Re: Bug: All git operations fail when .git contains a non-existent gitdir

[Atharva Raykar <raykar.ath@gmail.com>]

On 22-Jul-2021, at 18:43, Tom Cook <tom.k.cook@gmail.com> wrote:
> On Wed, Jul 21, 2021 at 11:59 PM brian m. carlson
> <sandals@crustytoothpaste.net> wrote:
>> 
>> On 2021-07-21 at 09:17:36, Tom Cook wrote:
>>> What did you do before the bug happened? (Steps to reproduce your issue)
>>> 
>>> Add a git submodule to a git repository.
>>> Overlay-mount that submodule to another place in the filesystem.
>>> Attempt any git operation in the overlay-mounted path.
>> 
>> I'm not sure about what you mean by an overlay-mount operation.  Can you
>> provide some specific commands that we can run at a shell that reproduce
>> the issue?
>> --
>> brian m. carlson (he/him or they/them)
>> Toronto, Ontario, CA
> 
> The easiest way to reproduce it is this:
> 
> $ mkdir test
> $ cd test
> $ echo "gitdir: /foo/bar" > .git
> $ git ls-remote https://github.com/torvalds/linux
> 
> We happen to use overlay mounts in our build system in a way that maps
> a git submodule from one place to another so that its "gitdir" is
> invalid and then attempt a `git ls-remote` from that location which
> unexpectedly fails.  But the above reproduces the problem well enough.

'ls-remote' needs a valid git directory for the case where the URL is not
explicitly supplied (to read the git config and learn the default remote).

Making a special case for when an explicit URL is not given is not as
straightforward as it seems, because by the time 'ls-remote' even knows about
its arguments, it already takes a worktree prefix and sets up the environment,
for which a valid Git repository path is required.

I am not too familiar with this area, and I don't know how feasible it is to
delay setting up the environment until after looking at the 'ls-remote'
arguments. At a cursory glance, it looks difficult to do without large
structural changes to the code.

This might have been less of a problem with old-form submodules, where '.git'
was an actual directory, rather than a text file pointer [1], but newer
versions of Git discourage their usage.

[1] https://git-scm.com/docs/gitsubmodules#_forms

PS: we prefer bottom posting or inline replies :)

---
Pointers for others who might be interested in looking into this:

The immediate cause of this seems to be 'setup.c:setup_gitdir_gently()' [2]
which calls 'setup_gitdir_gently_1()' with the 'die_on_error' argument set
to true. This function then calls 'read_gitfile_gently()' with the same flag,
which errors out when it runs 'is_git_directory()' [3], because the path in
the gitfile is not a valid repository.

[2] https://github.com/git/git/blob/eb27b338a3e71c7c4079fbac8aeae3f8fbb5c687/setup.c#L1234
[3] https://github.com/git/git/blob/eb27b338a3e71c7c4079fbac8aeae3f8fbb5c687/setup.c#L784-L799

---
Atharva Raykar
ಅಥರ್ವ ರಾಯ್ಕರ್
अथर्व रायकर