SELinux User-Based Access Control


Sven Vermeulen Mon 02 May 2011

Within the reference policy, support is given to a feature called UBAC constraints. Here, UBAC stands for User Based Access Control. The idea behind the constraint is that any activity between two types (say foo_t and bar_t) can be prohibited if the user contexts of the resources that are using those types are different. So even though foo_t can read files with label bar_t, a process running as user1:user_r:foo_t will not be able to read a file labeled user2:user_r:bar_t. The policy defines the constraint like so (taken from policy/constraints and rewritten in a more readable code):

Action is okay if
  user1 == user2, or
  user1 == system_u, or
  user2 == system_u, or
  type1 is not a UBAC constrained type, or
  type2 is not a UBAC constrained type

So the constraint only denies an activity if the users involved are not system_u (that would render your system useless), not the same, and both types are ubac constrained types. The latter is, within the policy, set using type attributes:

~$ seinfo -aubac_constrained_type -x

Some domains are also UBAC exempt (currently I know of the sysadm_t domain - cfr the ubacproc and ubacfile attributes), meaning that activities started from the sysadm_t domain will not trigger the constraint.

UBAC gives some additional control on information flow between resources. But it isn't perfect. One major downside is that the error you get when the constraint is hit is a simple AVC denial where most users would just check the inter-type privileges, without paying attention to the difference in SELinux user identities. Another is that it might be difficult for users or administrators that use different SELinux user identities to still work properly with UBAC constrained domains. Work is on the way in the SELinux development to improve the role-based access control (RBAC) by allowing files and directories to have a role as well (rather than the object_r placeholder used currently) and then work on those roles. You can then grant the users that need access to a particular resource the necessary role rather than requiring those users to use the same SELinux user id. This would take at least one major downside of UBAC away and I'm hoping that the logging will improve on this as well.

Of course, I do not ramble about UBAC here because it is fun (well yes, yes it is fun) but because in Gentoo, we've hit one UBAC-related issue. When a user starts vixie-cron, the root crontab would fail to be loaded. What gives? The root crontab has the SELinux identity of staff_u (as it is created by a regular staff user that su(do)'ed) whereas the cronjob_t process would have the SELinux identity of root. Bang. Dead. No error beyond what vixie-cron gives.

Of course this can be easily worked around. chcon -u root /var/spool/cron/crontabs/root works, or you can recreate the crontab as a console-logged-on root user. We could also change the default context used by cronjob_t to use staff_u:sysadm_r:cronjob_t for root. But we can also take a look at how other distributions do this. What gives: most distributions disable UBAC within the policy. Their reasons might vary, but manageability of the policy comes to mind, as well as reducing the number of (difficult to debug) problems. Most are keen to include the RBAC at some point in the future though. Some discussion on #gentoo-hardened and #selinux later, and I decided to use a USE flag called "ubac" to optionally enable UBAC within the policy. How very Gentoo, isn't it? At least users have the choice of using UBAC or not (I know I'm going to enable it) and when RBAC is available, we'll definitely make sure that support for RBAC is available too.

Currently in the hardened overlay, sec-policy/selinux-base-policy-2.20101213-r13. Take your pick on it, give it a try and report any bugs you have on Bugzilla. And if you enable USE="ubac", you get user based access control for free.

PS I'm also going to reapply for Gentoo developer-ship and, amongst other things, help out the hardened team with SELinux policies and documentation.