CPU models

Nova allows you to configure features of the virtual CPU that are exposed to instances. The combined set of CPU features is collectively referred to as the CPU model. Use cases include:

  • To maximize performance of instances by exposing new host CPU features to the guest

  • To ensure a consistent default behavior across all machines, removing reliance on system defaults.

To configure the virtual CPU, you can configure a CPU mode, configure one or more named CPU models, and explicitly request CPU feature flags.

The Effective Virtual CPU configuration in Nova presentation from the 2018 Berlin Summit provides a good overview of this topic.

Note

It is also possible to configure the topology of the CPU. This is discussed in CPU topologies.

Important

The functionality described below is currently only supported by the libvirt driver.

CPU modes

The first step in configuring the guest CPU is configuring the CPU mode. The CPU mode determines whether the CPU model is configured manually based on admin configuration or is automatically configured based on the host CPU. The CPU mode is configured using the libvirt.cpu_mode config option. This option can accepts one of the following values: none, host-passthrough, host-model, and custom.

Host model

If cpu_mode=host-model, libvirt requests the named CPU model that most closely matches the host and requests additional CPU flags to complete the match. This CPU model has a number of advantages:

  • It provides almost all of the host CPU features to the guest, thus providing close to the maximum functionality and performance possible.

  • It auto-adds critical guest CPU flags for mitigation from certain security flaws, provided the CPU microcode, kernel, QEMU, and libvirt are all updated.

  • It computes live migration compatibility, with the caveat that live migration in both directions is not always possible.

In general, using host-model is a safe choice if your compute node CPUs are largely identical. However, if your compute nodes span multiple processor generations, you may be better advised to select a custom CPU model.

The host-model CPU mode is the effective default for the KVM & QEMU hypervisors (libvirt.virt_type=kvm/qemu) on x86-64 hosts. This default is provided by libvirt itself.

Note

As noted above, live migration is not always possible in both directions when using host-model. During live migration, the source CPU model definition is transferred to the destination host as-is. This results in the migrated guest on the destination seeing exactly the same CPU model as on source even if the destination compute host is capable of providing more CPU features. However, shutting down and restarting the guest may result in a different hardware configuration for the guest, as per the new capabilities of the destination compute.

Host passthrough

If cpu_mode=host-passthrough, libvirt tells KVM to pass through the host CPU with no modifications. In comparison to host-model which simply matches feature flags, host-passthrough ensures every last detail of the host CPU is matched. This gives the best performance, and can be important to some apps which check low level CPU details, but it comes at a cost with respect to migration.

In host-passthrough mode, the guest can only be live-migrated to a target host that matches the source host extremely closely. This includes the physical CPU model and running microcode, and may even include the running kernel. Use this mode only if your compute nodes have a very large degree of homogeneity (i.e. substantially all of your compute nodes use the exact same CPU generation and model), and you make sure to only live-migrate between hosts with exactly matching kernel versions. Failure to do so will result in an inability to support any form of live migration.

Note

The reason for that it is necessary for the CPU microcode versions to match is that hardware performance counters are exposed to an instance and it is likely that they may vary between different CPU models. There may also be other reasons due to security fixes for some hardware security flaws being included in CPU microcode.

Custom

If cpu_mode=custom, you can explicitly specify an ordered list of one or more supported named CPU models using the libvirt.cpu_models configuration option. This accepts any named CPU model that is valid for the given host, as discussed in CPU models below. When more than one CPU model is provided, it is expected that the list will be ordered so that the more common and less advanced CPU models are listed first.

In selecting the custom mode, along with a named CPU model that matches the oldest of your compute node CPUs, you can ensure that live migration between compute nodes will always be possible. However, you should ensure that the CPU model you select passes the correct CPU feature flags to the guest.

If you need to further tweak your CPU feature flags in the custom mode, see CPU feature flags.

Note

If libvirt.cpu_models is configured, the CPU models in the list needs to be compatible with the host CPU. Also, if libvirt.cpu_model_extra_flags is configured, all flags needs to be compatible with the host CPU. If incompatible CPU models or flags are specified, nova service will raise an error and fail to start.

None

If cpu_mode=none, libvirt does not specify a CPU model. Instead, the hypervisor chooses the default model.

The none CPU model is the default for all non-KVM/QEMU hypervisors. (libvirt.virt_type!=``kvm``/qemu)

CPU models

When libvirt.cpu_mode is set to custom, it is possible to configure one or more explicit CPU models that should be used. These CPU model names are shorthand for a set of feature flags. The libvirt KVM driver provides a number of standard CPU model names. These models are defined in /usr/share/libvirt/cpu_map/*.xml. You can inspect these files to determine which models are supported by your local installation. For example, consider a host that provides the following (incomplete) set of CPU models:

$ ls /usr/share/libvirt/cpu_map/x86_*.xml -1
...
/usr/share/libvirt/cpu_map/x86_Broadwell-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Broadwell-noTSX-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Broadwell-noTSX.xml
/usr/share/libvirt/cpu_map/x86_Broadwell.xml
/usr/share/libvirt/cpu_map/x86_Haswell-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Haswell-noTSX-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Haswell-noTSX.xml
/usr/share/libvirt/cpu_map/x86_Haswell.xml
/usr/share/libvirt/cpu_map/x86_Icelake-Client-noTSX.xml
/usr/share/libvirt/cpu_map/x86_Icelake-Client.xml
/usr/share/libvirt/cpu_map/x86_Icelake-Server-noTSX.xml
/usr/share/libvirt/cpu_map/x86_Icelake-Server.xml
/usr/share/libvirt/cpu_map/x86_IvyBridge-IBRS.xml
/usr/share/libvirt/cpu_map/x86_IvyBridge.xml
/usr/share/libvirt/cpu_map/x86_SandyBridge-IBRS.xml
/usr/share/libvirt/cpu_map/x86_SandyBridge.xml
/usr/share/libvirt/cpu_map/x86_Skylake-Client-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Skylake-Client-noTSX-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Skylake-Client.xml
/usr/share/libvirt/cpu_map/x86_Skylake-Server-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Skylake-Server-noTSX-IBRS.xml
/usr/share/libvirt/cpu_map/x86_Skylake-Server.xml
...

Each of these files contains information about the feature set provided by the CPU model. For example:

$ cat /usr/share/libvirt/cpu_map/x86_SandyBridge-IBRS.xml
<cpus>
  <model name='SandyBridge-IBRS'>
    <decode host='on' guest='on'/>
    <signature family='6' model='42'/> <!-- 0206a0 -->
    <signature family='6' model='45'/> <!-- 0206d0 -->
    <vendor name='Intel'/>
    <feature name='aes'/>
    <feature name='apic'/>
    ...
  </model>
</cpus>

You can also list these CPU models using virsh cpu-models ARCH. For example:

$ virsh cpu-models x86_64
...
SandyBridge
SandyBridge-IBRS
IvyBridge
IvyBridge-IBRS
Haswell-noTSX
Haswell-noTSX-IBRS
Haswell
Haswell-IBRS
Broadwell-noTSX
Broadwell-noTSX-IBRS
Broadwell
Broadwell-IBRS
Skylake-Client
Skylake-Client-IBRS
Skylake-Client-noTSX-IBRS
Skylake-Server
Skylake-Server-IBRS
Skylake-Server-noTSX-IBRS
Icelake-Client
Icelake-Client-noTSX
Icelake-Server
Icelake-Server-noTSX
...

By settings cpu_mode=custom, it is possible to list one or more of these CPU models in the libvirt.cpu_models config option in nova.conf. For example:

[libvirt]
cpu_mode = custom
cpu_models = IvyBridge

Typically you will only need to list a single model here, but it can be useful to list multiple CPU models to support requesting CPU feature flags via traits. To do this, simply list the additional CPU models in order of oldest (and therefore most widely supported) to newest. For example:

[libvirt]
cpu_mode = custom
cpu_models = Penryn,IvyBridge,Haswell,Broadwell,Skylake-Client

More details on how to request CPU feature flags and why you might wish to specify multiple CPU models are provided in CPU feature flags below.

CPU feature flags

New in version 18.0.0: (Rocky)

Regardless of your configured libvirt.cpu_mode, it is also possible to selectively enable additional feature flags. This can be accomplished using the libvirt.cpu_model_extra_flags config option. For example, suppose you have configured a custom CPU model of IvyBridge, which normally does not enable the pcid feature flag, but you do want to pass pcid into your guest instances. In this case, you could configure the following in nova.conf to enable this flag.

[libvirt]
cpu_mode = custom
cpu_models = IvyBridge
cpu_model_extra_flags = pcid

An end user can also specify required CPU features through traits. When specified, the libvirt driver will select the first CPU model in the libvirt.cpu_models list that can provide the requested feature traits. If no CPU feature traits are specified then the instance will be configured with the first CPU model in the list.

Consider the following nova.conf:

[libvirt]
cpu_mode = custom
cpu_models = Penryn,IvyBridge,Haswell,Broadwell,Skylake-Client

These different CPU models support different feature flags and are correctly configured in order of oldest (and therefore most widely supported) to newest. If the user explicitly required the avx and avx2 CPU features, the latter of which is only found of Haswell-generation processors or newer, then they could request them using the trait{group}:HW_CPU_X86_AVX and trait{group}:HW_CPU_X86_AVX2 flavor extra specs. For example:

$ openstack flavor set $FLAVOR \
    --property trait:HW_CPU_X86_AVX=required \
    --property trait:HW_CPU_X86_AVX2=required

As Haswell is the first CPU model supporting both of these CPU features, the instance would be configured with this model.

Mitigation for MDS (“Microarchitectural Data Sampling”) Security Flaws

In May 2019, four new microprocessor flaws, known as MDS and also referred to as RIDL and Fallout or ZombieLoad, were discovered. These flaws affect unpatched Nova compute nodes and instances running on Intel x86_64 CPUs.

Resolution

To get mitigation for the said MDS security flaws, a new CPU flag, md-clear, needs to be exposed to the Nova instances. This can be done as follows.

  1. Update the following components to the versions from your Linux distribution that have fixes for the MDS flaws, on all compute nodes with Intel x86_64 CPUs:

    • microcode_ctl

    • kernel

    • qemu-system-x86

    • libvirt

  2. When using the libvirt driver, ensure that the CPU flag md-clear is exposed to the Nova instances. This can be done in one of three ways, depending on your configured CPU mode:

    1. libvirt.cpu_mode=host-model

      When using the host-model CPU mode, the md-clear CPU flag will be passed through to the Nova guests automatically.

      This mode is the default, when libvirt.virt_type=kvm|qemu is set in /etc/nova/nova-cpu.conf on compute nodes.

    2. libvirt.cpu_mode=host-passthrough

      When using the host-passthrough CPU mode, the md-clear CPU flag will be passed through to the Nova guests automatically.

    3. libvirt.cpu_mode=custom

      When using the custom CPU mode, you must explicitly enable the CPU flag md-clear to the Nova instances, in addition to the flags required for previous vulnerabilities, using the libvirt.cpu_model_extra_flags. For example:

      [libvirt]
      cpu_mode = custom
      cpu_models = IvyBridge
      cpu_model_extra_flags = spec-ctrl,ssbd,md-clear
      
  3. Reboot the compute node for the fixes to take effect.

    To minimize workload downtime, you may wish to live migrate all guests to another compute node first.

Once the above steps have been taken on every vulnerable compute node in the deployment, each running guest in the cluster must be fully powered down, and cold-booted (i.e. an explicit stop followed by a start), in order to activate the new CPU models. This can be done by the guest administrators at a time of their choosing.

Validation

After applying relevant updates, administrators can check the kernel’s sysfs interface to see what mitigation is in place, by running the following command on the host:

# cat /sys/devices/system/cpu/vulnerabilities/mds
Mitigation: Clear CPU buffers; SMT vulnerable

To unpack the message “Mitigation: Clear CPU buffers; SMT vulnerable”:

  • Mitigation: Clear CPU buffers means you have the “CPU buffer clearing” mitigation enabled, which is mechanism to invoke a flush of various exploitable CPU buffers by invoking a CPU instruction called “VERW”.

  • SMT vulnerable means, depending on your workload, you may still be vulnerable to SMT-related problems. You need to evaluate whether your workloads need SMT (also called “Hyper-Threading”) to be disabled or not. Refer to the guidance from your Linux distribution and processor vendor.

To see the other possible values for /sys/devices/system/cpu/vulnerabilities/mds, refer to the MDS system information section in Linux kernel’s documentation for MDS.

On the host, validate that KVM is capable of exposing the md-clear flag to guests:

# virsh domcapabilities kvm | grep md-clear
<feature policy='require' name='md-clear'/>

More information can be found on the ‘Diagnosis’ tab of this security notice document.

Performance Impact

Refer to this section titled “Performance Impact and Disabling MDS” from this security notice document, under the Resolve tab.

Note

Although the article referred to is from Red Hat, the findings and recommendations about performance impact apply for other distributions also.