Re: [PATCH V4 4/7] KVM, pkeys: add pkeys support for permission_fault logic

[Paolo Bonzini <pbonzini@redhat.com>]

On 06/03/2016 21:36, Paolo Bonzini wrote:
> 
> 
> On 06/03/2016 09:00, Xiao Guangrong wrote:
>>>
>>>       if (vcpu_match_mmio_gva(vcpu, gva)
>>>           && !permission_fault(vcpu, vcpu->arch.walk_mmu,
>>> -                 vcpu->arch.access, access)) {
>>> +                 vcpu->arch.access, 0, access)) {
>>
>> No. The pkey is not always 0.
>>
>> We should cache PKEY for the mmio access and use it here to check if the
>> right is adequate.
> 
> This is just an optimization I think, so it can have false negatives (it
> won't have many in practice because MMIO accesses are usually done in
> supervisor mode).  The actual check is done when
> vcpu->arch.walk_mmu->gva_to_gpa is called.

Duh, sorry, false _positives_ are okay (i.e. you can say something
faults even if it actually doesn't).

One thing you could do is:

- do not pass pte_pkeys to permission_fault, instead read PKRU in
walk_addr_generic and pass "(pkru >> (pte_pkeys * PKRU_ATTRS)) & 3" to
permission_fault.

- here, pass ~0.  So if CR0.PKE=0 or U=0 caching works properly, but if
CR0.PKE=1 and U=1 it is disabled gracefully and gva_to_gpa does the page
walk correctly.

Paolo