Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Fri, 15 Oct 2021 23:26:19 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Wanpeng Li <kernellwp@gmail.com>
Cc:     linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        Paolo Bonzini <pbonzini@redhat.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>
Subject: Re: [PATCH v2 3/3] KVM: vCPU kick tax cut for running vCPU
Message-ID: <YWoOG40Ap0Islpu2@google.com>
References: <1633770532-23664-1-git-send-email-wanpengli@tencent.com>
 <1633770532-23664-3-git-send-email-wanpengli@tencent.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1633770532-23664-3-git-send-email-wanpengli@tencent.com>
Precedence: bulk

On Sat, Oct 09, 2021, Wanpeng Li wrote:
> From: Wanpeng Li <wanpengli@tencent.com>
> 
> Sometimes a vCPU kick is following a pending request, even if @vcpu is 
> the running vCPU. It suffers from both rcuwait_wake_up() which has 
> rcu/memory barrier operations and cmpxchg(). Let's check vcpu->wait 
> before rcu_wait_wake_up() and whether @vcpu is the running vCPU before 
> cmpxchg() to tax cut this overhead.
> 
> We evaluate the kvm-unit-test/vmexit.flat on an Intel ICX box, most of the 
> scores can improve ~600 cpu cycles especially when APICv is disabled.
> 
> tscdeadline_immed
> tscdeadline
> self_ipi_sti_nop
> ..............
> x2apic_self_ipi_tpr_sti_hlt
> 
> Suggested-by: Sean Christopherson <seanjc@google.com>
> Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
> ---
> v1 -> v2:
>  * move checking running vCPU logic to kvm_vcpu_kick
>  * check rcuwait_active(&vcpu->wait) etc
> 
>  virt/kvm/kvm_main.c | 13 ++++++++++---
>  1 file changed, 10 insertions(+), 3 deletions(-)
> 
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index 7851f3a1b5f7..18209d7b3711 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -3314,8 +3314,15 @@ void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
>  {
>  	int me, cpu;
>  
> -	if (kvm_vcpu_wake_up(vcpu))
> -		return;
> +	me = get_cpu();
> +
> +	if (rcuwait_active(&vcpu->wait) && kvm_vcpu_wake_up(vcpu))

This needs to use kvm_arch_vcpu_get_wait(), not vcpu->wait, because PPC has some
funky wait stuff.

One potential issue I didn't think of before.  rcuwait_active() comes with the
below warning, which means we might be at risk of a false negative that could
result in a missed wakeup.  I'm not postive on that though.

/*
 * Note: this provides no serialization and, just as with waitqueues,
 * requires care to estimate as to whether or not the wait is active.
 */

> +		goto out;
> +
> +	if (vcpu == __this_cpu_read(kvm_running_vcpu)) {
> +		WARN_ON_ONCE(vcpu->mode == IN_GUEST_MODE);
> +		goto out;
> +	}
>  
>  	/*
>  	 * Note, the vCPU could get migrated to a different pCPU at any point
> @@ -3324,12 +3331,12 @@ void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
>  	 * IPI is to force the vCPU to leave IN_GUEST_MODE, and migrating the
>  	 * vCPU also requires it to leave IN_GUEST_MODE.
>  	 */
> -	me = get_cpu();
>  	if (kvm_arch_vcpu_should_kick(vcpu)) {
>  		cpu = READ_ONCE(vcpu->cpu);
>  		if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
>  			smp_send_reschedule(cpu);
>  	}
> +out:
>  	put_cpu();
>  }
>  EXPORT_SYMBOL_GPL(kvm_vcpu_kick);
> -- 
> 2.25.1
>