Currently, when we switch from L1 to L2 we do the following: - Re-initialize L1 MMU as shadow EPT MMU (nested_ept_init_mmu_context()) - Re-initialize 'nested' MMU (nested_vmx_load_cr3() -> init_kvm_nested_mmu()) - Reload MMU root upon guest entry. When we switch back we do: - Re-initialize L1 MMU (nested_vmx_load_cr3() -> init_kvm_tdp_mmu()) - Reload MMU root upon guest entry. This seems to be sub-optimal. Initializing MMU is expensive (thanks to update_permission_bitmask(), update_pkru_bitmask(),..) and reloading MMU root doesn't come for free. Try to approach the issue by splitting L1-normal and L1-nested MMUs and checking if MMU reset is really needed. This spares us about 1000 cpu cycles on nested vmexit. RFC part: - Does this look like a plausible solution? - SVM nested can probably be optimized in the same way. - Doesn mmu_update_needed() cover everything? Vitaly Kuznetsov (7): x86/kvm/mmu: make vcpu->mmu a pointer to the current MMU x86/kvm/mmu.c: set get_pdptr hook in kvm_init_shadow_ept_mmu() x86/kvm/mmu.c: add kvm_mmu parameter to kvm_mmu_free_roots() x86/kvm/mmu: introduce guest_mmu x86/kvm/mmu: get rid of redundant kvm_mmu_setup() x86/kvm/nVMX: introduce scache for kvm_init_shadow_ept_mmu x86/kvm/nVMX: optimize MMU switch from nested_vmx_load_cr3() arch/x86/include/asm/kvm_host.h | 36 ++++- arch/x86/kvm/cpuid.c | 2 +- arch/x86/kvm/mmu.c | 282 ++++++++++++++++++++++++++-------------- arch/x86/kvm/mmu.h | 2 +- arch/x86/kvm/mmu_audit.c | 12 +- arch/x86/kvm/paging_tmpl.h | 17 +-- arch/x86/kvm/svm.c | 20 +-- arch/x86/kvm/vmx.c | 52 +++++--- arch/x86/kvm/x86.c | 34 ++--- 9 files changed, 292 insertions(+), 165 deletions(-) -- 2.14.4
