On 30/04/19 19:36, Sean Christopherson wrote: > KVM's GPR caching logic is unconditionally emitted for all GPR accesses > (that go through the accessors), even when the register being accessed > is fixed and always available. This bloats KVM due to the instructions > needed to test and set the available/dirty bitmaps, and to conditionally > invoke the .cache_reg() callback. The latter is especially painful when > compiling with retpolines. > > Eliminate the unnecessary overhead by: > > - Adding dedicated accessors for every GPR > - Omitting the caching logic for GPRs that are always available > - Preventing use of the unoptimized versions for fixed accesses > > The last patch is an opportunistic clean up of VMX, which has gradually > acquired a bad habit of sprinkling in direct access to GPRs. Another related cleanup is to replace these with the direct accessors: arch/x86/kvm/vmx/nested.c: vmcs12->guest_rsp = kvm_register_read(vcpu, VCPU_REGS_RSP); arch/x86/kvm/vmx/nested.c: vmcs12->guest_rip = kvm_register_read(vcpu, VCPU_REGS_RIP); arch/x86/kvm/x86.c: regs->rsp = kvm_register_read(vcpu, VCPU_REGS_RSP); arch/x86/kvm/svm.c: kvm_register_write(&svm->vcpu, VCPU_REGS_RSP, hsave->save.rsp); arch/x86/kvm/svm.c: kvm_register_write(&svm->vcpu, VCPU_REGS_RIP, hsave->save.rip); arch/x86/kvm/svm.c: kvm_register_write(&svm->vcpu, VCPU_REGS_RSP, nested_vmcb->save.rsp); arch/x86/kvm/svm.c: kvm_register_write(&svm->vcpu, VCPU_REGS_RIP, nested_vmcb->save.rip); arch/x86/kvm/vmx/nested.c: kvm_register_write(vcpu, VCPU_REGS_RSP, vmcs12->guest_rsp); arch/x86/kvm/vmx/nested.c: kvm_register_write(vcpu, VCPU_REGS_RIP, vmcs12->guest_rip); arch/x86/kvm/vmx/nested.c: kvm_register_write(vcpu, VCPU_REGS_RSP, vmcs12->host_rsp); arch/x86/kvm/vmx/nested.c: kvm_register_write(vcpu, VCPU_REGS_RIP, vmcs12->host_rip); arch/x86/kvm/x86.c: kvm_register_write(vcpu, VCPU_REGS_RSP, regs->rsp); I can take care of this. I have applied patches 1 and 3. I didn't apply patch 2 for the reasons I mentioned in my reply, and because I am not sure if it works properly---it should have flagged the above occurrences, shouldn't it? Paolo
