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;
MIME-Version: 1.0
References: <1618542490-14756-1-git-send-email-wanpengli@tencent.com> <9c49c6ff-d896-e6a5-c051-b6707f6ec58a@redhat.com>
In-Reply-To: <9c49c6ff-d896-e6a5-c051-b6707f6ec58a@redhat.com>
From:   Wanpeng Li <kernellwp@gmail.com>
Date:   Mon, 19 Apr 2021 15:34:53 +0800
Message-ID: <CANRm+Cy-xmDRQoUfOYm+GGvWiS+qC_sBjyZmcLykbKqTF2YDxQ@mail.gmail.com>
Subject: Re: [PATCH] KVM: Boost vCPU candidiate in user mode which is
 delivering interrupt
To:     Paolo Bonzini <pbonzini@redhat.com>
Cc:     LKML <linux-kernel@vger.kernel.org>, kvm <kvm@vger.kernel.org>,
        Sean Christopherson <seanjc@google.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Sat, 17 Apr 2021 at 21:09, Paolo Bonzini <pbonzini@redhat.com> wrote:
>
> On 16/04/21 05:08, Wanpeng Li wrote:
> > From: Wanpeng Li <wanpengli@tencent.com>
> >
> > Both lock holder vCPU and IPI receiver that has halted are condidate for
> > boost. However, the PLE handler was originally designed to deal with the
> > lock holder preemption problem. The Intel PLE occurs when the spinlock
> > waiter is in kernel mode. This assumption doesn't hold for IPI receiver,
> > they can be in either kernel or user mode. the vCPU candidate in user mode
> > will not be boosted even if they should respond to IPIs. Some benchmarks
> > like pbzip2, swaptions etc do the TLB shootdown in kernel mode and most
> > of the time they are running in user mode. It can lead to a large number
> > of continuous PLE events because the IPI sender causes PLE events
> > repeatedly until the receiver is scheduled while the receiver is not
> > candidate for a boost.
> >
> > This patch boosts the vCPU candidiate in user mode which is delivery
> > interrupt. We can observe the speed of pbzip2 improves 10% in 96 vCPUs
> > VM in over-subscribe scenario (The host machine is 2 socket, 48 cores,
> > 96 HTs Intel CLX box). There is no performance regression for other
> > benchmarks like Unixbench spawn (most of the time contend read/write
> > lock in kernel mode), ebizzy (most of the time contend read/write sem
> > and TLB shoodtdown in kernel mode).
> >
> > +bool kvm_arch_interrupt_delivery(struct kvm_vcpu *vcpu)
> > +{
> > +     if (vcpu->arch.apicv_active && static_call(kvm_x86_dy_apicv_has_pending_interrupt)(vcpu))
> > +             return true;
> > +
> > +     return false;
> > +}
>
> Can you reuse vcpu_dy_runnable instead of this new function?

I have some concerns. For x86 arch, vcpu_dy_runnable() will add extra
vCPU candidates by KVM_REQ_EVENT and async pf(which has already
opportunistically made the guest do other stuff). For other arches,
kvm_arch_dy_runnale() is equal to kvm_arch_vcpu_runnable() except
powerpc which has too many events and is not conservative. In general,
 vcpu_dy_runnable() will loose the conditions and add more vCPU
candidates.

    Wanpeng