On Tue, Jul 30, 2024, Paolo Bonzini wrote: > On 7/30/24 07:32, flyingpenghao@xxxxxxxxx wrote: > > > > When tdp_mmu is enabled, invalid root calls kvm_tdp_mmu_zap_invalidated_roots > > to implement it, and kvm_zap_obsolete_pages is not used. > > > > Signed-off-by: Peng Hao<flyingpeng@xxxxxxxxxxx> > > --- > > arch/x86/kvm/mmu/mmu.c | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c > > index 901be9e420a4..e91586c2ef87 100644 > > --- a/arch/x86/kvm/mmu/mmu.c > > +++ b/arch/x86/kvm/mmu/mmu.c > > @@ -6447,7 +6447,8 @@ static void kvm_mmu_zap_all_fast(struct kvm *kvm) > > */ > > kvm_make_all_cpus_request(kvm, KVM_REQ_MMU_FREE_OBSOLETE_ROOTS); > > - kvm_zap_obsolete_pages(kvm); > > + if (!tdp_mmu_enabled) > > + kvm_zap_obsolete_pages(kvm); > > Can't you have obsolete pages from the shadow MMU that's used for nested > (nGPA->HPA) virtualization? Yep. And kvm_zap_obsolete_pages() is a relatively cheap nop if there are no pages on active_mmu_pages. E.g. we could check kvm_memslots_have_rmaps(), but I don't see any point in doing so, as the existing code should be blazing fast relative to the total cost of the zap.
