Re: [PATCH 05/17] Keep SMP VMs more in sync on unstable TSC

[Jason Wang <jasowang@redhat.com>]

Zachary Amsden wrote:
> SMP VMs on machines with unstable TSC have their TSC offset adjusted by the
> local offset delta from last measurement.  This does not take into account how
> long it has been since the measurement, leading to drift.  Minimize the drift
> by accounting for any time difference the kernel has observed.
>
> Signed-off-by: Zachary Amsden <zamsden@redhat.com>
> ---
>  arch/x86/include/asm/kvm_host.h |    1 +
>  arch/x86/kvm/x86.c              |   20 +++++++++++++++++++-
>  2 files changed, 20 insertions(+), 1 deletions(-)
>
> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> index 94f6ce8..1afecd7 100644
> --- a/arch/x86/include/asm/kvm_host.h
> +++ b/arch/x86/include/asm/kvm_host.h
> @@ -337,6 +337,7 @@ struct kvm_vcpu_arch {
>  	unsigned int time_offset;
>  	struct page *time_page;
>  	u64 last_host_tsc;
> +	u64 last_host_ns;
>  
>  	bool nmi_pending;
>  	bool nmi_injected;
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 618c435..b1bdf05 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -1810,6 +1810,19 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
>  		/* Make sure TSC doesn't go backwards */
>  		s64 tsc_delta = !vcpu->arch.last_host_tsc ? 0 : 
>  				native_read_tsc() - vcpu->arch.last_host_tsc;
> +
> +		/* Subtract elapsed cycle time from the delta computation */
> +		if (check_tsc_unstable() && vcpu->arch.last_host_ns) {
> +			s64 delta;
> +			struct timespec ts;
> +			ktime_get_ts(&ts);
> +			monotonic_to_bootbased(&ts);
> +			delta = timespec_to_ns(&ts) - vcpu->arch.last_host_ns;
> +			delta = delta * per_cpu(cpu_tsc_khz, cpu);
>   
This seems not work well on a cpu w/o CONSTANT_TSC.
> +			delta = delta / USEC_PER_SEC;
> +			tsc_delta -= delta;
> +		}
> +
>  		if (tsc_delta < 0 || check_tsc_unstable())
>  			kvm_x86_ops->adjust_tsc_offset(vcpu, -tsc_delta);
>  		kvm_migrate_timers(vcpu);
> @@ -1832,8 +1845,13 @@ void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
>  	 * vcpu->cpu != cpu can not detect this condition.  So set
>  	 * vcpu->cpu = -1 to force the recalibration above.
>  	 */
> -	if (check_tsc_unstable())
> +	if (check_tsc_unstable()) {
> +		struct timespec ts;
> +		ktime_get_ts(&ts);
> +		monotonic_to_bootbased(&ts);
> +		vcpu->arch.last_host_ns = timespec_to_ns(&ts);
>  		vcpu->cpu = -1;
> +	}
>  }
>  
>  static int is_efer_nx(void)
>