From: Phillip Wood <phillip.wood123@gmail.com>
Date: Wed, 04 Aug 2021 18:14:34 GMT
Subject: Re: git revert --continue --no-verify
Message-ID: <74443c64-efe9-ea47-e918-d62b8c976abc@gmail.com>
In-Reply-To: <YQm+PWAtc3rixqsw@nand.local>

Hi Cameron and Taylor

On 03/08/2021 23:07, Taylor Blau wrote:
> On Tue, Aug 03, 2021 at 04:33:09PM -0500, Cameron Steffen wrote:
>> Perhaps the issue then is that the pre-commit hook should not run for
>> `git revert --continue`? It does not run for `git revert`.

The general rule for cherry-pick, revert and rebase is that if the 
commit message is edited then the commit is made with --verify and if 
the commit message is not being edited then the commit is made with 
--no-verify. I think the reasoning for this is that if the user has 
altered the commit message or contents and they have a pre-commit hook 
set then it is reasonable to check the new commit is acceptable to the 
hook. There are some exceptions to this which are oversights (for 
example I think we always pass --no-verify when committing a conflict 
resolution with 'git rebase --continue' even though the editor opens 
when committing in that case - I should try and fix that)

For 'git revert --continue' the commit is made when sequencer_continue() 
calls continue_single_pick() which looks like

static int continue_single_pick(struct repository *r, struct replay_opts 
*opts)
{
	struct strvec argv = STRVEC_INIT;
	int ret;

	if (!refs_ref_exists(get_main_ref_store(r), "CHERRY_PICK_HEAD") &&
	    !refs_ref_exists(get_main_ref_store(r), "REVERT_HEAD"))
		return error(_("no cherry-pick or revert in progress"));

	strvec_push(&argv, "commit");

	/*
	 * continue_single_pick() handles the case of recovering from a
	 * conflict.  should_edit() doesn't handle that case; for a conflict,
	 * we want to edit if the user asked for it, or if they didn't specify
	 * and stdin is a tty.
	 */
	if (!opts->edit || (opts->edit < 0 && !isatty(0)))
		/*
		 * Include --cleanup=strip as well because we don't want the
		 * "# Conflicts:" messages.
		 */
		strvec_pushl(&argv, "--no-edit", "--cleanup=strip", NULL);

	ret = run_command_v_opt(argv.v, RUN_GIT_CMD);
	strvec_clear(&argv);
	return ret;
}

It runs git commit directly and never passes --no-verify. I wouldn't be 
opposed to someone adding support for --no-verify (and --no-edit) to 
"cherry-pick/revert/rebase --continue" on the understanding that it only 
applied when committing the conflict resolution. There is a possible 
confusion for users though who might expect that the options passed with 
'--continue' applied to all the commits made by the command.

> This does look like an oversight to me, but you'll have to bear with me
> since I am relatively unfamiliar with the sequencer code.
> 
> Ultimately `git revert` calls do_pick_commit()

This is what happens when we are picking commits, the commit made with 
'--continue' uses a different code path

> which either calls
> do_commit() or run_git_commit(). A couple of curiosities there:
> 
>    - do_commit() does fall back to run_git_commit() if it has the
>      VERIFY_MSG bit set in `flags`.
>    - run_git_commit() passes `-n` only when VERIFY_MSG *isn't* set, so
>      the VERIFY_MSG bit does imply that the pre-commit hook would be run
>      there.
>    - when do_pick_commit() does have to fall back to run_git_commit(), it
>      sets the VERIFY_MSG bit in flags.

Calling run_git_commit() there is not a fallback it, that call is used 
to reword commits once they have been picked (see 450efe2d53 ("rebase 
-i: always update HEAD before rewording", 2019-08-19) for the reasoning 
behind this) . Because the message is being edited there is no point in 
going through do_commit().

> But we never end up calling run_git_commit() (except in the case of
> errors) because do_pick_commit() special-cases `command == TODO_REVERT`
> (which is the case for `git revert`) and calls `do_commit()`.
> 
> But it gets weirder: do_commit() calls run_git_commit() itself, but
> before the caller in do_pick_commit() has had a chance to add VERIFY_MSG
> to the flags.

do_commit() does not change the flags that it is called with - callers 
that want VERIFY_MSG will set that before they call do_commit(). 
do_commit() is there to commit simple picks without forking 'git commit'

> So I suspect that this is an oversight, but perhaps somebody more
> familiar with this code could confirm my thinking.

I hope the above helps - basically the idea is "if the commit has been 
edited use VERIFY_MSG" and --continue unhelpfully uses a completely 
different code-path to the main commit picking/reverting loop.

Best Wishes

Phillip

> Thanks,
> Taylor
>