Re: git init BUG when gitconfig has includeIf

From: Patrick Steinhardt <ps@pks.im>
Date: May 22, 2024, 08:06 UTC
Message-ID: <Zk2ndVx4gMTV8ZUf@tanuki>
In-Reply-To: <D1FH40J6URHF.186U29CMRJ0N3@protonmail.com>

On Tue, May 21, 2024 at 04:46:23PM +0000, Heghedus Razvan wrote:

Show 24 quoted lines

> On Tue May 21, 2024 at 7:36 PM EEST, Junio C Hamano wrote:
> > Heghedus Razvan <heghedus.razvan@protonmail.com> writes:
> >
> > > Yesterday I stumble upon a bug when doing git init. I didn't
> > > find any references to it, so I don't know if is a known problem
> > > or not.
> > >
> > > Steps to reproduce:
> > > # git init .
> > > BUG: refs.c:2123: reference backend is unknown
> >
> > Patrick, this looks similar to an earlier one during "git clone"
> > that was discussed at
> >
> > https://lore.kernel.org/git/72771da0-a0ef-4fd9-8071-6467cd7b6a6b@kernel-space.org/
> >
> > that was fixed with 199f44cb (builtin/clone: allow remote helpers to
> > detect repo, 2024-02-27)?  The fix was about "git clone", but the
> > crux of the fix went to setup.c:initialize_repository_version()
> > which is also called by setup.c:init_db() that is the workhorse of
> > "git init", so it may already have been fixed (I didn't try).
> 
> I guess I forgot to mention, but I tested the current master 4365c6fcf9
> and the issue is still present.

I cannot reproduce the issue as-is, neither on Git v2.44 nor on the current master branch. So clearly, there must be something special to your setup. The following testcase and variants of it do not reproduce:

    test_expect_success 'init with includeIf.onbranch condition' '
        git config -f ./config foo.bar baz &&
        include=$(test-tool path-utils absolute_path config) &&
        test_when_finished "rm -rf repo" &&
        git -c includeIf.onbranch:main.path="$(<include)" init repo
    '

Now digging into the code, the condition gets evaluated in `include_by_branch()`. The call to `refs_resolve_ref_unsafe()` is guarded by `the_repository->gitdir`, which is `NULL` the first time it is called by git-init(1). It does get called a second time, but at that point we already initialized the refdb and configured the ref storage format as expected.

Aha! Seems like this only happens when re-initializing an already existent repository, that's what's missing. In that case we do already have `the_repository->gitdir` set even though we did not yet set up the ref storage format. I'll investigate and send a patch.

Can you confirm that this is what you see, or do you also see this when creating an entirely new repository?

> > Even if it is already fixed in the current version by the same
> > 199f44cb, we may want to follow up 0eab85b9 (t5601: exercise clones
> > with "includeIf.*.onbranch", 2024-03-12) with an additional test to
> > cover "git init".

Agreed.

Patrick