Sandboxing Phone Apps

As a follow up to Wayland [1]:

A difficult problem with Linux desktop systems (which includes phones and tablets) is restricting application access so that applications can’t mess with each other’s data or configuration but also allowing them to share data as needed. This has been mostly solved for Android but that involved giving up all “legacy” Linux apps. I think that we need to get phones capable of running a full desktop environment and having Android level security on phone apps and regular desktop apps. My interest in this is phones running Debian and derivatives such as PureOS. But everything I describe in this post should work equally well for all full featured Linux distributions for phones such as Arch, Gentoo, etc and phone based derivatives of those such as Manjaro. It may be slightly less applicable to distributions such as Alpine Linux and it’s phone derivative PostmarketOS, I would appreciate comments from contributors to PostmarketOS or Alpine Linux about this.

I’ve investigated some of the ways of solving these problems. Many of the ways of doing this involves namespaces. The LWN articles about namespaces are a good background to some of these technologies [2].

The LCA keynote lecture Containers aka crazy user space fun by Jess Frazelle has a good summary of some of the technology [3]. One part that I found particularly interesting was the bit about recognising the container access needed at compile time. This can also be part of recognising bad application design at compile time, it’s quite common for security systems to flag bad security design in programs.

Firejail

To sandbox applications you need to have some method of restricting what they do, this means some combination of namespaces and similar features. Here’s an article on sandboxing with firejail [4]. Firejail uses namespaces, seccomp-bpf, and capabilities to restrict programs. It allows bind mounts if run as root and if not run as root it can restrict file access by name and access to networking, system calls, and other features. It has a convenient learning mode that generates policy for you, so if you have a certain restricted set of tasks that an application is to perform you can run it once and then force it to do only the same operations in future. I recommend that everyone who is still reading at this point try out firejail. Here’s an example of what you can do:

# create a profile
firejail --build=xterm.profile xterm
# now this run can only do what the previous run did
firejail --profile=xterm.profile xterm

Note that firejail is SETUID root so can potentially reduce system security and it has had security issues in the past. In spite of that it can be good for allowing a trusted user to run programs with less access to the system. Also it is a good way to start learning about such things. I don’t think it’s a good solution for what I want to do. But I won’t rule out the possibility of using it at some future time for special situations.

Bubblewrap

I tried out firejail with the browser Epiphany (Debian package epiphany-browser) on my Librem5, but that didn’t work as Epiphany uses /usr/bin/bwrap (bubblewrap) for it’s internal sandboxing (here is an informative blog post about the history of bubblewrap AKA xdg-app-helper which was developed as part of flatpak [5]). The Epiphany bubblewrap sandbox is similar to the situation with Chrome/Chromium which have internal sandboxing that’s incompatible with firejail. The firejail man page notes that it’s not compatible with Snap, Flatpack, and similar technologies. One issue this raises is that we can’t have a namespace based sandboxing system applied to all desktop apps with no extra configuration as some desktop apps won’t work with it.

Bubblewrap requires setting kernel.unprivileged_userns_clone=1 to run as non-root (IE provide the normal and expected functionality) which potentially reduces system security. Here is an example of a past kernel bug that was exploitable by creating a user namespace with CAP_SYS_ADMIN [6]. But it’s the default in recent Debian kernels which means that the issues have been reviewed and determined to be a reasonable trade-off and also means that many programs will use the feature and break if it’s disabled.

Here is an example of how to use Bubblewrap on Debian, after installing the bubblewrap run the following command. Note that the –new-session option (to prevent injecting characters in the keyboard buffer with TIOCSTI) makes the session mostly unusable for a shell.

bwrap --ro-bind /usr /usr --symlink usr/lib64 /lib64 --symlink usr/lib /lib --proc /proc --dev /dev --unshare-pid --die-with-parent bash

Here is an example of using Bubblewrap to sandbox the game Warzone2100 running with Wayland/Vulkan graphics and Pulseaudio sound.

bwrap --bind $HOME/.local/share/warzone2100 $HOME/.local/share/warzone2100 --bind /run/user/$UID/pulse /run/user/$UID/pulse --bind /run/user/$UID/wayland-0 /run/user/$UID/wayland-0 --bind /run/user/$UID/wayland-0.lock /run/user/$UID/wayland-0.lock --ro-bind /usr /usr --symlink usr/bin /bin --symlink usr/lib64 /lib64 --symlink usr/lib /lib --proc /proc --dev /dev --unshare-pid --dev-bind /dev/dri /dev/dri --ro-bind $HOME/.pulse $HOME/.pulse --ro-bind $XAUTHORITY $XAUTHORITY --ro-bind /sys /sys --new-session --die-with-parent warzone2100

Here is an example of using Bubblewrap to sandbox the Debian bug reporting tool reportbug

bwrap --bind /tmp /tmp --ro-bind /etc /etc --ro-bind /usr /usr --ro-bind /var/lib/dpkg /var/lib/dpkg --symlink usr/sbin /sbin --symlink usr/bin /bin --symlink usr/lib64 /lib64 --symlink usr/lib /lib --symlink /usr/lib32 /lib32 --symlink /usr/libx32 /libx32 --proc /proc --dev /dev --die-with-parent --unshare-ipc --unshare-pid reportbug

Here is an example shell script to wrap the build process for Debian packages. This needs to run with –unshare-user and specifying the UID as 0 because fakeroot doesn’t work in the container, I haven’t worked out why but doing it through the container is a better method anyway. This script shares read-write the parent of the current directory as the Debian build process creates packages and metadata files in the parent directory. This will prevent the automatic signing scripts which is a feature not a bug, so after building packages you have to sign the .changes file with debsign. One thing I just learned is that the Debian build system Sbuild can use chroots for building packages for similar benefits [7]. Some people believe that sbuild is the correct way of doing it regardless of the chroot issue. I think it’s too heavy-weight for most of my Debian package building, but even if I had been convinced I’d still share the information about how to use bwrap as Debian is about giving users choice.

#!/bin/bash
set -e

BUILDDIR=$(realpath $(pwd)/..)
exec bwrap --bind /tmp /tmp --bind $BUILDDIR $BUILDDIR --ro-bind /etc /etc --ro-bind /usr /usr --ro-bind /var/lib/dpkg /var/lib/dpkg --symlink usr/bin /bin --symlink usr/lib64 /lib64 --symlink usr/lib /lib --proc /proc --dev /dev --die-with-parent --unshare-user --unshare-ipc --unshare-net --unshare-pid --new-session --uid 0 --gid 0 $@

Here is an informative blog post about using Bubblewrap with Seccomp (BPF) [8]. In a future post I’ll write about how to get this sort of thing going but I’ll just leave the URL here for people who want to do it on their own.

The source for the flatpak-run program is the only example I could find of using Seccomp with Bubblewrap [9]. A lot of that code is worth copying for application sandboxing, maybe the entire program.

Unshare

The unshare command from the util-linux package has a large portion of the Bubblewrap functionality. The things that it doesn’t do like creating a new session can be done by other utilities. Here is an example of creating a container with unshare and then using cgroups with it [10].

systemd --user

Recent distributions have systemd support for running a user session, the Arch Linux Wiki has a good description of how this works [11]. The units for a user are .service files stored in /usr/lib/systemd/user/ (distribution provided), ~/.local/share/systemd/user/ (user installed applications – in debian a link to ~/.config/systemd/user/), ~/.config/systemd/user/ (for manual user config), and /etc/systemd/user/ (local sysadmin provided)

Here are some example commands for manipulating this:

# show units running for the current user
systemctl --user
# show status of one unit
systemctl --user status kmail.service
# add an environment variable to the list for all user units
systemctl --user import-environment XAUTHORITY
# start a user unit
systemctl --user start kmail.service
# analyse security for all units for the current user
systemd-analyze --user security
# analyse security for one unit
systemd-analyze --user security kmail.service

Here is a test kmail.service file I wrote to see what could be done for kmail, I don’t think that kmail is the app most needing to be restricted it is in more need of being protected from other apps but it still makes a good test case. This service file took it from the default risk score of 9.8 (UNSAFE) to 6.3 (MEDIUM) even though I was getting the error “code=exited, status=218/CAPABILITIES” when I tried anything that used capabilities (apparently due to systemd having some issue talking to the kernel).

[Unit]
Description=kmail

[Service]
ExecStart=/usr/bin/kmail

# can not limit capabilities (code=exited, status=218/CAPABILITIES)
#CapabilityBoundingSet=~CAP_SYS_TIME CAP_SYS_PACCT CAP_KILL CAP_WAKE_ALARM CAP_DAC_OVERRIDE CAP_DAC_READ_SEARCH CAP_FOWNER CAP_IPC_OWNER CAP_LINUX_IMMUTABLE CAP_IPC_LOCK CAP_SYS_MODULE CAP_SYS_TTY_CONFIG CAP_SYS_BOOT CAP_SYS_CHROOT CAP_BLOCK_SUSPEND CAP_LEASE CAP_MKNOD CAP_CHOWN CAP_FSETID CAP_SETFCAP CAP_SETGID CAP_SETUID CAP_SETPCAP CAP_SYS_RAWIO CAP_SYS_PTRACE CAP_SYS_NICE CAP_SYS_RESOURCE CAP_NET_ADMIN CAP_NET_BIND_SERVICE CAP_NET_BROADCAST CAP_NET_RAW CAP_SYS_ADMIN CAP_SYSLOG

# also 218 for ProtectKernelModules PrivateDevices ProtectKernelLogs ProtectClock

# MemoryDenyWriteExecute stops it displaying message content (bad)

# needs @resources and @mount to startup
# needs @privileged to display message content
SystemCallFilter=~@cpu-emulation @debug @raw-io @reboot @swap @obsolete
SystemCallArchitectures=native
UMask=077
NoNewPrivileges=true
ProtectControlGroups=true
PrivateMounts=false
RestrictNamespaces=~user pid net uts mnt cgroup ipc

RestrictSUIDSGID=true
ProtectHostname=true
LockPersonality=true
ProtectKernelTunables=true
RestrictAddressFamilies=~AF_PACKET
RestrictRealtime=true

ProtectSystem=strict
ProtectProc=invisible
PrivateUsers=true

[Install]

When I tried to use the TemporaryFileSystem=%h” directive (to make the home directory a tmpfs – the most basic step in restricting what a regular user application can do) I got the error “(code=exited, status=226/NAMESPACE)”. So I don’t think the “systemd –user” setup competes with bubblewrap for restricting user processes. But if anyone else can start where I left off and go further then that will be interesting.

#!/bin/bash
# systemd-run Firefox with DynamicUser and home directory.
#
# Run as a non-root user.
# Or, run as root and change $USER below.
SANDBOX_MINIMAL=(
    --property=DynamicUser=1
    --property=StateDirectory=openstreetmap
    # --property=RootDirectory=/debian_sid
)
SANDBOX_X11=(
    # Sharing Xorg always defeats security, regardless of any sandboxing tech,
    # but the config is almost ready for Wayland, and there's Xephyr.
#    --property=LoadCredential=.ICEauthority:/home/$USER/.ICEauthority
    --property=LoadCredential=.Xauthority:/home/$USER/.Xauthority
    --property=Environment=DISPLAY=:0
)
SANDBOX_FIREFOX=(
    # hardware-accelerated rendering
    --property=BindPaths=/dev/dri
    # webcam
    # --property=SupplementaryGroups=video
)
systemd-run  \
    "${SANDBOX_MINIMAL[@]}"  "${SANDBOX_X11[@]}" "${SANDBOX_FIREFOX[@]}" \
    bash -c '
        export XAUTHORITY="$CREDENTIALS_DIRECTORY"/.Xauthority
        export ICEAUTHORITY="$CREDENTIALS_DIRECTORY"/.ICEauthority
        export HOME="$STATE_DIRECTORY"/home
        firefox --no-remote about:blank
    '