Documentation/virt/kvm/x86/errata.rst

.. SPDX-License-Identifier: GPL-2.0

=======================================
Known limitations of CPU virtualization
=======================================

Whenever perfect emulation of a CPU feature is impossible or too hard, KVM
has to choose between not implementing the feature at all or introducing
behavioral differences between virtual machines and bare metal systems.

This file documents some of the known limitations that KVM has in
virtualizing CPU features.

x86
===

``KVM_GET_SUPPORTED_CPUID`` issues
----------------------------------

x87 features
~~~~~~~~~~~~

Unlike most other CPUID feature bits, CPUID[EAX=7,ECX=0]:EBX[6]
(FDP_EXCPTN_ONLY) and CPUID[EAX=7,ECX=0]:EBX]13] (ZERO_FCS_FDS) are
clear if the features are present and set if the features are not present.

Clearing these bits in CPUID has no effect on the operation of the guest;
if these bits are set on hardware, the features will not be present on
any virtual machine that runs on that hardware.

**Workaround:** It is recommended to always set these bits in guest CPUID.
Note however that any software (e.g ``WIN87EM.DLL``) expecting these features
to be present likely predates these CPUID feature bits, and therefore
doesn't know to check for them anyway.

``KVM_SET_VCPU_EVENTS`` issue
-----------------------------

Invalid KVM_SET_VCPU_EVENTS input with respect to error codes *may* result in
failed VM-Entry on Intel CPUs.  Pre-CET Intel CPUs require that exception
injection through the VMCS correctly set the "error code valid" flag, e.g.
require the flag be set when injecting a #GP, clear when injecting a #UD,
clear when injecting a soft exception, etc.  Intel CPUs that enumerate
IA32_VMX_BASIC[56] as '1' relax VMX's consistency checks, and AMD CPUs have no
restrictions whatsoever.  KVM_SET_VCPU_EVENTS doesn't sanity check the vector
versus "has_error_code", i.e. KVM's ABI follows AMD behavior.

Nested virtualization features
------------------------------

On AMD CPUs, when GIF is cleared, #DB exceptions or traps due to a breakpoint
register match are ignored and discarded by the CPU. The CPU relies on the VMM
to fully virtualize this behavior, even when vGIF is enabled for the guest
(i.e. vGIF=0 does not cause the CPU to drop #DBs when the guest is running).
KVM does not virtualize this behavior as the complexity is unjustified given
the rarity of the use case. One way to handle this would be for KVM to
intercept the #DB, temporarily disable the breakpoint, single-step over the
instruction, then re-enable the breakpoint.

x2APIC
------
When KVM_X2APIC_API_USE_32BIT_IDS is enabled, KVM activates a hack/quirk that
allows sending events to a single vCPU using its x2APIC ID even if the target
vCPU has legacy xAPIC enabled, e.g. to bring up hotplugged vCPUs via INIT-SIPI
on VMs with > 255 vCPUs.  A side effect of the quirk is that, if multiple vCPUs
have the same physical APIC ID, KVM will deliver events targeting that APIC ID
only to the vCPU with the lowest vCPU ID.  If KVM_X2APIC_API_USE_32BIT_IDS is
not enabled, KVM follows x86 architecture when processing interrupts (all vCPUs
matching the target APIC ID receive the interrupt).

MTRRs
-----
KVM does not virtualize guest MTRR memory types.  KVM emulates accesses to MTRR
MSRs, i.e. {RD,WR}MSR in the guest will behave as expected, but KVM does not
honor guest MTRRs when determining the effective memory type, and instead
treats all of guest memory as having Writeback (WB) MTRRs.

CR0.CD
------
KVM does not virtualize CR0.CD on Intel CPUs.  Similar to MTRR MSRs, KVM
emulates CR0.CD accesses so that loads and stores from/to CR0 behave as
expected, but setting CR0.CD=1 has no impact on the cachaeability of guest
memory.

Note, this erratum does not affect AMD CPUs, which fully virtualize CR0.CD in
hardware, i.e. put the CPU caches into "no fill" mode when CR0.CD=1, even when
running in the guest.