Re: [PATCH 0/8] Re: git complains packed-refs is not a directory when used with GIT_DIR and GIT_WORK_TREE envvars.

From: Jeff King <peff@peff.net>
Date: Mar 28, 2018, 18:19 UTC
Message-ID: <20180328181932.GB16565@sigill.intra.peff.net>
In-Reply-To: <20180328175537.17450-1-pclouds@gmail.com>

On Wed, Mar 28, 2018 at 07:55:29PM +0200, Nguyễn Thái Ngọc Duy wrote:

Show 10 quoted lines

> >> The problem is switching relative paths relies on the old $CWD if I'm
> >> not mistaken and we need  getcwd() for this. I'd love to have one
> >> callback that says "$CWD has been switched from this path to that
> >> path, do whatever you need to" that can be called any time, before or
> >> after chdir(). I'll look more into it.
> >
> > I think it should be OK to save getcwd() and just construct the original
> > path after the fact. Here's some patches which do that in a nice way.
> 
> Heh.. I should have checked mails more often while coding ;-)

I was worried about that. I started this earlier as a "well, I could probably just knock this out quickly" task. Many hours later, I found there were quite a few subtleties. :)

> This is what I got, which is slightly different from your series
> because I want to call set_git_dir() just one time (by
> setup_git_directory) and never again. I think the API looks close
> enough.

Yeah, I actually think after this series we could enforce that set_git_dir() is never called twice.

I think we could even make sure that repo_set_gitdir() is never called twice, too, but it would involve a little more surgery (we'd have to teach it to set up a child_notify handler, and the way it interacts with the set_git_dir() one is subtle. I did try it and it worked, but I went with what you saw in patch 3 because it was simpler).

> I will probably rework on top of your chdir-notify instead (and let
> yours to be merged earlier)

Thanks. I like some of the related changes you made, like including this in the tracing output. That should be easy to do on top of mine, I think.

I didn't look for other spots that might need similar treatment (like the object-store bits).

-Peff