Linux capabilities support for user namespaces

Linux capabilities are a sparsely used kernel facility to add granularity to the set of privileges that a process can have. By using capabilities, an administrator can grant a process a limited set of privileges, rather than the usual, essentially binary, choice between granting all privileges via setuid() or granting just those of the user running the program. Combining capabilities with user namespaces will allow administrators to apply those fine-grained privileges to containers, which is just what a patch set proposed by Serge E. Hallyn sets out to do.

We have looked at capabilities several times in the past, most recently in the context of adding capability sets to files, though an earlier article provides more details on the rules that govern how capabilities are applied and inherited. With the addition of file capabilities, Linux systems have all the tools needed to eliminate most setuid programs though, in practice, that hasn't happened. There is an effort underway to eliminate most setuid programs for Fedora 15, however.

Namespaces are part of the Linux containers implementation, which is a lightweight virtualization technique that allows groups of processes to run in their own little world, separate from the rest of the processes running on the system. These containers must not be able to see or interact with things outside, so various global resources (things like process IDs, network devices, filesystems, and so on) need to be wrapped in a namespace layer that provides the illusion that the container is its own system. User namespaces provide a container with its own set of UIDs, completely separate from those in the parent. Each of the different kinds of namespaces can be created by using flags to the clone() system call.

The idea behind Hallyn's patches, the core of which was originally developed by Eric Biederman, is to eventually allow unprivileged users to create namespaces. In order to do that, the capabilities of processes in a namespace must not leak out to parent (or even sibling) namespaces. In the core patch, Hallyn says that the proposed changes accomplish 90% of the goal to allow unprivileged namespace creation, with some UID confusion issues still to be addressed.

In the initial user namespace—the "normal" namespace that is created at boot time—capabilities for a task are calculated in the usual way, using the permitted, effective, and inheritable capability sets associated with the task. The proposed changes will restrict any capabilities in a child user namespace to only act within that namespace or on any of its descendants.

Each capabilities set is contained in a structure that references the user it corresponds to, and those user structures have a namespace to which they are attached. When checking to determine whether a particular set of capabilities should be used, the code looks at whether the user is part of the target namespace. If so, its capabilities are used, if not, each parent namespace is checked all the way back to the initial user namespace. Since the capabilities can only be associated with one namespace (via a user in that namespace), they are only active in the namespace that contains them or any descendant from that namespace.

The user that creates the namespace will have all capabilities in that namespace, not just the set of capabilities they have in the parent. Essentially, the creator has the privileges of the root user in any namespace he or she creates.

In order to ensure that the namespace creator's capabilities don't leak out to the rest of the system, a new capability check is added in the patch:

    int ns_capable(struct user_namespace *ns, int cap);

In this way, kernel functionality that requires certain capabilities can be incrementally added to user namespaces while still protecting the rest of the kernel from being affected. Hallyn's patches enable three specific capabilities for user namespaces by making the change from capable() to ns_capable(). The first, and simplest, just allows the sethostname() system call to be successfully called if the user in the namespace has CAP_SYSADMIN. The second, which is slightly more complicated, but still a pretty small change, alters check_kill_permission() to allow CAP_KILL enabled tasks to send a signal to another task. The last patch allows CAP_SYS_PTRACE capable tasks to use ptrace() on other tasks in the user namespace.

This is an incremental approach that will allow each addition of user namespace capabilities to be reviewed and tested separately before adding them into the mainline. Hallyn notes his current plans for enabling some additional capabilities from user namespaces:

Capabilities are something of gnarly corner of the kernel, and one that has caused problems in the past (e.g. the "sendmail capabilities" bug). Combining them with namespaces is a bit of a delicate task. Clearly, if regular users are able to create these namespaces, it is imperative that any tricky interactions caused by capabilities in namespaces do not lead to privilege escalations. From that perspective, Hallyn's approach seems sound.

Note: you can avoid this step in the future by logging into your LWN account.