On 6/3/23 08:22, Chang S. Bae wrote:
> Document the overview of the feature along with relevant consideration
> when provisioning dm-crypt volumes with AES-KL instead of AES-NI.
>
> ---
> ---
> Documentation/arch/x86/index.rst | 1 +
> Documentation/arch/x86/keylocker.rst | 97 ++++++++++++++++++++++++++++
> 2 files changed, 98 insertions(+)
> create mode 100644 Documentation/arch/x86/keylocker.rst
>
> diff --git a/Documentation/arch/x86/keylocker.rst b/Documentation/arch/x86/keylocker.rst
> new file mode 100644
> index 000000000000..5557b8d0659a
> --- /dev/null
> +++ b/Documentation/arch/x86/keylocker.rst
> @@ -0,0 +1,97 @@
> +.. SPDX-License-Identifier: GPL-2.0
> +
> +==============
> +x86 Key Locker
> +==============
> +
> +Introduction
> +============
> +
> +Key Locker is a CPU feature to reduce key exfiltration opportunities
> +while maintaining a programming interface similar to AES-NI. It
> +converts the AES key into an encoded form, called the 'key handle'.
> +The key handle is a wrapped version of the clear-text key where the
> +wrapping key has limited exposure. Once converted, all subsequent data
> +encryption using new AES instructions (AES-KL) uses this key handle,
> +reducing the exposure of private key material in memory.
> +
> +CPU-internal Wrapping Key
> +=========================
> +
> +The CPU-internal wrapping key is an entity in a software-invisible CPU
> +state. On every system boot, a new key is loaded. So the key handle that
> +was encoded by the old wrapping key is no longer usable on system shutdown
> +or reboot.
> +
> +And the key may be lost on the following exceptional situation upon wakeup:
> +
> +Wrapping Key Restore Failure
> +----------------------------
> +
> +The CPU state is volatile with the ACPI S3/4 sleep states. When the system
> +supports those states, the key has to be backed up so that it is restored
> +on wake up. The kernel saves the key in non-volatile media.
> +
> +The event of a wrapping key restore failure upon resume from suspend, all
Upon the event of a ...
> +established key handles become invalid. In flight dm-crypt operations
> +receive error results from pending operations. In the likely scenario that
> +dm-crypt is hosting the root filesystem the recovery is identical to if a
> +storage controller failed to resume from suspend, reboot. If the volume
> +impacted by a wrapping key restore failure is a data-volume then it is
data volume
> +possible that I/O errors on that volume do not bring down the rest of the
> +system. However, a reboot is still required because the kernel will have
> +soft-disabled Key Locker. Upon the failure, the crypto library code will
> +return -ENODEV on every AES-KL function call. The Key Locker implementation
> +only loads a new wrapping key at initial boot, not any time after like
> +resume from suspend.
> +
> +Use Case and Non-use Cases
> +==========================
> +
> +Bare metal disk encryption is the only intended use case.
> +
> +Userspace usage is not supported because there is no ABI provided to
> +communicate and coordinate wrapping-key restore failure to userspace. For
> +now, key restore failures are only coordinated with kernel users. But the
> +kernel can not prevent userspace from using the feature's AES instructions
> +('AES-KL') when the feature has been enabled. So, the lack of userspace
> +support is only documented, not actively enforced.
> +
> +Key Locker is not expected to be advertised to guest VMs and the kernel
> +implementation ignores it even if the VMM enumerates the capability. The
> +expectation is that a guest VM wants private wrapping key state, but the
> +architecture does not provide that. An emulation of that capability, by
> +caching per-VM wrapping keys in memory, defeats the purpose of Key Locker.
> +The backup / restore facility is also not performant enough to be suitable
> +for guest VM context switches.
> +
> +AES Instruction Set
> +===================
> +
> +The feature accompanies a new AES instruction set. This instruction set is
> +analogous to AES-NI. A set of AES-NI instructions can be mapped to an
> +AES-KL instruction. For example, AESENC128KL is responsible for ten rounds
> +of transformation, which is equivalent to nine times AESENC and one
> +AESENCLAST in AES-NI.
> +
> +But they have some notable differences:
> +
> +* AES-KL provides a secure data transformation using an encrypted key.
> +
> +* If an invalid key handle is provided, e.g. a corrupted one or a handle
> + restriction failure, the instruction fails with setting RFLAGS.ZF. The
> + crypto library implementation includes the flag check to return -EINVAL.
> + Note that this flag is also set if the wrapping key is changed, e.g.,
> + because of the backup error.
> +
> +* AES-KL implements support for 128-bit and 256-bit keys, but there is no
> + AES-KL instruction to process an 192-bit key. The AES-KL cipher
> + implementation logs a warning message with a 192-bit key and then falls
> + back to AES-NI. So, this 192-bit key-size limitation is only documented,
Is it logged anywhere? i.e., a kernel log message?
> + not enforced. It means the key will remain in clear-text in memory. This
> + is to meet Linux crypto-cipher expectation that each implementation must
> + support all the AES-compliant key sizes.
> +
> +* Some AES-KL hardware implementation may have noticeable performance
> + overhead when compared with AES-NI instructions.
> +
--
~Randy
