答复: [PATCH] KVM: Clear pv eoi pending bit only when it is set

["Li,Rongqing" <lirongqing@baidu.com>]

> -----邮件原件-----
> 发件人: Vitaly Kuznetsov <vkuznets@redhat.com>
> 发送时间: 2021年10月19日 15:24
> 收件人: Li,Rongqing <lirongqing@baidu.com>
> 抄送: Li,Rongqing <lirongqing@baidu.com>; pbonzini@redhat.com;
> seanjc@google.com; wanpengli@tencent.com; jmattson@google.com;
> joro@8bytes.org; tglx@linutronix.de; mingo@redhat.com; bp@alien8.de;
> x86@kernel.org; hpa@zytor.com; kvm@vger.kernel.org
> 主题: Re: [PATCH] KVM: Clear pv eoi pending bit only when it is set
> 
> Li RongQing <lirongqing@baidu.com> writes:
> 
> > clear pv eoi pending bit only when it is set, to avoid calling
> > pv_eoi_put_user()
> >
> > and this can speed pv_eoi_clr_pending about 300 nsec on AMD EPYC most
> > of the time
> >
> > Signed-off-by: Li RongQing <lirongqing@baidu.com>
> > ---
> >  arch/x86/kvm/lapic.c |    7 ++++---
> >  1 files changed, 4 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c index
> > 76fb009..c434f70 100644
> > --- a/arch/x86/kvm/lapic.c
> > +++ b/arch/x86/kvm/lapic.c
> > @@ -694,9 +694,9 @@ static void pv_eoi_set_pending(struct kvm_vcpu
> *vcpu)
> >  	__set_bit(KVM_APIC_PV_EOI_PENDING, &vcpu->arch.apic_attention);  }
> >
> > -static void pv_eoi_clr_pending(struct kvm_vcpu *vcpu)
> > +static void pv_eoi_clr_pending(struct kvm_vcpu *vcpu, bool pending)
> 
> Nitpick (and probably a matter of personal taste): pv_eoi_clr_pending() has only
> one user and the change doesn't make its interface much nicer, I'd suggest we
> just inline in instead. (we can probably do the same to
> pv_eoi_get_pending()/pv_eoi_set_pending() too).
> 
> >  {
> > -	if (pv_eoi_put_user(vcpu, KVM_PV_EOI_DISABLED) < 0) {
> > +	if (pending && pv_eoi_put_user(vcpu, KVM_PV_EOI_DISABLED) < 0) {
> >  		printk(KERN_WARNING "Can't clear EOI MSR value: 0x%llx\n",
> >  			   (unsigned long long)vcpu->arch.pv_eoi.msr_val);
> >  		return;
> > @@ -2693,7 +2693,8 @@ static void apic_sync_pv_eoi_from_guest(struct
> kvm_vcpu *vcpu,
> >  	 * While this might not be ideal from performance point of view,
> >  	 * this makes sure pv eoi is only enabled when we know it's safe.
> >  	 */
> > -	pv_eoi_clr_pending(vcpu);
> > +	pv_eoi_clr_pending(vcpu, pending);
> > +
> >  	if (pending)
> >  		return;
> >  	vector = apic_set_eoi(apic);
> 
> Could you probably elaborate a bit (probably by enhancing the comment above
> pv_eoi_clr_pending()) why the race we have here (even before the
> patch) doesn't matter? As far as I understand it, the guest can change PV EOI
> status from a different CPU (it shouldn't do it but it still can) at any time: e.g.
> between pv_eoi_get_pending() and pv_eoi_clr_pending() but also right after we
> do pv_eoi_clr_pending() so the patch doesn't really change much in this regard.
> 

Is it reasonable that the guest change PV EOI status from a different CPU?  I think this can lead to guest error or stuck

And new function pv_eoi_test_and_clear_pending and kvm_test_and_clear_bit_guest_cached should be able to fix the race

I will send V2

Thanks

-Li



> --
> Vitaly