docs/CONTAINER_INTERFACE.md

   1 ---
   2 title: Container Interface
   3 category: Interfaces
   4 layout: default
   5 ---
   6
   7 # The Container Interface
   8
   9 Also consult [Writing Virtual Machine or Container
  10 Managers](http://www.freedesktop.org/wiki/Software/systemd/writing-vm-managers).
  11
  12 systemd has a number of interfaces for interacting with container managers,
  13 when systemd is used inside of an OS container. If you work on a container
  14 manager, please consider supporting the following interfaces.
  15
  16 ## Execution Environment
  17
  18 1. If the container manager wants to control the hostname for a container
  19    running systemd it may just set it before invoking systemd, and systemd will
  20    leave it unmodified when there is no hostname configured in `/etc/hostname`
  21    (that file overrides whatever is pre-initialized by the container manager).
  22
  23 2. Make sure to pre-mount `/proc/`, `/sys/`, and `/sys/fs/selinux/` before
  24    invoking systemd, and mount `/proc/sys/`, `/sys/`, and `/sys/fs/selinux/`
  25    read-only in order to prevent the container from altering the host kernel's
  26    configuration settings. (As a special exception, if your container has
  27    network namespaces enabled, feel free to make `/proc/sys/net/` writable).
  28    systemd and various other subsystems (such as the SELinux userspace) have
  29    been modified to behave accordingly when these file systems are read-only.
  30    (It's OK to mount `/sys/` as `tmpfs` btw, and only mount a subset of its
  31    sub-trees from the real `sysfs` to hide `/sys/firmware/`, `/sys/kernel/` and
  32    so on. If you do that, make sure to mark `/sys/` read-only, as that
  33    condition is what systemd looks for, and is what is considered to be the API
  34    in this context.)
  35
  36 3. Pre-mount `/dev/` as (container private) `tmpfs` for the container and bind
  37    mount some suitable TTY to `/dev/console`. Also, make sure to create device
  38    nodes for `/dev/null`, `/dev/zero`, `/dev/full`, `/dev/random`,
  39    `/dev/urandom`, `/dev/tty`, `/dev/ptmx` in `/dev/`. It is not necessary to
  40    create `/dev/fd` or `/dev/stdout`, as systemd will do that on its own. Make
  41    sure to set up a `BPF_PROG_TYPE_CGROUP_DEVICE` BPF program — on cgroupv2 —
  42    or the `devices` cgroup controller — on cgroupv1 — so that no other devices
  43    but these may be created in the container. Note that many systemd services
  44    use `PrivateDevices=`, which means that systemd will set up a private
  45    `/dev/` for them for which it needs to be able to create these device nodes.
  46    Dropping `CAP_MKNOD` for containers is hence generally not advisable, but
  47    see below.
  48
  49 4. `systemd-udevd` is not available in containers (and refuses to start), and
  50    hence device dependencies are unavailable. The `systemd-udevd` unit files
  51    will check for `/sys/` being read-only, as an indication whether device
  52    management can work. Therefore make sure to mount `/sys/` read-only in the
  53    container (see above). Various clients of `systemd-udevd` also check the
  54    read-only state of `/sys/`, including PID 1 itself and `systemd-networkd`.
  55
  56 5. If systemd detects it is run in a container it will spawn a single shell on
  57    `/dev/console`, and not care about VTs or multiple gettys on VTs. (But see
  58    `$container_ttys` below.)
  59
  60 6. Either pre-mount all cgroup hierarchies in full into the container, or leave
  61    that to systemd which will do so if they are missing. Note that it is
  62    explicitly *not* OK to just mount a sub-hierarchy into the container as that
  63    is incompatible with `/proc/$PID/cgroup` (which lists full paths). Also the
  64    root-level cgroup directories tend to be quite different from inner
  65    directories, and that distinction matters. It is OK however, to mount the
  66    "upper" parts read-only of the hierarchies, and only allow write-access to
  67    the cgroup sub-tree the container runs in. It's also a good idea to mount
  68    all controller hierarchies with exception of `name=systemd` fully read-only
  69    (this only applies to cgroupv1, of course), to protect the controllers from
  70    alteration from inside the containers. Or to turn this around: only the
  71    cgroup sub-tree of the container itself (on cgroupv2 in the unified
  72    hierarchy, and on cgroupv1 in the `name=systemd` hierarchy) may be writable
  73    to the container.
  74
  75 7. Create the control group root of your container by either running your
  76    container as a service (in case you have one container manager instance per
  77    container instance) or creating one scope unit for each container instance
  78    via systemd's transient unit API (in case you have one container manager
  79    that manages all instances. Either way, make sure to set `Delegate=yes` in
  80    it. This ensures that that the unit you created will be part of all cgroup
  81    controllers (or at least the ones systemd understands). The latter may also
  82    be done via `systemd-machined`'s `CreateMachine()` API. Make sure to use the
  83    cgroup path systemd put your process in for all operations of the container.
  84    Do not add new cgroup directories to the top of the tree. This will not only
  85    confuse systemd and the admin, but also prevent your implementation from
  86    being "stackable".
  87
  88 ## Environment Variables
  89
  90 1. To allow systemd (and other programs) to identify that it is executed within
  91    a container, please set the `$container` environment variable for PID 1 in
  92    the container to a short lowercase string identifying your
  93    implementation. With this in place the `ConditionVirtualization=` setting in
  94    unit files will work properly. Example: `container=lxc-libvirt`
  95
  96 2. systemd has special support for allowing container managers to initialize
  97    the UUID for `/etc/machine-id` to some manager supplied value. This is only
  98    enabled if `/etc/machine-id` is empty (i.e. not yet set) at boot time of the
  99    container. The container manager should set `$container_uuid` as environment
 100    variable for the container's PID 1 to the container UUID. (This is similar
 101    to the effect of `qemu`'s `-uuid` switch). Note that you should pass only a
 102    UUID here that is actually unique (i.e. only one running container should
 103    have a specific UUID), and gets changed when a container gets duplicated.
 104    Also note that systemd will try to persistently store the UUID in
 105    `/etc/machine-id` (if writable) when this option is used, hence you should
 106    always pass the same UUID here. Keeping the externally used UUID for a
 107    container and the internal one in sync is hopefully useful to minimize
 108    surprise for the administrator.
 109
 110 3. systemd can automatically spawn login gettys on additional ptys. A container
 111    manager can set the `$container_ttys` environment variable for the
 112    container's PID 1 to tell it on which ptys to spawn gettys. The variable
 113    should take a space separated list of pty names, without the leading `/dev/`
 114    prefix, but with the `pts/` prefix included. Note that despite the
 115    variable's name you may only specify ptys, and not other types of ttys. Also
 116    you need to specify the pty itself, a symlink will not suffice. This is
 117    implemented in
 118    [systemd-getty-generator(8)](https://www.freedesktop.org/software/systemd/man/systemd-getty-generator.html).
 119    Note that this variable should not include the pty that `/dev/console` maps
 120    to if it maps to one (see below). Example: if the container receives
 121    `container_ttys=pts/7 pts/8 pts/14` it will spawn three additional login
 122    gettys on ptys 7, 8, and 14.
 123
 124 ## Advanced Integration
 125
 126 1. Consider syncing `/etc/localtime` from the host file system into the
 127    container. Make it a relative symlink to the containers's zoneinfo dir, as
 128    usual. Tools rely on being able to determine the timezone setting from the
 129    symlink value, and making it relative looks nice even if people list the
 130    container's `/etc/` from the host.
 131
 132 2. Make the container journal available in the host, by automatically
 133    symlinking the container journal directory into the host journal directory.
 134    More precisely, link `/var/log/journal/<container-machine-id>` of the
 135    container into the same dir of the host. Administrators can then
 136    automatically browse all container journals (correctly interleaved) by
 137    issuing `journalctl -m`. The container machine ID can be determined from
 138    `/etc/machine-id` in the container.
 139
 140 3. If the container manager wants to cleanly shutdown the container, it might
 141    be a good idea to send `SIGRTMIN+3` to its init process. systemd will then
 142    do a clean shutdown. Note however, that since only systemd understands
 143    `SIGRTMIN+3` like this, this might confuse other init systems.
 144
 145 4. To support [Socket Activated
 146    Containers](http://0pointer.de/blog/projects/socket-activated-containers.html)
 147    the container manager should be capable of being run as a systemd
 148    service. It will then receive the sockets starting with FD 3, the number of
 149    passed FDs in `$LISTEN_FDS` and its PID as `$LISTEN_PID`. It should take
 150    these and pass them on to the container's init process, also setting
 151    $LISTEN_FDS and `$LISTEN_PID` (basically, it can just leave the FDs and
 152    `$LISTEN_FDS` untouched, but it needs to adjust `$LISTEN_PID` to the
 153    container init process). That's all that's necessary to make socket
 154    activation work. The protocol to hand sockets from systemd to services is
 155    hence the same as from the container manager to the container systemd. For
 156    further details see the explanations of
 157    [sd_listen_fds(1)](http://0pointer.de/public/systemd-man/sd_listen_fds.html)
 158    and the [blog story for service
 159    developers](http://0pointer.de/blog/projects/socket-activation.html).
 160
 161 5. Container managers should stay away from the cgroup hierarchy outside of the
 162    unit they created for their container. That's private property of systemd,
 163    and no other code should modify it.
 164
 165 ## Networking
 166
 167 1. Inside of a container, if a `veth` link is named `host0`, `systemd-networkd`
 168    running inside of the container will by default run DHCPv4, DHCPv6, and
 169    IPv4LL clients on it. It is thus recommended that container managers that
 170    add a `veth` link to a container name it `host0`, to get an automatically
 171    configured network, with no manual setup.
 172
 173 2. Outside of a container, if a `veth` link is prefixed "ve-", `systemd-networkd`
 174    will by default run DHCPv4 and DHCPv6 servers on it, as well as IPv4LL. It
 175    is thus recommended that container managers that add a `veth` link to a
 176    container name the external side `ve-` + the container name.
 177
 178 3. It is recommended to configure stable MAC addresses for container `veth`
 179    devices, for example hashed out of the container names. That way it is more
 180    likely that DHCP and IPv4LL will acquire stable addresses.
 181
 182 ## What You Shouldn't Do
 183
 184 1. Do not drop `CAP_MKNOD` from the container. `PrivateDevices=` is a commonly
 185    used service setting that provides a service with its own, private, minimal
 186    version of `/dev/`. To set this up systemd in the container needs this
 187    capability. If you take away the capability than all services that set this
 188    flag will cease to work. Use `BPF_PROG_TYPE_CGROUP_DEVICE` BPF programs — on
 189    cgroupv2 — or the `devices` controller — on cgroupv1 — to restrict what
 190    device nodes the container can create instead of taking away the capability
 191    wholesale. (Also see the section about fully unprivileged containers below.)
 192
 193 2. Do not drop `CAP_SYS_ADMIN` from the container. A number of the most
 194    commonly used file system namespacing related settings, such as
 195    `PrivateDevices=`, `ProtectHome=`, `ProtectSystem=`, `MountFlags=`,
 196    `PrivateTmp=`, `ReadWriteDirectories=`, `ReadOnlyDirectories=`,
 197    `InaccessibleDirectories=`, and `MountFlags=` need to be able to open new
 198    mount namespaces and the mount certain file systems into them. You break all
 199    services that make use of these options if you drop the capability. Also
 200    note that logind mounts `XDG_RUNTIME_DIR` as `tmpfs` for all logged in users
 201    and that won't work either if you take away the capability. (Also see
 202    section about fully unprivileged containers below.)
 203
 204 3. Do not cross-link `/dev/kmsg` with `/dev/console`. They are different things,
 205    you cannot link them to each other.
 206
 207 4. Do not pretend that the real VTs are available in the container. The VT
 208    subsystem consists of all the devices `/dev/tty*`, `/dev/vcs*`, `/dev/vcsa*`
 209    plus their `sysfs` counterparts. They speak specific `ioctl()`s and
 210    understand specific escape sequences, that other ptys don't understand.
 211    Hence, it is explicitly not OK to mount a pty to `/dev/tty1`, `/dev/tty2`,
 212    `/dev/tty3`. This is explicitly not supported.
 213
 214 5. Don't pretend that passing arbitrary devices to containers could really work
 215    well. For example, do not pass device nodes for block devices to the
 216    container. Device access (with the exception of network devices) is not
 217    virtualized on Linux. Enumeration and probing of meta information from
 218    `/sys/` and elsewhere is not possible to do correctly in a container. Simply
 219    adding a specific device node to a container's `/dev/` is *not* *enough* to
 220    do the job, as `systemd-udevd` and suchlike are not available at all, and no
 221    devices will appear available or enumerable, inside the container.
 222
 223 6. Don't mount only a sub-tree of the `cgroupfs` into the container. This will not
 224    work as `/proc/$PID/cgroup` lists full paths and cannot be matched up with
 225    the actual `cgroupfs` tree visible, then. (You may "prune" some branches
 226    though, see above.)
 227
 228 7. Do not make `/sys/` writable in the container. If you do,
 229    `systemd-udevd.service` is started to manage your devices — inside the
 230    container, but that will cause conflicts and errors given that the Linux
 231    device model is not virtualized for containers on Linux and thus the
 232    containers and the host would try to manage the same devices, fighting for
 233    ownership. Multiple other subsystems of systemd similarly test for `/sys/`
 234    being writable to decide whether to use `systemd-udevd` or assume that
 235    device management is properly available on the instance. Among them
 236    `systemd-networkd` and `systemd-logind`. The conditionalization on the
 237    read-only state of `/sys/` enables a nice automatism: as soon as `/sys/` and
 238    the Linux device model are changed to be virtualized properly the container
 239    payload can make use of that, simply by marking `/sys/` writable. (Note that
 240    as special exception, the devices in `/sys/class/net/` are virtualized
 241    already, if network namespacing is used. Thus it is OK to mount the relevant
 242    sub-directories of `/sys/` writable, but make sure to leave the root of
 243    `/sys/` read-only.)
 244
 245 ## Fully Unprivileged Container Payload
 246
 247 First things first, to make this clear: Linux containers are not a security
 248 technology right now. There are more holes in the model than in swiss cheese.
 249
 250 For example: if you do not use user namespacing, and share root and other users
 251 between container and host, the `struct user` structures will be shared between
 252 host and container, and hence `RLIMIT_NPROC` and so of the container users
 253 affect the host and other containers, and vice versa. This is a major security
 254 hole, and actually is a real-life problem: since Avahi sets `RLIMIT_NPROC` of
 255 its user to 2 (to effectively disallow `fork()`ing) you cannot run more than
 256 one Avahi instance on the entire system...
 257
 258 People have been asking to be able to run systemd without `CAP_SYS_ADMIN` and
 259 `CAP_SYS_MKNOD` in the container. This is now supported to some level in
 260 systemd, but we recommend against it (see above). If `CAP_SYS_ADMIN` and
 261 `CAP_SYS_MKNOD` are missing from the container systemd will now gracefully turn
 262 off `PrivateTmp=`, `PrivateNetwork=`, `ProtectHome=`, `ProtectSystem=` and
 263 others, because those capabilities are required to implement these options. The
 264 services using these settings (which include many of systemd's own) will hence
 265 run in a different, less secure environment when the capabilities are missing
 266 than with them around.
 267
 268 With user namespacing in place things get much better. With user namespaces the
 269 `struct user` issue described above goes away, and containers can keep
 270 `CAP_SYS_ADMIN` safely for the user namespace, as capabilities are virtualized
 271 and having capabilities inside a container doesn't mean one also has them
 272 outside.
 273
 274 ## Final Words
 275
 276 If you write software that wants to detect whether it is run in a container,
 277 please check `/proc/1/environ` and look for the `container=` environment
 278 variable. Do not assume the environment variable is inherited down the process
 279 tree. It generally is not. Hence check the environment block of PID 1, not your
 280 own. Note though that that file is only accessible to root. systemd hence early
 281 on also copies the value into `/run/systemd/container`, which is readable for
 282 everybody. However, that's a systemd-specific interface and other init systems
 283 are unlikely to do the same.
 284
 285 Note that it is our intention to make systemd systems work flawlessly and
 286 out-of-the-box in containers. In fact we are interested to ensure that the same
 287 OS image can be booted on a bare system, in a VM and in a container, and behave
 288 correctly each time. If you notice that some component in systemd does not work
 289 in a container as it should, even though the container manager implements
 290 everything documented above, please contact us.