Re: [PATCH 14/15] KVM: MTRR: do not map huage page for non-consistent range

[Xiao Guangrong <guangrong.xiao@linux.intel.com>]

On 06/03/2015 03:55 PM, Paolo Bonzini wrote:
>
>
> On 03/06/2015 04:56, Xiao Guangrong wrote:
>>
>>
>> On 06/01/2015 05:36 PM, Paolo Bonzini wrote:
>>>
>>>
>>> On 30/05/2015 12:59, Xiao Guangrong wrote:
>>>> Currently guest MTRR is completely prohibited if cache snoop is
>>>> supported on
>>>> IOMMU (!noncoherent_dma) and host does the emulation based on the
>>>> knowledge
>>>> from host side, however, host side is not the good point to know
>>>> what the purpose of guest is. A good example is that pass-throughed VGA
>>>> frame buffer is not always UC as host expected
>>>
>>> Can you explain how?  The original idea was that such a framebuffer
>>> would be kvm_is_reserved_pfn and thus be unconditionally UC.
>>
>> Yes, frame-buffer is always UC in current code, however, UC for
>> frame-buffer causes bad performance.
>
> Understood now, thanks.
>
>> So that guest will configure the range to MTRR, this patchset follows
>> guest MTRR and cooperates with guest PAT (ept.VMX_EPT_IPAT_BIT = 0) to
>> emulate guest cache type as guest expects.
>
> Unlike e.g. CR0.CD=1, UC memory does not snoop the cache to preserve
> coherency.  AMD, has special logic to do this, for example:
>
> - if guest PAT says "UC" and host MTRR says "WB", the processor will not
> cache the memory but will snoop the cache as if CR0.CD=1
>
> - if guest PAT says "WC" and host (nested page table) PAT says "WB" and
> host MTRR says "WB", the processor will still do write combining but
> also snoop the cache as if CR0.CD=1
>
> I am worried that the lack of this feature could cause problems if
> guests map QEMU's VGA framebuffer as uncached.  We have this problem on
> ARM, so it's not 100% theoretical.

CR0.CD is always 0 in both host and guest, i guess it's why we cleared
CR0.CD and CR0.NW in vmx_set_cr0().

>
> So, why do you need to always use IPAT=0?  Can patch 15 keep the current
> logic for RAM, like this:
>
> 	if (is_mmio || kvm_arch_has_noncoherent_dma(vcpu->kvm))
> 		ret = kvm_mtrr_get_guest_memory_type(vcpu, gfn) <<
> 		      VMX_EPT_MT_EPTE_SHIFT;
> 	else
> 		ret = (MTRR_TYPE_WRBACK << VMX_EPT_MT_EPTE_SHIFT)
> 			| VMX_EPT_IPAT_BIT;

Yeah, it's okay, actually we considered this way, however
- it's light enough, it did not hurt guest performance based on our
   benchmark.
- the logic has always used for noncherent_dma case, extend it to
   normal case should have low risk and also help us to check the logic.
- completely follow MTRRS spec would be better than host hides it.