Re: [Patch v5 1/4] Remove SMEP bit from CR4_RESERVED_BITS

[Ingo Molnar <mingo@elte.hu>]

* Avi Kivity <avi@redhat.com> wrote:

> On 05/30/2011 10:40 AM, Ingo Molnar wrote:
> >* Yang, Wei Y<wei.y.yang@intel.com>  wrote:
> >
> >>  This patch removes SMEP bit from CR4_RESERVED_BITS.
> >
> >I'm wondering, what is the best-practice way for tools/kvm/ to set
> >SMEP for the guest kernel automatically, even if the guest kernel
> >itsef has not requested SMEP?
> >
> > The portion i'm worried about are old KVM versions that have the 
> > SMEP bit in CR4_RESERVED_BITS and reject it. So we cannot just 
> > unilaterally add SMEP to every cr4 write of the guest.
> 
> tools/kvm doesn't see cr4 writes at all. [...]

I feared small complications like that! :-)

We can definitely use KVM_GET_SREGS, fiddle the SMEP bit in 
kvm_regs.cr4 and call KVM_SET_SREGS, once the fine patch above goes 
upstream.

> [...]  The only way to do this is in kvm itself.
> 
> > Is there a way to query whether the host KVM version supports 
> > SMEP setting in cr4?
> >
> 
> KVM_GET_SUPPORTED_CPUID (it returns whether both the host cpu and 
> kvm support smep; if one of them doesn't, you'll see smep 
> disabled).

That looks useful.

So the way to go appears to be to do a GET_SREGS/SET_SREGS sequence 
to enable SMEP in the guest, some time after it has booted and has 
enabled paging.

I'm wondering whether there's a suitable place to do that, when we 
are more or less guaranteed to exit the VM for some other reason - 
such as the first MMIO done with paging enabled?

This solution means that we'll slow down pre-paging MMIOs with a 
GET_SREGS call, but that's ok, they are rare and the pre-paging 
bootup phase is very short.

So the only worry would be where the guest sets cr4 itself - and 
since it does not know about SMEP it will probably disable it. Guest 
suspend/resume is one such place ...

Another option would be to try to set the SMEP bit *before* we enable 
paging. In theory this should not confuse a Linux guest - and while i 
have not tested it i *think* we let it survive in the 
saved_cr4_features shadow variable. That would make guest 
suspend/resume work out of box as well.

Thanks,

	Ingo