Re: [PATCH V4 7/7] KVM, pkeys: disable PKU feature without ept

[Yang Zhang <yang.zhang.wz@gmail.com>]

On 2016/3/9 15:21, Xiao Guangrong wrote:
>
>
> On 03/09/2016 02:37 PM, Yang Zhang wrote:
>> On 2016/3/9 13:51, Xiao Guangrong wrote:
>>>
>>>
>>> On 03/08/2016 06:02 PM, Paolo Bonzini wrote:
>>>>
>>>>
>>>> On 08/03/2016 10:32, Xiao Guangrong wrote:
>>>>> On 03/08/2016 04:47 PM, Paolo Bonzini wrote:
>>>>>> Some XSAVE features (currently only MPX, but in the future PKRU too)
>>>>>> will force eagerfpu on, see fpu__init_system_ctx_switch:
>>>>>>
>>>>>>           if (xfeatures_mask & XFEATURE_MASK_EAGER) {
>>>>>>                   if (eagerfpu == DISABLE) {
>>>>>>                           xfeatures_mask &= ~XFEATURE_MASK_EAGER;
>>>>>
>>>>> So if the kernel parameter, eagerfpu is set to "off", then eager is
>>>>> not
>>>>> enabled, so PKRU can not work in KVM?
>>>>
>>>> Yes.  Neither PKRU nor MPX.
>>>
>>> Er... I noticed fpregs is not switched if the CPU is running in KVM
>>> module
>>> (vcpu is not scheduled out and does not exit to userspace), that is why
>>> read_pkru() can be used to read guest's PKRU in the patch 4.
>>>
>>> However, then guest can fully control the access of userspace's
>>> memory if
>>> CR4.PKRU is enabled on host and KVM needs to access QEMU's memory to do
>>> some
>>> emulation anyway. Is it really safe？
>>
>> I think it depends on how we understand the guest uses Pkeys. From my
>> point, guest only wants to
>> protect the pages from guest's view, not kvm. So the access from KVM
>> should be totally transparent
>> to guest. And should not be aware by guest. For modification, i think
>> current KVM only touch the DMA
>> buffer which is setup by guest driver. It's guest responsibility to
>> ensure the pages he want to
>> protect are not used as DMA buffer.
>>
>
> No really. A example is KVM need to read guest memory to emulate some
> instructions.

This is the access case. I don't think guest is interesting in such access.

>
> More worse, the pkey-bits on pte entry is different between guest and
> host (they
> are using different page tables) so guest can not know what index in
> PKRU will be
> used by KVM. A evil guest can use it to attack host...

Can you give a example how a evil guest can attack host?

One question is if host and guest are both using pkeys, how to handle 
this case? I don't see current patch consider this case?

-- 
best regards
yang