Also consult Writing Virtual Machine or Container Managers.
systemd has a number of interfaces for interacting with container managers, when systemd is used inside of an OS container. If you work on a container manager, please consider supporting the following interfaces.
If the container manager wants to control the hostname for a container running systemd it may just set it before invoking systemd, and systemd will leave it unmodified when there is no hostname configured in /etc/hostname (that file overrides whatever is pre-initialized by the container manager).
Make sure to pre-mount /proc/, /sys/, and /sys/fs/selinux/ before invoking systemd, and mount /proc/sys/, /sys/, and /sys/fs/selinux/ read-only in order to prevent the container from altering the host kernel‘s configuration settings. (As a special exception, if your container has network namespaces enabled, feel free to make /proc/sys/net/ writable). systemd and various other subsystems (such as the SELinux userspace) have been modified to behave accordingly when these file systems are read-only. (It’s OK to mount /sys/ as tmpfs btw, and only mount a subset of its sub-trees from the real sysfs to hide /sys/firmware/, /sys/kernel/ and so on. If you do that, make sure to mark /sys/ read-only, as that condition is what systemd looks for, and is what is considered to be the API in this context.)
Pre-mount /dev/ as (container private) tmpfs for the container and bind mount some suitable TTY to /dev/console. Also, make sure to create device nodes for /dev/null, /dev/zero, /dev/full, /dev/random, /dev/urandom, /dev/tty, /dev/ptmx in /dev/. It is not necessary to create /dev/fd or /dev/stdout, as systemd will do that on its own. Make sure to set up a BPF_PROG_TYPE_CGROUP_DEVICE BPF program — on cgroupv2 — or the devices cgroup controller — on cgroupv1 — so that no other devices but these may be created in the container. Note that many systemd services use PrivateDevices=, which means that systemd will set up a private /dev/ for them for which it needs to be able to create these device nodes. Dropping CAP_MKNOD for containers is hence generally not advisable, but see below.
systemd-udevd is not available in containers (and refuses to start), and hence device dependencies are unavailable. The systemd-udevd unit files will check for /sys/ being read-only, as an indication whether device management can work. Therefore make sure to mount /sys/ read-only in the container (see above). Various clients of systemd-udevd also check the read-only state of /sys/, including PID 1 itself and systemd-networkd.
If systemd detects it is run in a container it will spawn a single shell on /dev/console, and not care about VTs or multiple gettys on VTs. (But see $container_ttys below.)
Either pre-mount all cgroup hierarchies in full into the container, or leave that to systemd which will do so if they are missing. Note that it is explicitly not OK to just mount a sub-hierarchy into the container as that is incompatible with /proc/$PID/cgroup (which lists full paths). Also the root-level cgroup directories tend to be quite different from inner directories, and that distinction matters. It is OK however, to mount the “upper” parts read-only of the hierarchies, and only allow write-access to the cgroup sub-tree the container runs in. It's also a good idea to mount all controller hierarchies with exception of name=systemd fully read-only (this only applies to cgroupv1, of course), to protect the controllers from alteration from inside the containers. Or to turn this around: only the cgroup sub-tree of the container itself (on cgroupv2 in the unified hierarchy, and on cgroupv1 in the name=systemd hierarchy) may be writable to the container.
Create the control group root of your container by either running your container as a service (in case you have one container manager instance per container instance) or creating one scope unit for each container instance via systemd's transient unit API (in case you have one container manager that manages all instances. Either way, make sure to set Delegate=yes in it. This ensures that that the unit you created will be part of all cgroup controllers (or at least the ones systemd understands). The latter may also be done via systemd-machined's CreateMachine() API. Make sure to use the cgroup path systemd put your process in for all operations of the container. Do not add new cgroup directories to the top of the tree. This will not only confuse systemd and the admin, but also prevent your implementation from being “stackable”.
To allow systemd (and other programs) to identify that it is executed within a container, please set the $container environment variable for PID 1 in the container to a short lowercase string identifying your implementation. With this in place the ConditionVirtualization= setting in unit files will work properly. Example: container=lxc-libvirt
systemd has special support for allowing container managers to initialize the UUID for /etc/machine-id to some manager supplied value. This is only enabled if /etc/machine-id is empty (i.e. not yet set) at boot time of the container. The container manager should set $container_uuid as environment variable for the container's PID 1 to the container UUID. (This is similar to the effect of qemu's -uuid switch). Note that you should pass only a UUID here that is actually unique (i.e. only one running container should have a specific UUID), and gets changed when a container gets duplicated. Also note that systemd will try to persistently store the UUID in /etc/machine-id (if writable) when this option is used, hence you should always pass the same UUID here. Keeping the externally used UUID for a container and the internal one in sync is hopefully useful to minimize surprise for the administrator.
systemd can automatically spawn login gettys on additional ptys. A container manager can set the $container_ttys environment variable for the container‘s PID 1 to tell it on which ptys to spawn gettys. The variable should take a space separated list of pty names, without the leading /dev/ prefix, but with the pts/ prefix included. Note that despite the variable’s name you may only specify ptys, and not other types of ttys. Also you need to specify the pty itself, a symlink will not suffice. This is implemented in systemd-getty-generator(8). Note that this variable should not include the pty that /dev/console maps to if it maps to one (see below). Example: if the container receives container_ttys=pts/7 pts/8 pts/14 it will spawn three additional login gettys on ptys 7, 8, and 14.
To allow applications to detect the OS version and other metadata of the host running the container manager, if this is considered desirable, please parse the host's /etc/os-release and set a $container_host_<key>=<VALUE> environment variable for the ID fields described by the os-release interface, eg: $container_host_id=debian $container_host_build_id=2020-06-15 $container_host_variant_id=server $container_host_version_id=10
Consider syncing /etc/localtime from the host file system into the container. Make it a relative symlink to the containers‘s zoneinfo dir, as usual. Tools rely on being able to determine the timezone setting from the symlink value, and making it relative looks nice even if people list the container’s /etc/ from the host.
Make the container journal available in the host, by automatically symlinking the container journal directory into the host journal directory. More precisely, link /var/log/journal/<container-machine-id> of the container into the same dir of the host. Administrators can then automatically browse all container journals (correctly interleaved) by issuing journalctl -m. The container machine ID can be determined from /etc/machine-id in the container.
If the container manager wants to cleanly shutdown the container, it might be a good idea to send SIGRTMIN+3 to its init process. systemd will then do a clean shutdown. Note however, that since only systemd understands SIGRTMIN+3 like this, this might confuse other init systems.
To support Socket Activated Containers the container manager should be capable of being run as a systemd service. It will then receive the sockets starting with FD 3, the number of passed FDs in $LISTEN_FDS and its PID as $LISTEN_PID. It should take these and pass them on to the container‘s init process, also setting $LISTEN_FDS and $LISTEN_PID (basically, it can just leave the FDs and $LISTEN_FDS untouched, but it needs to adjust $LISTEN_PID to the container init process). That’s all that's necessary to make socket activation work. The protocol to hand sockets from systemd to services is hence the same as from the container manager to the container systemd. For further details see the explanations of sd_listen_fds(1) and the blog story for service developers.
Container managers should stay away from the cgroup hierarchy outside of the unit they created for their container. That's private property of systemd, and no other code should modify it.
Inside of a container, if a veth link is named host0, systemd-networkd running inside of the container will by default run DHCPv4, DHCPv6, and IPv4LL clients on it. It is thus recommended that container managers that add a veth link to a container name it host0, to get an automatically configured network, with no manual setup.
Outside of a container, if a veth link is prefixed “ve-”, systemd-networkd will by default run DHCPv4 and DHCPv6 servers on it, as well as IPv4LL. It is thus recommended that container managers that add a veth link to a container name the external side ve- + the container name.
It is recommended to configure stable MAC addresses for container veth devices, for example hashed out of the container names. That way it is more likely that DHCP and IPv4LL will acquire stable addresses.
Do not drop CAP_MKNOD from the container. PrivateDevices= is a commonly used service setting that provides a service with its own, private, minimal version of /dev/. To set this up systemd in the container needs this capability. If you take away the capability than all services that set this flag will cease to work. Use BPF_PROG_TYPE_CGROUP_DEVICE BPF programs — on cgroupv2 — or the devices controller — on cgroupv1 — to restrict what device nodes the container can create instead of taking away the capability wholesale. (Also see the section about fully unprivileged containers below.)
Do not drop CAP_SYS_ADMIN from the container. A number of the most commonly used file system namespacing related settings, such as PrivateDevices=, ProtectHome=, ProtectSystem=, MountFlags=, PrivateTmp=, ReadWriteDirectories=, ReadOnlyDirectories=, InaccessibleDirectories=, and MountFlags= need to be able to open new mount namespaces and the mount certain file systems into them. You break all services that make use of these options if you drop the capability. Also note that logind mounts XDG_RUNTIME_DIR as tmpfs for all logged in users and that won't work either if you take away the capability. (Also see section about fully unprivileged containers below.)
Do not cross-link /dev/kmsg with /dev/console. They are different things, you cannot link them to each other.
Do not pretend that the real VTs are available in the container. The VT subsystem consists of all the devices /dev/tty*, /dev/vcs*, /dev/vcsa* plus their sysfs counterparts. They speak specific ioctl()s and understand specific escape sequences, that other ptys don't understand. Hence, it is explicitly not OK to mount a pty to /dev/tty1, /dev/tty2, /dev/tty3. This is explicitly not supported.
Don‘t pretend that passing arbitrary devices to containers could really work well. For example, do not pass device nodes for block devices to the container. Device access (with the exception of network devices) is not virtualized on Linux. Enumeration and probing of meta information from /sys/ and elsewhere is not possible to do correctly in a container. Simply adding a specific device node to a container’s /dev/ is not enough to do the job, as systemd-udevd and suchlike are not available at all, and no devices will appear available or enumerable, inside the container.
Don't mount only a sub-tree of the cgroupfs into the container. This will not work as /proc/$PID/cgroup lists full paths and cannot be matched up with the actual cgroupfs tree visible, then. (You may “prune” some branches though, see above.)
Do not make /sys/ writable in the container. If you do, systemd-udevd.service is started to manage your devices — inside the container, but that will cause conflicts and errors given that the Linux device model is not virtualized for containers on Linux and thus the containers and the host would try to manage the same devices, fighting for ownership. Multiple other subsystems of systemd similarly test for /sys/ being writable to decide whether to use systemd-udevd or assume that device management is properly available on the instance. Among them systemd-networkd and systemd-logind. The conditionalization on the read-only state of /sys/ enables a nice automatism: as soon as /sys/ and the Linux device model are changed to be virtualized properly the container payload can make use of that, simply by marking /sys/ writable. (Note that as special exception, the devices in /sys/class/net/ are virtualized already, if network namespacing is used. Thus it is OK to mount the relevant sub-directories of /sys/ writable, but make sure to leave the root of /sys/ read-only.)
First things first, to make this clear: Linux containers are not a security technology right now. There are more holes in the model than in swiss cheese.
For example: if you do not use user namespacing, and share root and other users between container and host, the struct user structures will be shared between host and container, and hence RLIMIT_NPROC and so of the container users affect the host and other containers, and vice versa. This is a major security hole, and actually is a real-life problem: since Avahi sets RLIMIT_NPROC of its user to 2 (to effectively disallow fork()ing) you cannot run more than one Avahi instance on the entire system...
People have been asking to be able to run systemd without CAP_SYS_ADMIN and CAP_SYS_MKNOD in the container. This is now supported to some level in systemd, but we recommend against it (see above). If CAP_SYS_ADMIN and CAP_SYS_MKNOD are missing from the container systemd will now gracefully turn off PrivateTmp=, PrivateNetwork=, ProtectHome=, ProtectSystem= and others, because those capabilities are required to implement these options. The services using these settings (which include many of systemd's own) will hence run in a different, less secure environment when the capabilities are missing than with them around.
With user namespacing in place things get much better. With user namespaces the struct user issue described above goes away, and containers can keep CAP_SYS_ADMIN safely for the user namespace, as capabilities are virtualized and having capabilities inside a container doesn't mean one also has them outside.
If you write software that wants to detect whether it is run in a container, please check /proc/1/environ and look for the container= environment variable. Do not assume the environment variable is inherited down the process tree. It generally is not. Hence check the environment block of PID 1, not your own. Note though that that file is only accessible to root. systemd hence early on also copies the value into /run/systemd/container, which is readable for everybody. However, that's a systemd-specific interface and other init systems are unlikely to do the same.
Note that it is our intention to make systemd systems work flawlessly and out-of-the-box in containers. In fact we are interested to ensure that the same OS image can be booted on a bare system, in a VM and in a container, and behave correctly each time. If you notice that some component in systemd does not work in a container as it should, even though the container manager implements everything documented above, please contact us.