From: Phillip Wood Date: Thu, 20 Jul 2023 09:42:31 GMT Subject: Re: [PATCH] sequencer: finish parsing the todo list despite an invalid first line Message-ID: <395274b4-37a9-8c95-203f-94178c99772a@gmail.com> In-Reply-To: <20230719144339.447852-1-alexhenrie24@gmail.com> Hi Alex Thanks for working on this. On 19/07/2023 15:43, Alex Henrie wrote: > ddb81e5072 (rebase-interactive: use todo_list_write_to_file() in > edit_todo_list(), 2019-03-05) made edit_todo_list more efficient by > replacing transform_todo_file with todo_list_parse_insn_buffer. > Unfortunately, that innocuous change caused a regression because > todo_list_parse_insn_buffer would stop parsing after encountering an > invalid 'fixup' line. If the user accidentally made the first line a > 'fixup' and tried to recover from their mistake with `git rebase > --edit-todo`, all of the commands after the first would be lost. I found this description a little confusing as transform_todo_file() also called todo_list_parse_insn_buffer(). transform_todo_file() does not exist anymore but it looked like static int transform_todo_file(unsigned flags) { const char *todo_file = rebase_path_todo(); struct todo_list todo_list = TODO_LIST_INIT; int res; if (strbuf_read_file(&todo_list.buf, todo_file, 0) < 0) return error_errno(_("could not read '%s'."), todo_file); if (todo_list_parse_insn_buffer(the_repository, todo_list.buf.buf, &todo_list)) { todo_list_release(&todo_list); return error(_("unusable todo list: '%s'"), todo_file); } res = todo_list_write_to_file(the_repository, &todo_list, todo_file, NULL, NULL, -1, flags); todo_list_release(&todo_list); if (res) return error_errno(_("could not write '%s'."), todo_file); return 0; } If it could not parse the todo list it did not try and write it to disc. After ddb81e5072 this changed as edit_todo_list() tries to shorten the OIDs in the todo list before it is edited even if it cannot be parsed. The fix below works around that by making sure we always try and parse the whole todo list even if the the first line is a fixup command. That is a worthwhile improvement because it means we notify the user of all the errors we find rather than just the first one and is in keeping with the way we handle other invalid lines. It does not however fix the root cause of this regression which is the change in behavior in edit_todo_list(). After the user edits the todo file we do not try to transform the OIDs if it cannot be parsed or has missing commits. Therefore it still contains the shortened OIDs and editing hints and there is no need for edit_todo_list() to call write_todo_list() when "incorrect" is true. > To avoid throwing away important parts of the todo list, change > todo_list_parse_insn_buffer to keep going and not return early on error. > > Signed-off-by: Alex Henrie > --- > sequencer.c | 2 +- > t/t3404-rebase-interactive.sh | 19 +++++++++++++++++++ > 2 files changed, 20 insertions(+), 1 deletion(-) > > diff --git a/sequencer.c b/sequencer.c > index cc9821ece2..adc9cfb4df 100644 > --- a/sequencer.c > +++ b/sequencer.c > @@ -2702,7 +2702,7 @@ int todo_list_parse_insn_buffer(struct repository *r, char *buf, > if (fixup_okay) > ; /* do nothing */ > else if (is_fixup(item->command)) > - return error(_("cannot '%s' without a previous commit"), > + res = error(_("cannot '%s' without a previous commit"), > command_to_string(item->command)); > else if (!is_noop(item->command)) > fixup_okay = 1; > diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase-interactive.sh > index ff0afad63e..d2801ffee4 100755 > --- a/t/t3404-rebase-interactive.sh > +++ b/t/t3404-rebase-interactive.sh > @@ -1596,6 +1596,25 @@ test_expect_success 'static check of bad command' ' > test C = $(git cat-file commit HEAD^ | sed -ne \$p) > ' > > +test_expect_success 'the first command cannot be a fixup' ' > + # When using `git rebase --edit-todo` to recover from this error, ensure > + # that none of the original todo list is lost > + rebase_setup_and_clean fixup-first && > + ( > + set_fake_editor && > + test_must_fail env FAKE_LINES="fixup 1 2 3 4 5" \ > + git rebase -i --root 2>actual && Thanks for taking the time to add a test. It is not worth a re-roll on its own, but there is no need to use "--root" here. It is confusing as it is not clear if we're refusing "fixup" as the first command because we're rewriting the root commit or if we always refuse to have "fixup" as the first command. As an aside this restriction is pretty easy to defeat. In fact I think we probably allow a todo list that starts with [new root] fixup which is a bug. We certainly allow todo lists starting with exec true / label