Many container runtime tools like
etc. focus on providing infrastructure for system administrators and
orchestration tools (e.g. Kubernetes) to run containers.
These tools are not suitable to give to unprivileged users, because it is trivial to turn such access into to a fully privileged root shell on the host.
There is an effort in the Linux kernel called user namespaces which attempts to allow unprivileged users to use container features. While significant progress has been made, there are still concerns about it, and it is not available to unprivileged users in several production distributions such as CentOS/Red Hat Enterprise Linux 7, Debian Jessie, etc.
Bubblewrap could be viewed as setuid implementation of a subset of user namespaces. Emphasis on subset - specifically relevant to the above CVE, bubblewrap does not allow control over iptables.
The maintainers of this tool believe that it does not, even when used in combination with typical software installed on that distribution, allow privilege escalation. It may increase the ability of a logged in user to perform denial of service attacks, however.
In particular, bubblewrap uses
PR_SET_NO_NEW_PRIVS to turn off
setuid binaries, which is the traditional way to get out of things
This program can be shared by all container tools which perform non-root operation, such as:
We would also like to see this be available in Kubernetes/OpenShift clusters. Having the ability for unprivileged users to use container features would make it significantly easier to do interactive debugging scenarios and the like.
bubblewrap works by creating a new, completely empty, mount namespace where the root is on a tmpfs that is invisible from the host, and will be automatically cleaned up when the last process exits. You can then use commandline options to construct the root filesystem and process environment and command to run in the namespace.
There's a larger demo script in the
source code, but here's a trimmed down version which runs
a new shell reusing the host's
bwrap --ro-bind /usr /usr --symlink usr/lib64 /lib64 --proc /proc --dev /dev --unshare-pid bash
This is an incomplete example, but useful for purposes of
illustration. More often, rather than creating a container using the
host's filesystem tree, you want to target a chroot. There, rather
than creating the symlink
lib64 -> usr/lib64 in the tmpfs, you might
have already created it in the target rootfs.
The goal of bubblewrap is to run an application in a sandbox, where it has restricted access to parts of the operating system or user data such as the home directory.
bubblewrap always creates a new mount namespace, and the user can specify
exactly what parts of the filesystem should be visible in the sandbox.
Any such directories you specify mounted
nodev by default, and can be made readonly.
Additionally you can use these kernel features:
User namespaces (CLONE_NEWUSER): This hides all but the current uid and gid from the sandbox. You can also change what the value of uid/gid should be in the sandbox.
IPC namespaces (CLONE_NEWIPC): The sandbox will get its own copy of all the different forms of IPCs, like SysV shared memory and semaphores.
PID namespaces (CLONE_NEWPID): The sandbox will not see any processes outside the sandbox. Additionally, bubblewrap will run a trivial pid1 inside your container to handle the requirements of reaping children in the sandbox. This avoids what is known now as the Docker pid 1 problem.
Network namespaces (CLONE_NEWNET): The sandbox will not see the network. Instead it will have its own network namespace with only a loopback device.
UTS namespace (CLONE_NEWUTS): The sandbox will have its own hostname.
Seccomp filters: You can pass in seccomp filters that limit which syscalls can be done in the sandbox. For more information, see Seccomp.
Related project comparison: Firejail
Firejail is similar to Flatpak before bubblewrap was split out in that it combines a setuid tool with a lot of desktop-specific sandboxing features. For example, Firejail knows about Pulseaudio, whereas bubblewrap does not.
The bubblewrap authors believe it's much easier to audit a small setuid program, and keep features such as Pulseaudio filtering as an unprivileged process, as now occurs in Flatpak.
Also, @cgwalters thinks trying to
whitelist file paths
is a bad idea given the myriad ways users have to manipulate paths,
and the myriad ways in which system administrators may configure a
system. The bubblewrap approach is to only retain a few specific
Linux capabilities such as
CAP_SYS_ADMIN, but to always access the
filesystem as the invoking uid. This entirely closes
TOCTTOU attacks and
Related project comparison: Sandstorm.io
Sandstorm.io requries unprivileged user namespaces to set up its sandbox, though it could easily be adapted to operate in a setuid mode as well. @cgwalters believes their code is fairly good, but it could still make sense to unify on bubblewrap. However, @kentonv (of Sandstorm) feels that while this makes sense in principle, the switching cost outweighs the practical benefits for now. This decision could be re-evaluated in the future, but it is not being actively pursued today.
Related project comparison: runc/binctr
runC is currently working on
supporting rootless containers,
setuid or any other privileges during installation of
runC (using unprivileged user namespaces rather than
creation, and management of containers. However, the standard mode of
using runC is similar to systemd nspawn
in that it is tooling intended to be invoked by root.
The bubblewrap authors believe that runc and systemd-nspawn are not designed to be made setuid, and are distant from supporting such a mode. However with rootless containers, runC will be able to fulfill certain usecases that bubblewrap supports (with the added benefit of being a standardised and complete OCI runtime).
binctr is just a wrapper for runC, so inherits all of its design tradeoffs.
Whats with the name ?!
The name bubblewrap was chosen to convey that this tool runs as the parent of the application (so wraps it in some sense) and creates a protective layer (the sandbox) around it.