Re: Enforcing clone/fetch to use references.

From: Jeff King <peff@peff.net>
Date: Sep 21, 2010, 21:31 UTC
Message-ID: <20100921213135.GB1255@sigill.intra.peff.net>
In-Reply-To: <20100921204456.GA24357@huya.qualcomm.com>

On Tue, Sep 21, 2010 at 01:44:56PM -0700, David Brown wrote:

Show 7 quoted lines

> Suppose I want to publish some changes to a tree.  I have a server
> available where I can run a git daemon, but for one reason or another
> I want to force people to use the another git repo as a reference.
> The reason could be one of bandwidth, or someone who isn't comfortable
> making all of the other source available.  Ideally, someone who
> already has the other git repo cloned, and just adds mine as a remote
> wouldn't notice the difference.

I think the gentoo people were talking about doing something like this. They wanted you to use some faster and/or restartable protocol to clone initially, and so they wanted to reject initial clones. I'm not sure if they are doing that, and how (from the thread below, I suspect they run a patched git).

The simplest thing would be a pre-upload-pack hook. There was some discussion of that in this thread:

  http://article.gmane.org/gmane.comp.version-control.git/137007

but there are some security implications. For a remote site, where the user running upload-pack trusts the hook, it should be fine.

Once you have that hook, then you basically need to just check whether they are requesting objects in your "you should have this in your reference repository" set (probably by using merge-base to see if it is contained in some "everything our alternate has" ref), and then just

  echo >&2 Sorry, here's how to clone properly...
  exit 1

which should go to the user over the sideband.

And that would cover both the "clone --reference" and "git remote add" cases, as it is not looking at what the user invoked, but which objects they're requesting.

-Peff