Hello,
I have implemented support for exposing and enabling APERF/MPERF
MSR passthrough for x86 QEMU guests when running on KVM.
As of last year, the Linux kernel supports providing a capability
to disable APERF/MPERF read intercepts [1]. However, there is
currently no native way in QEMU to instruct KVM to activate this
capability or expose the feature bit cleanly via CPUID.
My patch introduces the `aperfmperf` feature flag via `FEAT_6_ECX`
(CPUID.06H:ECX[bit 0]). It ties into the existing power management
framework, so when a user enables power management via
`-overcommit cpu-pm=on` and passes `+aperfmperf` to the CPU, QEMU
requests KVM to drop the MSR intercepts.
For testing, I passed the feature via Libvirt XML like so:
<qemu:arg value="-overcommit"/>
<qemu:arg value="cpu-pm=on"/>
<qemu:arg value="-cpu"/>
<qemu:arg value="+aperfmperf,+invtsc"/>
A FreeBSD 16 guest is successfully able to read the changing
values of the MSRs directly without exiting to the hypervisor:
root@freebsd16development:/home/user # cpucontrol -i 6 /dev/cpuctl0
cpuid level 0x6: 0x00000004 0x00000000 0x00000001 0x00000000
root@freebsd16development:/home/user # cpucontrol -m 0xe7 /dev/cpuctl0
MSR 0xe7: 0x00007fdf 0x22480f90
root@freebsd16development:/home/user # cpucontrol -m 0xe7 /dev/cpuctl0
MSR 0xe7: 0x00007fdf 0x2b8b48b0
root@freebsd16development:/home/user # cpucontrol -m 0xe8 /dev/cpuctl0
MSR 0xe8: 0x000050de 0x49270831
root@freebsd16development:/home/user # cpucontrol -m 0xe8 /dev/cpuctl0
MSR 0xe8: 0x000050de 0x4ad66ae6
root@freebsd16development:/home/user #
I am looking for feedback on the approach, specifically if tying
this directly into `cpu-pm` makes the most sense to the
maintainers.
Based-on: QEMU v11.0.50 (commit 5611a9268d)
Anderson Nascimento (1):
target/i386: Add support for KVM APERF/MPERF passthrough
target/i386/cpu.c | 17 ++++++++++++++++-
target/i386/cpu.h | 2 ++
target/i386/kvm/kvm.c | 5 ++++-
3 files changed, 22 insertions(+), 2 deletions(-)
--
2.54.0
