git daemon DoS

[Todd Zullinger]

Greetings,

I'm seeking help and input in dealing with a git-daemon bug that was
recently fixed upstream.  Discussion of the bug and the patch for
git's maint branch is in the git list archives:

http://article.gmane.org/gmane.comp.version-control.git/120724

The patch in the first mail of the thread isn't what was queued in
maint.  What was committed is at:

http://git.kernel.org/?p=git/git.git;a=commitdiff;h=73bb33a9

Of the active Fedora/EPEL branches, only devel and F-11 are recent
enough for this to apply cleanly.  The other branches required a small
amount of reworking to account for changes made to git-daemon since
the releases those branches were based upon.  I don't think the
backporting is all that difficult, but I am not a strong C coder.  Any
extra eyes on my backported patches would be most helpful.

A simple way to test this against a git server, taken from the initial
patch in the git list thread above:

$ perl -e '
    $s="git-upload-pack git\0user=me\0host=localhost\0";
    printf "%4.4x%s",4+length $s,$s
' | nc $GITHOST 9418

This will cause the git-daemon process spawned via xinetd to enter an
infinite loop.  New requests will still be handled, as xinetd will
spawn a new git-daemon process.  But, of course, an attacker can
easily cause many git-daemon processes to be started that will not
exit.

I've created patched packages with the backported patch for EL-{4,5}
and F-10 (F-9 is in sync with F-10, so the same spec/srpm should work
there).  These packages and patches against current CVS are at:

http://tmz.fedorapeople.org/tmp/git-daemon-extra-args/

(Apologies for the minor non-related changes in some of the diffs, as
those were changes I had slated for release soon and didn't want to
revert at the last minute.)

Thanks for any time the good folks on this list could give to checking
and/or testing these packages.  I've not had time to see if any other
distros have backported this fix to compare.  If anyone has a moment
to do so, it would be much appreciated.

If it would be better to file this in bugzilla rather than discuss it
here, let me know and I'll file it accordingly.

[Tomas Hoger]

Hi Todd!

On Fri, 12 Jun 2009 00:02:06 -0400 Todd Zullinger  wrote:
> http://git.kernel.org/?p=git/git.git;a=commitdiff;h=73bb33a9
>
> Of the active Fedora/EPEL branches, only devel and F-11 are recent
> enough for this to apply cleanly.  The other branches required a small
> amount of reworking to account for changes made to git-daemon since
> the releases those branches were based upon.  I don't think the
> backporting is all that difficult, but I am not a strong C coder.  Any
> extra eyes on my backported patches would be most helpful.

Your backported patches seem to do the same thing as the upstream
commit.  Look good to me and seem to fix the issue in both F-10 and
EL-4 version I quick-tested.
> A simple way to test this against a git server, taken from the initial
> patch in the git list thread above:
>
> $ perl -e '
>     $s="git-upload-pack git\0user=me\0host=localhost\0";
>     printf "%4.4x%s",4+length $s,$s
> ' | nc $GITHOST 9418

Or | git-daemon --inetd --base-path=`pwd` --export-all as suggested in
the upstream mailing list thread.
> Thanks for any time the good folks on this list could give to checking
> and/or testing these packages.  I've not had time to see if any other
> distros have backported this fix to compare.  If anyone has a moment
> to do so, it would be much appreciated.

I'm not sure any distro already worked on backports, the issue does not
seem to be fixed in any released git version yet.
> If it would be better to file this in bugzilla rather than discuss it
> here, let me know and I'll file it accordingly.

Feel free to file in BZ, also for Bodhi update request reference.

[Todd Zullinger]

Hello Tomas,

Tomas Hoger wrote:
> Your backported patches seem to do the same thing as the upstream
> commit.  Look good to me and seem to fix the issue in both F-10 and
> EL-4 version I quick-tested.

Great, thanks for checking and testing them.
> > A simple way to test this against a git server, taken from the
> > initial patch in the git list thread above:
> >
> > $ perl -e '
> >     $s="git-upload-pack git\0user=me\0host=localhost\0";
> >     printf "%4.4x%s",4+length $s,$s
> > ' | nc $GITHOST 9418
>
> Or | git-daemon --inetd --base-path=`pwd` --export-all as suggested
> in the upstream mailing list thread.

Yeah.  My reason for using netcat was to show how the git-daemon
processes built up and consumed CPU. :)
> I'm not sure any distro already worked on backports, the issue does
> not seem to be fixed in any released git version yet.

Correct.  I don't when a 1.6.3.3 is slated.  Usually git doesn't go
too long between releases, but Junio said he's been busier than normal
in the last "What's cooking in git.git" note to the git list.
> Feel free to file in BZ, also for Bodhi update request reference.

Filed as https://bugzilla.redhat.com/show_bug.cgi?id=505761

Thanks,