@@ -38,6 +38,7 @@ Supported mechanisms
Currently supported confidential guest mechanisms are:
* AMD Secure Encrypted Virtualization (SEV) (see :doc:`i386/amd-memory-encryption`)
+* Intel Trust Domain Extension (TDX) (see :doc:`i386/tdx`)
* POWER Protected Execution Facility (PEF) (see :ref:`power-papr-protected-execution-facility-pef`)
* s390x Protected Virtualization (PV) (see :doc:`s390x/protvirt`)
new file mode 100644
@@ -0,0 +1,143 @@
+Intel Trusted Domain eXtension (TDX)
+====================================
+
+Intel Trusted Domain eXtensions (TDX) refers to an Intel technology that extends
+Virtual Machine Extensions (VMX) and Multi-Key Total Memory Encryption (MKTME)
+with a new kind of virtual machine guest called a Trust Domain (TD). A TD runs
+in a CPU mode that is designed to protect the confidentiality of its memory
+contents and its CPU state from any other software, including the hosting
+Virtual Machine Monitor (VMM), unless explicitly shared by the TD itself.
+
+Prerequisites
+-------------
+
+To run TD, the physical machine needs to have TDX module loaded and initialized
+while KVM hypervisor has TDX support and has TDX enabled. If those requirements
+are met, the ``KVM_CAP_VM_TYPES`` will report the support of ``KVM_X86_TDX_VM``.
+
+Trust Domain Virtual Firmware (TDVF)
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Trust Domain Virtual Firmware (TDVF) is required to provide TD services to boot
+TD Guest OS. TDVF needs to be copied to guest private memory and measured before
+the TD boots.
+
+KVM vcpu ioctl ``KVM_MEMORY_MAPPING`` can be used to populates the TDVF content
+into its private memory.
+
+Since TDX doesn't support readonly memslot, TDVF cannot be mapped as pflash
+device and it actually works as RAM. "-bios" option is chosen to load TDVF.
+
+OVMF is the opensource firmware that implements the TDVF support. Thus the
+command line to specify and load TDVF is ``-bios OVMF.fd``
+
+KVM private memory
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+TD's memory (RAM) needs to be able to be transformed between private and shared.
+Its BIOS (OVMF/TDVF) needs to be mapped as private as well. Thus QEMU needs to
+allocate private guest memfd for them via KVM's IOCTL (KVM_CREATE_GUEST_MEMFD),
+which requires KVM is newer enough that reports KVM_CAP_GUEST_MEMFD.
+
+Feature Control
+---------------
+
+Unlike non-TDX VM, the CPU features (enumerated by CPU or MSR) of a TD is not
+under full control of VMM. VMM can only configure part of features of a TD on
+``KVM_TDX_INIT_VM`` command of VM scope ``MEMORY_ENCRYPT_OP`` ioctl.
+
+The configurable features have three types:
+
+- Attributes:
+ - PKS (bit 30) controls whether Supervisor Protection Keys is exposed to TD,
+ which determines related CPUID bit and CR4 bit;
+ - PERFMON (bit 63) controls whether PMU is exposed to TD.
+
+- XSAVE related features (XFAM):
+ XFAM is a 64b mask, which has the same format as XCR0 or IA32_XSS MSR. It
+ determines the set of extended features available for use by the guest TD.
+
+- CPUID features:
+ Only some bits of some CPUID leaves are directly configurable by VMM.
+
+What features can be configured is reported via TDX capabilities.
+
+TDX capabilities
+~~~~~~~~~~~~~~~~
+
+The VM scope ``MEMORY_ENCRYPT_OP`` ioctl provides command ``KVM_TDX_CAPABILITIES``
+to get the TDX capabilities from KVM. It returns a data structure of
+``struct kvm_tdx_capabilites``, which tells the supported configuration of
+attributes, XFAM and CPUIDs.
+
+TD attestation
+--------------
+
+In TD guest, the attestation process is used to verify the TDX guest
+trustworthiness to other entities before provisioning secrets to the guest.
+
+TD attestation is initiated first by calling TDG.MR.REPORT inside TD to get the
+REPORT. Then the REPORT data needs to be converted into a remotely verifiable
+Quote by SGX Quoting Enclave (QE).
+
+A host daemon, Quote Generation Service (QGS), provides the functionality of
+SGX GE. It provides a socket address, to which a TD guest can connect via
+"quote-generation-socket" property. On the request of <GETQUOTE> from TD guest,
+QEMU sends the TDREPORT to QGS via "quote-generation-socket" socket, and gets
+the returning Quoting and return it back to TD guest.
+
+Though "quote-generation-socket" is optional for booting the TD guest, it's a
+must for supporting TD guest atteatation.
+
+Launching a TD (TDX VM)
+-----------------------
+
+To launch a TDX guest, below are new added and required:
+
+.. parsed-literal::
+
+ |qemu_system_x86| \\
+ -object tdx-guest,id=tdx0 \\
+ -machine ...,kernel-irqchip=split,confidential-guest-support=tdx0 \\
+ -bios OVMF.fd \\
+
+If TD attestation support is wanted:
+
+.. parsed-literal::
+
+ |qemu_system_x86| \\
+ -object '{"qom-type":"tdx-guest","id":"tdx0","quote-generation-socket":{"type": "vsock", "cid":"1","port":"1234"}}' \\
+ -machine ...,kernel-irqchip=split,confidential-guest-support=tdx0 \\
+ -bios OVMF.fd \\
+
+Debugging
+---------
+
+Bit 0 of TD attributes, is DEBUG bit, which decides if the TD runs in off-TD
+debug mode. When in off-TD debug mode, TD's VCPU state and private memory are
+accessible via given SEAMCALLs. This requires KVM to expose APIs to invoke those
+SEAMCALLs and resonponding QEMU change.
+
+It's targeted as future work.
+
+restrictions
+------------
+
+ - kernel-irqchip must be split;
+
+ - No readonly support for private memory;
+
+ - No SMM support: SMM support requires manipulating the guset register states
+ which is not allowed;
+
+Live Migration
+--------------
+
+TODO
+
+References
+----------
+
+- `TDX Homepage <https://www.intel.com/content/www/us/en/developer/articles/technical/intel-trust-domain-extensions.html>`__
+
+- `SGX QE <https://github.com/intel/SGXDataCenterAttestationPrimitives/tree/master/QuoteGeneration>`__
@@ -29,6 +29,7 @@ Architectural features
i386/kvm-pv
i386/sgx
i386/amd-memory-encryption
+ i386/tdx
OS requirements
~~~~~~~~~~~~~~~
Add docs/system/i386/tdx.rst for TDX support, and add tdx in confidential-guest-support.rst Signed-off-by: Xiaoyao Li <xiaoyao.li@intel.com> --- Changes in v5: - Add TD attestation section and update the QEMU parameter; Changes since v1: - Add prerequisite of private gmem; - update example command to launch TD; Changes since RFC v4: - add the restriction that kernel-irqchip must be split --- docs/system/confidential-guest-support.rst | 1 + docs/system/i386/tdx.rst | 143 +++++++++++++++++++++ docs/system/target-i386.rst | 1 + 3 files changed, 145 insertions(+) create mode 100644 docs/system/i386/tdx.rst