qemu-e2k/docs/hyperv.txt

215 lines
9.3 KiB
Plaintext
Raw Normal View History

Hyper-V Enlightenments
======================
1. Description
===============
In some cases when implementing a hardware interface in software is slow, KVM
implements its own paravirtualized interfaces. This works well for Linux as
guest support for such features is added simultaneously with the feature itself.
It may, however, be hard-to-impossible to add support for these interfaces to
proprietary OSes, namely, Microsoft Windows.
KVM on x86 implements Hyper-V Enlightenments for Windows guests. These features
make Windows and Hyper-V guests think they're running on top of a Hyper-V
compatible hypervisor and use Hyper-V specific features.
2. Setup
=========
No Hyper-V enlightenments are enabled by default by either KVM or QEMU. In
QEMU, individual enlightenments can be enabled through CPU flags, e.g:
qemu-system-x86_64 --enable-kvm --cpu host,hv_relaxed,hv_vpindex,hv_time, ...
Sometimes there are dependencies between enlightenments, QEMU is supposed to
check that the supplied configuration is sane.
When any set of the Hyper-V enlightenments is enabled, QEMU changes hypervisor
identification (CPUID 0x40000000..0x4000000A) to Hyper-V. KVM identification
and features are kept in leaves 0x40000100..0x40000101.
3. Existing enlightenments
===========================
3.1. hv-relaxed
================
This feature tells guest OS to disable watchdog timeouts as it is running on a
hypervisor. It is known that some Windows versions will do this even when they
see 'hypervisor' CPU flag.
3.2. hv-vapic
==============
Provides so-called VP Assist page MSR to guest allowing it to work with APIC
more efficiently. In particular, this enlightenment allows paravirtualized
(exit-less) EOI processing.
3.3. hv-spinlocks=xxx
======================
Enables paravirtualized spinlocks. The parameter indicates how many times
spinlock acquisition should be attempted before indicating the situation to the
hypervisor. A special value 0xffffffff indicates "never to retry".
3.4. hv-vpindex
================
Provides HV_X64_MSR_VP_INDEX (0x40000002) MSR to the guest which has Virtual
processor index information. This enlightenment makes sense in conjunction with
hv-synic, hv-stimer and other enlightenments which require the guest to know its
Virtual Processor indices (e.g. when VP index needs to be passed in a
hypercall).
3.5. hv-runtime
================
Provides HV_X64_MSR_VP_RUNTIME (0x40000010) MSR to the guest. The MSR keeps the
virtual processor run time in 100ns units. This gives guest operating system an
idea of how much time was 'stolen' from it (when the virtual CPU was preempted
to perform some other work).
3.6. hv-crash
==============
Provides HV_X64_MSR_CRASH_P0..HV_X64_MSR_CRASH_P5 (0x40000100..0x40000105) and
HV_X64_MSR_CRASH_CTL (0x40000105) MSRs to the guest. These MSRs are written to
by the guest when it crashes, HV_X64_MSR_CRASH_P0..HV_X64_MSR_CRASH_P5 MSRs
contain additional crash information. This information is outputted in QEMU log
and through QAPI.
Note: unlike under genuine Hyper-V, write to HV_X64_MSR_CRASH_CTL causes guest
to shutdown. This effectively blocks crash dump generation by Windows.
3.7. hv-time
=============
Enables two Hyper-V-specific clocksources available to the guest: MSR-based
Hyper-V clocksource (HV_X64_MSR_TIME_REF_COUNT, 0x40000020) and Reference TSC
page (enabled via MSR HV_X64_MSR_REFERENCE_TSC, 0x40000021). Both clocksources
are per-guest, Reference TSC page clocksource allows for exit-less time stamp
readings. Using this enlightenment leads to significant speedup of all timestamp
related operations.
3.8. hv-synic
==============
Enables Hyper-V Synthetic interrupt controller - an extension of a local APIC.
When enabled, this enlightenment provides additional communication facilities
to the guest: SynIC messages and Events. This is a pre-requisite for
implementing VMBus devices (not yet in QEMU). Additionally, this enlightenment
is needed to enable Hyper-V synthetic timers. SynIC is controlled through MSRs
HV_X64_MSR_SCONTROL..HV_X64_MSR_EOM (0x40000080..0x40000084) and
HV_X64_MSR_SINT0..HV_X64_MSR_SINT15 (0x40000090..0x4000009F)
Requires: hv-vpindex
3.9. hv-stimer
===============
Enables Hyper-V synthetic timers. There are four synthetic timers per virtual
CPU controlled through HV_X64_MSR_STIMER0_CONFIG..HV_X64_MSR_STIMER3_COUNT
(0x400000B0..0x400000B7) MSRs. These timers can work either in single-shot or
periodic mode. It is known that certain Windows versions revert to using HPET
(or even RTC when HPET is unavailable) extensively when this enlightenment is
not provided; this can lead to significant CPU consumption, even when virtual
CPU is idle.
Requires: hv-vpindex, hv-synic, hv-time
3.10. hv-tlbflush
==================
Enables paravirtualized TLB shoot-down mechanism. On x86 architecture, remote
TLB flush procedure requires sending IPIs and waiting for other CPUs to perform
local TLB flush. In virtualized environment some virtual CPUs may not even be
scheduled at the time of the call and may not require flushing (or, flushing
may be postponed until the virtual CPU is scheduled). hv-tlbflush enlightenment
implements TLB shoot-down through hypervisor enabling the optimization.
Requires: hv-vpindex
3.11. hv-ipi
=============
Enables paravirtualized IPI send mechanism. HvCallSendSyntheticClusterIpi
hypercall may target more than 64 virtual CPUs simultaneously, doing the same
through APIC requires more than one access (and thus exit to the hypervisor).
Requires: hv-vpindex
3.12. hv-vendor-id=xxx
=======================
This changes Hyper-V identification in CPUID 0x40000000.EBX-EDX from the default
"Microsoft Hv". The parameter should be no longer than 12 characters. According
to the specification, guests shouldn't use this information and it is unknown
if there is a Windows version which acts differently.
Note: hv-vendor-id is not an enlightenment and thus doesn't enable Hyper-V
identification when specified without some other enlightenment.
3.13. hv-reset
===============
Provides HV_X64_MSR_RESET (0x40000003) MSR to the guest allowing it to reset
itself by writing to it. Even when this MSR is enabled, it is not a recommended
way for Windows to perform system reboot and thus it may not be used.
3.14. hv-frequencies
============================================
Provides HV_X64_MSR_TSC_FREQUENCY (0x40000022) and HV_X64_MSR_APIC_FREQUENCY
(0x40000023) allowing the guest to get its TSC/APIC frequencies without doing
measurements.
3.15 hv-reenlightenment
========================
The enlightenment is nested specific, it targets Hyper-V on KVM guests. When
enabled, it provides HV_X64_MSR_REENLIGHTENMENT_CONTROL (0x40000106),
HV_X64_MSR_TSC_EMULATION_CONTROL (0x40000107)and HV_X64_MSR_TSC_EMULATION_STATUS
(0x40000108) MSRs allowing the guest to get notified when TSC frequency changes
(only happens on migration) and keep using old frequency (through emulation in
the hypervisor) until it is ready to switch to the new one. This, in conjunction
with hv-frequencies, allows Hyper-V on KVM to pass stable clocksource (Reference
TSC page) to its own guests.
Recommended: hv-frequencies
3.16. hv-evmcs
===============
The enlightenment is nested specific, it targets Hyper-V on KVM guests. When
enabled, it provides Enlightened VMCS feature to the guest. The feature
implements paravirtualized protocol between L0 (KVM) and L1 (Hyper-V)
hypervisors making L2 exits to the hypervisor faster. The feature is Intel-only.
Note: some virtualization features (e.g. Posted Interrupts) are disabled when
hv-evmcs is enabled. It may make sense to measure your nested workload with and
without the feature to find out if enabling it is beneficial.
Requires: hv-vapic
3.17. hv-stimer-direct
=======================
Hyper-V specification allows synthetic timer operation in two modes: "classic",
when expiration event is delivered as SynIC message and "direct", when the event
is delivered via normal interrupt. It is known that nested Hyper-V can only
use synthetic timers in direct mode and thus 'hv-stimer-direct' needs to be
enabled.
Requires: hv-vpindex, hv-synic, hv-time, hv-stimer
3.17. hv-no-nonarch-coresharing=on/off/auto
===========================================
This enlightenment tells guest OS that virtual processors will never share a
physical core unless they are reported as sibling SMT threads. This information
is required by Windows and Hyper-V guests to properly mitigate SMT related CPU
vulnerabilities.
When the option is set to 'auto' QEMU will enable the feature only when KVM
reports that non-architectural coresharing is impossible, this means that
hyper-threading is not supported or completely disabled on the host. This
setting also prevents migration as SMT settings on the destination may differ.
When the option is set to 'on' QEMU will always enable the feature, regardless
of host setup. To keep guests secure, this can only be used in conjunction with
exposing correct vCPU topology and vCPU pinning.
4. Development features
========================
In some cases (e.g. during development) it may make sense to use QEMU in
'pass-through' mode and give Windows guests all enlightenments currently
supported by KVM. This pass-through mode is enabled by "hv-passthrough" CPU
flag.
Note: enabling this flag effectively prevents migration as supported features
may differ between target and destination.
4. Useful links
================
Hyper-V Top Level Functional specification and other information:
https://github.com/MicrosoftDocs/Virtualization-Documentation