On Wed, Aug 25, 2021, Sean Christopherson wrote:

On Wed, Aug 25, 2021, Like Xu wrote:

quoted

On 24/8/2021 3:37 am, Sean Christopherson wrote:

quoted

@@ -11061,6 +11061,8 @@ int kvm_arch_hardware_setup(void *opaque)
  	memcpy(&kvm_x86_ops, ops->runtime_ops, sizeof(kvm_x86_ops));
  	kvm_ops_static_call_update();
+	if (ops->intel_pt_intr_in_guest && ops->intel_pt_intr_in_guest())
+		kvm_guest_cbs.handle_intel_pt_intr = kvm_handle_intel_pt_intr;

Emm, it's still buggy.

The guest "unknown NMI" from the host Intel PT can still be reproduced
after the following operation:

	rmmod kvm_intel
	modprobe kvm-intel pt_mode=1 ept=1
	rmmod kvm_intel
	modprobe kvm-intel pt_mode=1 ept=0

Since the handle_intel_pt_intr is not reset to NULL in kvm_arch_hardware_unsetup(),
and the previous function pointer still exists in the generic KVM data structure.

Ooof, good catch.  Any preference between nullifying handle_intel_pt_intr in
setup() vs. unsetup()?  I think I like the idea of "unwinding" during unsetup(),
even though it splits the logic a bit.

Never mind, I figured out a way to clean all this up and land the PT interrupt
handler in vmx.c where it belongs.  Getting there is a bit of a journey, but it's
very doable.  That means unwinding in unsetup() is the preferred approach,
otherwise there would be potential for leaving a dangling pointer if a different
vendor module was succesfully loaded.