Re: IO on guest is 20 times slower than host

[Avi Kivity <avi@redhat.com>]

Joerg Roedel wrote:
>> index 1fcbc17..d9774e9 100644
>> --- a/kernel/x86/kvm/svm.c
>> +++ b/kernel/x86/kvm/svm.c
>> @@ -575,7 +575,7 @@ static void init_vmcb(struct vcpu_svm *svm)
>>  						INTERCEPT_CR3_MASK);
>>  		control->intercept_cr_write &= ~(INTERCEPT_CR0_MASK|
>>  						 INTERCEPT_CR3_MASK);
>> -		save->g_pat = 0x0007040600070406ULL;
>> +		save->g_pat = 0x0606060606060606ULL;
>>  		/* enable caching because the QEMU Bios doesn't enable it */
>>  		save->cr0 = X86_CR0_ET;
>>  		save->cr3 = 0;
>>     
>
> Yeah, that patch makes sense. But I think we need some more work on this
> because the guest may change the pat msr afterwards. Best would be a simple
> shadow of the pat msr. Last question is how this will effect pci passthrough.
>
>   

I've noticed that Windows (and likely Linux, didn't test) maps the 
cirrus framebuffer with PWT=1, which should slow down the emulated 
framebuffer.  So this patch should speed up things.

If a device is assigned, we must respect the guest PAT, so cirrus 
performance will be low.  On Intel there's an 'ignore PAT' bit which can 
be set on an ept pte for the framebuffer.  Any trick we can do on AMD to 
achieve a similar result?


-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.