Subject: mmap_min_addr and your local LSM (ok, just SELinux)
From: Eric Paris <eparis@redhat.com>
To: linux-kernel@vger.kernel.org, linux-security-module@vger.kernel.org,
       selinux@tycho.nsa.gov
Cc: sds@tycho.nsa.gov, jmorris@namei.org, spender@grsecurity.net,
       dwalsh@redhat.com, cl@linux-foundation.org, arjan@infradead.org,
       alan@lxorguk.ukuu.org.uk
Content-Type: text/plain
Date: Mon, 20 Jul 2009 19:23:43 -0400
Message-Id: <1248132223.2654.278.camel@localhost>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3943
Lines: 80

Brad Spengler recently pointed out that the SELinux decision on how to
handle mmap_min_addr in some ways weakens system security vs on a system
without SELinux (and in other ways can be stronger).  There is a trade
off and a reason I did what I did but I would like ideas and discussion
on how to get the best of both worlds.

With SELinux mapping the 0 page requires an SELinux policy permission,
mmap_zero.  Without SELinux mapping the 0 page requires CAP_SYS_RAWIO.
Note that CAP_SYS_RAWIO roughly translates to uid=0 since noone really
does interesting things with capabilities.

The main problem is WINE.  I'm told that WINE needs to map the 0 page to
support 16bit applications.  On distros without SELinux users must
disable the mmap_min_addr protections for the ENTIRE system if they want
to run WINE.

http://wiki.winehq.org/PreloaderPageZeroProblem

I believe (from reading mailing lists) if you install WINE on ubuntu it
automatically disables these protections.  Thus installing wine on
ubuntu disables ALL hardening gains of the mmap_min_addr.

On Fedora, with SELinux, we allow users to run WINE in a domain that has
the SELinux mmap_zero permission and thus other programs/domains, do not
have security weakened.  Your daemons, like the web server, are still
unable to map the 0 page.  This is different than distros without
SELinux, remember they have to disable protection globally.

But logged in users (by default), under SELinux, are 'unconfined' and
can by their very nature run their program in a domain that allows
mmap_zero.  Trying to 'confine' the 'unconfined' user with SELinux is an
open problem which we don't currently even reasonably attempt address on
a broad scale.  It's like besieging the user in a gentle mist of water
hoping they won't try to escape.

So in Fedora your web server is a harder entry point to exploit kernel
NULL pointer bugs, but you have no protections against a malicious user.
On Ubuntu if you install WINE your web server and your logged in users
have no hardening.  If you do not install WINE non-root is hardened,
anything running as root is not (aka suid apps, aka pulseaudio).

So I was thinking today, wondering how to get the best (or at least
better) of both worlds on an SELinux system.  I was considering adding a
second mmap_min_addr_lsm which would typically be equal to
mmap_min_addr.  The purpose would be to allow the sysadmin to
individually control DAC/LSM protections.  The security checks would
turn (sort of) into

if (addr < mmap_min_addr)
   ret |= capable(CAP_SYS_RAWIO);
if (addr < mmap_min_addr_lsm)
   ret |= [insert LSM check here]

So on a non-SELinux system users would end up with exactly what they
have today.  if you want to run WINE as a normal user you have to set
mmap_min_addr = 0 and then you no longer need CAP_SYS_RAWIO.  Not much
else we can do if your distro down support fine grained permissions.

On an SELinux system what this lets me do is default to a stricter
setup, one in which you have to have both CAP_SYS_RAWIO and the selinux
mmap_zero permission.  You, out of the box, get protection for both your
malicious logged in user and your web server.  Then if a user decides to
run WINE they would turn down mmap_min_addr.  This would remove the
requirement that they are root, and leave the system vulnerable to a
malicious user, but would still allow SELinux to protect confined
domains and daemons.

Does anyone see a better way to let users continue to be users while
protecting most people?  Yes SELinux is stronger in some areas than
without confining the ability to map the 0 page, but as has be rightly
pointed out it's foolish an broken that SELinux can weaken any
protections.

-Eric

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/