On 2013-09-27 08:37, Jan Kiszka wrote: > On 2013-09-26 22:44, Paolo Bonzini wrote: >> Il 26/09/2013 19:47, Paolo Bonzini ha scritto: >>> >>> If I only apply this hunk, which disables the preemption timer while >>> in L1: >>> >>> @@ -8396,6 +8375,8 @@ static void nested_vmx_vmexit(struct kvm_vcpu *vcpu) >>> >>> load_vmcs12_host_state(vcpu, vmcs12); >>> >>> + vmcs_write32(PIN_BASED_VM_EXEC_CONTROL, vmx_pin_based_exec_ctrl(vmx)); >>> + >>> /* Update TSC_OFFSET if TSC was changed while L2 ran */ >>> vmcs_write64(TSC_OFFSET, vmx->nested.vmcs01_tsc_offset); >>> >>> then the testcase works for somewhat larger values of the preemption timer >>> (up to ~1500000 TSC cycles), but then fails. > > Err, does this mean we run L1 with PIN_BASED_VM_EXEC_CONTROL of L2? > Ouch. Should be fixed independently. No, it doesn't mean this. L1 and L2 run on different VMCS, thus should be able to set their PIN_BASED_VM_EXEC_CONTROL independently. I have no clue ATM what that hunk can make a difference for you. Will have a closer look. BTW, aren't many VMCS fields written redundantly in prepare_vmcs02? I mean, those that weren't changed by L1 in the shadow VMCS or require updates for other reasons. There seems to be some room for saving a few cycles. Jan
Attachment:
signature.asc
Description: OpenPGP digital signature
