This document discusses threat models associated with the Kata Containers project. Kata was designed to provide additional isolation of container workloads, protecting the host infrastructure from potentially malicious container users or workloads. Since Kata Containers adds a level of isolation on top of traditional containers, the focus is on the additional layer provided, not on traditional container security.
This document provides a brief background on containers and layered security, describes the interface to Kata from CRI runtimes, a review of utilized virtual machine interfaces, and then a review of threats.
Kata seeks to prevent an untrusted container workload or user of that container workload to gain control of, obtain information from, or tamper with the host infrastructure.
In our scenario, an asset is anything on the host system, or elsewhere in the cluster infrastructure. The attacker is assumed to be either a malicious user or the workload itself running within the container. The goal of Kata is to prevent attacks which would allow any access to the defined assets.
Traditional containers leverage several key Linux kernel features to provide isolation and
a view that the container workload is the only entity running on the host. Key features include
Namespaces, cgroups, capablities, SELinux and seccomp. The canonical runtime for creating such
a container is runc. In the remainder of the document, the term traditional-container will be used
to describe a container workload created by runc.
Kata Containers provides a second layer of isolation on top of those provided by traditional-containers. The hardware virtualization interface is the basis of this additional layer. Kata launches a lightweight virtual machine, and uses the guest’s Linux kernel to create a container workload, or workloads in the case of multi-container pods. In Kubernetes and in the Kata implementation, the sandbox is carried out at the pod level. In Kata, this sandbox is created using a virtual machine.
A typical Kata Containers deployment uses Kubernetes with a CRI implementation.
On every node, Kubelet will interact with a CRI implementor, which will in turn interface with
an OCI based runtime, such as Kata Containers. Typical CRI implementors are cri-o and containerd.
The CRI API, as defined at the Kubernetes CRI-API repo, results in a few constructs being supported by the CRI implementation, and ultimately in the OCI runtime creating the workloads.
In order to run a container inside of the Kata sandbox, several virtual machine devices and interfaces are required. Kata translates sandbox and container definitions to underlying virtualization technologies provided by a set of virtual machine monitors (VMMs) and hypervisors. These devices and their underlying implementations are discussed in detail in the following section.
In case of Kata, today the devices which we need in the guest are:
- Storage: In the current design of Kata Containers, we are reliant on the CRI implementor to
assist in image handling and volume management on the host. As a result, we need to support a way of passing to the sandbox the container rootfs, volumes requested
by the workload, and any other volumes created to facilitate sharing of secrets and
configmapswith the containers. Depending on how these are managed, a block based device or file-system sharing is required. Kata Containers does this by way ofvirtio-blkand/orvirtio-fs. - Networking: A method for enabling network connectivity with the workload is required. Typically this will be done providing a
TAPdevice to the VMM, and this will be exposed to the guest as avirtio-netdevice. It is feasible to pass in a NIC device directly, in which caseVFIOis leveraged and the device itself will be exposed to the guest. - Control: In order to interact with the guest agent and retrieve
STDIOfrom containers, a medium of communication is required. This is available viavirtio-vsock. - Devices:
VFIOis utilized when devices are passed directly to the virtual machine and exposed to the container. - Dynamic Resource Management:
ACPIis utilized to allow for dynamic VM resource management (for example: CPU, memory, device hotplug). This is required when containers are resized, or more generally when containers are added to a pod.
How these devices are utilized varies depending on the VMM utilized. We clarify the default settings provided when integrating Kata with the QEMU, Firecracker and Cloud Hypervisor VMMs in the following sections.
Each virtio device is implemented by a backend, which may execute within userspace on the host (vhost-user), the VMM itself, or within the host kernel (vhost). While it may provide enhanced performance,
vhost devices are often seen as higher risk since an exploit would be already running within the kernel space. While VMM and vhost-user are both in userspace on the host, vhost-user generally allows for the back-end process to require less system calls and capabilities compared to a full VMM.
The backend for virtio-blk and virtio-scsi are based in the VMM itself (ring3 in the context of x86) by default for Cloud Hypervisor, Firecracker and QEMU.
While vhost based back-ends are available for QEMU, it is not recommended. vhost-user back-ends are being added for Cloud Hypervisor, they are not utilized in Kata today.
virtio-fs is supported in Cloud Hypervisor and QEMU. virtio-fs's interaction with the host filesystem is done through a vhost-user daemon, virtiofsd.
The virtio-fs client, running in the guest, will generate requests to access files. virtiofsd will receive requests, open the file, and request the VMM
to mmap it into the guest. When DAX is utilized, the guest will access the host's page cache, avoiding the need for copy and duplication. DAX is still an experimental feature,
and is not enabled by default.
From the virtiofsd documentation:
This program must be run as the root user. Upon startup the program will switch into a new file system namespace with the shared directory tree as its root. This prevents “file system escapes” due to symlinks and other file system objects that might lead to files outside the shared directory. The program also sandboxes itself using seccomp(2) to prevent ptrace(2) and other vectors that could allow an attacker to compromise the system after gaining control of the virtiofsd process.
DAX-less support for virtio-fs is available as of the 5.4 Linux kernel. QEMU VMM supports virtio-fs as of v4.2. Cloud Hypervisor
supports virtio-fs.
virtio-net has many options, depending on the VMM and Kata configurations.
While QEMU has options for vhost, virtio-net and vhost-user, the virtio-net backend
for Kata defaults to vhost-net for performance reasons. The default configuration is being
reevaluated.
For Firecracker, the virtio-net backend is within Firecracker's VMM.
For Cloud Hypervisor, the current backend default is within the VMM. vhost-user-net support
is being added (written in rust, Cloud Hypervisor specific).
In QEMU, vsock is backed by vhost_vsock, which runs within the kernel itself.
In Firecracker and Cloud Hypervisor, vsock is backed by a unix-domain-socket in the hosts userspace.
Utilizing VFIO, devices can be passed through to the virtual machine. We will assess this separately. Exposure to host is limited to gaps in device pass-through handling. This is supported in QEMU and Cloud Hypervisor, but not Firecracker.
ACPI is necessary for hotplug of CPU, memory and devices. ACPI is available in QEMU and Cloud Hypervisor. Device, CPU and memory hotplug are not available in Firecracker.
