On Wed, 2021-11-10 at 16:01 +0100, Paolo Bonzini wrote:
> On 11/10/21 11:00, Maxim Levitsky wrote:
> > + /*
> > + * The vcpu might currently contain L2's IA32_EFER, due to the way
> > + * some userspace kvm users (e.g qemu) restore nested state.
> > + *
> > + * To fix this, restore its IA32_EFER to the value it would have
> > + * after VM exit from the nested guest.
> > + *
> > + */
> > +
> > + vcpu->arch.efer = nested_vmx_get_vmcs12_host_efer(vcpu, vmcs12);
> > +
>
> In principle the value of LOAD_HOST_EFER on exit need not be the same as
> on entry. But you don't need all of EFER, only EFER.LME/EFER.LMA, and
> those two bits must match ("the values of the LMA and LME bits in the
> field must each be that of the “host address-space size” VM-exit
> control" from the "Checks on Host Control Registers, MSRs, and SSP";
> plus the "Checks Related to Address-Space Size").
>
> At least it's worth adjusting the comment to explain that. But the root
> cause of the issue is just nested_vmx_check_* accessing vcpu->arch. So
> you can instead:
>
> - split out of nested_vmx_check_host_state a new function
> nested_vmx_check_address_state_size that does
>
> #ifdef CONFIG_X86_64
> if (CC(!!(vmcs12->vm_exit_controls & VM_EXIT_HOST_ADDR_SPACE_SIZE) !=
> !!(vcpu->arch.efer & EFER_LMA)))
> return -EINVAL;
> #endif
> return 0;
>
> - call it from vmentry but not from migration
>
> - in nested_vmx_check_host_state, assign ia32e from
> vmcs12->vm_exit_controls instead of vcpu->arch.efer
I agree with you. I was thinking do something like that but wasn't sure at all.
Best regards,
Maxim Levitsky
>
> Paolo
>
