[Bug 2093146] Re: KVM: Cache CPUID at KVM.ko module init to reduce latency of VM-Enter and VM-Exit

Sapphire Rapids Intel(R) Xeon(R) Platinum 8468
24.04 LTS Noble

6.8.0-51-generic

Outside VM
cpuid 3527
cpuid 3384
cpuid 3467
cpuid 3300
cpuid 3544

Inside VM
cpuid 68395
cpuid 58364
cpuid 65254
cpuid 68554
cpuid 66905

6.8.0-51-generic+TEST403286v20250107b1

Outside VM
cpuid 3253
cpuid 3416
cpuid 3447
cpuid 3260
cpuid 3281

Inside VM
cpuid 22852
cpuid 22890
cpuid 18168
cpuid 23462
cpuid 23281

--
You received this bug notification because you are subscribed to linux
in Ubuntu.
Matching subscriptions: Bgg, Bmail, Nb
https://bugs.launchpad.net/bugs/2093146

Title:
KVM: Cache CPUID at KVM.ko module init to reduce latency of VM-Enter
and VM-Exit

Status in linux package in Ubuntu:
Fix Released
Status in linux source package in Jammy:
In Progress
Status in linux source package in Noble:
In Progress
Status in linux source package in Oracular:
In Progress

Bug description:
BugLink: https://bugs.launchpad.net/bugs/2093146

[Impact]

The CPUID instruction is particularly slow on newer generation processors,
with Intel's Emerald Rapids processor taking significantly longer to execute
CPUID than Skylake or Icelake.

This introduces significant latency into the KVM subsystem, as it frequently
calls CPUID when recomputing XSTATE offsets, and especially XSAVE values, as
they need to call CPUID twice for each XSAVE call.

CPUID.0xD.[1..n] are constant and do not change during runtime, as they don't
depend on XCR0 or XSS values, and can be saved and cached for future usage.

By caching CPUID.0xD.[1..n] at kvm.ko module load, latency decreases by up to
400%.

For a round trip transition between VM-Enter and VM-Exit figures from the
commit log are:

Skylake 11650
Icelake 22350
Emerald 28850

When you add the caching in:

Skylake 6850
Icelake 9000
Emerald 7900

That's a saving of 170% for Skylake, 248% for Icelake and 365% for
Emerald Rapids.

[Fix]

The fix is part of a 5 patch series. We will only SRU patch 1 for the moment, as
it is the only one in mainline, and provides a 400% latency improvement, doing
the brunt of the work. Patches 2-5 are refactors and smaller performance
improvements, not yet mainlined due to needing rework, and only account for
about 2.5% latency improvement, quite small, compared to what patch 1 does.

The fix is:

commit 1201f226c863b7da739f7420ddba818cedf372fc
Author: Sean Christopherson <seanjc@google.com>
Date: Tue Dec 10 17:32:58 2024 -0800
Subject: KVM: x86: Cache CPUID.0xD XSTATE offsets+sizes during module init
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=1201f226c863b7da739f7420ddba818cedf372fc

This landed in 6.13-rc3, and is currently queued up for upstream -stable 6.12,
6.6 and 6.1.

This applies cleanly to noble, oracular. For jammy, it requires the below
dependency, and a small backport to fix some minor context mismatches.

commit cc04b6a21d431359eceeec0d812b492088b04af5
Author: Jing Liu <jing2.liu@intel.com>
Date: Wed Jan 5 04:35:14 2022 -0800
Subject: kvm: x86: Fix xstate_required_size() to follow XSTATE alignment rule
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cc04b6a21d431359eceeec0d812b492088b04af5

1) Install KVM Stack on Baremetal host
$ sudo apt-get install qemu-kvm libvirt-daemon-system libvirt-clients bridge-utils
2) Enable nested virt
$ vim /etc/modprobe.d/kvm.conf
options kvm-intel nested=1
$ sudo reboot
3) Start a VM.
4) In guest, run:
$ sudo apt-get install qemu-kvm libvirt-daemon-system libvirt-clients bridge-utils
5) In host and the guest, run:
$ sudo apt install build-essential
6) Install kvm-unit-tests and run x86/vmexit/cpuid testcase.
https://gitlab.com/kvm-unit-tests/kvm-unit-tests
$ git clone https://gitlab.com/kvm-unit-tests/kvm-unit-tests.git
$ ./configure
$ make standalone
$ cd tests
$ sudo -s
# ACCEL=kvm ./vmexit_cpuid
BUILD_HEAD=0ed2cdf3
timeout -k 1s --foreground 90s /usr/bin/qemu-system-x86_64 --no-reboot -nodefaults -device pc-testdev -device isa-debug-exit,iobase=0xf4,iosize=0x4 -vnc none -serial stdio -device pci-testdev -machine accel=kvm -kernel /tmp/tmp.GMVjItBglu -smp 1 -append cpuid # -initrd /tmp/tmp.uaD4VVyIqc
qemu-system-x86_64: warning: host doesn't support requested feature: CPUID.80000001H:ECX.svm [bit 2]
enabling apic
smp: waiting for 0 APs
paging enabled
cr0 = 80010011
cr3 = 1007000
cr4 = 20
pci-testdev at 0x10 membar febff000 iobar c000
cpuid 66485
PASS vmexit_cpuid

The numbers next to cpuid, is the (t2 = rdtsc) - (t1 = rdtsc) count. Smaller is
better.

A test kernel is available in the following ppa:

https://launchpad.net/~mruffell/+archive/ubuntu/sf403286-test

Test data for the above test kernel is:

Sapphire Rapids Intel(R) Xeon(R) Platinum 8468
24.04 LTS Noble

6.8.0-51-generic

Outside VM
cpuid 3527
cpuid 3384
cpuid 3467
cpuid 3300
cpuid 3544

Inside VM
cpuid 68395
cpuid 58364
cpuid 65254
cpuid 68554
cpuid 66905

6.8.0-51-generic+TEST403286v20250107b1

Outside VM
cpuid 3253
cpuid 3416
cpuid 3447
cpuid 3260
cpuid 3281

Inside VM
cpuid 22852
cpuid 22890
cpuid 18168
cpuid 23462
cpuid 23281

The number of cycles of rdtsc in a nested VM is of the same order of magnitude
smaller as we are expecting.

[Where problems could occur]

This fix is related to nested virtualisation in the KVM subsystem. We are adding
a new function, called on KVM module load, which caches the CPUID instead of
fetching it every time XSAVE needs to be recomputed, which can be multiple times
on VM-Enter and VM-Exit on nested guests.

CPUID is static and should never change, so there should be no issues in saving
a value and reusing it later.

If a regression were to occur, it would affect all KVM users, and there would
be no workarounds.

[Other info]

Full mailing list series:
https://lore.kernel.org/kvm/20241211013302.1347853-1-seanjc@google.com/T/#u

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2093146/+subscriptions