On Mon, Dec 14, 2020 at 11:38 AM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > +Andy, who provided a lot of feedback on v1. > > > > > static unsigned long svm_get_rflags(struct kvm_vcpu *vcpu) > > @@ -3507,14 +3503,8 @@ static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu, > > > > __svm_vcpu_run(svm->vmcb_pa, (unsigned long *)&svm->vcpu.arch.regs); > > Tying in with avoiding svm->host_save_area, what about passing in the PA of the > save area and doing the vmload in __svm_vcpu_run()? One less instance of inline > assembly to stare at... One potential side benefit is that we wouldn't execute any C code with the wrong MSR_GS_BASE, which avoids any concerns about instrumentation, stack protector, or some *SAN feature exploding due to a percpu memory not working. --Andy
