On 07/31/2013 05:48 PM, Gleb Natapov wrote:
> From: Nadav Har'El
>
> kvm_set_cr3() attempts to check if the new cr3 is a valid guest physical
> address. The problem is that with nested EPT, cr3 is an *L2* physical
> address, not an L1 physical address as this test expects.
>
> As the comment above this test explains, it isn't necessary, and doesn't
> correspond to anything a real processor would do. So this patch removes it.
>
> Note that this wrong test could have also theoretically caused problems
> in nested NPT, not just in nested EPT. However, in practice, the problem
> was avoided: nested_svm_vmexit()/vmrun() do not call kvm_set_cr3 in the
> nested NPT case, and instead set the vmcb (and arch.cr3) directly, thus
> circumventing the problem. Additional potential calls to the buggy function
> are avoided in that we don't trap cr3 modifications when nested NPT is
> enabled. However, because in nested VMX we did want to use kvm_set_cr3()
> (as requested in Avi Kivity's review of the original nested VMX patches),
> we can't avoid this problem and need to fix it.
>
> Reviewed-by: Xiao Guangrong
> Signed-off-by: Nadav Har'El
> Signed-off-by: Jun Nakajima
> Signed-off-by: Xinhao Xu
> Signed-off-by: Yang Zhang
> Signed-off-by: Gleb Natapov
> ---
> arch/x86/kvm/x86.c | 11 ---
> 1 file changed, 11 deletions(-)
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index d2caeb9..e2fef8b 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -682,17 +682,6 @@ int kvm_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3)
>*/
> }
>
> - /*
> - * Does the new cr3 value map to physical memory? (Note, we
> - * catch an invalid cr3 even in real-mode, because it would
> - * cause trouble later on when we turn on paging anyway.)
> - *
> - * A real CPU would silently accept an invalid cr3 and would
> - * attempt to use it - with largely undefined (and often hard
> - * to debug) behavior on the guest side.
> - */
> - if (unlikely(!gfn_to_memslot(vcpu->kvm, cr3 >> PAGE_SHIFT)))
> - return 1;
> vcpu->arch.cr3 = cr3;
> __set_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail);
> vcpu->arch.mmu.new_cr3(vcpu);
>
Reviewed-by: Orit Wasserman
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
