Thread (5 messages) 5 messages, 4 authors, 2017-10-30

Re: [PATCH] rebase: exec leaks GIT_DIR to environment

From: Phillip Wood <hidden>
Date: 2017-10-29 18:34:39

On 28/10/17 17:00, Johannes Schindelin wrote:
quoted hunk ↗ jump to hunk
Hi Jake,

On Fri, 27 Oct 2017, Jacob Keller wrote:
quoted
From: Jacob Keller <redacted>

I noticed a failure with git rebase interactive mode which causes "exec"
commands to be run with GIT_DIR set. When GIT_DIR is in the environment,
then any command which results in running a git command in
a subdirectory will fail because GIT_DIR=".git".

This unfortunately breaks one of my project's Makefiles, which uses
git-describe to find the version information, but does so from within
a sub directory.

I'm in the process of running a bisect to find where this got
introduced, but I suspect it's part of the rebase--helper changes that
happened a while ago.
A safe assumption. I do not know how the shell code managed that GIT_DIR
reset, though:

-- snip from v2.12.0's git-rebase--interactive.sh --
        x|"exec")
                read -r command rest < "$todo"
                mark_action_done
                eval_gettextln "Executing: \$rest"
                "${SHELL:-@SHELL_PATH@}" -c "$rest" # Actual execution
-- snap --

Maybe you can spot it?

The fix should be as easy as

-- snip --
diff --git a/sequencer.c b/sequencer.c
index f2c84c2fa62..018ba8d27e2 100644
--- a/sequencer.c
+++ b/sequencer.c
@@ -1859,11 +1859,13 @@ static int error_failed_squash(struct commit *commit,
 static int do_exec(const char *command_line)
 {
 	const char *child_argv[] = { NULL, NULL };
+	const char *child_env[] = { "GIT_DIR", NULL };
 	int dirty, status;
 
 	fprintf(stderr, "Executing: %s\n", command_line);
 	child_argv[0] = command_line;
-	status = run_command_v_opt(child_argv, RUN_USING_SHELL);
+	status = run_command_v_opt_cd_env(child_argv, RUN_USING_SHELL, NULL,
+					  child_env);
 
 	/* force re-reading of the cache */
 	if (discard_cache() < 0 || read_cache() < 0)
-- snap --
Just clearing GIT_DIR does not match the behavior of the shell version
(tested by passing -p to avoid rebase--helper) as that passes GIT_DIR to
exec commands if it has been explicitly set. I think that users that set
GIT_DIR on the command line would expect it to be propagated to exec
commands.

$ git rebase -px'echo $GIT_DIR' @

                                                            Merge commit
'7c2f1abd64' into phil
Executing: echo $GIT_DIR

Successfully rebased and updated refs/heads/phil.

$ env GIT_DIR=.git git rebase -px'echo $GIT_DIR' @

                                                            Merge commit
'7c2f1abd64' into phil
Executing: echo $GIT_DIR
/home/phil/Documents/src/git/.git/worktrees/git-next
Successfully rebased and updated refs/heads/phil.
quoted hunk ↗ jump to hunk
*However*, your test still fails with this, as

- your added test tries to remove the directory with -ff instead of -rf

- it tries to run `git rebase --abort` afterwards, which fails with my fix
  because there is no rebase in progress

- instead of `cd subdir && ...`, it calls `>cd subdir && ...`, which
  causes it to abort with a "subdir: not fonud"

So I need this, too, to make it all work:

-- snip --
diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase-interactive.sh
index 60ab5136f70..967caab222a 100755
--- a/t/t3404-rebase-interactive.sh
+++ b/t/t3404-rebase-interactive.sh
@@ -108,13 +108,13 @@ test_expect_success 'rebase -i with the exec command runs from tree root' '
 	rm -fr subdir
 '
 
-test_expect_failure 'rebase -i with the exec git commands in subdirs still work' '
-	test_when_finished "rm -ff subdir" &&
-	test_when_finished "git rebase --abort" &&
+test_expect_success 'rebase -i with the exec git commands in subdirs still work' '
+	test_when_finished "rm -rf subdir" &&
+	test_when_finished "git rebase --abort || :" &&
 	git checkout master &&
 	mkdir subdir && (cd subdir &&
 	set_fake_editor &&
-	FAKE_LINES="1 exec_>cd_subdir_&&_git_rev-parse_--is-inside-work-tree" \
+	FAKE_LINES="1 exec_cd_subdir_&&_git_rev-parse_--is-inside-work-tree" \
 		git rebase -i HEAD^
 	)
 '
-- snap --
I only had time to write these two patches, and to verify that t3404
passes now, but not that anything else passes, neither to write a proper
commit message.

Maybe you can take it from there?

Ciao,
Dscho
  
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help