Xen ARM with Virtualization Extensions

For the ARM port supporting paravirtualized guests on processors without the virtualization extensions see Xen ARM (PV).

The ARM v7-A and ARM v8-A architectures include optional virtualization extensions that allow a hypervisor to manage fully hardware virtualized guests. These extensions are currently available in some ARM v7 processors such as the Cortex A15 and Cortex A7.

Introduction

If you are interested in the Xen on ARM architecture and how it compares to Xen on x86, read the Xen on ARM whitepaper.

Status

Both the 32-bit (arm32) and the 64-bit (arm64) ports of Xen boot dom0 and unprivileged guests can be created and destroyed using xl. See below for information on the hardware and models. You may also want to read our whitepaper that explains the basic architecture and terminology of Xen on ARM.

Guest ABI

The guest (including dom0) visible ABI exposed by the hypervisor has been declared stable as of the 4.4 release and will now be maintained in a backwards compatible manner.

Contributing

Please email xen-devel with comments, questions and patches. Please see the list info page for subscription information and the archives. For patches please see Submitting Xen Patches.

The xen-arm mailing list is focused on Xen ARM (PV) port.

Hardware

The arm32 port of Xen currently runs on:

• The Cortex A15 Real-time System Model's provided by ARM. See Xen ARM with Virtualization Extensions/FastModels for more information on obtaining and building the necessary model.
• A Cortex-A15 Processor running on the Versatile Express. See Xen ARM with Virtualization Extensions/Vexpress.
• The Arndale board. See Getting Xen on the Arndale.
• The Calxeda ECX-2000 (aka Midway) server. See Getting Xen running on Midway.
• Allwinner A20/A31 boards. See Getting Xen on the Allwinner A20/A31

Early work was also done to support the Google Chromebook, however this is no longer continued since it is not a good development platform.

The arm64 port of Xen currently runs on:

• The AEMv8 Real-time System Models by ARM. See Xen ARM with Virtualization Extensions/FastModels.
• The ARMv8 Foundation Model, is a freely available emulator. Use is very similar to the Fast Models, see Xen ARM with Virtualization Extensions/FastModels.
• APM X-Gene "Mustang". See Xen ARM with Virtualization Extensions/APMXGeneMustang.

Requirements

ARM Hardware or Software Model

See above for details of the hardware and models which are supported.

Firmware

Xen requires certain functionality from the system firmware. See below for details.

Device Tree

A device tree in the flat device tree format (.dtb). The host platform must be described in a DTB binary passed to Xen at boot time. This will be used by Xen and Dom0. Normally the regular device tree used when booting natively on the platform should be used.

Xen

All current work is now merged into the current development branch git://xenbits.xen.org/xen.git. It is recommended to use the latest Xen master branch.

Linux kernel for dom0

The patches necessary to boot Linux as dom0 under Xen were merged upstream in v3.7. In order to actually start guests a few additional patches were required however these patches have now been included in the v3.8 Linux release. The latest Linus' tree has everything needed to run on Xen on ARM as dom0 and domU. It is recommended to use the latest Linux release where possible.

dom0 userspace

The developers are using the armhf port of Debian Wheezy.

domU kernel

The patches necessary to boot Linux as a guest under Xen were merged upstream in v3.7.

Hypervisor ABI Compatibility

The ABI for Xen on ARM was declared stable from Xen 4.4 onwards.

This ABI is implemented by Linux mainline v3.9-rc1 onwards.

Although not a hypervisor ABI change Linux versions prior to v3.13-rc5 (specifically 380108d891ac "xen/block: Correctly define structures in public headers on ARM32 and ARM64") incorrectly defined the PV block protocol on ARM. This means that versions of Linux prior to this fix are only able to interoperate (as either front or backend) with other versions of Linux prior to this commit. After this fix Linux correctly uses the same protocol as other ports (e.g. BSD).

Firmware Requirements

The primary requirement is that the hypervisor must be launched in Non-Secure Hypervisor mode only. If the stock firmware on a platform does not obey this (most commonly by launching in Secure Supervisor mode) then a firmware update may be required. This support is present in u-boot 2014.01.

Booting secondary processors on an SMP system requires firmware support for the Power State Coordination Interfaces (PSCI). Initial u-boot support for this interface is available in https://git.kernel.org/cgit/linux/kernel/git/maz/u-boot.git/log/?h=wip/psci .

When running Xen under a FastModel (which typically have no firmware at all) it is sometimes necessary to use a boot-wrapper. See the Fast Model page for more information.

Xen previously included a workaround for firmware which lacked this functionality, however it was unmaintainable and interfered with proper support for other platforms and therefore as of September 2013 it has been removed.

Booting Natively

Before starting to load Xen it is highly recommended to get the kernel you intend to use as dom0 booting natively (i.e. without Xen underneath). This will let you iron out any driver issues and figure out the necessary kernel command line etc before adding Xen into the mix.

Booting Xen

Boot Protocol

Xen's boot requirements are spelled out in docs/misc/arm/booting.txt in the Xen tree, which references the Linux arm and arm64 booting documentation.

Device Trees

Xen needs the device trees to be in the flat device tree format (the device tree blob or DTB).

It is no longer necessary to build a specific DTB for use with Xen. The Device Tree files shipped with Linux or from the Split Device Tree Repository can be used.

Boot Modules

At boot time Xen must be provided with a dom0 kernel blob and an optional dom0 initramfs blob. The bootloader must load these into memory and describe their location in the Device Tree Blob using the bindings specified in docs/misc/arm/device-tree/booting.txt.

These nodes can either be added by hand (by editing and recompiling the .dts file) or by using u-boot's "fdt" command to add them dyanmically at boot time:

   fdt addr ${fdt_addr}
   fdt resize
   
   fdt set /chosen \#address-cells <1>
   fdt set /chosen \#size-cells <1>
   
   fdt mknod /chosen module@0
   fdt set /chosen/module@0 compatible "xen,linux-zimage" "xen,multiboot-module"
   fdt set /chosen/module@0 reg <${kernel_addr_r} 0x${filesize} >
   fdt set /chosen/module@0 bootargs "<DOMAIN 0 COMMAND LINE>"

(this assumes the device tree blob is loaded at ${fdt_addr} and the kernel at ${kernel_addr_r}).

See e.g. Xen_ARM_with_Virtualization_Extensions/Allwinner for a more concrete example of this.

Command Lines

docs/misc/arm/device-tree/booting.txt describes where Xen looks for both its own command line and the command line to pass to domain 0.

Getting Xen output

To get output log on the UART, Xen needs to know which UART to use. This should be passed in the hypervisor command line using the "dtuart" parameter. e.g.:

console=dtuart dtuart=myserial

where myserial is either an alias to the UART in the device tree (aliases are found in the aliases device tree node) or a full DTB path to the device. As Xen already uses it the UART will be disabled from the point of view of domain 0.

For instance, this is a dummy device tree (won't work) to use the uart0 in Xen:

/ {
    choosen {
       bootargs = "console=dtuart dtuart=myserial";
    }
    aliases {
        myserial = &myserial_0;
    }
    myserial_0: uart0 {
       ... configuration of your UART ...
    }
  }

Here dtuart is configured using the myserial alias. Alternatively /uart0 (the full path to the device) could have been used.

Note: If you don't see output from Xen, you can enable early printk. This option will turn on platform specific UART and output information before the console is initialized.

Dom0 kernel

In general the same kernel configuration as used to boot natively, plus turning on the Xen specific options should work. A good starting point is often the "multi_v7_defconfig" + Xen options.

If ARM_APPENDED_DTB is enabled then any appended DTB will be used instead of one supplied by Xen and the kernel will crash unless the memory in the DTB matches that location/size supplied by Xen. It is strongly recommended not to append a DTB to your dom0 kernel (or to disable APPENDED_DTB).

DomU kernel and DTS

Unprivileged guests can be created using xl. A simple VM config file would look like this:

kernel = "/root/image"
memory = 128
name = "guest"
vcpus = 1
disk = [ 'phy:/dev/loop0,xvda,w' ]
extra = "earlyprintk=xenboot console=hvc0 root=/dev/xvda debug rw init=/bin/sh"

where "/root/image" is a Linux zImage.

Common DomU Pitfalls

Enabling CONFIG_DEBUG_LL in the guest kernel configuration.

Although this option can work for dom0 if configured appropriately for the host it does not work for domU (which cannot see the host UART). The symptoms of this are that the guest console will be silent because the kernel has taken a fault accessing the early UART. This can be confirmed by using the xenctx tool (found in $PREFIX/lib/xen/bin/). The tool takes a numeric domid (not a name, use xl list or xl domid $name) and dumps the VCPU state. A PC of 0x0000000c will usually indicate that an early trap has occurred.

Building Xen on ARM

At least gcc version 4.7.3 is known to miscompile certain parts of Xen, most often leading to a segmentation fault in the xl toolstack while starting a guest. We recommend using 4.8.x or later

There are two major components which need to be built for a Xen system. The Xen hypervisor binary itself and the Xen toolstack.

Cross Compiling

Cross compiling the Xen hypervisor is relatively simple.

Linaro supply cross compilers for both arm32 (arm-linux-gnueabihf-) and arm64 (aarch64-linux-gnu-) via linaro-toolchain-binaries. Alternatively, for 32-bit at least, you can download the arm-unknown-linux-gnueabi compiler from kernel.org.

Once you have a suitable cross compiler you can compile Xen with:

   $ make dist-xen XEN_TARGET_ARCH=arm32 CROSS_COMPILE=arm-unknown-linux-gnueabihf-

or:

   $ make dist-xen XEN_TARGET_ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu-

This assumes that the command prefix for you cross compiler is arm-unknown-linux-gnueabihf- or aarch64-linux-gnu- and that the appropriate arm-unknown-linux-gnueabihf-gcc or aarch64-linux-gnu-gcc and friends are in your $PATH.

Cross compilation of the Xen toolstack is described in Xen ARM with Virtualization Extensions/CrossCompiling.

Native Building

In order to build the tools a native build environment is required. For 32-bit the developers mainly use the armhf port of Debian, which is present in Wheezy running on an IMX53 based development board, although any ARMv7 development board would do. Note that the build hardware does not need to support the virtualisation extensions, since you don't have to run Xen on the same system as where you build it.

It may also be possible to build using a cross-architecture chroot running on an X86 system as described in this blog post, although this hasn't been tried yet.

Platform specific configuration

• Xen ARM with Virtualization Extensions/Allwinner
• Xen ARM with Virtualization Extensions/Arndale
• Xen ARM with Virtualization Extensions/Chromebook
• Xen ARM with Virtualization Extensions/FastModels
• Xen ARM with Virtualization Extensions/Midway
• Xen ARM with Virtualization Extensions/Vexpress
• Xen ARM with Virtualization Extensions/APMXGeneMustang
• Xen ARM with Virtualization Extensions/OMAP5432_uEVM
• Xen ARM with Virtualization Extensions/OdroidXU

Open issues, known problems and workarounds

See this page for a full list of bugs and projects.