On Thu, 2017-11-30 at 13:55 +0100, Paolo Bonzini wrote: > Rik, > > it seems to me that your FPU switching patches would fix the > backtrace > below. We have get_user_pages ending up using the FPU, while KVM is > running (and so it owns the kernel FPU). > > Can you double check? Looks like that may be the problem, indeed. vcpu_enter_guest() calls kvm_load_guest_fpu() before running the guest, but it does not call kvm_put_guest_fpu() before calling ->handle_exit(). Note that the xcr0 get and put are symmetrical inside vcpu_enter_guest(). -- All Rights Reversed.
Attachment:
signature.asc
Description: This is a digitally signed message part
