[RFC] single-level refs vs push --all/--mirror

From: Robin H. Johnson <robbat2@gentoo.org>
Date: Nov 20, 2021, 00:03 UTC
Message-ID: <robbat2-20211119T234158-903098822Z@orbis-terrarum.net>
In-Reply-To: <robbat2-20211116T061554-223495439Z@orbis-terrarum.net>

On Tue, Nov 16, 2021 at 06:28:28AM +0000, Robin H. Johnson wrote:

Show 12 quoted lines

> > Perhaps there's a bug, but it is meant to and does push all refs in a
> > simple case:
> ...
> > There I used an explicit refspec naming it. But if I used "refs/*"
> > (which is what --mirror is doing under the hood), then it doesn't even
> > try sending it, as the client filters it out from the wildcard
> > (otherwise, everyone would get server-side errors from refs/stash).
> > 
> > So you probably want to choose a different refname to store your data.
> > 
> > I do think the status of these single-level refs is not well documented.
> > The rules in git-check-ref-format(1) imply that they are not valid:

..

Show 15 quoted lines

> >   They must contain at least one /. This enforces the presence of a
> >   category like heads/, tags/ etc. but the actual names are not
> >   restricted.
> > 
> > but that rule is not enforced internally, as "refs/" is sufficient for
> > the internal check_refname_format() to allow it. But receive-pack has,
> > since 1a7141ff28 (Ignore funny refname sent from remote, 2005-10-13),
> > implemented the format check after stripping refs/. And then the client
> > side followed that lead in 30affa1e9a (send-pack: do not send out
> > single-level refs such as refs/stash, 2008-10-29).
> 
> And the format of the ref wasn't noticed in the hook was previously
> posted back in 2014 :-(.
> https://www.spinics.net/lists/git/msg244053.html
> https://groups.google.com/forum/#!topic/gitolite/7cSrU6JorEY

Hi,

I didn't see response from Jeff King, so I'm wondering good next steps here.

I'm esp. surprised that git-stash ends up using single-level refs when git-check-ref-format(1) says they aren't valid.

I think Git should change git-stash and start issuing warnings to users for single-level refs.

Therein leads a question: What should be done with refs/stash?

Being able to work on stashes between multiple systems MIGHT be useful to some users, so maybe they do want it pushed with --mirror/--all.

Choices (not mutually exclusive):
1. Push the refs anyway
2. Make a configurable way to exclude refs from being pushed w/ --mirror/--all
   e.g. move refs/stash to the exclude list, then push everything not
   that is not excluded.
3. Change the rules to accept single-level refs?
4. Cleanup phases of warnings before errors and good migration
   mechanisms?
5. Something else?

-- 
Robin Hugh Johnson
Gentoo Linux: Dev, Infra Lead, Foundation Treasurer
E-Mail   : robbat2@gentoo.org
GnuPG FP : 11ACBA4F 4778E3F6 E4EDF38E B27B944E 34884E85
GnuPG FP : 7D0B3CEB E9B85B1F 825BCECF EE05E6F6 A48F6136