On Mon, Feb 20, 2017 at 8:51 AM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > > > On 20/02/2017 17:46, Andy Lutomirski wrote: >> On Mon, Feb 20, 2017 at 3:05 AM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: >>> >>> >>> On 18/02/2017 04:29, Andy Lutomirski wrote: >>>> There's no code here because the patch is trivial, but I want to run >>>> the idea by you all first to see if there are any issues. >>>> >>>> VMX is silly and forces the TSS limit to the minimum on VM exits. KVM >>>> wastes lots of cycles bumping it back up to accomodate the io bitmap. >>> >>> Actually looked at the code now... >>> >>> reload_tss is only invoked for userspace exits, so it is a nice-to-have >>> but it wouldn't show on most workloads. Still it does save 150-200 >>> clock cycles to remove it (I just commented out reload_tss() from >>> __vmx_load_host_state to test). >> >> That's for anything involving userspace or preemption, right? > > Yes. But 150-200 clock cycles are nothing compared to the cache misses > you get from preemption, so I'd ignore that. Saving 300 clock cycles on > userspace exits from TR+GSBASE would be about 5% on my Haswell. That's still 5% :) > >> That being said, vmx_save_host_state() is, um, poorly optimized. > > Yeah, but again it doesn't run that often in practical cases. But it will if someone ever tries to upstream Google's approach of not emulating anything in the kernel. :)
