libvirt/docs/kbase/launch_security_sev.html.in
Daniel P. Berrangé 3918fb0dc1 docs: introduce a "knowledge base" for task oriented guides
The previously added AMD SEV doc was not linked from anywhere on the
website. Address this by introducing a new "Knowledge base" section
that can hold task oriented guide to various features. Moving the SEV,
disk locking and secure usage guides under this section.

Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
2019-07-09 18:43:12 +01:00

522 lines
18 KiB
XML

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE html>
<html xmlns="http://www.w3.org/1999/xhtml">
<body>
<h1>Launch security with AMD SEV</h1>
<ul id="toc"></ul>
<p>
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.
</p>
<h2>AMD SEV</h2>
<p>
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
<a href="https://developer.amd.com/sev/">AMD's developer portal</a>.
</p>
<h2><a id="Host">Enabling SEV on the host</a></h2>
<p>
Before VMs can make use of the SEV feature you need to make sure your
AMD CPU does support SEV. You can check whether SEV is among the CPU
flags with:
</p>
<pre>
$ cat /proc/cpuinfo | grep sev
...
sme ssbd sev ibpb</pre>
<p>
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:
</p>
<pre>
mem_encrypt=on kvm_amd.sev=1
</pre>
<p>
To make the changes persistent, append the above to the variable holding
parameters of the kernel command line in
<code>/etc/default/grub</code> to preserve SEV settings across reboots
</p>
<pre>
$ cat /etc/default/grub
...
GRUB_CMDLINE_LINUX="... mem_encrypt=on kvm_amd.sev=1"
$ grub2-mkconfig -o /boot/efi/EFI/&lt;distro&gt;/grub.cfg</pre>
<p>
<code>mem_encrypt=on</code> turns on the SME memory encryption feature on
the host which protects against the physical attack on the hypervisor
memory. The <code>kvm_amd.sev</code> parameter actually enables SEV in
the kvm module. It can be set on the command line alongside
<code>mem_encrypt</code> like shown above, or it can be put into a
module config under <code>/etc/modprobe.d/</code>
</p>
<pre>
$ cat /etc/modprobe.d/sev.conf
options kvm_amd sev=1
</pre>
<p>
After rebooting the host, you should see SEV being enabled in the kernel:
</p>
<pre>
$ cat /sys/module/kvm_amd/parameters/sev
1
</pre>
<h2><a id="Virt">Checking SEV support in the virt stack</a></h2>
<p>
<b>Note: All of the commands bellow need to be run with root privileges.</b>
</p>
<p>
First make sure you have the following packages in the specified versions:
</p>
<ul>
<li>
libvirt >= 4.5.0 (>5.1.0 recommended due to additional SEV bugfixes)
</li>
<li>
QEMU >= 2.12.0
</li>
</ul>
<p>
To confirm that the virtualization stack supports SEV, run the following:
</p>
<pre>
# virsh domcapabilities
&lt;domainCapabilities&gt;
...
&lt;features&gt;
...
&lt;sev supported='yes'&gt;
&lt;cbitpos&gt;47&lt;/cbitpos&gt;
&lt;reducedPhysBits&gt;1&lt;/reducedPhysBits&gt;
&lt;/sev&gt;
...
&lt;/features&gt;
&lt;/domainCapabilities&gt;</pre>
<p>
Note that if libvirt 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:
</p>
<pre>
# systemctl stop libvirtd.service
</pre>
<p>
Now you need to clean the capabilities cache:
</p>
<pre>
# rm -f /var/cache/libvirt/qemu/capabilities/*
</pre>
<p>
If you now restart libvirtd, it will re-probe the capabilities and if
you now run:
</p>
<pre>
# virsh domcapabilities
</pre>
<p>
SEV should be listed as supported. If you still see:
</p>
<pre>
&lt;sev supported='no'/&gt;
</pre>
<p>
it means one of two things:
<ol>
<li>
libvirt does support SEV, but either QEMU or the host does not
</li>
<li>
you have libvirt &lt;=5.1.0 which suffered from getting a
<code>'Permission denied'</code> on <code>/dev/sev</code> 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
</li>
</ol>
</p>
<h2><a id="Configuration">VM Configuration</a></h2>
<p>
SEV is enabled in the XML by specifying the
<a href="https://libvirt.org/formatdomain.html#launchSecurity">&lt;launchSecurity&gt; </a> element. However, specifying <code>launchSecurity</code> isn't
enough to boot an SEV VM. Further configuration requirements are discussed
below.
</p>
<h3><a id="Machine">Machine type</a></h3>
<p>
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.
</p>
<h5>Q35</h5>
<pre>
...
&lt;os&gt;
&lt;type arch='x86_64' machine='pc-q35-3.0'&gt;hvm&lt;/type&gt;
...
&lt;/os&gt;
...</pre>
<h5>i440fx (discouraged)</h5>
<pre>
...
&lt;os&gt;
&lt;type arch='x86_64' machine='pc-i440fx-3.0'&gt;hvm&lt;/type&gt;
...
&lt;/os&gt;
...
</pre>
<h3><a id="Boot">Boot loader</a></h3>
<p>
SEV is only going to work with OVMF (UEFI), so you'll need to point libvirt to
the correct OVMF binary.
</p>
<pre>
...
&lt;os&gt;
&lt;type arch='x86_64' machine='pc-q35-3.0'&gt;hvm&lt;/type&gt;
&lt;loader readonly='yes' type='pflash'&gt;/usr/share/edk2/ovmf/OVMF_CODE.fd&lt;/loader&gt;
&lt;/os&gt;
...</pre>
<h3><a id="Memory">Memory</a></h3>
<p>
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 <code>&lt;currentMemory&gt;</code> 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.
<p>
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:
</p>
</p>
<pre>
# virsh edit &lt;domain&gt;
&lt;domain&gt;
...
&lt;currentMemory unit='KiB'&gt;4194304&lt;/currentMemory&gt;
&lt;memtune&gt;
&lt;hard_limit unit='KiB'&gt;4456448&lt;/hard_limit&gt;
&lt;/memtune&gt;
...
&lt;/domain&gt;</pre>
<p>
There's another, preferred method of taking care of the limits by
using the<code>&lt;memoryBacking&gt;</code> element along with the
<code>&lt;locked/&gt;</code> subelement:
</p>
<pre>
&lt;domain&gt;
...
&lt;memoryBacking&gt;
&lt;locked/&gt;
&lt;/memoryBacking&gt;
...
&lt;/domain&gt;</pre>
<p>
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 <code>machine.slice</code>.
</p>
<pre>
# systemctl set-property machine.slice MemoryHigh=&lt;value&gt;</pre>
<p>
To put even stricter measures in place which would involve the OOM killer, use
<pre>
# systemctl set-property machine.slice MemoryMax=&lt;value&gt;</pre>
instead. Alternatively, you can create a systemd config (don't forget
to reload systemd configuration in this case):
<pre>
# cat &lt;&lt; EOF &gt; /etc/systemd/system.control/machine.slice.d/90-MemoryMax.conf
MemoryMax=&lt;value&gt;
EOF</pre>
The trade-off to keep in mind with the second approach is that the VMs
can still perform DoS on each other.
</p>
<h3><a id="Virtio">Virtio</a></h3>
<p>
In order to make virtio devices work, we need to enable emulated IOMMU
on the devices so that virtual DMA can work.
</p>
<pre>
# virsh edit &lt;domain&gt;
&lt;domain&gt;
...
&lt;controller type='virtio-serial' index='0'&gt;
&lt;driver iommu='on'/&gt;
&lt;/controller&gt;
&lt;controller type='scsi' index='0' model='virtio-scsi'&gt;
&lt;driver iommu='on'/&gt;
&lt;/controller&gt;
...
&lt;memballoon model='virtio'&gt;
&lt;driver iommu='on'/&gt;
&lt;/memballoon&gt;
&lt;rng model='virtio'&gt;
&lt;backend model='random'&gt;/dev/urandom&lt;/backend&gt;
&lt;driver iommu='on'/&gt;
&lt;/rng&gt;
...
&lt;domain&gt;</pre>
<p>
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 &lt;1.0. In
libvirt, this is handled by using the virtio-non-transitional device model
(libvirt &gt;= 5.2.0 required).
<p>
Note: some devices like video devices don't
support non-transitional model, which means that virtio GPU cannot be used.
</p>
</p>
<pre>
&lt;domain&gt;
...
&lt;devices&gt;
...
&lt;memballoon model='virtio-non-transitional'&gt;
&lt;driver iommu='on'/&gt;
&lt;/memballoon&gt;
&lt;/devices&gt;
...
&lt;/domain&gt;</pre>
<h2><a id="Limitations">Limitations</a></h2>
<p>
Currently, the boot disk cannot be of type virtio-blk, instead, virtio-scsi
needs to be used if virtio is desired. This limitation is expected to be lifted
with future releases of kernel (the kernel used at the time of writing the
article is 5.0.14).
If you still cannot start an SEV VM, it could be because of wrong SELinux label on the <code>/dev/sev</code> device with selinux-policy &lt;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 <code>audit2allow</code> on how to do that) or putting SELinux into permissive mode (discouraged).
</p>
<h2><a id="Examples">Full domain XML examples</a></h2>
<h5>Q35 machine</h5>
<pre>
&lt;domain type='kvm'&gt;
&lt;name&gt;sev-dummy&lt;/name&gt;
&lt;memory unit='KiB'&gt;4194304&lt;/memory&gt;
&lt;currentMemory unit='KiB'&gt;4194304&lt;/currentMemory&gt;
&lt;memoryBacking&gt;
&lt;locked/&gt;
&lt;/memoryBacking&gt;
&lt;vcpu placement='static'&gt;4&lt;/vcpu&gt;
&lt;os&gt;
&lt;type arch='x86_64' machine='pc-q35-3.0'&gt;hvm&lt;/type&gt;
&lt;loader readonly='yes' type='pflash'&gt;/usr/share/edk2/ovmf/OVMF_CODE.fd&lt;/loader&gt;
&lt;nvram&gt;/var/lib/libvirt/qemu/nvram/sev-dummy_VARS.fd&lt;/nvram&gt;
&lt;/os&gt;
&lt;features&gt;
&lt;acpi/&gt;
&lt;apic/&gt;
&lt;vmport state='off'/&gt;
&lt;/features&gt;
&lt;cpu mode='host-model' check='partial'&gt;
&lt;model fallback='allow'/&gt;
&lt;/cpu&gt;
&lt;clock offset='utc'&gt;
&lt;timer name='rtc' tickpolicy='catchup'/&gt;
&lt;timer name='pit' tickpolicy='delay'/&gt;
&lt;timer name='hpet' present='no'/&gt;
&lt;/clock&gt;
&lt;on_poweroff&gt;destroy&lt;/on_poweroff&gt;
&lt;on_reboot&gt;restart&lt;/on_reboot&gt;
&lt;on_crash&gt;destroy&lt;/on_crash&gt;
&lt;pm&gt;
&lt;suspend-to-mem enabled='no'/&gt;
&lt;suspend-to-disk enabled='no'/&gt;
&lt;/pm&gt;
&lt;devices&gt;
&lt;emulator&gt;/usr/bin/qemu-kvm&lt;/emulator&gt;
&lt;disk type='file' device='disk'&gt;
&lt;driver name='qemu' type='qcow2'/&gt;
&lt;source file='/var/lib/libvirt/images/sev-dummy.qcow2'/&gt;
&lt;target dev='sda' bus='scsi'/&gt;
&lt;boot order='1'/&gt;
&lt;/disk&gt;
&lt;controller type='virtio-serial' index='0'&gt;
&lt;driver iommu='on'/&gt;
&lt;/controller&gt;
&lt;controller type='scsi' index='0' model='virtio-scsi'&gt;
&lt;driver iommu='on'/&gt;
&lt;/controller&gt;
&lt;interface type='network'&gt;
&lt;mac address='52:54:00:cc:56:90'/&gt;
&lt;source network='default'/&gt;
&lt;model type='virtio'/&gt;
&lt;driver iommu='on'/&gt;
&lt;/interface&gt;
&lt;graphics type='spice' autoport='yes'&gt;
&lt;listen type='address'/&gt;
&lt;gl enable='no'/&gt;
&lt;/graphics&gt;
&lt;video&gt;
&lt;model type='qxl'/&gt;
&lt;/video&gt;
&lt;memballoon model='virtio'&gt;
&lt;driver iommu='on'/&gt;
&lt;/memballoon&gt;
&lt;rng model='virtio'&gt;
&lt;driver iommu='on'/&gt;
&lt;/rng&gt;
&lt;/devices&gt;
&lt;launchSecurity type='sev'&gt;
&lt;cbitpos&gt;47&lt;/cbitpos&gt;
&lt;reducedPhysBits&gt;1&lt;/reducedPhysBits&gt;
&lt;policy&gt;0x0003&lt;/policy&gt;
&lt;/launchSecurity&gt;
&lt;/domain&gt;</pre>
<h5>PC-i440fx machine:</h5>
<pre>
&lt;domain type='kvm'&gt;
&lt;name&gt;sev-dummy-legacy&lt;/name&gt;
&lt;memory unit='KiB'&gt;4194304&lt;/memory&gt;
&lt;currentMemory unit='KiB'&gt;4194304&lt;/currentMemory&gt;
&lt;memtune&gt;
&lt;hard_limit unit='KiB'&gt;5242880&lt;/hard_limit&gt;
&lt;/memtune&gt;
&lt;vcpu placement='static'&gt;4&lt;/vcpu&gt;
&lt;os&gt;
&lt;type arch='x86_64' machine='pc-i440fx-3.0'&gt;hvm&lt;/type&gt;
&lt;loader readonly='yes' type='pflash'&gt;/usr/share/edk2/ovmf/OVMF_CODE.fd&lt;/loader&gt;
&lt;nvram&gt;/var/lib/libvirt/qemu/nvram/sev-dummy_VARS.fd&lt;/nvram&gt;
&lt;boot dev='hd'/&gt;
&lt;/os&gt;
&lt;features&gt;
&lt;acpi/&gt;
&lt;apic/&gt;
&lt;vmport state='off'/&gt;
&lt;/features&gt;
&lt;cpu mode='host-model' check='partial'&gt;
&lt;model fallback='allow'/&gt;
&lt;/cpu&gt;
&lt;clock offset='utc'&gt;
&lt;timer name='rtc' tickpolicy='catchup'/&gt;
&lt;timer name='pit' tickpolicy='delay'/&gt;
&lt;timer name='hpet' present='no'/&gt;
&lt;/clock&gt;
&lt;on_poweroff&gt;destroy&lt;/on_poweroff&gt;
&lt;on_reboot&gt;restart&lt;/on_reboot&gt;
&lt;on_crash&gt;destroy&lt;/on_crash&gt;
&lt;pm&gt;
&lt;suspend-to-mem enabled='no'/&gt;
&lt;suspend-to-disk enabled='no'/&gt;
&lt;/pm&gt;
&lt;devices&gt;
&lt;emulator&gt;/usr/bin/qemu-kvm&lt;/emulator&gt;
&lt;disk type='file' device='disk'&gt;
&lt;driver name='qemu' type='qcow2'/&gt;
&lt;source file='/var/lib/libvirt/images/sev-dummy-seabios.qcow2'/&gt;
&lt;target dev='sda' bus='sata'/&gt;
&lt;/disk&gt;
&lt;interface type='network'&gt;
&lt;mac address='52:54:00:d8:96:c8'/&gt;
&lt;source network='default'/&gt;
&lt;model type='virtio-non-transitional'/&gt;
&lt;/interface&gt;
&lt;serial type='pty'&gt;
&lt;target type='isa-serial' port='0'&gt;
&lt;model name='isa-serial'/&gt;
&lt;/target&gt;
&lt;/serial&gt;
&lt;console type='pty'&gt;
&lt;target type='serial' port='0'/&gt;
&lt;/console&gt;
&lt;input type='tablet' bus='usb'&gt;
&lt;address type='usb' bus='0' port='1'/&gt;
&lt;/input&gt;
&lt;input type='mouse' bus='ps2'/&gt;
&lt;input type='keyboard' bus='ps2'/&gt;
&lt;graphics type='spice' autoport='yes'&gt;
&lt;listen type='address'/&gt;
&lt;gl enable='no'/&gt;
&lt;/graphics&gt;
&lt;video&gt;
&lt;model type='qxl' ram='65536' vram='65536' vgamem='16384' heads='1' primary='yes'/&gt;
&lt;/video&gt;
&lt;memballoon model='virtio-non-transitional'&gt;
&lt;driver iommu='on'/&gt;
&lt;/memballoon&gt;
&lt;rng model='virtio-non-transitional'&gt;
&lt;driver iommu='on'/&gt;
&lt;/rng&gt;
&lt;/devices&gt;
&lt;launchSecurity type='sev'&gt;
&lt;cbitpos&gt;47&lt;/cbitpos&gt;
&lt;reducedPhysBits&gt;1&lt;/reducedPhysBits&gt;
&lt;policy&gt;0x0003&lt;/policy&gt;
&lt;/launchSecurity&gt;
&lt;/domain&gt;</pre>
</body>
</html>