In my previous post, a very valid question was raised by Alexander E. Patrakov: why still use chroot if you have SELinux?
Both chroot (especially with the additional restrictions that grSecurity enables on chroots that make it more difficult to break out of a chroot) and SELinux try to isolate an application so it only has access to those resources it needs. Chroot does this on file-level basis (and a bit more with grSecurity), SELinux on more general resources. However, things that make SELinux strong (flexible and detailed policy language, fine-grained authorizations) are also its weakness (consolidating files into groups having the same file label), and chroot does have an advantage on this.
Suppose that a flaw exists in BIND through which an attacker can read files on the host (through BIND). With SELinux, the domain in which BIND runs is prohibited from accessing and reading files whose label is not one of the labels that the policy thinks BIND should be able to read. More specifically, the BIND policy in the reference policy (which is what both Gentoo and RedHat base their policies on, and generally policies are only enlarged, never really shrinked):
- etc_runtime_t (read) means access to the files in /etc that are modified at runtime (like mtab, profile.env, gentoo's /etc/env.d)
- named_var_run_t (read) is access to /var/run/bind and /var/run/named (and a few other related locations)
- named_checkconf_exec_t (read/execute) is access to read and execute /usr/sbin/named-checkconf
- named_conf_t (read) to read the BIND-related configuration files
- dnssec_t (read) to read the DNSSEC keyfiles
- locale_t (read) to access /etc/localtime, /usr/share/locale/*, /usr/share/zoneinfo/*
- etc_t (read) to read the general configuration files in /etc (including passwd, fstab, ...)
- proc_t (read), proc_net_t (read) and sysfs_t (read) to access those pseudo filesystems
- udev_tbl_t (read) to access /dev/.udev and /var/run/udev (but I have no idea yet why this is in)
- named_log_t (read/write) for the log files of BIND
- net_conf_t (read) to access /etc/hosts (including deny/allow), resolv.conf, ...
- named_exec_t (read/execute) the BIND executables
- named_zone_t (read) to access the zone files, also write access in case of slave system
- cert_t (read) to read certificate information
- named_cache_t (read/write) to access its cache
- named_tmp_t (read/write) to work with temporary files
Isolation provided by SELinux is as powerful as the width of its labeling. For instance, by giving the named daemon read access to /etc files like passwd, fstab, group, hosts, resolv.conf and more, a malicious user who can exploit this hypothetical vulnerability can obtain information that might help him in his further attempts. By chrooting BIND, the files placed in the chroot itself should not offer the information he might be looking for (for instance, the passwd file, if needed at all, is limited to just the named and root accounts, etc.)
Chrooting, but not enabling SELinux, could lead to escalation. A chroot cannot restrict what a process is allowed to do beyond the regular access privileges that are given on the user. If a user can upload an exploit through BIND and have BIND execute it, he can use this as an attack vector for further activities. SELinux here prohibits BIND to write stuff it can also execute (there is no write and execute privilege defined here). It also ensures that the BIND daemon never exists his security domain (transitioning towards another domain with perhaps other privileges) as there are no transition rules from named_t to any other domain.
Another MAC system that would be better suited to fit both is grSecurity's RBAC model. Iirc, it uses path definitions to say which files are allowed to access and which not. The weakness SELinux here has (aggregation into sets of files with the same label) doesn't exist for grSecurity. This debate on path-based versus label-based access controls have been going on for very long time now - just google it ;-)
So, Alexander, in short: chroot further limits the SELinux-allowed privileges to a more fine-grained set of file system resources (files/directories).