On Mon, Sep 23, 2019 at 07:43:07PM -0400, Andrea Arcangeli wrote: > On Mon, Sep 23, 2019 at 02:24:35PM -0700, Sean Christopherson wrote: > > An extra CALL+RET isn't going to be noticeable, especially on modern > > hardware as the high frequency VMWRITE/VMREAD fields should hit the > > shadow VMCS. > > In your last email with regard to the inlining optimizations made > possible by the monolithic KVM model you said "That'd likely save a > few CALL/RET/JMP instructions", that kind of directly contradicts the > above. I think neither one if taken at face value can be possibly > measured. However the above only is relevant for nested KVM so I'm > fine if there's an agreement that it's better to hide the nested vmx > handlers in nested.c at the cost of some call/ret. For the immediate exit case, eliminating the CALL/RET/JMP instructions is a bonus. The real goal was to eliminate the oddity of bouncing through vendor code to invoke a one-line x86 function. Having a separate __kvm_request_immediate_exit() made sense when overwriting kvm_ops_x86, but not so much when using direct calls.
