On 11/08/2017 07:59, Wanpeng Li wrote:
> From: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
>
> Commit c016004494b0 (KVM: x86: Avoid guest page table walk when gpa_available
> is set) avoids the page table walk when cr2 has already contained a valid GPA.
> However, that is not the truth if ept == 0 and shadow page table is used. In
> this scenario cr2 can just contains a valid GVA. The commit results in guest
> stuck during boot due to read/write emulation against GVA instead of GPA of
> the guest.
>
> This patch fixes it by not setting the gpa_available flag if the shadow page table
> is used.
Yup, I got these on my overnight tests too. A better fix that I was
going to commit today is:
if (vcpu->arch.mmu.direct_map) {
vcpu->arch.gpa_available = true;
vcpu->arch.gpa_val = fault_address;
}
because AMD gets here even if npt=1.
Paolo
> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> Cc: Radim Krčmář <rkrcmar@xxxxxxxxxx>
> Cc: Brijesh Singh <brijesh.singh@xxxxxxx>
> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
> ---
> arch/x86/kvm/mmu.c | 2 --
> 1 file changed, 2 deletions(-)
>
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index 849312d..e4ce20b 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -3791,8 +3791,6 @@ int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
>
> if (need_unprotect && kvm_event_needs_reinjection(vcpu))
> kvm_mmu_unprotect_page_virt(vcpu, fault_address);
> - vcpu->arch.gpa_available = true;
> - vcpu->arch.gpa_val = fault_address;
> r = kvm_mmu_page_fault(vcpu, fault_address, error_code, insn,
> insn_len);
> break;
>
