Re: [PATCH v3] x86: clear incorrectly forced X86_FEATURE_LAHF_LM flag

[Kevin Winchester <kjwinchester@gmail.com>]

Borislav Petkov wrote:
> On Tue, Aug 11, 2009 at 09:15:50PM -0300, Kevin Winchester wrote:
>> Borislav Petkov wrote:
>>> On Tue, Aug 11, 2009 at 12:55:03PM -0300, Kevin Winchester wrote:
>>>> 2009/8/11 Borislav Petkov <borislav.petkov@amd.com>:
>>>>> On Tue, Aug 11, 2009 at 04:37:56PM +0200, Mikael Pettersson wrote:
>>>>>> Since the BIOS apparently wrote some MSR to get LAHF_LM incorrectly
>>>>>> reported by CPUID, would it be possible to also correct that MSR so
>>>>>> that applications that execute CPUID get the correct feature flags?
>>>>> That's a good catch, actually. We have to turn off that bit in the cpuid
>>>>> leaf too if the CPU doesn't support the instructions so that cpuid info
>>>>> is consistent. LAHF/SAHF support in 64bit mode has to be cpuid-checked
>>>>> prior to using them so that info has to be correct.
>>>>>
>>>>> @Kevin: willing to try a patch or two?
>>>>>
>>>> Sure, I'll give it a try this evening.  I assume that since Erratum 110 says:
>>>>
>>>> --------------------------
>>>> Suggested Workaround
>>>> For processors which support the feature (as determined by the
>>>> processor revision ID), BIOS should
>>>> write a one to:
>>>> • MSR C001_100Dh, bit 32 for revision D silicon.
>>>> • MSR C001_1005h, bit 32 for revision E and later silicon.
>>>> This will cause the extended feature flag in ECX[0] to be set.
>>>> --------------------------
>>>>
>>>> That writing a zero to those same MSRs would clear the feature flag?
>>> Yep :). Patch coming up...
>>>
>> I have been attempting to read those MSRs through the /dev/cpu/0/msr device
>> file, without any success.  Is it possible that my CPU will not have those
>> MSRs?  And if so, then maybe my original assumption about the BIOS forcing
>> on the LAHF_LM feature is wrong.
>>
>> In any case, clearing the feature flag (and thus fixing /proc/cpuinfo) is
>> still the right thing to do.
>>
>> Do you have any other suggestions for how I would affect that CPUID flag?
> 
> Before we do that though I'd like to verify that the BIOS is falsely
> setting that bit. Can you run this small c program on your machine and
> send me the result:
> 
> #include <stdio.h>
> 
> int main()
> {
> 
> 	unsigned int ecx = 0;
> 
> 	asm volatile("cpuid"
> 		     : "=c" (ecx)
> 		     : "a" (0x80000001));
> 
> 	printf("0x8000_0001_ecx = 0x%08x\n", ecx);
> 
> 	return 0;
> }
> 
> 

$ ./a.out 
0x8000_0001_ecx = 0x00000001

So that feature is definitely set.  I guess the questions would be:
Did the BIOS set it? If so, how? If not, who did?

Is there anything else I can do to help answer these questions?

-- 
Kevin Winchester