| 1 |
.. SPDX-License-Identifier: GPL-2.0 |
| 2 |
|
| 3 |
======================================= |
| 4 |
Known limitations of CPU virtualization |
| 5 |
======================================= |
| 6 |
|
| 7 |
Whenever perfect emulation of a CPU feature is impossible or too hard, KVM |
| 8 |
has to choose between not implementing the feature at all or introducing |
| 9 |
behavioral differences between virtual machines and bare metal systems. |
| 10 |
|
| 11 |
This file documents some of the known limitations that KVM has in |
| 12 |
virtualizing CPU features. |
| 13 |
|
| 14 |
x86 |
| 15 |
=== |
| 16 |
|
| 17 |
``KVM_GET_SUPPORTED_CPUID`` issues |
| 18 |
---------------------------------- |
| 19 |
|
| 20 |
x87 features |
| 21 |
~~~~~~~~~~~~ |
| 22 |
|
| 23 |
Unlike most other CPUID feature bits, CPUID[EAX=7,ECX=0]:EBX[6] |
| 24 |
(FDP_EXCPTN_ONLY) and CPUID[EAX=7,ECX=0]:EBX]13] (ZERO_FCS_FDS) are |
| 25 |
clear if the features are present and set if the features are not present. |
| 26 |
|
| 27 |
Clearing these bits in CPUID has no effect on the operation of the guest; |
| 28 |
if these bits are set on hardware, the features will not be present on |
| 29 |
any virtual machine that runs on that hardware. |
| 30 |
|
| 31 |
**Workaround:** It is recommended to always set these bits in guest CPUID. |
| 32 |
Note however that any software (e.g ``WIN87EM.DLL``) expecting these features |
| 33 |
to be present likely predates these CPUID feature bits, and therefore |
| 34 |
doesn't know to check for them anyway. |
| 35 |
|
| 36 |
``KVM_SET_VCPU_EVENTS`` issue |
| 37 |
----------------------------- |
| 38 |
|
| 39 |
Invalid KVM_SET_VCPU_EVENTS input with respect to error codes *may* result in |
| 40 |
failed VM-Entry on Intel CPUs. Pre-CET Intel CPUs require that exception |
| 41 |
injection through the VMCS correctly set the "error code valid" flag, e.g. |
| 42 |
require the flag be set when injecting a #GP, clear when injecting a #UD, |
| 43 |
clear when injecting a soft exception, etc. Intel CPUs that enumerate |
| 44 |
IA32_VMX_BASIC[56] as '1' relax VMX's consistency checks, and AMD CPUs have no |
| 45 |
restrictions whatsoever. KVM_SET_VCPU_EVENTS doesn't sanity check the vector |
| 46 |
versus "has_error_code", i.e. KVM's ABI follows AMD behavior. |
| 47 |
|
| 48 |
Nested virtualization features |
| 49 |
------------------------------ |
| 50 |
|
| 51 |
On AMD CPUs, when GIF is cleared, #DB exceptions or traps due to a breakpoint |
| 52 |
register match are ignored and discarded by the CPU. The CPU relies on the VMM |
| 53 |
to fully virtualize this behavior, even when vGIF is enabled for the guest |
| 54 |
(i.e. vGIF=0 does not cause the CPU to drop #DBs when the guest is running). |
| 55 |
KVM does not virtualize this behavior as the complexity is unjustified given |
| 56 |
the rarity of the use case. One way to handle this would be for KVM to |
| 57 |
intercept the #DB, temporarily disable the breakpoint, single-step over the |
| 58 |
instruction, then re-enable the breakpoint. |
| 59 |
|
| 60 |
x2APIC |
| 61 |
------ |
| 62 |
When KVM_X2APIC_API_USE_32BIT_IDS is enabled, KVM activates a hack/quirk that |
| 63 |
allows sending events to a single vCPU using its x2APIC ID even if the target |
| 64 |
vCPU has legacy xAPIC enabled, e.g. to bring up hotplugged vCPUs via INIT-SIPI |
| 65 |
on VMs with > 255 vCPUs. A side effect of the quirk is that, if multiple vCPUs |
| 66 |
have the same physical APIC ID, KVM will deliver events targeting that APIC ID |
| 67 |
only to the vCPU with the lowest vCPU ID. If KVM_X2APIC_API_USE_32BIT_IDS is |
| 68 |
not enabled, KVM follows x86 architecture when processing interrupts (all vCPUs |
| 69 |
matching the target APIC ID receive the interrupt). |
| 70 |
|
| 71 |
MTRRs |
| 72 |
----- |
| 73 |
KVM does not virtualize guest MTRR memory types. KVM emulates accesses to MTRR |
| 74 |
MSRs, i.e. {RD,WR}MSR in the guest will behave as expected, but KVM does not |
| 75 |
honor guest MTRRs when determining the effective memory type, and instead |
| 76 |
treats all of guest memory as having Writeback (WB) MTRRs. |
| 77 |
|
| 78 |
CR0.CD |
| 79 |
------ |
| 80 |
KVM does not virtualize CR0.CD on Intel CPUs. Similar to MTRR MSRs, KVM |
| 81 |
emulates CR0.CD accesses so that loads and stores from/to CR0 behave as |
| 82 |
expected, but setting CR0.CD=1 has no impact on the cachaeability of guest |
| 83 |
memory. |
| 84 |
|
| 85 |
Note, this erratum does not affect AMD CPUs, which fully virtualize CR0.CD in |
| 86 |
hardware, i.e. put the CPU caches into "no fill" mode when CR0.CD=1, even when |
| 87 |
running in the guest. |