CVE-2009-2695: Buffer Overflow

Description of problem: Dan Walsh's blog post mentioned in the previous comment details selinux-policy part of the issue tracked via bug #512284. As noted in the the blog, selinux-policy provides a boolean - allowunconfinedmmaplow - which controls whether mmapminaddr restriction is applied to the process. This boolean, however, did not work as expected, as unconfinedt domain (default domain for logged-in unprivileged users) was always permitted to map low memory pages regardless of the boolean setting. This problem with the boolean is being fixed and the fix will be included in future selinux-policy updates in RHEL5 and Fedora (see Dan's blog for NVRs).

Few notes specific to Red Hat Enterprise Linux 5: Support for mmapminaddr sysctl was not included in the GA version of RHEL5. It was only added in kernel update in 5.2 (it first appeared upstream in 2.6.24). selinux-policy was, to avoid breaking applications needing low memory pages mapping on upgrade from 5.1 to 5.2, configured to allow mmapzero in unconfined domains and only disallow it in confined domains (e.g. various network facing services). In 5.3, allowunconfinedmmaplow boolean was added, but it's default value for targeted policy was changed to on, i.e. allowing mmapzero in all unconfined domains by default. This default boolean value is planned to remain unchanged in 5.4.

It should also be noted, that even with allowunconfinedmmaplow boolean set to off, it is still possible for unconfinedt user to transition to other domain that is permitted to mmapzero, as details in the Dan's blog post. Upstream discussion on how to best address this issue is still ongoing.

Eric's proposed patches moving mmapminaddr check out of security were submitted. If they are accepted upstream, mmapminaddr will be checked before LSM hooks are called. Security modules will only be consulted if mmapminaddr check has passed (e.g. when mmapminaddr is 0), so SELinux may still be able to restrict mapping of the low pages / zero page for confined domains and permit mapping in unconfined, even when mmapminaddr is 0.

Eric's patches with further discussion: http://patchwork.kernel.org/patch/36540/ http://patchwork.kernel.org/patch/36539/

Updated version: http://patchwork.kernel.org/patch/36650/ http://patchwork.kernel.org/patch/36649/

Further discussion of the proposed change: http://thread.gmane.org/gmane.linux.kernel.lsm/9075

mmapminaddr on SELinux and non-SELinux systems http://eparis.livejournal.com/606.html

Confining the unconfined. Oxymoron? http://danwalsh.livejournal.com/30084.html

Kbase: http://kbase.redhat.com/faq/docs/DOC-18042