With git-1.7, I regularly get this when pushing to a non-bare repo:

remote: error: failed to lock refs/remotes/origin/master
...
 ! [remote rejected] origin/master -> origin/master (failed to lock)
error: failed to push some refs to 'ssh://remote/repo'

I see this pushing from a Mac, to either a Mac or to Cygwin. I have
seen it on various versions of the 1.7 series and possibly before. In
the instance above:

local git version 1.7.0.3.436.g2b878
remote git version 1.7.0.3

In addition, the ref is (apparently) updated just fine. If I push a
second time, no update occurs.

The problem is reproducible. Any idea where I should start debugging it?

j.
--

[ message continues ]

Ah, this is the problem:

$ git push mirror
Total 0 (delta 0), reused 0 (delta 0)
error: Ref refs/remotes/origin/master is at
81358fbe72926d74bdeda85669d655e144572c48 but expected
3c0a87afc2e9248890dd6de40b5039bcb48c8516
remote: error: failed to lock refs/remotes/origin/master
To ssh://mirror/repo
   3c0a87a..81358fb  origin/HEAD -> origin/HEAD
 ! [remote rejected] origin/master -> origin/master (failed to lock)
error: failed to push some refs to 'ssh://mirror/repo'

The origin/HEAD symref and --mirror do not get along together. Hmm.

j.
--

[ message continues ]

Proposal: receive-pack should look through the list of heads it has
received and check whether each is locally a symref. If so, and if the
destination of the symref is also among the list of heads, it should
ignore the symref update.

In the example above, receive-pack would determine that
refs/remotes/origin/HEAD locally is a symref to
refs/remotes/origin/master. Since refs/remotes/origin/master is in the
list of heads, it would ignore the update for
refs/remotes/origin/HEAD.

If that sounds reasonable, I can work on a patch.

In case you're wondering how I got into this situation:

server$ git clone --bare /shared/project.git ~/backup/project.git
client$ git clone server:/shared/project.git
client$ cd project
client$ git remote add --mirror backup server:~/backup/project.git
client$ work, commit, work, commit
client$ git push backup

j.
--

[ message continues ]

I would rather think that "git push" should notice that */HEAD is a symref
and refrain from pushing it.  Stated more precisely,

 - It is perfectly fine to do this:

   $ git push there HEAD:somebranch

 - But it does not make sense to push a symref via

   [remote "there"]
   	push = refs/*:refs/*

   to a bare repository that is used as a back-up mirror by including
   refs/HEAD:refs/HEAD in the result of wildcard expansion.


--

[ message continues ]

But nothing prevents the destination repo from having arbitrary
symrefs. And there is no way for the source to know which destination
refs may be symrefs and which are not.

So it seems safer to me that the destination should ignore duplicates
in the received list.

No?

j.
--

[ message continues ]

We probably need it on both ends.

On the receiving end, it probably is a better idea to make verify_lock()
succeed if the ref has already been updated to the value we are going to
update with, as I suspect that the issue is not limited to receive-pack.
It needs some restructuring of lock/verify/commit primitives, though.
--

[ message continues ]