On Thu, Feb 24, 2022, Maxim Levitsky wrote: > On Fri, 2022-02-18 at 21:00 +0000, Sean Christopherson wrote: > > On Fri, Feb 18, 2022, Paolo Bonzini wrote: > > > On 2/17/22 22:03, Paolo Bonzini wrote: > > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > > > > index adcee7c305ca..9800c8883a48 100644 > > > > --- a/arch/x86/kvm/x86.c > > > > +++ b/arch/x86/kvm/x86.c > > > > @@ -1189,7 +1189,7 @@ int kvm_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3) > > > > return 1; > > > > if (cr3 != kvm_read_cr3(vcpu)) > > > > - kvm_mmu_new_pgd(vcpu, cr3); > > > > + kvm_mmu_update_root(vcpu); > > > > vcpu->arch.cr3 = cr3; > > > > kvm_register_mark_dirty(vcpu, VCPU_EXREG_CR3); > > > > > > Uh-oh, this has to become: > > > > > > vcpu->arch.cr3 = cr3; > > > kvm_register_mark_dirty(vcpu, VCPU_EXREG_CR3); > > > if (!is_pae_paging(vcpu)) > > > kvm_mmu_update_root(vcpu); > > > > > > The regression would go away after patch 16, but this is more tidy apart > > > from having to check is_pae_paging *again*. > > > > > > Incremental patch: > > > > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > > > index adcee7c305ca..0085e9fba372 100644 > > > --- a/arch/x86/kvm/x86.c > > > +++ b/arch/x86/kvm/x86.c > > > @@ -1188,11 +1189,11 @@ int kvm_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3) > > > if (is_pae_paging(vcpu) && !load_pdptrs(vcpu, cr3)) > > > return 1; > > > - if (cr3 != kvm_read_cr3(vcpu)) > > > - kvm_mmu_update_root(vcpu); > > > - > > > vcpu->arch.cr3 = cr3; > > > kvm_register_mark_dirty(vcpu, VCPU_EXREG_CR3); > > > + if (!is_pae_paging(vcpu)) > > > + kvm_mmu_update_root(vcpu); > > > + > > > /* Do not call post_set_cr3, we do not get here for confidential guests. */ > > > > > > An alternative is to move the vcpu->arch.cr3 update in load_pdptrs. > > > Reviewers, let me know if you prefer that, then I'll send v3. > > > > c) None of the above. > > > > MOV CR3 never requires a new root if TDP is enabled, and the guest_mmu is used if > > and only if TDP is enabled. Even when KVM intercepts CR3 when EPT=1 && URG=0, it > > does so only to snapshot vcpu->arch.cr3, there's no need to get a new PGD. > > > > Unless I'm missing something, your original suggestion of checking tdp_enabled is > > the way to go. > > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > > index 6e0f7f22c6a7..2b02029c63d0 100644 > > --- a/arch/x86/kvm/x86.c > > +++ b/arch/x86/kvm/x86.c > > @@ -1187,7 +1187,7 @@ int kvm_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3) > > if (is_pae_paging(vcpu) && !load_pdptrs(vcpu, cr3)) > > return 1; > > > > - if (cr3 != kvm_read_cr3(vcpu)) > > + if (!tdp_enabled && cr3 != kvm_read_cr3(vcpu)) > > kvm_mmu_new_pgd(vcpu, cr3); > > > > vcpu->arch.cr3 = cr3; > > > > > > Is this actually related to the discussion? The original issue that Paolo > found in his patch was that kvm_mmu_update_root now reads _current_ cr3, thus > it has to be set before calling it. Yes, if we instead do the above, then replacing kvm_mmu_new_pgd() with kvm_mmu_update_root() is unnecessary. Paolo is trying to fix the case where kvm_mmu_new_pgd() does the wrong thing for guest_mmu. My point is that we should never call kvm_mmu_new_pgd() if mmu == guest_mmu in the first place, and adding the tdp_enabled checks fixes that bug. I'm ok with kvm_mmu_new_pgd() acting on a pre-computed role, assuming we actually get sanity checks. Deliberately ignoring the pgd/cr3 we already have is silly.
