Race condition in git push --mirror can cause silent ref rewinding

[Alex Vandiver <alex@chmrr.net>]

Heya,

We recently ran into a particularly troubling race condition, discovered
in git 2.0.0.  The setup for it is as follows:

The repository is a bare repository, which developers push to via ssh;
it mirrors its changes out onto github.  In its config:

    [remote "github"]
        url = git@github.com:bestpractical/rt.git
        fetch = +refs/*:refs/*
        mirror = yes

It has a post-receive hook which does:

    sudo -u git -H /usr/bin/git push github


We recently saw a situation where a push of a new branch caused a
simultaneous update of a different branch (by a different user) to be
rewound.  From the reflog of the created branch (4.2/html-gumbo-loading):

    0000000000000000000000000000000000000000
1aefd600fcbb5ded14376f77d77a14758668fb39 Wallace Reis
<wreis@bestpractical.com> 1404326443 -0400       push

And the updated branch (4.2-trunk), which was rewound:

    44dc8ad0e4603e3f674b7c00deacc122ca52707a
1e743b6225d502ad1a265929fb873f4c0bf4f8a5 Kevin Falcone
<falcone@bestpractical.com> 1404326446 -0400    push
    1e743b6225d502ad1a265929fb873f4c0bf4f8a5
44dc8ad0e4603e3f674b7c00deacc122ca52707a git <git@bestpractical.com>
1404326446 -0400        update by push

It is my belief that this comes because the "--mirror" argument causes
the local refs to be treated as tracking refs -- and thus updates all of
them during the push.  I believe the race condition is thus:

  1. User A starts a push --mirror; git records the values of the refs

  2. User B updates a ref, commit mail goes out, etc

  3. User A's push completes, updates "tracking" branch to value at (1).


Needless to say, silently losing commits which appeared for all purposes
to be pushed successfully (neither User A nor User B sees anything out
of the ordinary) is extremely troubling.

 - Alex