Re: [PATCH v2 00/11] KVM: Support guest MAXPHYADDR < host MAXPHYADDR

[Paolo Bonzini]

On 22/06/20 18:33, Tom Lendacky wrote:
> I'm not a big fan of trapping #PF for this. Can't this have a performance
> impact on the guest? If I'm not mistaken, Qemu will default to TCG
> physical address size (40-bits), unless told otherwise, causing #PF to now
> be trapped. Maybe libvirt defaults to matching host/guest CPU MAXPHYADDR?

Yes, this is true.  We should change it similar to how we handle TSC
frequency (and having support for guest MAXPHYADDR < host MAXPHYADDR is
a prerequisite).

> In bare-metal, there's no guarantee a CPU will report all the faults in a
> single PF error code. And because of race conditions, software can never
> rely on that behavior. Whenever the OS thinks it has cured an error, it
> must always be able to handle another #PF for the same access when it
> retries because another processor could have modified the PTE in the
> meantime.

I agree, but I don't understand the relation to this patch.  Can you
explain?

> What's the purpose of reporting RSVD in the error code in the
> guest in regards to live migration?
>
>> - if the page is accessible to the guest according to the permissions in
>> the page table, it will cause a #NPF.  Again, we need to trap it, check
>> the guest physical address and inject a P|RSVD #PF if the guest physical
>> address has any guest-reserved bits.
>>
>> The AMD specific issue happens in the second case.  By the time the NPF
>> vmexit occurs, the accessed and/or dirty bits have been set and this
>> should not have happened before the RSVD page fault that we want to
>> inject.  On Intel processors, instead, EPT violations trigger before
>> accessed and dirty bits are set.  I cannot find an explicit mention of
>> the intended behavior in either the
>> Intel SDM or the AMD APM.
> 
> Section 15.25.6 of the AMD APM volume 2 talks about page faults (nested vs
> guest) and fault ordering. It does talk about setting guest A/D bits
> during the walk, before an #NPF is taken. I don't see any way around that
> given a virtual MAXPHYADDR in the guest being less than the host MAXPHYADDR.

Right you are...  Then this behavior cannot be implemented on AMD.

Paolo