Re: [ANNOUNCE] Git v2.44.0

From: Patrick Steinhardt <ps@pks.im>
Date: Feb 24, 2024, 06:36 UTC
Message-ID: <ZdmOZRjJ-mClBR02@framework>
In-Reply-To: <20240224051040.ftuo24smozqugbde@glandium.org>

On Sat, Feb 24, 2024 at 02:10:40PM +0900, Mike Hommey wrote:

Show 35 quoted lines

> Hi,
> 
> On Fri, Feb 23, 2024 at 09:17:07AM -0800, Junio C Hamano wrote:
> > Patrick Steinhardt (139):
> >       builtin/clone: create the refdb with the correct object format
> 
> I haven't analyzed how/why exactly yet, but I've bisected a regression
> in the behavior of is_git_directory() during a clone to originate from
> this change.
> 
> Here's a way to reproduce the problem:
> 
> ```
> $ cat > git-remote-foo <<EOF
> #!/bin/sh
> git config --local -l >&2
> exit 1
> EOF
> $ chmod +x git-remote-foo
> $ PATH=$PWD:$PATH git clone foo::bar
> ```
> 
> With versions < 2.44.0, it displays the local configuration, e.g.:
> ```
> core.repositoryformatversion=0
> core.filemode=true
> core.bare=false
> core.logallrefupdates=true
> remote.origin.url=foo::bar
> ```
> 
> but with 2.44.0, it fails with:
> ```
> fatal: --local can only be used inside a git repository
> ```

Thanks for your report!

This has to be because we now initialize the refdb at a later point. The problem here was that before my change, we initialized the refdb at a point when it wasn't clear what the remote actually used as the object format. The consequence was twofold:

  - Cloning a repository with bundles was broken in case the remote uses
    the SHA256 object format.

  - Cloning into a repository that uses reftables when the remote uses
    the SHA256 object format was broken, too.

Both of these have the same root cause: because we didn't connect to the remote yet we had no idea what object format the remote uses. And as we initialized the refdb early, it was then initialized with the default object format, which is SHA1.

The change was to move initialization of the refdb to a later point in time where we know what object format the remote uses. By necessity, this has to be _after_ we have connected to the remote, because there is no way to learn about it without connecting to it.

One consequence of initializing the refdb at a later point in time is that we have no "HEAD" yet, and a repo without the "HEAD" file is not considered to be a repo. Thus, git-config(1) would now rightfully fail.

I assume that you discovered it via a remote helper that does something more interesting than git-config(1). I have to wonder whether we ever really specified what the environment of a remote helper should look like when used during cloning. Conceptually it doesn't feel _wrong_ to have a not-yet-initialized repo during clone.

But on the other hand, regressing functionality like this is of course bad. I was wondering whether we can get around this issue by setting e.g. GIT_DIR explicitly when spawning the remote helper, but I don't think it's as easy as that.

Another idea would be to simply pre-create HEAD regardless of the ref format, pointing to an invalid ref "refs/heads/.invalid". This is the same trick we use for the reftable backend, and should likely address your issue.

I will have a deeper look on Tuesday and send a patch.

Patrick