rev-parse: fix several options when running in a subdirectoryjs/git-path-in-subdir

In addition to making git_path() aware of certain file names that need
to be handled differently e.g. when running in worktrees, the commit
557bd833bb (git_path(): be aware of file relocation in $GIT_DIR,
2014-11-30) also snuck in a new option for `git rev-parse`:
`--git-path`.

On the face of it, there is no obvious bug in that commit's diff: it
faithfully calls git_path() on the argument and prints it out, i.e. `git
rev-parse --git-path <filename>` has the same precise behavior as
calling `git_path("<filename>")` in C.

The problem lies deeper, much deeper. In hindsight (which is always
unfair), implementing the .git/ directory discovery in
`setup_git_directory()` by changing the working directory may have
allowed us to avoid passing around a struct that contains information
about the current repository, but it bought us many, many problems.

In this case, when being called in a subdirectory, `git rev-parse`
changes the working directory to the top-level directory before calling
`git_path()`. In the new working directory, the result is correct. But
in the working directory of the calling script, it is incorrect.

Example: when calling `git rev-parse --git-path HEAD` in, say, the
Documentation/ subdirectory of Git's own source code, the string
`.git/HEAD` is printed.

Side note: that bug is hidden when running in a subdirectory of a
worktree that was added by the `git worktree` command: in that case, the
(correct) absolute path of the `HEAD` file is printed.

In the interest of time, this patch does not go the "correct" route to
introduce a struct with repository information (and removing global
state in the process), instead this patch chooses to detect when the
command was called in a subdirectory and forces the result to be an
absolute path.

While at it, we are also fixing the output of --git-common-dir and
--shared-index-path.

Lastly, please note that we reuse the same strbuf for all of the
relative_path() calls; this avoids frequent allocation (and duplicated
code), and it does not risk memory leaks, for two reasons: 1) the
cmd_rev_parse() function does not return anywhere between the use of
the new strbuf instance and its final release, and 2) git-rev-parse is
one of these "one-shot" programs in Git, i.e. it exits after running
for a very short time, meaning that all allocated memory is released
with the exit() call anyway.

Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

1 files changed, 2 insertions, 2 deletions

diff --git a/t/t2027-worktree-list.sh b/t/t2027-worktree-list.sh
index c1a072348e..848da5f368 100755
--- a/t/t2027-worktree-list.sh
+++ b/t/t2027-worktree-list.sh
@@ -8,7 +8,7 @@ test_expect_success 'setup' '
 	test_commit init
 '
 
-test_expect_failure 'rev-parse --git-common-dir on main worktree' '
+test_expect_success 'rev-parse --git-common-dir on main worktree' '
 	git rev-parse --git-common-dir >actual &&
 	echo .git >expected &&
 	test_cmp expected actual &&
@@ -18,7 +18,7 @@ test_expect_failure 'rev-parse --git-common-dir on main worktree' '
 	test_cmp expected2 actual2
 '
 
-test_expect_failure 'rev-parse --git-path objects linked worktree' '
+test_expect_success 'rev-parse --git-path objects linked worktree' '
 	echo "$(git rev-parse --show-toplevel)/.git/objects" >expect &&
 	test_when_finished "rm -rf linked-tree && git worktree prune" &&
 	git worktree add --detach linked-tree master &&