Re: [PATCH] checkout: fix BUG() case in 9081a421a6

[Junio C Hamano <gitster@pobox.com>]

Ævar Arnfjörð Bjarmason  <avarab@gmail.com> writes:

> Fix a regression in my 9081a421a6d (checkout: fix "branch info" memory
> leaks, 2021-11-16) where I'd assumed that the old_branch_info.path
> would have to start with refs/heads/*, but as has been reported[1]
> that's not the case.
>
> As a test case[2] to reproduce this shows the second "git checkout"
> here runs into the BUG() in the pre-image. The test being added is
> amended from[2] and will pass both with this change, and before
> 9081a421a6. I.e. our behavior now is again the same as before that
> commit.

> +test_expect_success REFFILES 'checkout a branch without refs/heads/* prefix' '
> +	git clone --no-tags . repo-odd-prefix &&
> +	(
> +		cd repo-odd-prefix &&
> +
> +		cp .git/refs/remotes/origin/HEAD .git/refs/heads/a-branch &&

I am not sure if this is a sensible test case to begin with.

It sets up recursive symbolic ref in this way:

	HEAD points at refs/heads/a-branch
	refs/heads/a-branch points at refs/remotes/origin/HEAD
	refs/remotes/origin/HEAD points at refs/remotes/origin/branch1

The checked out branch (i.e. what HEAD points at) is nominally a
local branch, but it totally violates the spirit of the safety valve
that says "HEAD" MUST point at a local branch or otherwise it is
detached.  Creating a commit while "a-branch" is checked out would
not affect *ANY* local branch state and instead makes an update to
the remote tracking branch that does not reflect *any* past states
at the remote repository.  Even worse, a "git fetch" that updates
the remote tracking branches will make the HEAD, the index and the
working tree into an inconsistent state.

Simply put, I think the BUG() is catching a case where we should
have been diagnosing as a broken repository.

So from my point of view, BUG() is indeed inappropriate because what
the conditional statement noticed was a broken repository, and not a
programming bug.

What we should never do is to promise this "only kosher in letter
but not in spirit" violation of "HEAD must point at a local branch"
rule will be supported.

So, unless I hear more convincing arguments (and Todd's example or
anything similar that makes "git commit" from that state update a
ref outside local branches is *not*), I am hesitant to call the new
behaviour and 9081a421a6d a regression.

What did the code before that BUG() do when faced with this nonsense
configuration?  If forbidding outright broke a sensible workflow
that happened to have been "working", I am OK to demote it to
warning() and restore the previous behaviour temporarily, whatever
it was (I think it was just old_branch_info.name was left unset
because we were not on local branch, but I don't know if the missing
.name was making any irrecoverable damage).  But the longer term
direction should be that we treat the "update HEAD ends up updating
some ref outside refs/heads/" a longstanding bug that needs to be
fixed.

Thanks.