Demonstrate and fix a rebase --autostash bug with dirty submodules

This bug report came in via Git for Windows (already with version 2.19.0, but I misread the reporter's enthusiasm to take matters into his own hands).

The culprit is, in a nutshell, that the built-in rebase tries to run git stash only when the worktree is dirty, but it includes submodules in that. However, git stash cannot do anything about submodules, and if the only changes are in submodules, then it won't even give us back an OID, and the built-in rebase acts surprised.

The solution is easy: simply exclude the submodules from the question whether the worktree is dirty.

What is surprisingly not simple is to get the regression test right. For that reason, and because I firmly believe that it is easier to verify a fix for a regression when the regression test is introduced separately (i.e. making it simple to verify that there is a regression), I really want to keep the first patch separate from the second one.

Since this bug concerns the built-in rebase, I based the patches on top of next.

Johannes Schindelin (2):
rebase --autostash: demonstrate a problem with dirty submodules rebase --autostash: fix issue with dirty submodules

 builtin/rebase.c            |  2 +-
 t/t3420-rebase-autostash.sh | 10 ++++++++++
 2 files changed, 11 insertions(+), 1 deletion(-)

base-commit: 209f214ca4ae4e301fc32e59ab26f937082f3ea3
Published-As: https://github.com/gitgitgadget/git/releases/tags/pr-56%2Fdscho%2Ffix-built-in-rebase-autostash-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-56/dscho/fix-built-in-rebase-autostash-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/56

gitgitgadget

From: Johannes Schindelin <[email protected]>

It has been reported that dirty submodules cause problems with the built-in rebase when it is asked to autostash. The symptom is:

fatal: Unexpected stash response: ''

This patch adds a regression test that demonstrates that bug.

Original report: https://github.com/git-for-windows/git/issues/1820

---
 t/t3420-rebase-autostash.sh | 10 ++++++++++
 1 file changed, 10 insertions(+)

index 0c4eefec7..7eb9f9041 100755
@@ -351,4 +351,14 @@ test_expect_success 'autostash is saved on editor failure with conflict' '
 	test_cmp expected file0
 '
 
+test_expect_failure 'autostash with dirty submodules' '
+	test_when_finished "git reset --hard && git checkout master" &&
+	git checkout -b with-submodule &&
+	git submodule add ./ sub &&
+	test_tick &&
+	git commit -m add-submodule &&
+	echo changed >sub/file0 &&
+	git rebase -i --autostash HEAD
+'
+
 test_done

gitgitgadget

From: Johannes Schindelin <[email protected]>

Since we cannot stash dirty submodules, there is no use in requiring them to be clean (or stash them when they are not).

This brings the built-in rebase in line with the previous, scripted version, which also did not care about dirty submodules (but it was admittedly not very easy to figure that out).

This fixes https://github.com/git-for-windows/git/issues/1820

---
 builtin/rebase.c            | 2 +-
 t/t3420-rebase-autostash.sh | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

index 09f55bfb9..1dd24301f 100644
@@ -1349,7 +1349,7 @@ int cmd_rebase(int argc, const char **argv, const char *prefix)
 			update_index_if_able(&the_index, &lock_file);
 		rollback_lock_file(&lock_file);
 
-		if (has_unstaged_changes(0) || has_uncommitted_changes(0)) {
+		if (has_unstaged_changes(1) || has_uncommitted_changes(1)) {
 			const char *autostash =
 				state_dir_path("autostash", &options);
 			struct child_process stash = CHILD_PROCESS_INIT;

index 7eb9f9041..f355c6825 100755
@@ -351,7 +351,7 @@ test_expect_success 'autostash is saved on editor failure with conflict' '
 	test_cmp expected file0
 '
 
-test_expect_failure 'autostash with dirty submodules' '
+test_expect_success 'autostash with dirty submodules' '
 	test_when_finished "git reset --hard && git checkout master" &&
 	git checkout -b with-submodule &&
 	git submodule add ./ sub &&

gitgitgadget

There's another bug with rebase.autoStash in master (and next/pu) but not v2.19.0. I tried to bisect bit it just comes down to 5541bd5b8f ("rebase: default to using the builtin rebase", 2018-08-08).

Credit to a co-worker of mine who wishes to remain anonymous for discovering this. I narrowed down his test-case to (any repo will do):

    (
        rm -rf /tmp/todo &&
        git clone --single-branch --no-tags --branch=todo https://github.com/git/git.git /tmp/todo &&
        cd /tmp/todo &&
        rm Make &&
        git rev-parse --abbrev-ref HEAD &&
        git -c rebase.autoStash=true -c pull.rebase=true pull &&
        if test $(git rev-parse --abbrev-ref HEAD) != 'todo'
        then
            echo 'On detached head!' &&
            git status &&
            exit 1
        else
            echo 'We are still on our todo branch!'
        fi
    )

That's just flipping the config default. If you add "-c rebase.usebuiltin=true" to your invocation here:

you can bisect further. However, the results weren't super useful, as I had to skip a bunch of commits (ones that did die("TODO") or just complained that the working tree wasn't clean; if you treat the latter as "bad", then it just bisects to e0333e5c63 (builtin rebase: require a clean worktree, 2018-09-04).

-Peff

Hi Ævar,

I found the culprit. Patch forthcoming,
Dscho