Re: [PATCH] VMX: Fix and improve guest state validity checks

[Avi Kivity <avi@redhat.com>]

On 05/11/2010 07:52 PM, Mohammed Gamal wrote:
> - Add 's' and 'g' field checks on segment registers
> - Correct SS checks for request and descriptor privilege levels
>
> Signed-off-by: Mohammed Gamal<m.gamal005@gmail.com>
> ---
>   arch/x86/kvm/vmx.c |   73 +++++++++++++++++++++++++++++++++++++++++++++++----
>   1 files changed, 67 insertions(+), 6 deletions(-)
>
> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
> index 777e00d..9805c2a 100644
> --- a/arch/x86/kvm/vmx.c
> +++ b/arch/x86/kvm/vmx.c
> @@ -2121,16 +2121,30 @@ static bool stack_segment_valid(struct kvm_vcpu *vcpu)
>   	vmx_get_segment(vcpu,&ss, VCPU_SREG_SS);
>   	ss_rpl = ss.selector&  SELECTOR_RPL_MASK;
>
> -	if (ss.unusable)
> +	if (ss.dpl != ss_rpl) /* DPL != RPL */
> +		return false;
> +
> +	if (ss.unusable) /* Short-circuit */
>   		return true;
>    

If ss.unusable, do the dpl and rpl have any meaning?

>   	if (!ss.present)
>   		return false;
> +	if (ss.limit&  0xfff00000) {
> +                if ((ss.limit&  0xfff)<  0xfff)
> +                        return false;
> +                if (!ss.g)
> +                        return false;
> +        } else {
> +                if ((ss.limit&  0xfff) == 0xfff)
> +                        return false;
> +                if (ss.g)
> +                        return false;
> +        }
>    

There is no architectural way to break this.  That is, without 
virtualization, there is no way a real cpu will ever have a limit of 
0x12345678.

We need to distinguish between big real mode and real mode that can be 
virtualized using vm86, but we don't need to consider impossible setups.


> @@ -2143,8 +2157,15 @@ static bool data_segment_valid(struct kvm_vcpu *vcpu, int seg)
>   	vmx_get_segment(vcpu,&var, seg);
>   	rpl = var.selector&  SELECTOR_RPL_MASK;
>
> -	if (var.unusable)
> +	if (var.unusable)  /* Short-circuit */
>   		return true;
> +	if (!(var.type&  AR_TYPE_ACCESSES_MASK))
> +		return false;
>    

Again, there is no architectural way for a segment not to have the 
accessed bit set.

> +	if (var.type&  AR_TYPE_CODE_MASK) {
> +		if (!(var.type&  AR_TYPE_READABLE_MASK))
> +			return false;
> +	}
>    

About this, I'm not sure.

> +
>   	if (!var.s)
>   		return false;
>   	if (!var.present)
> @@ -2154,6 +2175,18 @@ static bool data_segment_valid(struct kvm_vcpu *vcpu, int seg)
>   			return false;
>   	}
>
> +	if (var.limit&  0xfff00000) {
> +		if ((var.limit&  0xfff)<  0xfff)
> +			return false;
> +		if (!var.g)
> +			return false;
> +	} else {
> +		if ((var.limit&  0xfff) == 0xfff)
> +			return false;
> +		if (var.g)
> +			return false;
> +	}
>    

Even disregarding the incorrectness, you shouldn't duplicate code like this.

> @@ -2192,6 +2240,20 @@ static bool ldtr_valid(struct kvm_vcpu *vcpu)
>   		return false;
>   	if (!ldtr.present)
>   		return false;
> +	if (ldtr.s)
> +		return false;
>    

Architecturally impossible.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.