Re: [PATCH v2 2/3] x86/AMD: support further feature masking MSRs

[Aravind Gopalakrishnan <aravind.gopalakrishnan@amd.com>]

On 4/8/2014 10:14 AM, Boris Ostrovsky wrote:
> On 04/08/2014 10:33 AM, Jan Beulich wrote:
>>>>> On 08.04.14 at 16:14, <boris.ostrovsky@oracle.com> wrote:
>>> On 04/08/2014 10:02 AM, Jan Beulich wrote:
>>>>>>> On 08.04.14 at 15:50, <boris.ostrovsky@oracle.com> wrote:
>>>>> On 04/08/2014 03:15 AM, Jan Beulich wrote:
>>>>>>>>> On 07.04.14 at 17:21, <boris.ostrovsky@oracle.com> wrote:
>>>>>>> On 04/07/2014 05:43 AM, Jan Beulich wrote:
>>>>>>>
>>>>>>> @@ -179,11 +196,39 @@ static void __devinit set_cpuidmask(cons
>>>>>>>          printk("Writing CPUID extended feature mask ECX:EDX -> 
>>>>>>> %08Xh:%08Xh\n",
>>>>>>>                 extfeat_ecx, extfeat_edx);
>>>>>>>      +    if (c->cpuid_level >= 7)
>>>>>>> +        cpuid_count(7, 0, &eax, &ebx, &ecx, &edx);
>>>>>>> +    else
>>>>>>> +        ebx = eax = 0;
>>>>>>> +    if ((eax | ebx) && ~(l7s0_eax & l7s0_ebx)) {
>>>>>>> +        if (l7s0_eax > eax)
>>>>>>> +            l7s0_eax = eax;
>>>>>>> +        l7s0_ebx &= ebx;
>>>>>>>
>>>>>>>
>>>>>>> Can you explain why eax is treated differently here (i.e. not 
>>>>>>> ANDing it
>>>>>>> as is done with ebx)?
>>>>>> Generally I think code like this implies that you know the 
>>>>>> specification:
>>>>>> eax here represents the maximum supported subleaf, and hence
>>>>>> needs to be limited rather than masked.
>>>>> All specs that I have say that bits of CPUID Fn0000_0007_EAX_x0 are
>>>>> reserved.
>>>> Intel's SDM Vol 2 rev 49 (325383-049US) says "07H Sub-leaf 0 (Input
>>>> ECX = 0). EAX Bits 31-00: Reports the maximum input value for
>>>> supported leaf 7 sub-leaves." Not sure what other specs you might
>>>> be looking at.
>>> This is a patch to xen/arch/x86/cpu/amd.c so I was looking at AMD's
>>> BKDGs.  Why is Intel's definition relevant here?
>> Because leaf 7 is (largely) Intel-defined (AMD adds their extensions
>> usually at leaves 800000xx).
>
> I don't know whether we can assume that AMD will follow the same 
> definitions. They usually try not to deliberately do something 
> differently there but this is never guaranteed.
>


Hmm. Boris is right.. All BKDG's I can refer to also say cpuid[7,0].eax 
is reserved.

So, we should be OK with allowing user to mask only cpuid[7,0].ebx ?


-Aravind.