Re: [PATCH 3/3] rename_ref: replace empty deletion message in HEAD's log

From: Jeff King <peff@peff.net>
Date: Feb 17, 2017, 23:53 UTC
Message-ID: <20170217235324.n3jabck23bvd4cs2@sigill.intra.peff.net>
In-Reply-To: <871suwqtar.fsf@kyleam.com>

On Fri, Feb 17, 2017 at 06:41:32PM -0500, Kyle Meyer wrote:

Show 17 quoted lines

> Junio C Hamano <gitster@pobox.com> writes:
> 
> [...]
> 
> > Do we even want these "internal" delete_ref() invocations to be
> > logged in HEAD's reflog?  I understand that this is inside the
> > implementation of renaming an old ref to a new ref, and reflog
> > message given to delete_ref() would matter only if the HEAD happens
> > to be pointing at old ref---but then HEAD will be repointed to the
> > new ref by somebody else [*1*] that called this function to rename
> > old to new and it _will_ log it.
> 
> I know the discussion has developed further, but just a note that I
> think the last statement is inaccurate: currently, a rename will not be
> recorded in HEAD's log.  "git branch -m" will show a renaming event in
> the new branch's log, but the only trace of the event in HEAD's log is
> the deletion entry with an empty message.

Right. I assumed Junio was talking about the hypothetical behavior we'd want.

Your response did make me think of one other option: if we updated HEAD _before_ writing the new ref, then it would automatically get the "create" half of the rename. IOW, if the rename process were:

  1. delete old; this writes a reflog to HEAD, per the magic
     HEAD-detection in commit_ref_update().

  2. update HEAD to point to new

  3. re-create new; this uses the same magic to write a HEAD reflog.

That's probably a crazy path to go, though. Right now steps (1) and (3) happen in a low-level function, and step (2) happens outside of it completely. Arguably it would be good to change that, but I think we want to think of (1) and (3) as an atomic operation. Putting more things which might fail between them is a bad idea.

So I think your existing patches (modulo the review comments), plus the "something like this" that I sent on top are probably a better end state.

-Peff