Re: [PATCH v5 4/5] KVM: x86: emulation: Apply LAM mask when emulating data access in 64-bit mode

[Binbin Wu <binbin.wu@linux.intel.com>]

On 3/2/2023 4:55 PM, Chao Gao wrote:
> On Mon, Feb 27, 2023 at 04:45:46PM +0800, Robert Hoo wrote:
>> Emulate HW LAM masking when doing data access under 64-bit mode.
>>
>> kvm_lam_untag_addr() implements this: per CR4/CR3 LAM bits configuration,
>> firstly check the linear addr conforms LAM canonical, i.e. the highest
>> address bit matches bit 63. Then mask out meta data per LAM configuration.
>> If failed in above process, emulate #GP to guest.
>>
>> Signed-off-by: Robert Hoo <robert.hu@linux.intel.com>
>> ---
>> arch/x86/kvm/emulate.c | 13 ++++++++
>> arch/x86/kvm/x86.h     | 70 ++++++++++++++++++++++++++++++++++++++++++
>> 2 files changed, 83 insertions(+)
>>
>> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
>> index 5cc3efa0e21c..77bd13f40711 100644
>> --- a/arch/x86/kvm/emulate.c
>> +++ b/arch/x86/kvm/emulate.c
>> @@ -700,6 +700,19 @@ static __always_inline int __linearize(struct x86_emulate_ctxt *ctxt,
>> 	*max_size = 0;
>> 	switch (mode) {
>> 	case X86EMUL_MODE_PROT64:
>> +		/* LAM applies only on data access */
>> +		if (!fetch && guest_cpuid_has(ctxt->vcpu, X86_FEATURE_LAM)) {
>> +			enum lam_type type;
>> +
>> +			type = kvm_vcpu_lam_type(la, ctxt->vcpu);
>> +			if (type == LAM_ILLEGAL) {
>> +				*linear = la;
>> +				goto bad;
>> +			} else {
>> +				la = kvm_lam_untag_addr(la, type);
>> +			}
>> +		}
>> +
>> 		*linear = la;
>> 		va_bits = ctxt_virt_addr_bits(ctxt);
>> 		if (!__is_canonical_address(la, va_bits))
> ...
>
>> +static inline u64 kvm_lam_untag_addr(u64 addr, enum lam_type type)
>> +{
>> +	switch (type) {
>> +	case LAM_U57:
>> +	case LAM_S57:
>> +		addr = __canonical_address(addr, 57);
>> +		break;
>> +	case LAM_U48:
>> +	case LAM_S48:
>> +		addr = __canonical_address(addr, 48);
>> +		break;
>> +	case LAM_NONE:
>> +	default:
>> +		break;
>> +	}
>> +
>> +	return addr;
>> +}
> LAM's change to canonicality check is:
> before performing the check, software metadata in pointers is masked by
> sign-extending the value of bit 56/47.
>
> so, to emulate this behavior, in kvm_lam_untag_addr(), we can simply:
> 1. determine which LAM configuration is enabled, LAM57 or LAM48.
> 2. mask software metadata by sign-extending the bit56/47, i.e.,
>
> 	addr = (sign_extern64(addr, X) & ~BIT_ULL(63)) |
> 	       (addr & BIT_ULL(63));
>
> 	where X=56 for LAM57 and X=47 for LAM48.
>
> Note that this doesn't ensure the resulting @addr is canonical. It
> isn't a problem because the original canonicality check
> (__is_canonical_address() above) can identify non-canonical addresses
> and raise #GP/#SS to the guest.

Thanks for your suggestion. It's much simpler.