home/git/gitweb.projects"; $export_auth_hook = undef; # give up on the git-daemon-export-ok files.
Next, setup our git-daemon xinetd configuration to call our access hook script.
# /etc/xinetd.d/git-daemon service git { disable = no type = UNLISTED port = 9418 socket_type = stream wait = no user = git server = /usr/local/bin/git server_args = daemon --inetd --base-path=/home/git/repositories --export-all --access-hook=/home/git/git-daemon-access.sh /home/git/repositories log_on_failure += USERID }
After a bit of experimentation I found that you need to --export-all as well as have the access-hook.
Finally, the access hook script.
#!/bin/bash # # Called up from git-daemon to check for access to a given repository. # We check that the repo is listed in the gitweb.projects file as these # are the repos we consider publically readable. # # Args are: service-name repo-path hostname canonical-hostname ipaddr port # To deny, exit with non-zero status logger -p daemon.debug -t git-daemon "access: $*" [ -r /home/git/gitweb.projects ] \ && /bin/grep -q -E "^${2#/home/git/repositories/}(\\.git)?\$" /home/git/gitweb.projects
Here we log access attempts and fail if the gitweb.projects file is missing. Then trim off the base-path prefix and check for the repository name being listed in the file.
Now I can have GitLab and git-daemon and gitweb available.
In the future it might be nice to use this mechanism to have GitLab provide some control of this from it's user interface. That way I will avoid having to occasionally update this project file. However, thats something for another day - especially as the most recent release of GitLab has dropped the gitolite back-end in favour of an internal package (git-shell). I'm waiting to see how that turns out.