Re: Bug: relative core.worktree is resolved from symlink and not its target

From: Duy Nguyen <pclouds@gmail.com>
Date: Feb 17, 2014, 12:30 UTC
Message-ID: <CACsJy8B=34f71uW27J_qkq5auqwRdgHrraunon-H0eZvpKDbnA@mail.gmail.com>
In-Reply-To: <5301D835.1060301@thequod.de>

On Mon, Feb 17, 2014 at 4:36 PM, Daniel Hahler <genml+git-2014@thequod.de> wrote:

Show 22 quoted lines

> On 09.02.2014 10:08, Duy Nguyen wrote:
>> On Tue, Feb 04, 2014 at 11:20:39AM +0100, Daniel Hahler wrote:
>
> Thanks for looking into this.
>
>>> when using a submodule "sm", there is a relative worktree in its config:
>>>
>>>    .git/modules/sm/config:
>>>    [core]
>>>     worktree = ../../../smworktree
>>>
>>> git-new-worktree (from contrib) symlinks this config the new worktree.
>>>
>>> From inside the new worktree, git reads the config, but resolves the
>>> relative worktree setting based on the symlink's location.
>>
>> Hmm.. core.worktree is relative to $GIT_DIR. Whether "config" is a
>> symlink should have no effects.
>
> If "config" is a symlink, the relative path for worktree is meant to be
> resolved based on the config file's location, and not from the symlink
> ($GIT_DIR).

I think you started with a wrong assumption. See config.txt it says

-- 8< -- core.worktree:: Set the path to the root of the working tree. This can be overridden by the GIT_WORK_TREE environment variable and the '--work-tree' command line option. The value can be an absolute path or relative to the path to the .git directory, which is either specified by --git-dir or GIT_DIR, or automatically discovered. -- 8< --

So I think it fails "correctly" (by the book).

Show 59 quoted lines

> Here is a test case to reproduce it:
>
>   # Create a submodule repo
>   mkdir /tmp/t-sm
>   cd /tmp/t-sm
>   git init
>   touch foo
>   git add foo
>   git commit -m init
>
>   # Create the root repo
>   mkdir /tmp/t-root
>   cd /tmp/t-root
>   git init
>   mkdir submodules
>   git submodule add /tmp/t-sm submodules/sm
>   git commit -m init
>
>   # Create a new worktree from the submodule
>   cd /tmp/t-root/submodules/sm
>   git-new-workdir . /tmp/new-workdir
>
> This then fails when checking out:
> + git checkout -f
> fatal: Could not chdir to '../../../../submodules/sm': No such file or directory
>
> % ls -l /tmp/new-workdir/.git/config
> […] /tmp/new-workdir/.git/config -> /tmp/t-root/.git/modules/submodules/sm/config
>
> % cat /tmp/new-workdir/.git/config
> [core]
>         repositoryformatversion = 0
>         filemode = true
>         bare = false
>         logallrefupdates = true
>         worktree = ../../../../submodules/sm
>
>
> From inside of /tmp/new-workdir `git rev-parse --git-dir` fails already
> (with the same "cannot chdir" error).
>
> The problem appears to be that it tries to chdir based on
> /tmp/new-workdir/.git, but should do so based on
> $(readlink -f .git/config).
>
> I recognize that this case is constructed anyway, because even if
> `worktree` would get resolved correctly, it would not be what you'd
> expect: the point of git-new-workdir is to get a separate worktree, and
> not use the existing one.
>
> Therefore I see two problems here:
> 1. worktree is not resolved correctly by git itself (with .git/config
>    being a symlink)
> 2. git-new-workdir should handle this better, e.g. by creating a copy of
>    the "config" file with the worktree setting removed and printing a
>    warning about it.
>
> The workaround appears to be to explicitly set
> GIT_WORK_TREE=/tmp/new-workdir.

No if you copy "config" out then it may be not what you want anymore (e.g. new remotes not reflected in new worktree). A better solution is move core.worktree out of "config". Notice t-root/submodules/sm is a file that contains the path to the true .git directory. We could have stored worktree path in that file instead of in "config".

-- 
Duy