Re: [PATCH] Fix premature call to git_config() causing t1020-subdirectory to fail

From: Junio C Hamano <gitster@pobox.com>
Date: Feb 27, 2008, 01:34 UTC
Message-ID: <7vy79718tn.fsf@gitster.siamese.dyndns.org>
In-Reply-To: <alpine.LNX.1.00.0802261933551.19665@iabervon.org>

Daniel Barkalow <barkalow@iabervon.org> writes:

Show 6 quoted lines

> There's nothing in the documentation to suggest that you can use 
> GIT_CONFIG to affect how the old repository is read, or that GIT_CONFIG 
> doesn't affect the new repository. Actually, as far as I can tell, the 
> configuration of a repository you're cloning (local or remote) doesn't 
> matter at all. Note that GIT_DIR and GIT_WORK_TREE refer to the new repo, 
> so it would be surprising for GIT_CONFIG to refer to the old one.

There was a bit of confusion in this discussion.

GIT_DIR the user may have in the environment may refer to the old reopsitory before "git clone" is invoked, but it should not matter at all, as the origin of the cloning comes from the command line and that is where we will read from. The scripted version sets GIT_DIR for our own use to point at the new repository upfront and exports it, so we are safe from bogus GIT_DIR value the user may have in the environment.

GIT_WORK_TREE naming the new repository feels Ok, as you do not care about the work tree of the original tree when cloning, and you may want to have a say in where the work tree associated with the new repository should go.

GIT_CONFIG the user may have will refer to the old repository before "git clone" is invoked, as there is no new repository built yet. But clone does not read from the old config, so "you can use GIT_CONFIG to read from old repository" may be true, but it does not matter. We won't use it (we do _not_ want to use it) to read from the old configuration file.

We would however want to make sure that we write to the correct configuration file of the new repository and not some random other place, and that's where the environment variable in the scripted version comes into the picture.

In the scripted version, the only way to make sure which exact configuration file is updated is to set and export GIT_CONFIG when running "git config", so there are a few places that does exactly that (e.g. call to git-init and setting of core.bare). Unfortunately many codepaths in the scripted version are utterly careless (e.g. setting of remote."$origin".fetch); they should make sure that they protect themselves against GIT_CONFIG the user may have in the environment that point at random places.

> My current design is to collect some initial information, create 
> directories, and then set the work tree and git dir. Then we run fetch, 
> configure things, etc., in the new context.

That sounds like a clean design to me.