Re: [PATCH 15/17] KVM: retpolines: x86: eliminate retpoline from vmx.c exit handlers

From: Andrea Arcangeli
Date: Tue Sep 24 2019 - 17:47:02 EST


On Mon, Sep 23, 2019 at 03:05:14PM -0400, Andrea Arcangeli wrote:
> On Mon, Sep 23, 2019 at 11:57:57AM +0200, Paolo Bonzini wrote:
> > On 23/09/19 11:31, Vitaly Kuznetsov wrote:
> > > +#ifdef CONFIG_RETPOLINE
> > > + if (exit_reason == EXIT_REASON_MSR_WRITE)
> > > + return handle_wrmsr(vcpu);
> > > + else if (exit_reason == EXIT_REASON_PREEMPTION_TIMER)
> > > + return handle_preemption_timer(vcpu);
> > > + else if (exit_reason == EXIT_REASON_PENDING_INTERRUPT)
> > > + return handle_interrupt_window(vcpu);
> > > + else if (exit_reason == EXIT_REASON_EXTERNAL_INTERRUPT)
> > > + return handle_external_interrupt(vcpu);
> > > + else if (exit_reason == EXIT_REASON_HLT)
> > > + return handle_halt(vcpu);
> > > + else if (exit_reason == EXIT_REASON_PAUSE_INSTRUCTION)
> > > + return handle_pause(vcpu);
> > > + else if (exit_reason == EXIT_REASON_MSR_READ)
> > > + return handle_rdmsr(vcpu);
> > > + else if (exit_reason == EXIT_REASON_CPUID)
> > > + return handle_cpuid(vcpu);
> > > + else if (exit_reason == EXIT_REASON_EPT_MISCONFIG)
> > > + return handle_ept_misconfig(vcpu);
> > > +#endif
> > > return kvm_vmx_exit_handlers[exit_reason](vcpu);
> >
> > Most of these, while frequent, are already part of slow paths.
> >
> > I would keep only EXIT_REASON_MSR_WRITE, EXIT_REASON_PREEMPTION_TIMER,
> > EXIT_REASON_EPT_MISCONFIG and add EXIT_REASON_IO_INSTRUCTION.
>
> Intuition doesn't work great when it comes to CPU speculative
> execution runtime. I can however run additional benchmarks to verify
> your theory that keeping around frequent retpolines will still perform
> ok.

On one most recent CPU model there's no measurable difference with
your list or my list with a hrtimer workload (no cpuid). It's
challenging to measure any difference below 0.5%.

An artificial cpuid loop takes 1.5% longer to compute if I don't add
CPUID to the list, but that's just the measurement of the cost of
hitting a frequent retpoline in the exit reason handling code.

Thanks,
Andrea