On Tue, 10 Dec 2024 17:32:57 -0800, Sean Christopherson wrote:
> Fix a hilarious/revolting performance regression (relative to older CPU
> generations) in xstate_required_size() that pops up due to CPUID _in the
> host_ taking 3x-4x longer on Emerald Rapids than Skylake.
>
> The issue rears its head on nested virtualization transitions, as KVM
> (unnecessarily) performs runtime CPUID updates, including XSAVE sizes,
> multiple times per transition. And calculating XSAVE sizes, especially
> for vCPUs with a decent number of supported XSAVE features and compacted
> format support, can add up to thousands of cycles.
>
> [...]
Applied 2-5 to kvm-x86 misc, with a changelog that doesn't incorrectly state
that CPUID is a mandatory intercept on AMD.
[1/5] KVM: x86: Cache CPUID.0xD XSTATE offsets+sizes during module init
(no commit info)
[2/5] KVM: x86: Use for-loop to iterate over XSTATE size entries
https://github.com/kvm-x86/linux/commit/aa93b6f96f64
[3/5] KVM: x86: Apply TSX_CTRL_CPUID_CLEAR if and only if the vCPU has RTM or HLE
https://github.com/kvm-x86/linux/commit/7e9f735e7ac4
[4/5] KVM: x86: Query X86_FEATURE_MWAIT iff userspace owns the CPUID feature bit
https://github.com/kvm-x86/linux/commit/a487f6797c88
[5/5] KVM: x86: Defer runtime updates of dynamic CPUID bits until CPUID emulation
https://github.com/kvm-x86/linux/commit/93da6af3ae56
--
https://github.com/kvm-x86/linux/tree/next
