On Thu, 16 May 2019 at 01:42, Sean Christopherson
<sean.j.christopherson@xxxxxxxxx> wrote:
>
> On Wed, May 15, 2019 at 12:11:54PM +0800, Wanpeng Li wrote:
> > From: Wanpeng Li <wanpengli@xxxxxxxxxxx>
> >
> > Advance lapic timer tries to hidden the hypervisor overhead between the
> > host emulated timer fires and the guest awares the timer is fired. However,
> > it just hidden the time between apic_timer_fn/handle_preemption_timer ->
> > wait_lapic_expire, instead of the real position of vmentry which is
> > mentioned in the orignial commit d0659d946be0 ("KVM: x86: add option to
> > advance tscdeadline hrtimer expiration"). There is 700+ cpu cycles between
> > the end of wait_lapic_expire and before world switch on my haswell desktop,
> > it will be 2400+ cycles if vmentry_l1d_flush is tuned to always.
> >
> > This patch tries to narrow the last gap(wait_lapic_expire -> world switch),
> > it takes the real overhead time between apic_timer_fn/handle_preemption_timer
> > and before world switch into consideration when adaptively tuning timer
> > advancement. The patch can reduce 40% latency (~1600+ cycles to ~1000+ cycles
> > on a haswell desktop) for kvm-unit-tests/tscdeadline_latency when testing
> > busy waits.
> >
> > Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > Cc: Radim Krčmář <rkrcmar@xxxxxxxxxx>
> > Cc: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> > Cc: Liran Alon <liran.alon@xxxxxxxxxx>
> > Signed-off-by: Wanpeng Li <wanpengli@xxxxxxxxxxx>
> > ---
> > arch/x86/kvm/lapic.c | 3 ++-
> > arch/x86/kvm/lapic.h | 2 +-
> > arch/x86/kvm/svm.c | 4 ++++
> > arch/x86/kvm/vmx/vmx.c | 4 ++++
> > arch/x86/kvm/x86.c | 3 ---
> > 5 files changed, 11 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> > index af38ece..63513de 100644
> > --- a/arch/x86/kvm/lapic.c
> > +++ b/arch/x86/kvm/lapic.c
> > @@ -1531,7 +1531,7 @@ static inline void adaptive_tune_timer_advancement(struct kvm_vcpu *vcpu,
> > apic->lapic_timer.timer_advance_ns = timer_advance_ns;
> > }
> >
> > -void wait_lapic_expire(struct kvm_vcpu *vcpu)
> > +void kvm_wait_lapic_expire(struct kvm_vcpu *vcpu)
> > {
> > struct kvm_lapic *apic = vcpu->arch.apic;
> > u64 guest_tsc, tsc_deadline;
> > @@ -1553,6 +1553,7 @@ void wait_lapic_expire(struct kvm_vcpu *vcpu)
> > if (unlikely(!apic->lapic_timer.timer_advance_adjust_done))
> > adaptive_tune_timer_advancement(vcpu, apic->lapic_timer.advance_expire_delta);
> > }
> > +EXPORT_SYMBOL_GPL(kvm_wait_lapic_expire);
> >
> > static void start_sw_tscdeadline(struct kvm_lapic *apic)
> > {
> > diff --git a/arch/x86/kvm/lapic.h b/arch/x86/kvm/lapic.h
> > index 3e72a25..f974a3d 100644
> > --- a/arch/x86/kvm/lapic.h
> > +++ b/arch/x86/kvm/lapic.h
> > @@ -220,7 +220,7 @@ static inline int kvm_lapic_latched_init(struct kvm_vcpu *vcpu)
> >
> > bool kvm_apic_pending_eoi(struct kvm_vcpu *vcpu, int vector);
> >
> > -void wait_lapic_expire(struct kvm_vcpu *vcpu);
> > +void kvm_wait_lapic_expire(struct kvm_vcpu *vcpu);
> >
> > bool kvm_intr_is_single_vcpu_fast(struct kvm *kvm, struct kvm_lapic_irq *irq,
> > struct kvm_vcpu **dest_vcpu);
> > diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> > index 406b558..740fb3f 100644
> > --- a/arch/x86/kvm/svm.c
> > +++ b/arch/x86/kvm/svm.c
> > @@ -5646,6 +5646,10 @@ static void svm_vcpu_run(struct kvm_vcpu *vcpu)
> > */
> > x86_spec_ctrl_set_guest(svm->spec_ctrl, svm->virt_spec_ctrl);
> >
> > + if (lapic_in_kernel(vcpu) &&
> > + vcpu->arch.apic->lapic_timer.timer_advance_ns)
> > + kvm_wait_lapic_expire(vcpu);
> > +
> > local_irq_enable();
> >
> > asm volatile (
> > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > index 9663d41..1c49946 100644
> > --- a/arch/x86/kvm/vmx/vmx.c
> > +++ b/arch/x86/kvm/vmx/vmx.c
> > @@ -6437,6 +6437,10 @@ static void vmx_vcpu_run(struct kvm_vcpu *vcpu)
> > if (vcpu->arch.cr2 != read_cr2())
> > write_cr2(vcpu->arch.cr2);
> >
> > + if (lapic_in_kernel(vcpu) &&
> > + vcpu->arch.apic->lapic_timer.timer_advance_ns)
> > + kvm_wait_lapic_expire(vcpu);
>
> One potential hiccup with this approach is that we're now accessing more
> data after flushing the L1. Not sure if that's actually a problem here,
> but it probably should be explicitly addressed/considered.
Handle it in v3.
Regards,
Wanpeng Li
>
> > +
> > vmx->fail = __vmx_vcpu_run(vmx, (unsigned long *)&vcpu->arch.regs,
> > vmx->loaded_vmcs->launched);
> >
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 1d89cb9..0eb9549 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -7894,9 +7894,6 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
> > }
> >
> > trace_kvm_entry(vcpu->vcpu_id);
> > - if (lapic_in_kernel(vcpu) &&
> > - vcpu->arch.apic->lapic_timer.timer_advance_ns)
> > - wait_lapic_expire(vcpu);
> > guest_enter_irqoff();
> >
> > fpregs_assert_state_consistent();
> > --
> > 2.7.4
> >
