[1/3] rebase -r: do create merge commit after empty resolution

Commit Message

Philippe Blain March 28, 2025, 5:03 p.m. UTC

From: Philippe Blain <[e-mail]>

When a user runs 'git rebase --continue' to conclude a conflicted merge
during a 'git rebase -r' invocation, we do not create a merge commit if
the resolution was empty (i.e. if the index and HEAD are identical). We
simply continue the rebase as if no 'merge' instruction had been given.
This is confusing since all commits from the side branch are absent from
the rebased history. What's more, if that 'merge' is the last
instruction in the todo list, we fail to remove the merge state, such
that running 'git status' shows we are still merging after the rebase
has concluded.

This happens because in 'sequencer.c::commit_staged_changes', we exit
early before calling 'run_git_commit' if 'is_clean' is true, i.e. if
nothing is staged. Fix this by also checking for the presence of
MERGE_HEAD before exiting early, such that we do call 'run_git_commit'
when MERGE_HEAD is present. This also ensures that we unlink
git_path_merge_head later in 'commit_staged_changes' to clear the merge
state.

Make sure to also remove MERGE_HEAD when a merge command fails to start.
We already remove MERGE_MSG since e032abd5a0 (rebase: fix rewritten list
for failed pick, 2023-09-06). Removing MERGE_HEAD ensures that in this
situation, upon 'git rebase --continue' we still exit early in
'commit_staged_changes', without calling 'run_git_commit'. This is
already covered by t5407.11, which fails without this change because we
enter 'run_git_commit' and then fail to find 'rebase_path_message'.

Signed-off-by: Philippe Blain <[e-mail]>
---
 sequencer.c                |  3 ++-
 t/t3418-rebase-continue.sh | 24 ++++++++++++++++++++++++
 2 files changed, 26 insertions(+), 1 deletion(-)

Comments

Eric Sunshine March 28, 2025, 5:14 p.m. UTC | #1

On Fri, Mar 28, 2025 at 1:03 PM Philippe Blain via GitGitGadget
<[e-mail]> wrote:
> When a user runs 'git rebase --continue' to conclude a conflicted merge
> during a 'git rebase -r' invocation, we do not create a merge commit if
> the resolution was empty (i.e. if the index and HEAD are identical). We
> simply continue the rebase as if no 'merge' instruction had been given.
> This is confusing since all commits from the side branch are absent from
> the rebased history. What's more, if that 'merge' is the last
> instruction in the todo list, we fail to remove the merge state, such
> that running 'git status' shows we are still merging after the rebase
> has concluded.
> [...]
> Make sure to also remove MERGE_HEAD when a merge command fails to start.
> We already remove MERGE_MSG since e032abd5a0 (rebase: fix rewritten list
> for failed pick, 2023-09-06). Removing MERGE_HEAD ensures that in this
> situation, upon 'git rebase --continue' we still exit early in
> 'commit_staged_changes', without calling 'run_git_commit'. This is
> already covered by t5407.11, which fails without this change because we
> enter 'run_git_commit' and then fail to find 'rebase_path_message'.
>
> Signed-off-by: Philippe Blain <[e-mail]>
> ---
> diff --git a/t/t3418-rebase-continue.sh b/t/t3418-rebase-continue.sh
> +test_expect_success '--continue creates merge commit after empty resolution' '
> +       git reset --hard main &&
> +       git checkout -b rebase_i_merge &&
> +       test_commit unrelated &&
> +       git checkout -b rebase_i_merge_side &&
> +       test_commit side2 main.txt &&
> +       git checkout rebase_i_merge &&
> +       test_commit side1 main.txt &&
> +       PICK=$(git rev-parse --short rebase_i_merge) &&
> +       test_must_fail git merge rebase_i_merge_side &&
> +       echo side1 >main.txt &&
> +       git add main.txt &&
> +       test_tick &&
> +       git commit --no-edit &&
> +       FAKE_LINES="1 2 3 5 6 7 8 9 10 11" &&
> +       export FAKE_LINES &&
> +       test_must_fail git rebase -ir main &&

I don't think you want to be setting FAKE_LINES like this since doing
so will pollute the environment for all tests following this one. You
can find existing precedent in this script which demonstrates the
correct way to handle this case. Specifically, you'd want:

    test_must_fail env FAKE_LINES="1 2 3 5 6 7 8 9 10 11" \
        git rebase -ir main &&

> +       echo side1 >main.txt &&
> +       git add main.txt &&
> +       git rebase --continue &&
> +       git log --merges >out &&
> +       test_grep "Merge branch '\''rebase_i_merge_side'\''" out

You could take advantage of the SQ variable defined by t/test-lib.sh
to make this a bit easier to digest:

    test_grep "Merge branch ${SQ}rebase_i_merge_side${SQ}" out

Or, even simpler, you'll find that some test scripts just use regex
wildcard "." to make the needle even more readable:

    test_grep "Merge branch .rebase_i_merge_side." out

Eric Sunshine March 28, 2025, 5:23 p.m. UTC | #2

On Fri, Mar 28, 2025 at 1:14 PM Eric Sunshine <[e-mail]> wrote:
> On Fri, Mar 28, 2025 at 1:03 PM Philippe Blain via GitGitGadget
> <[e-mail]> wrote:
> > diff --git a/t/t3418-rebase-continue.sh b/t/t3418-rebase-continue.sh
> > +test_expect_success '--continue creates merge commit after empty resolution' '
> > +       [...]
> > +       git commit --no-edit &&
> > +       FAKE_LINES="1 2 3 5 6 7 8 9 10 11" &&
> > +       export FAKE_LINES &&
> > +       test_must_fail git rebase -ir main &&
>
> I don't think you want to be setting FAKE_LINES like this since doing
> so will pollute the environment for all tests following this one. You
> can find existing precedent in this script which demonstrates the
> correct way to handle this case. Specifically, you'd want:
>
>     test_must_fail env FAKE_LINES="1 2 3 5 6 7 8 9 10 11" \
>         git rebase -ir main &&

To clarify, by "pollute", I mean that it can impact subsequent tests
which don't take care to override FAKE_LINES as necessary. There
certainly are test scripts which use the:

    FAKE_LINES=... &&
    export FAKE_LINES &&

form successfully, but such scripts are careful to override/set
FAKE_LINES in every test. This particular script (t3418), on the other
hand, does not otherwise employ the form in which the variable is
exported, so introducing it in a test which is inserted into the
middle of the script feels dangerous.

Phillip Wood March 31, 2025, 3:37 p.m. UTC | #3

Hi Philippe

On 28/03/2025 17:03, Philippe Blain via GitGitGadget wrote:
> From: Philippe Blain <[e-mail]>
> 
> When a user runs 'git rebase --continue' to conclude a conflicted merge
> during a 'git rebase -r' invocation, we do not create a merge commit if
> the resolution was empty (i.e. if the index and HEAD are identical). We
> simply continue the rebase as if no 'merge' instruction had been given.
> This is confusing since all commits from the side branch are absent from
> the rebased history. What's more, if that 'merge' is the last
> instruction in the todo list, we fail to remove the merge state, such
> that running 'git status' shows we are still merging after the rebase
> has concluded.
> 
> This happens because in 'sequencer.c::commit_staged_changes', we exit
> early before calling 'run_git_commit' if 'is_clean' is true, i.e. if
> nothing is staged. Fix this by also checking for the presence of
> MERGE_HEAD before exiting early, such that we do call 'run_git_commit'
> when MERGE_HEAD is present. This also ensures that we unlink
> git_path_merge_head later in 'commit_staged_changes' to clear the merge
> state.
> 
> Make sure to also remove MERGE_HEAD when a merge command fails to start.
> We already remove MERGE_MSG since e032abd5a0 (rebase: fix rewritten list
> for failed pick, 2023-09-06). Removing MERGE_HEAD ensures that in this
> situation, upon 'git rebase --continue' we still exit early in
> 'commit_staged_changes', without calling 'run_git_commit'. This is
> already covered by t5407.11, which fails without this change because we
> enter 'run_git_commit' and then fail to find 'rebase_path_message'.

Thanks for fixing this.

> Signed-off-by: Philippe Blain <[e-mail]>
> ---
>   sequencer.c                |  3 ++-
>   t/t3418-rebase-continue.sh | 24 ++++++++++++++++++++++++
>   2 files changed, 26 insertions(+), 1 deletion(-)
> 
> diff --git a/sequencer.c b/sequencer.c
> index ad0ab75c8d4..2baaf716a3c 100644
> --- a/sequencer.c
> +++ b/sequencer.c
> @@ -4349,6 +4349,7 @@ static int do_merge(struct repository *r,
>   		error(_("could not even attempt to merge '%.*s'"),
>   		      merge_arg_len, arg);
>   		unlink(git_path_merge_msg(r));
> +		unlink(git_path_merge_head(r));


I think we want to clean up git_path_merge_mode() as well. Perhaps we 
should call remove_merge_branch_state() instead of deleting the 
individual files ourselves here.

> +test_expect_success '--continue creates merge commit after empty resolution' '
> +	git reset --hard main &&
> +	git checkout -b rebase_i_merge &&
> +	test_commit unrelated &&
> +	git checkout -b rebase_i_merge_side &&
> +	test_commit side2 main.txt &&
> +	git checkout rebase_i_merge &&
> +	test_commit side1 main.txt &&
> +	PICK=$(git rev-parse --short rebase_i_merge) &&
> +	test_must_fail git merge rebase_i_merge_side &&
> +	echo side1 >main.txt &&
> +	git add main.txt &&
> +	test_tick &&
> +	git commit --no-edit &&
> +	FAKE_LINES="1 2 3 5 6 7 8 9 10 11" &&
> +	export FAKE_LINES &&
> +	test_must_fail git rebase -ir main &&
> +	echo side1 >main.txt &&
> +	git add main.txt &&
> +	git rebase --continue &&
> +	git log --merges >out &&
> +	test_grep "Merge branch '\''rebase_i_merge_side'\''" out
> +'

I wonder if t3430 would be a better home for this as it already has the 
setup necessary to create a failing merge. It would be good to add a 
test to check that "git rebase --skip" does not create an empty merge as 
well.

Thanks

Phillip

>   test_expect_success '--skip after failed fixup cleans commit message' '
>   	test_when_finished "test_might_fail git rebase --abort" &&
>   	git checkout -b with-conflicting-fixup &&

Johannes Schindelin April 1, 2025, 4:17 p.m. UTC | #4

Hi Eric,

On Fri, 28 Mar 2025, Eric Sunshine wrote:

> On Fri, Mar 28, 2025 at 1:14 PM Eric Sunshine <[e-mail]> wrote:
> > On Fri, Mar 28, 2025 at 1:03 PM Philippe Blain via GitGitGadget
> > <[e-mail]> wrote:
> > > diff --git a/t/t3418-rebase-continue.sh b/t/t3418-rebase-continue.sh
> > > +test_expect_success '--continue creates merge commit after empty resolution' '
> > > +       [...]
> > > +       git commit --no-edit &&
> > > +       FAKE_LINES="1 2 3 5 6 7 8 9 10 11" &&
> > > +       export FAKE_LINES &&
> > > +       test_must_fail git rebase -ir main &&
> >
> > I don't think you want to be setting FAKE_LINES like this since doing
> > so will pollute the environment for all tests following this one. You
> > can find existing precedent in this script which demonstrates the
> > correct way to handle this case. Specifically, you'd want:
> >
> >     test_must_fail env FAKE_LINES="1 2 3 5 6 7 8 9 10 11" \
> >         git rebase -ir main &&
>
> To clarify, by "pollute", I mean that it can impact subsequent tests
> which don't take care to override FAKE_LINES as necessary. There
> certainly are test scripts which use the:
>
>     FAKE_LINES=... &&
>     export FAKE_LINES &&
>
> form successfully, but such scripts are careful to override/set
> FAKE_LINES in every test. This particular script (t3418), on the other
> hand, does not otherwise employ the form in which the variable is
> exported, so introducing it in a test which is inserted into the
> middle of the script feels dangerous.

The entire `FAKE_LINES` paradigm is broken, and since I suspect that it
was me who introduced it, I apologize.

A much better way to have done this would have been to write the string to
a certain file, say, $(git rev-parse --git-path sequencer.pick-lines), and
in the `fake-editor.sh`:

- test for the existence of that file, and if it exists
  - use its contents
  - delete that file

Ciao,
Johannes

Patch

diff --git a/sequencer.c b/sequencer.c
index ad0ab75c8d4..2baaf716a3c 100644
--- a/sequencer.c
+++ b/sequencer.c
@@ -4349,6 +4349,7 @@  static int do_merge(struct repository *r,
 		error(_("could not even attempt to merge '%.*s'"),
 		      merge_arg_len, arg);
 		unlink(git_path_merge_msg(r));
+		unlink(git_path_merge_head(r));
 		goto leave_merge;
 	}
 	/*
@@ -5364,7 +5365,7 @@  static int commit_staged_changes(struct repository *r,
 		flags |= AMEND_MSG;
 	}
 
-	if (is_clean) {
+	if (is_clean && !file_exists(git_path_merge_head(r))) {
 		if (refs_ref_exists(get_main_ref_store(r),
 				    "CHERRY_PICK_HEAD") &&
 		    refs_delete_ref(get_main_ref_store(r), "",
diff --git a/t/t3418-rebase-continue.sh b/t/t3418-rebase-continue.sh
index 127216f7225..f4b459fea16 100755
--- a/t/t3418-rebase-continue.sh
+++ b/t/t3418-rebase-continue.sh
@@ -111,6 +111,30 @@  test_expect_success 'rebase -r passes merge strategy options correctly' '
 	git rebase --continue
 '
 
+test_expect_success '--continue creates merge commit after empty resolution' '
+	git reset --hard main &&
+	git checkout -b rebase_i_merge &&
+	test_commit unrelated &&
+	git checkout -b rebase_i_merge_side &&
+	test_commit side2 main.txt &&
+	git checkout rebase_i_merge &&
+	test_commit side1 main.txt &&
+	PICK=$(git rev-parse --short rebase_i_merge) &&
+	test_must_fail git merge rebase_i_merge_side &&
+	echo side1 >main.txt &&
+	git add main.txt &&
+	test_tick &&
+	git commit --no-edit &&
+	FAKE_LINES="1 2 3 5 6 7 8 9 10 11" &&
+	export FAKE_LINES &&
+	test_must_fail git rebase -ir main &&
+	echo side1 >main.txt &&
+	git add main.txt &&
+	git rebase --continue &&
+	git log --merges >out &&
+	test_grep "Merge branch '\''rebase_i_merge_side'\''" out
+'
+
 test_expect_success '--skip after failed fixup cleans commit message' '
 	test_when_finished "test_might_fail git rebase --abort" &&
 	git checkout -b with-conflicting-fixup &&