Re: CVE Request -- kernel: sysctl: restrict write access to dmesg_restrict

[Solar Designer <solar () openwall com>]

On Wed, Oct 26, 2011 at 05:16:12PM +0200, Petr Matousek wrote:
> When dmesg_restrict is set to 1 CAP_SYS_ADMIN is needed to read the
> kernel ring buffer. But a root user without CAP_SYS_ADMIN is able
> to reset dmesg_restrict to 0.

FWIW, here's how this is implemented in OpenVZ, at least in RHEL5 branch
kernels:

suse114:/ # id
uid=0(root) gid=0(root) groups=0(root)
suse114:/ # sysctl -a|fgrep dmesg
kernel.dmesg_restrict = 1
suse114:/ # sysctl -w kernel.dmesg_restrict=0
error: "Operation not permitted" setting key "kernel.dmesg_restrict"
suse114:/ # uname -mrs
Linux 2.6.18-274.3.1.el5.028stab094.3.owl1 x86_64
suse114:/ # dmesg
klogctl: Operation not permitted

This is OpenSUSE 11.4 (arbitrary, whatever I happened to have as a
result of some other testing) inside an OpenVZ container on Owl. ;-)

I was actually thinking of making dmesg_restrict tri-state in OpenVZ, to
have a setting that would allow dmesg to work in containers but not on
host.  With OpenVZ, containers have their own dmesg buffers anyway, with
very little info getting in there (e.g., output from virtualized
iptables logging goes in there, but I've never seen any kernel pointer
exposed in an in-container dmesg).  Fully disabling dmesg in containers
just because we want it disabled for non-root on host seems overkill for
some uses, but that's what is currently implemented when a host admin
sets dmesg_restrict to 1.  There's room for improvement here, from a
usability standpoint.

As to security, this shows that OpenVZ is mature and LXC is not. ;-)

Alexander