On 10/08/21 19:19, Sean Christopherson wrote:
The goal of this series is to drop the vmx->secondary_exec_control cache without degrading nested VM-Enter performance. The cache is effective, e.g. saves ~1000 cycles on nested VM-Enter, but confusing. The worst of the confusion could be eliminated by returning the computed value from vmx_compute_secondary_exec_control() to make the calls to the "compute" helper more like the other controls. But, the nested VM-Enter path would still have special handling for secondary exec controls, and ideally all controls would benefit from caching, though the benefits are marginal for other controls and thus difficult to justify. Happily, vmcs01 already caches the calculated controls in the controls_shadow. The only issue is that the controls_shadow may have dynamically toggled bits set. However, that is not a fundamental problem, it's simply different than what is expected by the nested VM-Enter code and is easily remedied. TL;DR: Get KVM's (L0's) desires for vmcs02 controls from vmcs01's controls_shadow instead of recalculating the desired controls on every nested VM-Enter, thus eliminating the need to have a dedicated cache for the secondary exec controls calulation. Sean Christopherson (4): KVM: VMX: Use current VMCS to query WAITPKG support for MSR emulation KVM: nVMX: Pull KVM L0's desired controls directly from vmcs01 KVM: VMX: Drop caching of KVM's desired sec exec controls for vmcs01 KVM: VMX: Hide VMCS control calculators in vmx.c arch/x86/kvm/vmx/nested.c | 25 ++++++++++++-------- arch/x86/kvm/vmx/vmx.c | 48 +++++++++++++++++++++++++++------------ arch/x86/kvm/vmx/vmx.h | 35 +++++----------------------- 3 files changed, 56 insertions(+), 52 deletions(-)
Queued, thanks (patch 1 for 5.14, the rest for 5.15). Paolo
