Launch security with AMD SEV

Storage encryption in modern public cloud computing is a common practice. However, from the point of view of a user of these cloud workloads, a significant amount of trust needs to be put in the cloud platform security as well as integrity (was the hypervisor tampered?). For this reason there's ever rising demand for securing data in use, i.e. memory encryption. One of the solutions addressing this matter is AMD SEV.

AMD Secure Encrypted Virtualization (SEV)

SEV (Secure Encrypted Virtualization) is a feature extension of AMD's SME (Secure Memory Encryption) intended for KVM virtual machines which is supported primarily on AMD's EPYC CPU line. In contrast to SME, SEV uses a unique memory encryption key for each VM. The whole encryption of memory pages is completely transparent to the hypervisor and happens inside dedicated hardware in the on-die memory controller. Each controller includes a high-performance Advanced Encryption Standard (AES) engine that encrypts data when it is written to DRAM and decrypts it when read. For more details about the technology itself, you can visit AMD's developer portal.

Enabling SEV on the host

Before VMs can make use of the SEV feature you need to make sure your AMD CPU does support SEV. You can run virt-host-validate (libvirt >= 6.5.0) to check if your host supports secure guests or you can follow the manual checks below.

You can manually check whether SEV is among the CPU flags with:

$ grep -w sev /proc/cpuinfo
...
sme ssbd sev ibpb

Next step is to enable SEV in the kernel, because it is disabled by default. This is done by putting the following onto the kernel command line:

mem_encrypt=on kvm_amd.sev=1

To make the changes persistent, append the above to the variable holding parameters of the kernel command line in /etc/default/grub to preserve SEV settings across reboots

$ cat /etc/default/grub
...
GRUB_CMDLINE_LINUX="... mem_encrypt=on kvm_amd.sev=1"
$ grub2-mkconfig -o /boot/efi/EFI/<distro>/grub.cfg

mem_encrypt=on turns on the SME memory encryption feature on the host which protects against the physical attack on the hypervisor memory. The kvm_amd.sev parameter actually enables SEV in the kvm module. It can be set on the command line alongside mem_encrypt like shown above, or it can be put into a module config under /etc/modprobe.d/

$ cat /etc/modprobe.d/sev.conf
options kvm_amd sev=1

After rebooting the host, you should see SEV being enabled in the kernel:

$ cat /sys/module/kvm_amd/parameters/sev
1

Checking SEV support in the virt stack

Note: All of the commands below need to be run with root privileges.

First make sure you have the following packages in the specified versions:

  • libvirt >= 4.5.0 (>5.1.0 recommended due to additional SEV bugfixes)

  • QEMU >= 2.12.0

To confirm that the virtualization stack supports SEV, run the following:

# virsh domcapabilities
<domainCapabilities>
...
  <features>
    ...
    <sev supported='yes'>
      <cbitpos>47</cbitpos>
      <reducedPhysBits>1</reducedPhysBits>
    </sev>
    ...
  </features>
</domainCapabilities>

Note that if libvirt (<6.5.0) was already installed and libvirtd running before enabling SEV in the kernel followed by the host reboot you need to force libvirtd to re-probe both the host and QEMU capabilities. First stop libvirtd:

# systemctl stop libvirtd.service

Now you need to clean the capabilities cache:

# rm -f /var/cache/libvirt/qemu/capabilities/*

If you now restart libvirtd, it will re-probe the capabilities and if you now run:

# virsh domcapabilities

SEV should be listed as supported. If you still see:

<sev supported='no'/>

it means one of two things:

  1. libvirt does support SEV, but either QEMU or the host does not

  2. you have libvirt <=5.1.0 which suffered from getting a 'Permission denied' on /dev/sev because of the default permissions on the character device which prevented QEMU from opening it during capabilities probing - you can either manually tweak the permissions so that QEMU has access to it or preferably install libvirt 5.1.0 or higher

VM Configuration

SEV is enabled in the XML by specifying the <launchSecurity> element. However, specifying launchSecurity isn't enough to boot an SEV VM. Further configuration requirements are discussed below.

Machine type

Even though both Q35 and legacy PC machine types (for PC see also "virtio") can be used with SEV, usage of the legacy PC machine type is strongly discouraged, since depending on how your OVMF package was built (e.g. including features like SecureBoot or SMM) Q35 may even be required.

Q35

...
<os>
  <type arch='x86_64' machine='pc-q35-3.0'>hvm</type>
  ...
</os>
...

i440fx (discouraged)

...
<os>
  <type arch='x86_64' machine='pc-i440fx-3.0'>hvm</type>
  ...
</os>
...

Boot loader

SEV is only going to work with OVMF (UEFI), so you'll need to point libvirt to the correct OVMF binary.

...
<os>
  <type arch='x86_64' machine='pc-q35-3.0'>hvm</type>
  <loader readonly='yes' type='pflash'>/usr/share/edk2/ovmf/OVMF_CODE.fd</loader>
</os>
...

If intending to attest the boot measurement, it is required to use a firmware binary that is stateless, as persistent NVRAM can undermine the trust of the secure guest. This is achieved by telling libvirt that a stateless binary is required

...
<os type='efi'>
  <type arch='x86_64' machine='q35'>hvm</type>
  <loader stateless='yes'/>
</os>
...

Memory

Internally, SEV expects that the encrypted memory pages won't be swapped out or move around so the VM memory needs to be pinned in physical RAM which will be handled by QEMU. Apart from that, certain memory regions allocated by QEMU itself (UEFI pflash, device ROMs, video RAM, etc.) have to be encrypted as well. This causes a conflict in how libvirt tries to protect the host. By default, libvirt enforces a memory hard limit on each VM's cgroup in order to protect the host from malicious QEMU to allocate and lock all the available memory. This limit corresponds to the total memory allocation for the VM given by <currentMemory> element. However, trying to account for the additional memory regions QEMU allocates when calculating the limit in an automated manner is non-deterministic. One way to resolve this is to set the hard limit manually.

Note: Figuring out the right number so that your guest boots and isn't killed is challenging, but 256MiB extra memory over the total guest RAM should suffice for most workloads and may serve as a good starting point. For example, a domain with 4GB memory with a 256MiB extra hard limit would look like this:

# virsh edit <domain>
<domain>
  ...
  <currentMemory unit='KiB'>4194304</currentMemory>
  <memtune>
    <hard_limit unit='KiB'>4456448</hard_limit>
  </memtune>
  ...
</domain>

There's another, preferred method of taking care of the limits by using the<memoryBacking> element along with the <locked/> subelement:

<domain>
  ...
  <memoryBacking>
    <locked/>
  </memoryBacking>
  ...
</domain>

What that does is that it tells libvirt not to force any hard limit (well, unlimited) upon the VM cgroup. The obvious advantage is that one doesn't need to determine the hard limit for every single SEV-enabled VM. However, there is a significant security-related drawback to this approach. Since no hard limit is applied, a malicious QEMU could perform a DoS attack by locking all of the host's available memory. The way to avoid this issue and to protect the host is to enforce a bigger hard limit on the master cgroup containing all of the VMs - on systemd this is machine.slice.

# systemctl set-property machine.slice MemoryHigh=<value>

To put even stricter measures in place which would involve the OOM killer, use

# systemctl set-property machine.slice MemoryMax=<value>

instead. Alternatively, you can create a systemd config (don't forget to reload systemd configuration in this case):

# cat << EOF > /etc/systemd/system.control/machine.slice.d/90-MemoryMax.conf
MemoryMax=<value>
EOF

The trade-off to keep in mind with the second approach is that the VMs can still perform DoS on each other.

Virtio

In order to make virtio devices work, we need to use <driver iommu='on'/> inside the given device XML element in order to enable DMA API in the virtio driver.

Starting with QEMU 6.0.0 QEMU will set this for us by default. For earlier versions though, you will need to explicitly enable this in the device XML as follows:

# virsh edit <domain>
<domain>
  ...
  <controller type='virtio-serial' index='0'>
    <driver iommu='on'/>
  </controller>
  <controller type='scsi' index='0' model='virtio-scsi'>
    <driver iommu='on'/>
  </controller>
  ...
  <memballoon model='virtio'>
    <driver iommu='on'/>
  </memballoon>
  <rng model='virtio'>
    <backend model='random'>/dev/urandom</backend>
    <driver iommu='on'/>
  </rng>
  ...
<domain>

If you for some reason want to use the legacy PC machine type, further changes to the virtio configuration is required, because SEV will not work with Virtio <1.0. In libvirt, this is handled by using the virtio-non-transitional device model (libvirt >= 5.2.0 required).

Note: some devices like video devices don't support non-transitional model, which means that virtio GPU cannot be used.

<domain>
  ...
  <devices>
    ...
    <memballoon model='virtio-non-transitional'>
      <driver iommu='on'/>
    </memballoon>
  </devices>
  ...
</domain>

Virtio-net

With virtio-net it's also necessary to disable the iPXE option ROM as iPXE is not aware of SEV (at the time of this writing). This translates to the following XML:

<domain>
  ...
  <interface type='network'>
     ...
    <model type='virtio'/>
    <driver iommu='on'/>
    <rom enabled='no'/>
  </interface>
  ...
<domain>

Checking SEV from within the guest

After making the necessary adjustments discussed in VM Configuration, the VM should now boot successfully with SEV enabled. You can then verify that the guest has SEV enabled by running:

# dmesg | grep -i sev
AMD Secure Encrypted Virtualization (SEV) active

Guest attestation for SEV/SEV-ES from a trusted host

Before a confidential guest is used, it may be desirable to attest the boot measurement. To be trustworthy the attestation process needs to be performed from a machine that is already trusted. This would typically be a physical machine that the guest owner controls, or could be a previously launched confidential guest that has already itself been attested. Most notably, it is not possible to securely attest a guest from the hypervisor host itself, as the goal of the attestation process is to detect whether the hypervisor is malicious.

Performing an attestation requires that the <launchSecurity> element is configured with a guest owner Diffie-Hellman (DH) certificate, and a session data blob. These must be unique for every guest launch attempt. Any reuse will open avenues of attack for the hypervisor to fake the measurement. Unique data can be generated using the sevctl tool.

First of all the Platform Diffie-Hellman key (PDH) for the hypervisor host needs to be obtained. The PDH is used to negotiate a master secret between the SEV firmware and external entities.

The admin of the hypervisor can extract the PDH using:

$ sevctl export --full ${hostname}.pdh

Upon receiving the PDH associated with the hypervisor, the guest owner should validate its integrity:

$ sevctl verify --sev ${hostname}.pdh
PDH EP384 D256 008cec87d6bd9df67a35e7d6057a933463cd8a02440f60c5df150821b5662ee0
 ⬑ PEK EP384 E256 431ba88424378200d58b6fb5db9657268c599b1be25f8047ac2e2981eff667e6
   •⬑ OCA EP384 E256 b4f1d0a2144186d1aa9c63f19039834e729f508000aa05a76ba044f8e1419765
    ⬑ CEK EP384 E256 22c27ee3c1c33287db24d3c06869a5ae933eb44148fdb70838019e267077c6b8
       ⬑ ASK R4096 R384 d8cd9d1798c311c96e009a91552f17b4ddc4886a064ec933697734965b9ab29db803c79604e2725658f0861bfaf09ad4
         •⬑ ARK R4096 R384 3d2c1157c29ef7bd4207fc0c8b08db080e579ceba267f8c93bec8dce73f5a5e2e60d959ac37ea82176c1a0c61ae203ed

 • = self signed, ⬑ = signs, •̷ = invalid self sign, ⬑̸ = invalid signs

Assuming this is successful, it is now possible to generate a unique launch data for the guest boot attempt:

$ sevctl session --name ${myvmname} ${hostname}.pdh ${policy}

This will generate four files

  • ${myvmname}_tik.bin

  • ${myvmname}_tek.bin

  • ${myvmname}_godh.bin

  • ${myvmname}_session.bin

The tik.bin and tek.bin files will be needed to perform the boot attestation, and must be kept somewhere secure, away from the hypervisor host.

The godh.bin file contents should be copied into the <dhCert> field in the <launchSecurity> configuration, while the session.bin file contents should be copied into the <session> field.

When launching the guest, it should be set to remain in the paused state with no vCPUs running:

$ virsh start --paused ${myvmname}

With it launched, it is possible to query the launch measurement:

$ virsh domlaunchsecinfo ${myvmname}
sev-measurement: LMnv8i8N2QejezMPkscShF0cyPYCslgUoCxGWRqQuyt0Q0aUjVkH/T6NcmkwZkWp
sev-api-major  : 0
sev-api-minor  : 24
sev-build-id   : 15
sev-policy     : 3

The techiques required to validate the measurement reported are beyond the scope of this document. Fortunately, libvirt provides a tool that can be used to perform this validation:

$ virt-qemu-sev-validate \
    --measurement LMnv8i8N2QejezMPkscShF0cyPYCslgUoCxGWRqQuyt0Q0aUjVkH/T6NcmkwZkWp
    --api-major 0
    --api-minor 24
    --build-id 15
    --policy 3
    --tik ${myvmname}_tik.bin
    --tek ${myvmname}_tek.bin
OK: Looks good to me

The man page for virt-qemu-sev-validate outlines a great many other ways to invoke this tool.

Limitations

With older kernels (kernel <5.1) the boot disk cannot not be of type virtio-blk, instead, virtio-scsi needs to be used if virtio is desired.

If you still cannot start an SEV VM, it could be because of wrong SELinux label on the /dev/sev device with selinux-policy <3.14.2.40 which prevents QEMU from touching the device. This can be resolved by upgrading the package, tuning the selinux policy rules manually to allow svirt_t to access the device (see audit2allow on how to do that) or putting SELinux into permissive mode (discouraged).

Full domain XML examples

Q35 machine

<domain type='kvm'>
  <name>sev-dummy</name>
  <memory unit='KiB'>4194304</memory>
  <currentMemory unit='KiB'>4194304</currentMemory>
  <memoryBacking>
    <locked/>
  </memoryBacking>
  <vcpu placement='static'>4</vcpu>
  <os>
    <type arch='x86_64' machine='pc-q35-3.0'>hvm</type>
    <loader readonly='yes' type='pflash'>/usr/share/edk2/ovmf/OVMF_CODE.fd</loader>
    <nvram>/var/lib/libvirt/qemu/nvram/sev-dummy_VARS.fd</nvram>
  </os>
  <features>
    <acpi/>
    <apic/>
    <vmport state='off'/>
  </features>
  <cpu mode='host-model' check='partial'>
    <model fallback='allow'/>
  </cpu>
  <clock offset='utc'>
    <timer name='rtc' tickpolicy='catchup'/>
    <timer name='pit' tickpolicy='delay'/>
    <timer name='hpet' present='no'/>
  </clock>
  <on_poweroff>destroy</on_poweroff>
  <on_reboot>restart</on_reboot>
  <on_crash>destroy</on_crash>
  <pm>
    <suspend-to-mem enabled='no'/>
    <suspend-to-disk enabled='no'/>
  </pm>
  <devices>
    <emulator>/usr/bin/qemu-kvm</emulator>
    <disk type='file' device='disk'>
      <driver name='qemu' type='qcow2'/>
      <source file='/var/lib/libvirt/images/sev-dummy.qcow2'/>
      <target dev='sda' bus='scsi'/>
      <boot order='1'/>
    </disk>
    <controller type='virtio-serial' index='0'>
      <driver iommu='on'/>
    </controller>
    <controller type='scsi' index='0' model='virtio-scsi'>
      <driver iommu='on'/>
    </controller>
    <interface type='network'>
      <mac address='52:54:00:cc:56:90'/>
      <source network='default'/>
      <model type='virtio'/>
      <driver iommu='on'/>
      <rom enabled='no'/>
    </interface>
    <graphics type='spice' autoport='yes'>
      <listen type='address'/>
      <gl enable='no'/>
    </graphics>
    <video>
      <model type='qxl'/>
    </video>
    <memballoon model='virtio'>
      <driver iommu='on'/>
    </memballoon>
    <rng model='virtio'>
      <driver iommu='on'/>
    </rng>
  </devices>
  <launchSecurity type='sev'>
    <cbitpos>47</cbitpos>
    <reducedPhysBits>1</reducedPhysBits>
    <policy>0x0003</policy>
  </launchSecurity>
</domain>

PC-i440fx machine

<domain type='kvm'>
  <name>sev-dummy-legacy</name>
  <memory unit='KiB'>4194304</memory>
  <currentMemory unit='KiB'>4194304</currentMemory>
  <memtune>
    <hard_limit unit='KiB'>5242880</hard_limit>
  </memtune>
  <vcpu placement='static'>4</vcpu>
  <os>
    <type arch='x86_64' machine='pc-i440fx-3.0'>hvm</type>
    <loader readonly='yes' type='pflash'>/usr/share/edk2/ovmf/OVMF_CODE.fd</loader>
    <nvram>/var/lib/libvirt/qemu/nvram/sev-dummy_VARS.fd</nvram>
    <boot dev='hd'/>
  </os>
  <features>
  <acpi/>
  <apic/>
  <vmport state='off'/>
  </features>
  <cpu mode='host-model' check='partial'>
    <model fallback='allow'/>
  </cpu>
  <clock offset='utc'>
    <timer name='rtc' tickpolicy='catchup'/>
    <timer name='pit' tickpolicy='delay'/>
    <timer name='hpet' present='no'/>
  </clock>
  <on_poweroff>destroy</on_poweroff>
  <on_reboot>restart</on_reboot>
  <on_crash>destroy</on_crash>
  <pm>
    <suspend-to-mem enabled='no'/>
    <suspend-to-disk enabled='no'/>
  </pm>
  <devices>
    <emulator>/usr/bin/qemu-kvm</emulator>
    <disk type='file' device='disk'>
      <driver name='qemu' type='qcow2'/>
      <source file='/var/lib/libvirt/images/sev-dummy-seabios.qcow2'/>
      <target dev='sda' bus='sata'/>
    </disk>
    <interface type='network'>
      <mac address='52:54:00:d8:96:c8'/>
      <source network='default'/>
      <model type='virtio-non-transitional'/>
      <driver iommu='on'/>
      <rom enabled='no'/>
    </interface>
    <serial type='pty'>
      <target type='isa-serial' port='0'>
        <model name='isa-serial'/>
      </target>
    </serial>
    <console type='pty'>
      <target type='serial' port='0'/>
    </console>
    <input type='tablet' bus='usb'>
      <address type='usb' bus='0' port='1'/>
    </input>
    <input type='mouse' bus='ps2'/>
    <input type='keyboard' bus='ps2'/>
    <graphics type='spice' autoport='yes'>
      <listen type='address'/>
      <gl enable='no'/>
    </graphics>
    <video>
      <model type='qxl' ram='65536' vram='65536' vgamem='16384' heads='1' primary='yes'/>
    </video>
    <memballoon model='virtio-non-transitional'>
      <driver iommu='on'/>
    </memballoon>
      <rng model='virtio-non-transitional'>
    <driver iommu='on'/>
    </rng>
  </devices>
  <launchSecurity type='sev'>
    <cbitpos>47</cbitpos>
    <reducedPhysBits>1</reducedPhysBits>
    <policy>0x0003</policy>
  </launchSecurity>
</domain>