Skip to content

Latest commit

 

History

History
258 lines (167 loc) · 11.3 KB

Limitations.md

File metadata and controls

258 lines (167 loc) · 11.3 KB
Raw

Overview

A Kata Container utilizes a Virtual Machine (VM) to enhance security and isolation of container workloads. As a result, the system has a number of differences and limitations when compared with the default Docker* runtime, runc.

Some of these limitations have potential solutions, whereas others exist due to fundamental architectural differences generally related to the use of VMs.

The Kata Container runtime launches each container within its own hardware isolated VM, and each VM has its own kernel. Due to this higher degree of isolation, certain container capabilities cannot be supported or are implicitly enabled through the VM.

Definition of a limitation

The Open Container Initiative Runtime Specification ("OCI spec") defines the minimum specifications a runtime must support to interoperate with container managers such as Docker. If a runtime does not support some aspect of the OCI spec, it is by definition a limitation.

However, the OCI runtime reference implementation (runc) does not perfectly align with the OCI spec itself.

Further, since the default OCI runtime used by Docker is runc, Docker expects runtimes to behave as runc does. This implies that another form of limitation arises if the behavior of a runtime implementation does not align with that of runc. Having two standards complicates the challenge of supporting a Docker environment since a runtime must support the official OCI spec and the non-standard extensions provided by runc.

Scope

Each known limitation is captured in a separate GitHub issue that contains detailed information about the issue. These issues are tagged with the limitation label. This document is a curated summary of important known limitations and provides links to the relevant GitHub issues.

The following link shows the latest list of limitations:

Contributing

If you would like to work on resolving a limitation, please refer to the contributors guide. If you wish to raise an issue for a new limitation, either raise an issue directly on the runtime or see the project table of contents for advice on which repository to raise the issue against.

Pending items

This section lists items that might be possible to fix.

Runtime commands

checkpoint and restore

The runtime does not provide checkpoint and restore commands. There are discussions about using VM save and restore to give us a [criu](https://github.com/checkpoint-restore/criu)-like functionality, which might provide a solution.

Note that the OCI standard does not specify checkpoint and restore commands.

See issue kata-containers/runtime#184 for more information.

events command

The runtime does not fully implement the events command. OOM notifications and Intel RDT stats are not fully supported.

Note that the OCI standard does not specify an events command.

See issue kata-containers/runtime#308 and kata-containers/runtime#309 for more information.

update command

Currently, only block I/O weight is not supported. All other configurations are supported and are working properly.

Networking

Docker swarm and compose support

The newest version of Docker supported is specified by the externals.docker.version variable in the versions database.

Basic Docker swarm support works. However, if you want to use custom networks with Docker's swarm, an older version of Docker is required. This is specified by the externals.docker.meta.swarm-version variable in the versions database.

See issue kata-containers/runtime#175 for more information.

Docker compose normally uses custom networks, so also has the same limitations.

Resource management

Due to the way VMs differ in their CPU and memory allocation, and sharing across the host system, the implementation of an equivalent method for these commands is potentially challenging.

See issue clearcontainers/runtime#341 and the constraints challenge for more information.

For CPUs resource management see CPU constraints.

docker run and shared memory

The runtime does not implement the docker run --shm-size command to set the size of the /dev/shm tmpfs within the container. It is possible to pass this configuration value into the VM container so the appropriate mount command happens at launch time.

See issue kata-containers#21 for more information.

docker run and sysctl

The docker run --sysctl feature is not implemented. At the runtime level, this equates to the linux.sysctl OCI configuration. Docker allows configuring the sysctl settings that support namespacing. From a security and isolation point of view, it might make sense to set them in the VM, which isolates sysctl settings. Also, given that each Kata Container has its own kernel, we can support setting of sysctl settings that are not namespaced. In some cases, we might need to support configuring some of the settings on both the host side Kata Container namespace and the Kata Containers kernel.

See issue kata-containers/runtime#185 for more information.

Docker daemon features

Some features enabled or implemented via the dockerd daemon configuration are not yet implemented.

SELinux support

The dockerd configuration option "selinux-enabled": true is not presently implemented in Kata Containers. Enabling this option causes an OCI runtime error.

See issue kata-containers/runtime#784 for more information.

The consequence of this is that the Docker --security-opt is only partially supported.

Kubernetes SELinux labels will also not be applied.

Architectural limitations

This section lists items that might not be fixed due to fundamental architectural differences between "soft containers" (i.e. traditional Linux* containers) and those based on VMs.

Networking limitations

Support for joining an existing VM network

Docker supports the ability for containers to join another containers namespace with the docker run --net=containers syntax. This allows multiple containers to share a common network namespace and the network interfaces placed in the network namespace. Kata Containers does not support network namespace sharing. If a Kata Container is setup to share the network namespace of a runc container, the runtime effectively takes over all the network interfaces assigned to the namespace and binds them to the VM. Consequently, the runc container loses its network connectivity.

docker --net=host

Docker host network support (docker --net=host run) is not supported. It is not possible to directly access the host networking configuration from within the VM.

The --net=host option can still be used with runc containers and inter-mixed with running Kata Containers, thus enabling use of --net=host when necessary.

It should be noted, currently passing the --net=host option into a Kata Container may result in the Kata Container networking setup modifying, re-configuring and therefore possibly breaking the host networking setup. Do not use --net=host with Kata Containers.

docker run --link

The runtime does not support the docker run --link command. This command is now deprecated by docker and we have no intention of adding support. Equivalent functionality can be achieved with the newer docker networking commands.

See more documentation at docs.docker.com.

Storage limitations

Kubernetes volumeMounts.subPaths

Kubernetes volumeMount.subPath is not supported by Kata Containers at the moment.

See this issue for more details. Another issue focuses on the case of emptyDir.

Host resource sharing

docker run --privileged

Privileged support in Kata is essentially different from runc containers. Kata does support docker run --privileged command, but in this case full access to the guest VM is provided in addition to some host access.

The container runs with elevated capabilities within the guest and is granted access to guest devices instead of the host devices. This is also true with using securityContext privileged=true with Kubernetes.

The container may also be granted full access to a subset of host devices (kata-containers/runtime#1568).

See Privileged Kata Containers for how to configure some of this behavior.

Miscellaneous

This section lists limitations where the possible solutions are uncertain.

Docker --security-opt option partially supported

The --security-opt= option used by Docker is partially supported. We only support --security-opt=no-new-privileges and --security-opt seccomp=/path/to/seccomp/profile.json option as of today.

Note: The --security-opt apparmor=your_profile is not yet supported. See kata-containers/runtime#707.

Appendices

The constraints challenge

Applying resource constraints such as cgroup, CPU, memory, and storage to a workload is not always straightforward with a VM based system. A Kata Container runs in an isolated environment inside a virtual machine. This, coupled with the architecture of Kata Containers, offers many more possibilities than are available to traditional Linux containers due to the various layers and contexts.

In some cases it might be necessary to apply the constraints to multiple levels. In other cases, the hardware isolated VM provides equivalent functionality to the the requested constraint.

The following examples outline some of the various areas constraints can be applied:

  • Inside the VM

    Constrain the guest kernel. This can be achieved by passing particular values through the kernel command line used to boot the guest kernel. Alternatively, sysctl values can be applied at early boot.

  • Inside the container

    Constrain the container created inside the VM.

  • Outside the VM:

    • Constrain the hypervisor process by applying host-level constraints.

    • Constrain all processes running inside the hypervisor.

      This can be achieved by specifying particular hypervisor configuration options.

Note that in some circumstances it might be necessary to apply particular constraints to more than one of the previous areas to achieve the desired level of isolation and resource control.