New in version 20.0.0: (Train)
Secure Encrypted Virtualization (SEV) is a technology from AMD which enables the memory for a VM to be encrypted with a key unique to the VM. SEV is particularly applicable to cloud computing since it can reduce the amount of trust VMs need to place in the hypervisor and administrator of their host system.
First the operator will need to ensure the following prerequisites are met:
Currently SEV is only supported when using the libvirt compute driver with a
libvirt.virt_type
of kvm
or qemu
.
At least one of the Nova compute hosts must be AMD hardware capable of supporting SEV. It is entirely possible for the compute plane to be a mix of hardware which can and cannot support SEV, although as per the section on Permanent limitations below, the maximum number of simultaneously running guests with SEV will be limited by the quantity and quality of SEV-capable hardware available.
In order for users to be able to use SEV, the operator will need to perform the following steps:
Ensure that sufficient memory is reserved on the SEV compute hosts for host-level services to function correctly at all times. This is particularly important when hosting SEV-enabled guests, since they pin pages in RAM, preventing any memory overcommit which may be in normal operation on other compute hosts.
It is recommended to achieve this by configuring an rlimit
at
the /machine.slice
top-level cgroup
on the host, with all VMs
placed inside that. (For extreme detail, see this discussion on the
spec.)
An alternative approach is to configure the
reserved_host_memory_mb
option in the
[DEFAULT]
section of nova.conf
, based on the expected
maximum number of SEV guests simultaneously running on the host, and
the details provided in an earlier version of the AMD SEV spec
regarding memory region sizes, which cover how to calculate it
correctly.
See the Memory Locking and Accounting section of the AMD SEV spec and previous discussion for further details.
A cloud administrator will need to define one or more SEV-enabled flavors as described below, unless it is sufficient for users to define SEV-enabled images.
Additionally the cloud operator should consider the following optional steps:
Configure the libvirt.num_memory_encrypted_guests
option in nova.conf
to represent the number of guests an SEV
compute node can host concurrently with memory encrypted at the
hardware level. For example:
[libvirt]
num_memory_encrypted_guests = 15
This option exists because on AMD SEV-capable hardware, the memory controller has a fixed number of slots for holding encryption keys, one per guest. For example, at the time of writing, earlier generations of hardware only have 15 slots, thereby limiting the number of SEV guests which can be run concurrently to 15. Nova needs to track how many slots are available and used in order to avoid attempting to exceed that limit in the hardware.
At the time of writing (September 2019), work is in progress to allow QEMU and libvirt to expose the number of slots available on SEV hardware; however until this is finished and released, it will not be possible for Nova to programmatically detect the correct value.
So this configuration option serves as a stop-gap, allowing the cloud operator the option of providing this value manually. It may later be demoted to a fallback value for cases where the limit cannot be detected programmatically, or even removed altogether when Nova’s minimum QEMU version guarantees that it can always be detected.
Note
When deciding whether to use the default of None
or manually
impose a limit, operators should carefully weigh the benefits
vs. the risk. The benefits of using the default are a) immediate
convenience since nothing needs to be done now, and b) convenience
later when upgrading compute hosts to future versions of Nova,
since again nothing will need to be done for the correct limit to
be automatically imposed. However the risk is that until
auto-detection is implemented, users may be able to attempt to
launch guests with encrypted memory on hosts which have already
reached the maximum number of guests simultaneously running with
encrypted memory. This risk may be mitigated by other limitations
which operators can impose, for example if the smallest RAM
footprint of any flavor imposes a maximum number of simultaneously
running guests which is less than or equal to the SEV limit.
Configure ram_allocation_ratio
on all SEV-capable
compute hosts to 1.0
. Use of SEV requires locking guest memory, meaning
it is not possible to overcommit host memory.
Alternatively, you can explicitly configure small pages for instances using
the hw:mem_page_size
flavor extra spec and equivalent
image metadata property. For more information, see Huge pages.
Configure libvirt.hw_machine_type
on all
SEV-capable compute hosts to include x86_64=q35
, so that all
x86_64 images use the q35
machine type by default. (Currently
Nova defaults to the pc
machine type for the x86_64
architecture, although it is expected that this will change in the
future.)
Changing the default from pc
to q35
makes the creation and
configuration of images by users more convenient by removing the
need for the hw_machine_type
property to be set to q35
on
every image for which SEV booting is desired.
Caution
Consider carefully whether to set this option. It is
particularly important since a limitation of the implementation
prevents the user from receiving an error message with a helpful
explanation if they try to boot an SEV guest when neither this
configuration option nor the image property are set to select
a q35
machine type.
On the other hand, setting it to q35
may have other
undesirable side-effects on other images which were expecting to
be booted with pc
, so it is suggested to set it on a single
compute node or aggregate, and perform careful testing of typical
images before rolling out the setting to all SEV-capable compute
hosts.
Once an operator has covered the above steps, users can launch SEV
instances either by requesting a flavor for which the operator set the
hw:mem_encryption
extra spec to True
, or by using an
image with the hw_mem_encryption
property set to True
. For example, to
enable SEV for a flavor:
$ openstack flavor set FLAVOR-NAME \
--property hw:mem_encryption=true
These do not inherently cause a preference for SEV-capable hardware,
but for now SEV is the only way of fulfilling the requirement for
memory encryption. However in the future, support for other
hardware-level guest memory encryption technology such as Intel MKTME
may be added. If a guest specifically needs to be booted using SEV
rather than any other memory encryption technology, it is possible to
ensure this by setting the trait{group}:HW_CPU_X86_AMD_SEV
extra spec or equivalent image metadata property to required
.
In all cases, SEV instances can only be booted from images which have
the hw_firmware_type
property set to uefi
, and only when the
machine type is set to q35
. This can be set per image by setting
the image property hw_machine_type=q35
, or per compute node by
the operator via libvirt.hw_machine_type
as
explained above.
The following limitations may be removed in the future as the hardware, firmware, and various layers of software receive new features:
SEV-encrypted VMs cannot yet be live-migrated or suspended, therefore they will need to be fully shut down before migrating off an SEV host, e.g. if maintenance is required on the host.
SEV-encrypted VMs cannot contain directly accessible host devices (PCI passthrough). So for example mdev vGPU support will not currently work. However technologies based on vhost-user should work fine.
The boot disk of SEV-encrypted VMs can only be virtio
.
(virtio-blk
is typically the default for libvirt disks on x86,
but can also be explicitly set e.g. via the image property
hw_disk_bus=virtio
). Valid alternatives for the disk
include using hw_disk_bus=scsi
with
hw_scsi_model=virtio-scsi
, or hw_disk_bus=sata
.
QEMU and libvirt cannot yet expose the number of slots available for
encrypted guests in the memory controller on SEV hardware. Until
this is implemented, it is not possible for Nova to programmatically
detect the correct value. As a short-term workaround, operators can
optionally manually specify the upper limit of SEV guests for each
compute host, via the new
libvirt.num_memory_encrypted_guests
configuration option described above.
The following limitations are expected long-term:
The number of SEV guests allowed to run concurrently will always be limited. On the first generation of EPYC machines it will be limited to 15 guests; however this limit becomes much higher with the second generation (Rome).
The operating system running in an encrypted virtual machine must contain SEV support.
For the sake of eliminating any doubt, the following actions are not expected to be limited when SEV encryption is used:
Cold migration or shelve, since they power off the VM before the operation at which point there is no encrypted memory (although this could change since there is work underway to add support for PMEM)
Snapshot, since it only snapshots the disk
nova evacuate
(despite the name, more akin to resurrection than
evacuation), since this is only initiated when the VM is no longer
running
Attaching any volumes, as long as they do not require attaching via an IDE bus
Use of spice / VNC / serial / RDP consoles
Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.