d2017-11-23 0:56 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>: > On 22/11/2017 10:43, Liran Alon wrote: >>>> >>>> I think we should also set "vmx->nested.smm.vmxon = false;" >>>> after "vmx->nested.vmxon = false;" to correctlyhandle the case >>>> VMXOFF is executed from SMI handler. Otherwise, when SMI handler >>>> executes RSM, we will reach vmx_pre_leave_smm() which will set >>>> again "vmx->nested.vmxon = true;" which I think shouldn't >>>> happen. >>> >>> I didn't see a real scenario for this. >> >> Actually I later saw that handle_vmoff() calls >> nested_vmx_check_permission() which indeed won't allow to continue >> executing if running from SMI because vmx->nested.vmxon=false; and >> therefore this will raise a #UD. So you are right. :) > > Still, not clearing the flag is wrong. free_nested is also called by > vmx_leave_nested when the host writes 0 to MSR_IA32_FEATURE_CONTROL with > KVM_SET_MSRS. Do it in v2. :) Regards, Wanpeng Li
