A few weeks ago, I wrote a blog on removing capabilities from a container. But what if you want to add capabilities?

While I recommend that people remove capabilities, in certain situations users need to add capabilities in order to get their container to run.

One example is when you have a app that needs a single capability, like an Network Time Protocol (NTP) daemon container that resets the system time on a machine. So if you wanted to run a container for an ntp daemon, you would need to do a --cap-add SYS_TIME. Sadly, many users don't think this through, or understand what it means to add a capability.

The most dangerous capability: SYS_ADMIN

Whenever I present on container security I explain that the power of root was originally broken into 32 separate capabilities. These capabilities were originally fairly fine grained. But two things conspired to make some of them become very powerful.

  • As number of available capability slots were used up, the cost of using up a new capability versus just assigning a permission check to an existing capability increased.
  • Kernel developers being lazy and not wanting to go through the hassle of allocating a new capability began to say stuff like: "This permission check is an admin function, I have a capability called SYS_ADMIN, I will just use that."

This led the SYS_ADMIN capability to need two pages of the capability man page just to list the features given to a process with the SYS_ADMIN check.

The main feature that containers take advantage of when removing the ability for processes inside of the container to mount new file systems.

Recently I received an email from a co-worker showing me what they could do with the SYS_ADMIN capability.

Talking about selinuxfs, there is also an quick exploit that allows disabling the host selinux from a container, although so far it needs CAP_SYS_ADMIN (to mount selinuxfs) and without --selinux-enabled (like in previous overlayfs) in docker.


# docker run -it --rm --cap-add SYS_ADMIN rhel7 bash
[root@d2b1fb6bd57c /] mount -t selinuxfs none /mnt
[root@d2b1fb6bd57c /] getenforce
Enforcing
[root@d2b1fb6bd57c /] setenforce 0
[root@d2b1fb6bd57c /] getenforce
Permissive
[root@d2b1fb6bd57c /] exit
# getenforce
Permissive

(And we all know that disabling SELinux this makes me cry.)

Use Sparingly!

Bottom line: giving a process SYS_ADMIN capability is pretty close to removing all isolation. As a rule, you shouldn't do this, but it's helpful to know that you can if you need to for some reason.


About the author

Joe Brockmeier is the editorial director of the Red Hat Blog. He also acts as Vice President of Marketing & Publicity for the Apache Software Foundation.

Read full bio