Add support for emulated virtual TPM

Include the URL of your launchpad blueprint:

There are a class of applications which expect to use a TPM device to store secrets. In order to run these applications in a virtual machine, it would be useful to expose a virtual TPM device within the guest. Accordingly, the suggestion is to add a placement trait which could be requested in the flavor or image which would cause such a device to be added to the VM by the relevent virt driver.

Problem description

Currently there is no way to create virtual machines within nova that provide a virtual TPM device to the guest.

Use Cases

Support the virtualizing of existing applications and operating systems which expect to make use of physical TPM devices. At least one hypervisor (libvirt/qemu) currently supports the creation of an emulated TPM device which is associated with a per-VM “swtpm” process on the host, but there is no way to tell nova to enable it.

Proposed change

In recent libvirt and qemu (and possibly other hypervisors as well) there is support for an emulated vTPM device. We propose to modify nova to make use of this capability.

For the libvirt virt driver in particular, there is support for vTPM as of libvirt 4.5. The desired libvirt XML arguments are something like this:

  <tpm model='tpm-tis'>
    <backend type='emulator' version='2.0'>

Support for emulated TPM requires qemu 2.11 at a minimum, though qemu 2.12 is recommended by the author. The virt driver code should add suitable version checks (in the case of LibvirtDriver, this would include checks for both libvirt and qemu). Currently emulated TPM is only supported for x86, though this is an implementation detail rather than an architectural limitation.

Support for emulated TPM also requires the “swtpm” binary and libraries to be available on the host. If there is no way to check whether this is available from the hypervisor, we may need to add a hypervisor-specific nova.conf flag indicating that we want to enable emulated TPM support. This would presumably default to false for minimal surprise on upgrades.

In order to request this functionality (and to allow scheduling to nodes that provide this functionality) we propose to define two new traits, COMPUTE_SECURITY_TPM_1_2 and COMPUTE_SECURITY_TPM_2_0. (The emulated TPM is just a process running on the host, so the concept of inventory doesn’t apply.) The two traits represent the two different versions of the TPM spec that are currently supported. (A summary of the differences between the two versions is currently available here.) The flavor extra-specs or image properties could then specify something like trait:COMPUTE_SECURITY_TPM_1_2=required to indicate that they wish to have access to a TPM. Virt drivers which could provide a TPM to their instances would be responsible for setting either (or both) of the two traits on the compute nodes. If an instance has specified one of the traits in the flavor or image, the virt driver will do whatever is needed to provide a TPM to the instance. If for any reason this is not possible, the instance creation will fail.

When using COMPUTE_SECURITY_TPM_2_0, there are two possible device models for the emulated TPM device, TIS and CRB. By default the TIS model will be used, but it can also be explicitly specified by setting hw:tpm_model=TIS in the image or hw_tpm_model=TIS in the image properties. The CRB option can be specified by setting hw:tpm_model=CRB in the flavor (or via the equivalent image property). In the case of libvirt/qemu, the version of libvirt that supports TPM 2.0 (v4.5.0) also supports the CRB device model.

If both the flavor and the image specify a TPM trait or device model and the two values do not match, an exception will be raised. If the CRB model is specified with COMPUTE_SECURITY_TPM_1_2 the hypervisor will fail to create the instance.

As a future enhancement beyond the scope of the immediate work, it would be possible to extend this to support physical TPM passthrough. In this case the virt driver would also advertise an inventory with a resource class of PTPM with total=1 (since current hardware only has a single TPM), and the image or flavor could request it by specifying resources:PTPM=1. The trait would not be necessary in this case, as the desire for a TPM in the instance is implied by the resource request. Also, for TPM passthrough the device model is controlled by the actual hardware device.

As part of implementing the this feature, the nova cold migration code will need to copy over the directory containing the emulated TPM files. For libvirt this would mean copying the file under /var/lib/libvirt/swtpm/<instance> from within LibvirtDriver.migrate_disk_and_power_off().

Shelve/unshelve could be supported by saving the persistent TPM data as a glance image during the shelve operation, and recreating it (and deleting the image) during unshelve.

Resizing will result in a reschedule, so shouldn’t be a problem. If the admin resizes from a flavor with TPM to a flavor without TPM nova won’t care, but it might cause problems in the guest.

Rebuilding to a new image is problematic if the new image specifies a TPM trait and the current host cannot provide TPM support. This will cause the rebuild to fail. In this case, the user would need to rebuild with a suitable image.

It should be noted that if a compute node goes down and the VM has to be rebuilt on another compute node then we’re going to lose any emulated TPM data. In the shared-storage case this is exactly analogous to taking the hard drive out of one physical machine and putting it into another physical machine.


Rather than using a trait, we could instead use a resource with a large inventory.

Data model impact


REST API impact


Security impact

The guest will be able to use the emulated TPM for all the security enhancing functionality that a physical TPM provides, in order to protect itself against attacks from within the guest. The guest must still trust the host.

Notifications impact


Other end user impact

There are no immediate plans to make emulated TPM work over shelve/unshelve. To make this work reliably would require saving the persistent TPM data file to a glance image or swift object on “shelve” and then recover the data on “unshelve”.

Instances which use UEFI NVRAM are currently in a similar position, as the NVRAM is not persisted over shelve/unshelve.

Performance Impact


Other deployer impact


Developer impact

The various virt drivers would be able to implement the emulated vTPM as desired.

Upgrade impact

If a config option is needed to opt-in to emulated TPM support, the operator would need to set the config option appropriately after an upgrade.



Primary assignee:


Other contributors:


Work Items

  • Support for new placement traits

  • Libvirt driver changes to report traits to placement

  • Libvirt driver changes to enable specifying libvirt XML

  • Libvirt driver changes to copy vTPM files on cold migration.


  • Up-to-date qemu/libvirt

  • “swtpm” binary and libraries


Unit and functional testing will be added.

Documentation Impact

Operations Guide and End User Guide will be updated appropriately. Feature support matrix will be updated.




Release Name