Contents
The libvirt LXC driver manages "Linux Containers". At their simplest, containers can just be thought of as a collection of processes, separated from the main host processes via a set of resource namespaces and constrained via control groups resource tunables. The libvirt LXC driver has no dependency on the LXC userspace tools hosted on sourceforge.net. It directly utilizes the relevant kernel features to build the container environment. This allows for sharing of many libvirt technologies across both the QEMU/KVM and LXC drivers. In particular sVirt for mandatory access control, auditing of operations, integration with control groups and many other features.
In order to control the resource usage of processes inside containers, the libvirt LXC driver requires that certain cgroups controllers are mounted on the host OS. The minimum required controllers are 'cpuacct', 'memory' and 'devices', while recommended extra controllers are 'cpu', 'freezer' and 'blkio'. Libvirt will not mount the cgroups filesystem itself, leaving this up to the init system to take care of. Systemd will do the right thing in this respect, while for other init systems the cgconfig init service will be required. For further information, consult the general libvirt cgroups documentation.
In order to separate processes inside a container from those in the primary "host" OS environment, the libvirt LXC driver requires that certain kernel namespaces are compiled in. Libvirt currently requires the 'mount', 'ipc', 'pid', and 'uts' namespaces to be available. If separate network interfaces are desired, then the 'net' namespace is required. If the guest configuration declares a UID or GID mapping, the 'user' namespace will be enabled to apply these. A suitably configured UID/GID mapping is a pre-requisite to making containers secure, in the absence of sVirt confinement.
When the container "init" process is started, it will typically not be given any command line arguments (eg the equivalent of the bootloader args visible in /proc/cmdline). If any arguments are desired, then must be explicitly set in the container XML configuration via one or more initarg elements. For example, to run systemd --unit emergency.service would use the following XML
<os> <type arch='x86_64'>exe</type> <init>/bin/systemd</init> <initarg>--unit</initarg> <initarg>emergency.service</initarg> </os>
When the container "init" process is started, it will be given several useful environment variables. The following standard environment variables are mandated by systemd container interface to be provided by all container technologies on Linux.
The fixed string libvirt-lxc to identify libvirt as the creator
The UUID assigned to the container by libvirt
The fixed string /bin:/usr/bin
The fixed string linux
The fixed string /
In addition to the standard variables, the following libvirt specific environment variables are also provided
The name assigned to the container by libvirt
The UUID assigned to the container by libvirt
The unparsed command line arguments specified in the container configuration. Use of this is discouraged, in favour of passing arguments directly to the container init process via the initarg config element.
In the absence of any explicit configuration, the container will inherit the host OS filesystem mounts. A number of mount points will be made read only, or re-mounted with new instances to provide container specific data. The following special mounts are setup by libvirt
/dev a new "tmpfs" pre-populated with authorized device nodes
/dev/pts a new private "devpts" instance for console devices
/sys the host "sysfs" instance remounted read-only
/proc a new instance of the "proc" filesystem
/proc/sys the host "/proc/sys" bind-mounted read-only
/sys/fs/selinux the host "selinux" instance remounted read-only
/sys/fs/cgroup/NNNN the host cgroups controllers bind-mounted to only expose the sub-tree associated with the container
/proc/meminfo a FUSE backed file reflecting memory limits of the container
The container init process will be started with CAP_MKNOD capability removed and blocked from re-acquiring it. As such it will not be able to create any device nodes in /dev or anywhere else in its filesystems. Libvirt itself will take care of pre-populating the /dev filesystem with any devices that the container is authorized to use. The current devices that will be made available to all containers are
/dev/zero
/dev/null
/dev/full
/dev/random
/dev/urandom
/dev/stdin symlinked to /proc/self/fd/0
/dev/stdout symlinked to /proc/self/fd/1
/dev/stderr symlinked to /proc/self/fd/2
/dev/fd symlinked to /proc/self/fd
/dev/ptmx symlinked to /dev/pts/ptmx
/dev/console symlinked to /dev/pts/0
In addition, for every console defined in the guest configuration, a symlink will be created from /dev/ttyN symlinked to the corresponding /dev/pts/M pseudo TTY device. The first console will be /dev/tty1, with further consoles numbered incrementally from there.
Since /dev/ttyN and /dev/console are linked to the pts devices. The tty device of login program is pts device. The pam module securetty may prevent root user from logging in container. If you want root user to log in container successfully, add the pts device to the file /etc/securetty of container.
Further block or character devices will be made available to containers depending on their configuration.
The libvirt LXC driver is fairly flexible in how it can be configured, and as such does not enforce a requirement for strict security separation between a container and the host. This allows it to be used in scenarios where only resource control capabilities are important, and resource sharing is desired. Applications wishing to ensure secure isolation between a container and the host must ensure that they are writing a suitable configuration.
If the guest configuration does not list any network interfaces, the network namespace will not be activated, and thus the container will see all the host's network interfaces. This will allow apps in the container to bind to/connect from TCP/UDP addresses and ports from the host OS. It also allows applications to access UNIX domain sockets associated with the host OS, which are in the abstract namespace. If access to UNIX domains sockets in the abstract namespace is not wanted, then applications should set the <privnet/> flag in the <features>....</features> element.
If the guest configuration does not list any filesystems, then the container will be set up with a root filesystem that matches the host's root filesystem. As noted earlier, only a few locations such as /dev, /proc and /sys will be altered. This means that, in the absence of restrictions from sVirt, a process running as user/group N:M inside the container will be able to access almost exactly the same files as a process running as user/group N:M in the host.
There are multiple options for restricting this. It is possible to simply map the existing root filesystem through to the container in read-only mode. Alternatively a completely separate root filesystem can be configured for the guest. In both cases, further sub-mounts can be applied to customize the content that is made visible. Note that in the absence of sVirt controls, it is still possible for the root user in a container to unmount any sub-mounts applied. The user namespace feature can also be used to restrict access to files based on the UID/GID mappings.
Sharing the host filesystem tree, also allows applications to access UNIX domains sockets associated with the host OS, which are in the filesystem namespaces. It should be noted that a number of init systems including at least systemd and upstart have UNIX domain socket which are used to control their operation. Thus, if the directory/filesystem holding their UNIX domain socket is exposed to the container, it will be possible for a user in the container to invoke operations on the init service in the same way it could if outside the container. This also applies to other applications in the host which use UNIX domain sockets in the filesystem, such as DBus, Libvirtd, and many more. If this is not desired, then applications should either specify the UID/GID mapping in the configuration to enable user namespaces and thus block access to the UNIX domain socket based on permissions, or should ensure the relevant directories have a bind mount to hide them. This is particularly important for the /run or /var/run directories.
If the guest configuration does not list any ID mapping, then the user and group IDs used inside the container will match those used outside the container. In addition, the capabilities associated with a process in the container will infer the same privileges they would for a process in the host. This has obvious implications for security, since a root user inside the container will be able to access any file owned by root that is visible to the container, and perform more or less any privileged kernel operation. In the absence of additional protection from sVirt, this means that the root user inside a container is effectively as powerful as the root user in the host. There is no security isolation of the root user.
The ID mapping facility was introduced to allow for stricter control over the privileges of users inside the container. It allows apps to define rules such as "user ID 0 in the container maps to user ID 1000 in the host". In addition the privileges associated with capabilities are somewhat reduced so that they cannot be used to escape from the container environment. A full description of user namespaces is outside the scope of this document, however LWN has a good write-up on the topic. From the libvirt point of view, the key thing to remember is that defining an ID mapping for users and groups in the container XML configuration causes libvirt to activate the user namespace feature.
The LXC driver comes with sane default values. However, during its initialization it reads a configuration file which offers system administrator to override some of that default. The file is located under /etc/libvirt/lxc.conf
The libvirt LXC driver provides the ability to pass across pre-opened file descriptors when starting LXC guests. This allows for libvirt LXC to support systemd's socket activation capability, where an incoming client connection in the host OS will trigger the startup of a container, which runs another copy of systemd which gets passed the server socket, and then activates the actual service handler in the container.
Let us assume that you already have a LXC guest created, running a systemd instance as PID 1 inside the container, which has an SSHD service configured. The goal is to automatically activate the container when the first SSH connection is made. The first step is to create a couple of unit files for the host OS systemd instance. The /etc/systemd/system/mycontainer.service unit file specifies how systemd will start the libvirt LXC container
[Unit] Description=My little container [Service] ExecStart=/usr/bin/virsh -c lxc:///system start --pass-fds 3 mycontainer ExecStop=/usr/bin/virsh -c lxc:///system destroy mycontainer Type=oneshot RemainAfterExit=yes KillMode=none
The --pass-fds 3 argument specifies that the file descriptor number 3 that virsh inherits from systemd, is to be passed into the container. Since virsh will exit immediately after starting the container, the RemainAfterExit and KillMode settings must be altered from their defaults.
Next, the /etc/systemd/system/mycontainer.socket unit file is created to get the host systemd to listen on port 23 for TCP connections. When this unit file is activated by the first incoming connection, it will cause the mycontainer.service unit to be activated with the FD corresponding to the listening TCP socket passed in as FD 3.
[Unit] Description=The SSH socket of my little container [Socket] ListenStream=23
Port 23 was picked here so that the container doesn't conflict with the host's SSH which is on the normal port 22. That's it in terms of host side configuration.
Inside the container, the /etc/systemd/system/sshd.socket unit file must be created
[Unit] Description=SSH Socket for Per-Connection Servers [Socket] ListenStream=23 Accept=yes
The ListenStream value listed in this unit file, must match the value used in the host file. When systemd in the container receives the pre-opened FD from libvirt during container startup, it looks at the ListenStream values to figure out which FD to give to which service. The actual service to start is defined by a correspondingly named /etc/systemd/system/sshd@.service
[Unit] Description=SSH Per-Connection Server for %I [Service] ExecStart=-/usr/sbin/sshd -i StandardInput=socket
Finally, make sure this SSH service is set to start on boot of the container, by running the following command inside the container:
# mkdir -p /etc/systemd/system/sockets.target.wants/ # ln -s /etc/systemd/system/sshd.socket /etc/systemd/system/sockets.target.wants/
This example shows how to activate the container based on an incoming SSH connection. If the container was also configured to have an httpd service, it may be desirable to activate it upon either an httpd or a sshd connection attempt. In this case, the mycontainer.socket file in the host would simply list multiple socket ports. Inside the container a separate xxxxx.socket file would need to be created for each service, with a corresponding ListenStream value set.
In the absence of the "user" namespace being used, containers cannot be considered secure against exploits of the host OS. The sVirt SELinux driver provides a way to secure containers even when the "user" namespace is not used. The cost is that writing a policy to allow execution of arbitrary OS is not practical. The SELinux sVirt policy is typically tailored to work with a simpler application confinement use case, as provided by the "libvirt-sandbox" project.
The LXC driver is integrated with libvirt's auditing subsystem, which causes audit messages to be logged whenever there is an operation performed against a container which has impact on host resources. So for example, start/stop, device hotplug will all log audit messages providing details about what action occurred and any resources associated with it. There are the following 3 types of audit messages
VIRT_MACHINE_ID - details of the SELinux process and image security labels assigned to the container.
VIRT_CONTROL - details of an action / operation performed against a container. There are the following types of operation
op=start - a container has been started. Provides the machine name, uuid and PID of the libvirt_lxc controller process
op=init - the init PID of the container has been started. Provides the machine name, uuid and PID of the libvirt_lxc controller process and PID of the init process (in the host PID namespace)
op=stop - a container has been stopped. Provides the machine name, uuid
VIRT_RESOURCE - details of a host resource associated with a container action.
All containers are launched with the CAP_MKNOD capability cleared and removed from the bounding set. Libvirt will ensure that the /dev filesystem is pre-populated with all devices that a container is allowed to use. In addition, the cgroup "device" controller is configured to block read/write/mknod from all devices except those that a container is authorized to use.
<domain type='lxc'>
<name>vm1</name>
<memory>500000</memory>
<os>
<type>exe</type>
<init>/bin/sh</init>
</os>
<vcpu>1</vcpu>
<clock offset='utc'/>
<on_poweroff>destroy</on_poweroff>
<on_reboot>restart</on_reboot>
<on_crash>destroy</on_crash>
<devices>
<emulator>/usr/libexec/libvirt_lxc</emulator>
<interface type='network'>
<source network='default'/>
</interface>
<console type='pty' />
</devices>
</domain>
In the <emulator> element, be sure you specify the correct path to libvirt_lxc, if it does not live in /usr/libexec on your system.
The next example assumes there is a private root filesystem (perhaps hand-crafted using busybox, or installed from media, debootstrap, whatever) under /opt/vm-1-root:
<domain type='lxc'>
<name>vm1</name>
<memory>32768</memory>
<os>
<type>exe</type>
<init>/init</init>
</os>
<vcpu>1</vcpu>
<clock offset='utc'/>
<on_poweroff>destroy</on_poweroff>
<on_reboot>restart</on_reboot>
<on_crash>destroy</on_crash>
<devices>
<emulator>/usr/libexec/libvirt_lxc</emulator>
<filesystem type='mount'>
<source dir='/opt/vm-1-root'/>
<target dir='/'/>
</filesystem>
<interface type='network'>
<source network='default'/>
</interface>
<console type='pty' />
</devices>
</domain>
By default the libvirt LXC driver drops some capabilities among which CAP_MKNOD. However since 1.2.6 libvirt can be told to keep or drop some capabilities using a domain configuration like the following:
...
<features>
<capabilities policy='default'>
<mknod state='on'/>
<sys_chroot state='off'/>
</capabilities>
</features>
...
The capabilities children elements are named after the capabilities as defined in man 7 capabilities. An off state tells libvirt to drop the capability, while an on state will force to keep the capability even though this one is dropped by default.
The policy attribute can be one of default, allow or deny. It defines the default rules for capabilities: either keep the default behavior that is dropping a few selected capabilities, or keep all capabilities or drop all capabilities. The interest of allow and deny is that they guarantee that all capabilities will be kept (or removed) even if new ones are added later.
The following example, drops all capabilities but CAP_MKNOD:
...
<features>
<capabilities policy='deny'>
<mknod state='on'/>
</capabilities>
</features>
...
Note that allowing capabilities that are normally dropped by default can seriously affect the security of the container and the host.
Libvirt allows you to inherit the namespace from container/process just like lxc tools or docker provides to share the network namespace. The following can be used to share required namespaces. If we want to share only one then the other namespaces can be ignored. The netns option is specific to sharenet. It can be used in cases we want to use existing network namespace rather than creating new network namespace for the container. In this case privnet option will be ignored.
<domain type='lxc' xmlns:lxc='http://libvirt.org/schemas/domain/lxc/1.0'> ... <lxc:namespace> <lxc:sharenet type='netns' value='red'/> <lxc:shareuts type='name' value='container1'/> <lxc:shareipc type='pid' value='12345'/> </lxc:namespace> </domain>
The use of namespace passthrough requires libvirt >= 1.2.19
As with any libvirt virtualization driver, LXC containers can be managed via a wide variety of libvirt based tools. At the lowest level the virsh command can be used to perform many tasks, by passing the -c lxc:///system argument. As an alternative to repeating the URI with every command, the LIBVIRT_DEFAULT_URI environment variable can be set to lxc:///system. The examples that follow outline some common operations with virsh and LXC. For further details about usage of virsh consult its manual page.
The virsh define command takes an XML configuration document and loads it into libvirt, saving the configuration on disk
# virsh -c lxc:///system define myguest.xml
The virsh dumpxml command can be used to view the current XML configuration of a container. By default the XML output reflects the current state of the container. If the container is running, it is possible to explicitly request the persistent configuration, instead of the current live configuration using the --inactive flag
# virsh -c lxc:///system dumpxml myguest
The virsh start command can be used to start a container from a previously defined persistent configuration
# virsh -c lxc:///system start myguest
It is also possible to start so called "transient" containers, which do not require a persistent configuration to be saved by libvirt, using the virsh create command.
# virsh -c lxc:///system create myguest.xml
The virsh shutdown command can be used to request a graceful shutdown of the container. By default this command will first attempt to send a message to the init process via the /dev/initctl device node. If no such device node exists, then it will send SIGTERM to PID 1 inside the container.
# virsh -c lxc:///system shutdown myguest
If the container does not respond to the graceful shutdown request, it can be forcibly stopped using the virsh destroy
# virsh -c lxc:///system destroy myguest
The virsh reboot command can be used to request a graceful shutdown of the container. By default this command will first attempt to send a message to the init process via the /dev/initctl device node. If no such device node exists, then it will send SIGHUP to PID 1 inside the container.
# virsh -c lxc:///system reboot myguest
The virsh undefine command can be used to delete the persistent configuration of a container. If the guest is currently running, this will turn it into a "transient" guest.
# virsh -c lxc:///system undefine myguest
The virsh console command can be used to connect to the text console associated with a container.
# virsh -c lxc:///system console myguest
If the container has been configured with multiple console devices, then the --devname argument can be used to choose the console to connect to. In LXC, multiple consoles will be named as 'console0', 'console1', 'console2', etc.
# virsh -c lxc:///system console myguest --devname console1
The virsh lxc-enter-namespace command can be used to enter the namespaces and security context of a container and then execute an arbitrary command.
# virsh -c lxc:///system lxc-enter-namespace myguest -- /bin/ls -al /dev
The virt-top command can be used to monitor the activity and resource utilization of all containers on a host
# virt-top -c lxc:///system
The virsh domxml-from-native command can be used to convert most of the LXC container configuration into a domain XML fragment
# virsh -c lxc:///system domxml-from-native lxc-tools /var/lib/lxc/myguest/config
This conversion has some limitations due to the fact that the domxml-from-native command output has to be independent of the host. Here are a few things to take care of before converting:
Replace the fstab file referenced by lxc.mount by the corresponding lxc.mount.entry lines.
Replace all relative sizes of tmpfs mount entries to absolute sizes. Also make sure that tmpfs entries all have a size option (default is 50%).
Define lxc.cgroup.memory.limit_in_bytes to properly limit the memory available to the container. The conversion will use 64MiB as the default.