config: only read .git/config from configured repos

When git_config() runs, it looks in the system, user-wide,
and repo-level config files. It gets the latter by calling
git_pathdup(), which in turn calls get_git_dir(). If we
haven't set up the git repository yet, this may simply
return ".git", and we will look at ".git/config".  This
seems like it would be helpful (presumably we haven't set up
the repository yet, so it tries to find it), but it turns
out to be a bad idea for a few reasons:

  - it's not sufficient, and therefore hides bugs in a
    confusing way. Config will be respected if commands are
    run from the top-level of the working tree, but not from
    a subdirectory.

  - it's not always true that we haven't set up the
    repository _yet_; we may not want to do it at all. For
    instance, if you run "git init /some/path" from inside
    another repository, it should not load config from the
    existing repository.

  - there might be a path ".git/config", but it is not the
    actual repository we would find via setup_git_directory().
    This may happen, e.g., if you are storing a git
    repository inside another git repository, but have
    munged one of the files in such a way that the
    inner repository is not valid (e.g., by removing HEAD).

We have at least two bugs of the second type in git-init,
introduced by ae5f677 (lazily load core.sharedrepository,
2016-03-11). It causes init to use git_configset(), which
loads all of the config, including values from the current
repo (if any).  This shows up in two ways:

  1. If we happen to be in an existing repository directory,
     we'll read and respect core.sharedrepository from it,
     even though it should have no bearing on the new
     repository. A new test in t1301 covers this.

  2. Similarly, if we're in an existing repo that sets
     core.logallrefupdates, that will cause init to fail to
     set it in a newly created repository (because it thinks
     that the user's templates already did so). A new test
     in t0001 covers this.

We also need to adjust an existing test in t1302, which
gives another example of why this patch is an improvement.

That test creates an embedded repository with a bogus
core.repositoryformatversion of "99". It wants to make sure
that we actually stop at the bogus repo rather than
continuing upward to find the outer repo. So it checks that
"git config core.repositoryformatversion" returns 99. But
that only works because we blindly read ".git/config", even
though we _know_ we're in a repository whose vintage we do
not understand.

After this patch, we avoid reading config from the unknown
vintage repository at all, which is a safer choice.  But we
need to tweak the test, since core.repositoryformatversion
will not return 99; it will claim that it could not find the
variable at all.

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

1 files changed, 7 insertions, 0 deletions

diff --git a/environment.c b/environment.c
index bccdee9d0b..20a870c3d7 100644
--- a/environment.c
+++ b/environment.c
@@ -194,6 +194,13 @@ int is_bare_repository(void)
 	return is_bare_repository_cfg && !get_git_work_tree();
 }
 
+int have_git_dir(void)
+{
+	return startup_info->have_repository
+		|| git_dir
+		|| getenv(GIT_DIR_ENVIRONMENT);
+}
+
 const char *get_git_dir(void)
 {
 	if (!git_dir)