On 08/03/19 16:50, Sean Christopherson wrote: > On Fri, Mar 08, 2019 at 10:15:15AM +0100, Paolo Bonzini wrote: >> On 04/03/19 19:56, Sean Christopherson wrote: >>> On Tue, Jan 29, 2019 at 04:19:34PM +0100, Paolo Bonzini wrote: >>>> On 28/01/19 11:33, Yang Weijiang wrote: >>>>>> There is no code in this series to pass these fields to and from >>>>>> userspace, and also to save/restore U_CET, INT_SSP_TAB, PL0_SSP and >>>>>> PL3_SSP across context switches. >>>>>> >>>>> The kernel consumes these MSRs, please see kernel CET patch: >>>>> https://lkml.org/lkml/fancy/2018/11/20/225 >>>> >>>> Still, even if the kernel saves these fields across context switch in >>>> XSAVE areas, KVM must support accesses to the MSRs from userspace, for >>>> example in order to perform live migration. >>>> >>>> For example, when reading/writing these in kvm_set_msr or >>>> kvm_get_msr_common, you would have to do a read/write from the host >>>> MSRs. You also have to put kvm_load_guest_fpu/kvm_put_guest_fpu calls >>>> in __msr_io. >>> >>> Paolo, can you elaborate on why KVM would read the host MSRs? Wouldn't >>> kvm_{get,set}_msr() pull the values from the VMCS when necessary? >> >> Not all MSRs are in the VMCS; IA32_U_CET and IA32_PL*_SSP are not. > > Ah, "host MSRs" confused me. I though you meant the host's version of > the MSRs, but you're saying do an XRSTORS to load the guest's FPU state > and then {RD,WR}MSR to pull the guest's value from hardware. Yes, that's what Weijang's patches are already doing in the next version. Paolo