Re: BUG: fetch in certain repo always gives "did not send all necessary objects"

From: Jeff King <peff@peff.net>
To: Elijah Newren <newren@gmail.com>
Cc: "Stefan Beller" <sbeller@google.com>,
	"Git Mailing List" <git@vger.kernel.org>,
	"Nguyễn Thái Ngọc" <pclouds@gmail.com>
Subject: Re: BUG: fetch in certain repo always gives "did not send all necessary objects"
Date: Wed, 7 Feb 2018 13:17:06 -0500	[thread overview]
Message-ID: <20180207181706.GA4227@sigill.intra.peff.net> (raw)
In-Reply-To: <CABPp-BH0rSELVJV11MJM=KFU2jAEuhVTzUUBskghgku+y_ozDQ@mail.gmail.com>

On Wed, Feb 07, 2018 at 09:25:42AM -0800, Elijah Newren wrote:

> > So other_head_refs knows that it's looking at the worktrees. And it
> > passes the alternate ref-store to refs_head_ref(), with "add_one_ref" as
> > the callback. But the knowledge that we're not talking about the real
> > "HEAD" is lost as we cross that callback boundary. We'd need to either
> > add another parameter to the callback, or have some way of talking about
> > "HEAD in this worktree" as a refname (which AFAIK we don't have).
> 
> Can we use "worktrees/${WORKTREE}/HEAD"?  It already satisfies all the
> necessary rev-parse rules...

True, but it's mostly an accident that it works. And once we have ref
backends besides the filesystem, it will probably stop working.

I think there was discussion at some point of embedding worktree refs
into the normal ref namespace, but I don't know what came of it (it's
not a feature I've followed very closely).

> (And on a slight tangent...do we want to start disallowing the
> creation of branches/tags whose name starts with "worktrees/",
> "refs/", "hooks/", or other paths that exists under gitdir?  Making a
> branch named "refs/heads/foo" so that it fully-qualifies as
> "refs/heads/refs/heads/foo" is always fun)

We recently taught the porcelain to disallow a branch named "HEAD".
Though I think there are actually two related problems with different
solutions. One is saying something like:

  git checkout -b HEAD

or:

  git checkout -b refs/heads/foo

both of which will not do what you want, and leave you with a
funnily-named branch in the ref namespace.

But that's separate from the fact that:

  git rev-parse info/refs

will look at a file that is not a ref at all. Long-term I think the
solution is storage formats that don't mingle with other files. But we
could probably teach even the files-backend that any ref at the
top-level is supposed to be either in refs/, or to consist only of
"[A-Z_]".

-Peff