cloud-hypervisor/docs/iommu.md

Virtual IOMMU

Rationales

Having the possibility to expose a virtual IOMMU to the guest can be interesting to support specific use cases. That being said, it is always important to keep in mind a virtual IOMMU can impact the performance of the attached devices, which is the reason why one should be careful when enabling this feature.

Protect nested virtual machines

The first reason why one might want to expose a virtual IOMMU to the guest is to increase the security regarding the memory accesses performed by the virtual devices (VIRTIO devices), on behalf of the guest drivers.

With a virtual IOMMU, the VMM stands between the guest driver and its device counterpart, validating and translating every address before to try accessing the guest memory. This is standard interposition that is performed here by the VMM.

The increased security does not apply for a simple case where we have one VM per VMM. Because the guest cannot be trusted, as we always consider it could be malicious and gain unauthorized privileges inside the VM, preventing some devices from accessing the entire guest memory is pointless.

But let's take the interesting case of nested virtualization, and let's assume we have a VMM running a first layer VM. This L1 guest is fully trusted as the user intends to run multiple VMs from this L1. We can end up with multiple L2 VMs running on a single L1 VM. In this particular case, and without exposing a virtual IOMMU to the L1 guest, it would be possible for any L2 guest to use the device implementation from the host VMM to access the entire guest L1 memory. The virtual IOMMU prevents from this kind of trouble as it will validate the addresses the device is authorized to access.

Achieve VFIO nested

Another reason for having a virtual IOMMU is to allow passing physical devices from the host through multiple layers of virtualization. Let's take as example a system with a physical IOMMU running a VM with a virtual IOMMU. The implementation of the virtual IOMMU is responsible for updating the physical DMA Remapping table (DMAR) everytime the DMA mapping changes. This must happen through the VFIO framework on the host as this is the only userspace interface to interact with a physical IOMMU.

Relying on this update mechanism, it is possible to attach physical devices to the virtual IOMMU, which allows these devices to be passed from L1 to another layer of virtualization.

Why virtio-iommu?

The Cloud Hypervisor project decided to implement the brand new virtio-iommu device in order to provide a virtual IOMMU to its users. The reason being the simplicity brought by the paravirtualization solution. By having one side handled from the guest itself, it removes the complexity of trapping memory page accesses and shadowing them. This is why the project will not try to implement a full emulation of a physical IOMMU.

Pre-requisites

Kernel

Since virtio-iommu has landed partially into the version 5.3 of the Linux kernel, a special branch is needed to get things working with Cloud Hypervisor. By partially, we are talking about x86 specifically, as it is already fully functional for ARM architectures.

Usage

In order to expose a virtual IOMMU to the guest, it is required to create a virtio-iommu device and expose it through the ACPI IORT table. This can be simply achieved by attaching at least one device to the virtual IOMMU.

The way to expose to the guest a specific device as sitting behind this IOMMU is to explicitly tag it from the command line with the option iommu=on.

Not all devices support this extra option, and the default value will always be off since we want to avoid the performance impact for most users who don't need this.

Refer to the command line --help to find out which device support to be attached to the virtual IOMMU.

Below is a simple example exposing the virtio-blk device as attached to the virtual IOMMU:

./cloud-hypervisor \
    --cpus boot=1 \
    --memory size=512M \
    --disk path=focal-server-cloudimg-amd64.raw,iommu=on \
    --kernel custom-vmlinux \
    --cmdline "console=ttyS0 console=hvc0 root=/dev/vda1 rw" \

From a guest perspective, it is easy to verify if the device is protected by the virtual IOMMU. Check the directories listed under /sys/kernel/iommu_groups:

ls /sys/kernel/iommu_groups
0

In this case, only one IOMMU group should be created. Under this group, it is possible to find out the b/d/f of the device(s) part of this group.

ls /sys/kernel/iommu_groups/0/devices/
0000:00:03.0

And you can validate the device is the one we expect running lspci:

lspci
00:00.0 Host bridge: Intel Corporation Device 0d57
00:01.0 Unassigned class [ffff]: Red Hat, Inc. Device 1057
00:02.0 Unassigned class [ffff]: Red Hat, Inc. Virtio console
00:03.0 Mass storage controller: Red Hat, Inc. Virtio block device
00:04.0 Unassigned class [ffff]: Red Hat, Inc. Virtio RNG

Work with FDT on AArch64

On AArch64 architecture, the virtual IOMMU can still be used even if ACPI is not enabled. But the effect is different with what the aforementioned test showed.

When ACPI is disabled, virtual IOMMU is supported through Flattened Device Tree (FDT). In this case, the guest kernel can not tell which device should be IOMMU-attached and which should not. No matter how many devices you attached to the virtual IOMMU by setting iommu=on option, all the devices on the PCI bus will be attached to the virtual IOMMU (except the IOMMU itself). Each of the devices will be added into a IOMMU group.

As a result, the directory content of /sys/kernel/iommu_groups would be:

ls /sys/kernel/iommu_groups/0/devices/
0000:00:02.0
ls /sys/kernel/iommu_groups/1/devices/
0000:00:03.0
ls /sys/kernel/iommu_groups/2/devices/
0000:00:04.0

Faster mappings

By default, the guest memory is mapped with 4k pages and no huge pages, which causes the virtual IOMMU device to be asked for 4k mappings only. This configuration slows down the setup of the physical IOMMU as an important number of requests need to be issued in order to create large mappings.

One use case is even more impacted by the slowdown, the nested VFIO case. When passing a device through a L2 guest, the VFIO driver running in L1 will update the DMAR entries for the specific device. Because VFIO pins the entire guest memory, this means the entire mapping of the L2 guest need to be stored into multiple 4k mappings. Obviously, the bigger the L2 guest RAM is, the longer the update of the mappings will last. There is an additional problem happening in this case, if the L2 guest RAM is quite large, it will require a large number of mappings, which might exceed the VFIO limit set on the host. The default value is 65536, which can simply be reached with a 256MiB sized RAM.

The way to solve both problems, the slowdown and the limit being exceeded, is to reduce the amount of requests to describe those same large mappings. This can be achieved by using 2MiB pages, known as huge pages. By seeing the guest RAM as larger pages, and because the virtual IOMMU device supports it, the guest will require less mappings, which will prevent the limit from being exceeded, but also will take less time to process them on the host. That's how using huge pages as much as possible can speed up VM boot time.

Basic usage

Let's look at an example of how to run a guest with huge pages.

First, make sure your system has enough pages to cover the entire guest RAM:

# This example creates 4096 hugepages
echo 4096 > /proc/sys/vm/nr_hugepages

Next step is simply to create the VM. Two things are important, first we want the VM RAM to be mapped on huge pages by backing it with /dev/hugepages. And second thing, we need to create some huge pages in the guest itself so they can be consumed.

./cloud-hypervisor \
    --cpus boot=1 \
    --memory size=8G,hugepages=on \
    --disk path=focal-server-cloudimg-amd64.raw \
    --kernel custom-vmlinux \
    --cmdline "console=ttyS0 console=hvc0 root=/dev/vda1 rw hugepagesz=2M hugepages=2048" \
    --net tap=,mac=,iommu=on

Nested usage

Let's now look at the specific example of nested virtualization. In order to reach optimized performances, the L2 guest also need to be mapped based on huge pages. Here is how to achieve this, assuming the physical device you are passing through is 0000:00:01.0.

./cloud-hypervisor \
    --cpus boot=1 \
    --memory size=8G,hugepages=on \
    --disk path=focal-server-cloudimg-amd64.raw \
    --kernel custom-vmlinux \
    --cmdline "console=ttyS0 console=hvc0 root=/dev/vda1 rw kvm-intel.nested=1 vfio_iommu_type1.allow_unsafe_interrupts rw hugepagesz=2M hugepages=2048" \
    --device path=/sys/bus/pci/devices/0000:00:01.0,iommu=on

Once the L1 VM is running, unbind the device from the default driver in the guest, and bind it to VFIO (it should appear as 0000:00:04.0).

echo 0000:00:04.0 > /sys/bus/pci/devices/0000\:00\:04.0/driver/unbind
echo 8086 1502 > /sys/bus/pci/drivers/vfio-pci/new_id
echo 0000:00:04.0 > /sys/bus/pci/drivers/vfio-pci/bind

Last thing is to start the L2 guest with the huge pages memory backend.

./cloud-hypervisor \
    --cpus boot=1 \
    --memory size=4G,hugepages=on \
    --disk path=focal-server-cloudimg-amd64.raw \
    --kernel custom-vmlinux \
    --cmdline "console=ttyS0 console=hvc0 root=/dev/vda1 rw" \
    --device path=/sys/bus/pci/devices/0000:00:04.0

Dedicated IOMMU PCI segments

To facilitate hotplug of devices that require being behind an IOMMU it is possible to mark entire PCI segments as behind the IOMMU.

This is accomplished through --platform num_pci_segments=<number_of_segments>,iommu_segments=<range of segments> or via the equivalents in PlatformConfig for the API.

e.g.

./cloud-hypervisor \
    --api-socket=/tmp/api \
    --cpus boot=1 \
    --memory size=4G,hugepages=on \
    --disk path=focal-server-cloudimg-amd64.raw \
    --kernel custom-vmlinux \
    --cmdline "console=ttyS0 console=hvc0 root=/dev/vda1 rw" \
    --platform num_pci_segments=2,iommu_segments=1

This adds a second PCI segment to the platform behind the IOMMU. A VFIO device requiring the IOMMU then may be hotplugged:

e.g.

./ch-remote --api-socket=/tmp/api add-device path=/sys/bus/pci/devices/0000:00:04.0,iommu=on,pci_segment=1

Devices that cannot be placed behind an IOMMU (e.g. lacking an iommu= option) cannot be placed on the IOMMU segments.