Re: [PATCH v3 2/2] rebase-interactive: warn if commit is dropped with `rebase --edit-todo'

Hi Alban

On 09/01/2020 21:13, Alban Gruin wrote:

Hi Phillip,

Le 09/12/2019 à 17:00, Phillip Wood a écrit :

quoted

quoted

quoted

diff --git a/rebase-interactive.c b/rebase-interactive.c
index ad5dd49c31..80b6a2e7a6 100644
--- a/rebase-interactive.c
+++ b/rebase-interactive.c

@@ -97,7 +97,8 @@ int edit_todo_list(struct repository *r, struct

todo_list *todo_list,
              struct todo_list *new_todo, const char *shortrevisions,
              const char *shortonto, unsigned flags)
   {
-    const char *todo_file = rebase_path_todo();
+    const char *todo_file = rebase_path_todo(),
+        *todo_backup = rebase_path_todo_backup();
       /* If the user is editing the todo list, we first try to parse

@@ -110,9 +111,9 @@ int edit_todo_list(struct repository *r, struct

todo_list *todo_list,
                       -1, flags | TODO_LIST_SHORTEN_IDS |
TODO_LIST_APPEND_TODO_HELP))
           return error_errno(_("could not write '%s'"), todo_file);
   -    if (initial && copy_file(rebase_path_todo_backup(), todo_file,
0666))
-        return error(_("could not copy '%s' to '%s'."), todo_file,
-                 rebase_path_todo_backup());
+    unlink(todo_backup);
+    if (copy_file(todo_backup, todo_file, 0666))
+        return error(_("could not copy '%s' to '%s'."), todo_file,
todo_backup);

We used to copy ONLY when initial is set and we left old todo_backup
intact when !initial.  That is no longer true after this change, but
it is intended---we create an exact copy of what we would hand out
to the end-user, so that we can compare it with the edited result
to figure out what got changed.

I think it would be better to only create a new copy if the last edit
was successful. As it stands if I edit the todo list and accidentally
delete some lines and then edit the todo list again to try and fix it
the second edit will succeed whether or not I reinserted the deleted lines.

We could add this to the tests to check that a subsequent edit that does
not fix the problem fails

diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase-interactive.sh
index 969e12d281..8544d8ab2c 100755
--- a/t/t3404-rebase-interactive.sh
+++ b/t/t3404-rebase-interactive.sh

@@ -1416,6 +1416,7 @@ test_expect_success 'rebase --edit-todo respects

rebase.missingCommitsCheck = er
                 test_i18ncmp expect actual &&
                 test_must_fail git rebase --continue 2>actual &&
                 test_i18ncmp expect actual &&
+               test_must_fail git rebase --edit-todo &&
                 cp orig .git/rebase-merge/git-rebase-todo &&
                 test_must_fail env FAKE_LINES="1 2 3 4" \
                         git rebase --edit-todo 2>actual &&

In which case, if the check did not pass at the previous edit, the new
todo list should be compared to the backup.  As sequencer_continue()
already does this, extract this to its own function in
rebase-interactive.c.  To keep track of this, a file is created on the
disk (as you suggested in your other email.)  At the next edit, if this
file exists and no errors were found, it is deleted.  The backup is only
created if there is no errors in `todo_list' and in `new_todo'.

This would guarantee that there is no errors in the backup, and that the
edited list is always compared to a list exempt of errors.

This approach also has the benefit to detect if a commit part of a
badcmd was dropped.

After some tweaks (ie. `expect' now lists 2 commits instead of one),
this passes the test with the change you suggested, and the one you sent
in your other email.

That sounds good. I'm not sure how it passes the test in my other email though, if sequencer_continue() compares the todo list to the backup wont it still fail when continuing after conflicts as the backup is out of date?

Best Wishes

Phillip

quoted

quoted

We unlink(2) unconditionally because the only effect we want to see
here is that todo_backup does not exist before we call copy_file()
that wants to do O_CREAT|O_EXCL.  I wonder if we want to avoid
unlink() when initial, and also if we want to do unlink_or_warn()
when !initial (read: this is just "wondering" without thinking long
enough to suggest that doing so would be better)

quoted

diff --git a/t/t3404-rebase-interactive.sh

b/t/t3404-rebase-interactive.sh
index 29a35840ed..9051c1e11d 100755

--- a/t/t3404-rebase-interactive.sh
+++ b/t/t3404-rebase-interactive.sh

@@ -1343,6 +1343,89 @@ test_expect_success 'rebase -i respects

rebase.missingCommitsCheck = error' '
       test B = $(git cat-file commit HEAD^ | sed -ne \$p)
   '
   +test_expect_success 'rebase --edit-todo respects
rebase.missingCommitsCheck = ignore' '
+    test_config rebase.missingCommitsCheck ignore &&
+    rebase_setup_and_clean missing-commit &&
+    set_fake_editor &&
+    FAKE_LINES="break 1 2 3 4 5" git rebase -i --root &&
+    FAKE_LINES="1 2 3 4" git rebase --edit-todo 2>actual &&

OK, so we lost "pick 5" but with missing-check disabled, that should
not trigger any annoying warning or error.

quoted

+    git rebase --continue 2>actual &&

This clobbers actual which hasn't been used yet

Good catch.

quoted

quoted

quoted

+    test D = $(git cat-file commit HEAD | sed -ne \$p) &&

quoted

+    test_i18ngrep \
+        "Successfully rebased and updated refs/heads/missing-commit" \
+        actual
+'
+
+test_expect_success 'rebase --edit-todo respects
rebase.missingCommitsCheck = warn' '
+    cat >expect <<-EOF &&
+    error: invalid line 1: badcmd $(git rev-list --pretty=oneline
--abbrev-commit -1 master~4)
+    Warning: some commits may have been dropped accidentally.
+    Dropped commits (newer to older):
+     - $(git rev-list --pretty=oneline --abbrev-commit -1 master)
+    To avoid this message, use "drop" to explicitly remove a commit.
+    EOF
+    tail -n4 expect >expect.2 &&
+    test_config rebase.missingCommitsCheck warn &&
+    rebase_setup_and_clean missing-commit &&
+    set_fake_editor &&
+    test_must_fail env FAKE_LINES="bad 1 2 3 4 5" \
+        git rebase -i --root &&
+    cp .git/rebase-merge/git-rebase-todo.backup orig &&
+    FAKE_LINES="2 3 4" git rebase --edit-todo 2>actual.2 &&
+    head -n5 actual.2 >actual &&
+    test_i18ncmp expect actual &&

OK, so we lost "pick 1" while discarding "bad", and we should notice
the lossage?  I see "head -n5" there, which means we are still
getting "invalid line 1: badcmd", even though FAKE_LINES now got rid
of "bad"?  Puzzled...

Is the bad there to stop the rebase so we can edit the todo list? If so
it would be better to use 'break' instead.

No, it was here to show that we can detect dropped commits, even if the
todo list has an error.

quoted

Best Wishes

Phillip

Cheers,
Alban