Bug in test-lib.sh: test_create_repo() / RFC

Hi there,

running the test suite with -v for the upcoming release exposed a
certain problem with test_create_repo() whose consequences I can't quite
fathom at the moment. That means: I don't know whether it's maint
material or forbidden fruits during rc-cycle...

Problem:
Since a6d63b7 (test-lib: avoid assuming that templates/ are in the
GIT_EXEC_PATH, 2009-02-04), test_create_repo() assumes to be called from
a directory such that `pwd`/../templates/blt/ contains templates for
git-init.

Several tests (see below) call test_create_repo() from a different
directory, which means the repo is created without any of the default
files (and that a mv .git/hooks .git/hooks-disabled later in the
function errors out). Now, for most tests this probably doesn't matter
at all but it's not nice.

RFC:
I see several possible solutions:

- Make sure all tests use test_create_repo() from t/. Cumbersome and
fragile.

- Simply use $(TEST_DIRECTORY)/../templates/blt/. Nice and easy. But
uses the templates from the git repo containing t/ even when testing
against and installed git (just like now, for most of the tests).

- Teach git a "--templates-dir" option similar to "--html-path" and use
that (from the git actually being tested). Means we use the templates
belonging to the tested git; but also means we can test only git
versions containing that new option.

What do you think?

Michael

Affected tests:
t0050-filesystem.sh
t1007-hash-object.sh
t1302-repo-version.sh
t2103-update-index-ignore-missing.sh
t4027-diff-submodule.sh
t5300-pack-object.sh
t5513-fetch-track.sh
t5600-clone-fail-cleanup.sh
t5601-clone.sh
t5700-clone-reference.sh
t5710-info-alternate.sh
t6026-merge-attr.sh
t7001-mv.sh
t7010-setup.sh
t7401-submodule-summary.sh
t7506-status-submodule.sh
t7508-status.sh

On Mon, Apr 20, 2009 at 04:51:18PM +0200, Michael J Gruber wrote:

Problem:
Since a6d63b7 (test-lib: avoid assuming that templates/ are in the
GIT_EXEC_PATH, 2009-02-04), test_create_repo() assumes to be called from
a directory such that `pwd`/../templates/blt/ contains templates for
git-init.

Several tests (see below) call test_create_repo() from a different
directory, which means the repo is created without any of the default
files (and that a mv .git/hooks .git/hooks-disabled later in the
function errors out). Now, for most tests this probably doesn't matter
at all but it's not nice.

If I am understanding it correctly, I think this is simply a bug
introduced by a6d63b7, and the fix is maint-worthy.

And I think the right fix is:

- Simply use $(TEST_DIRECTORY)/../templates/blt/. Nice and easy. But
uses the templates from the git repo containing t/ even when testing
against and installed git (just like now, for most of the tests).

The original code (before a6d63b7) used $(GIT_EXEC_PATH), since it
handily pointed to "$(TEST_DIRECTORY)/.." (and I didn't check, but I
suspect that line predates $(TEST_DIRECTORY) entirely). So I think this
is just doing what the a6d63b7 should have done in the first place,
instead of using `pwd` to guess at the top-level location.

One thing to consider: we now have code to test an installed version of
git. Should it be using the vanilla hooks from the source directory, or
should it be using the regular templates? I think that question is
orthogonal to the bug you mention, though -- the current code is trying
to use the hooks from the source directory, but it is just failing to do
so; by fixing it, you certainly won't be breaking the
test-installed-version case.

-Peff