Re: [PATCH] kvm: x86: move tracepoints outside extended quiescent state

[Paolo Bonzini <pbonzini@redhat.com>]

On 11/12/2015 11:22, Borislav Petkov wrote:
> On Thu, Dec 10, 2015 at 07:15:19PM +0100, Paolo Bonzini wrote:
>> Yeah, wait_lapic_expire also have to be moved before __kvm_guest_enter.
> 
> Yeah, v2 doesn't splat on the Intel box anymore but the AMD box still
> has, and it is a different problem. With the v2 applied, it still
> explodes, see below.

Yes, I didn't expect it to fix anything.  I just wanted to pinpoint it
to kvm-amd.

> And I'm willing to bet good money on that shadow pages fun.

You can disable it (well, make it take a few days to appear) with this:

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 484079efea5b..a9070e260c72 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -496,7 +496,7 @@ static struct kvm_memslots *kvm_alloc_memslots(void)
 	 * Init kvm generation close to the maximum to easily test the
 	 * code of handling generation number wrap-around.
 	 */
-	slots->generation = -150;
+	slots->generation = 0;
 	for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
 		slots->id_to_index[i] = slots->memslots[i].id = i;

but it would not be AMD-specific.

Anyway if this theory is true:

> [  959.466549] kernel tried to execute NX-protected page - exploit attempt? (uid: 1000)
> 
> line basically says that we're pagefaulting when trying to fetch
> instructions, i.e., we're trying to execute something from a page, rIP
> points to 0xffff8800b9f9bdf0 and that is most likely a page belonging to
> kvm, which, however, is for some reason not executable (anymore?).

It would be a kvm hypervisor page, not a kvm guest page, hence unrelated
to the zapping thing.

Can you grab the kallsyms before making it crash?  I will get to it next
week.

Paolo