Re: git clone (ssh://) skips detached HEAD

From: Dmitry Ivankov <divanorama@gmail.com>
Date: Jun 1, 2011, 22:18 UTC
Message-ID: <BANLkTi=pK4pu=XpdPWWCaQ2KdEQRXB29pw@mail.gmail.com>
In-Reply-To: <20110601220518.GA32681@sigill.intra.peff.net>

Show 11 quoted lines

>> git clone ssh://127.0.0.1/`pwd`/test test2
>> remote: Counting objects: 3, done.
>> remote: Total 3 (delta 0), reused 0 (delta 0)
>> Receiving objects: 100% (3/3), done.
>> error: Trying to write ref HEAD with nonexistant object
>> 91dbc2403853783f637744c31036f94a66084286
>> fatal: Cannot update the ref 'HEAD'.
>
> This is quite bad behavior. In addition to the ugly error messages, it
> actually aborts the clone. So it is impossible to clone a repo with a
> detached HEAD that is not otherwise referenced.

Which could be evil without resumable clone on a big repo. git clone --bare will work though.

Show 12 quoted lines

> We basically have two choices:
>
>  1. Fetch objects for HEAD on clone.
>
>  2. Don't checkout a detached HEAD if we don't have the object (or
>     possibly, don't checkout a detached HEAD at all; we already do
>     something similar for the case of a HEAD that points to a bogus
>     branch).
>
> I think (2) is more consistent with the refspec we set up, but (1) is
> probably more convenient to users (and better matches the case where the
> remote is on a detached HEAD that _does_ point to something we have).

I like (2) a bit more too, and give a hint about explicit HEAD fetching.

> I think it should just be a matter of adding HEAD to the list of things
> we fetch, although we don't actually want to write it to a ref. I'll
> take a look.

I tried to hack it through via adding HEAD:HEADtmp static const refspec to remote.{h,c} and hit "Ignoring funny ref 'HEAD' locally" in remote.c:get_fetch_map, commented out that ignore and got:

error: unable to resolve reference HEAD: No such file or directory fatal: Cannot lock the ref 'HEAD'.

from builtin/clone.c:
...
                        const char *head = skip_prefix(our_head_points_at->name,
                                                       "refs/heads/");

                        update_ref(reflog_msg.buf, "HEAD",
                                   our_head_points_at->old_sha1,
                                   NULL, 0, DIE_ON_ERR);
...
afaik from gdb HEAD ended up being a symlink to "" ref or so.