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: <1563961393-10301-1-git-send-email-wanpengli@tencent.com> <5ffaea5b-fb07-0141-cab8-6dce39071abe@redhat.com>
In-Reply-To: <5ffaea5b-fb07-0141-cab8-6dce39071abe@redhat.com>
From:   Lai Jiangshan <jiangshanlai+lkml@gmail.com>
Date:   Fri, 13 Aug 2021 16:34:28 +0800
Message-ID: <CAJhGHyAY-kN-CYwoq_R2v9067fgjZVOimPOXv_kxzq8aZfFteg@mail.gmail.com>
Subject: Re: [PATCH] KVM: X86: Boost queue head vCPU to mitigate lock waiter preemption
To:     Paolo Bonzini <pbonzini@redhat.com>
Cc:     Wanpeng Li <kernellwp@gmail.com>,
        LKML <linux-kernel@vger.kernel.org>, kvm@vger.kernel.org,
        =?UTF-8?B?UmFkaW0gS3LEjW3DocWZ?= <rkrcmar@redhat.com>,
        Waiman Long <longman@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Wed, Jul 24, 2019 at 9:26 PM Paolo Bonzini <pbonzini@redhat.com> wrote:
>
> On 24/07/19 11:43, Wanpeng Li wrote:
> > From: Wanpeng Li <wanpengli@tencent.com>
> >
> > Commit 11752adb (locking/pvqspinlock: Implement hybrid PV queued/unfair=
 locks)
> > introduces hybrid PV queued/unfair locks
> >  - queued mode (no starvation)
> >  - unfair mode (good performance on not heavily contended lock)
> > The lock waiter goes into the unfair mode especially in VMs with over-c=
ommit
> > vCPUs since increaing over-commitment increase the likehood that the qu=
eue
> > head vCPU may have been preempted and not actively spinning.
> >
> > However, reschedule queue head vCPU timely to acquire the lock still ca=
n get
> > better performance than just depending on lock stealing in over-subscri=
be
> > scenario.
> >
> > Testing on 80 HT 2 socket Xeon Skylake server, with 80 vCPUs VM 80GB RA=
M:
> > ebizzy -M
> >              vanilla     boosting    improved
> >  1VM          23520        25040         6%
> >  2VM           8000        13600        70%
> >  3VM           3100         5400        74%
> >
> > The lock holder vCPU yields to the queue head vCPU when unlock, to boos=
t queue
> > head vCPU which is involuntary preemption or the one which is voluntary=
 halt
> > due to fail to acquire the lock after a short spin in the guest.
>
> Clever!  I have applied the patch.

Hello

I think this patch is very very counter-intuition.  The current vCPU
can now still continue to run, but this patch puts it on hold for a while
via yield_to().  KVM_HC_KICK_CPU is used by spin_unlock() in guest,
what if the guest CPU is in irq or in irq-disabled section, or nested
in other spin_lock(). It could add more latency to these cases.

It is convinced that the test proved the patch.  But I think we need
stronger reasoning between the code and the test (and even more tests)
since it is counter-intuition.  Why the code can boost the tests in
detail. I don't think these:

> The lock holder vCPU yields to the queue head vCPU when unlock, to boost =
queue
> head vCPU which is involuntary preemption or the one which is voluntary h=
alt
> due to fail to acquire the lock after a short spin in the guest.

are enough to explain it to me.  But I'm Okay with if this short
reason can be added to the code to reduce shockness.

At least when I glanced kvm_sched_yield() in case KVM_HC_KICK_CPU, it made
me wonder due to there is no reasoning comment before kvm_sched_yield().

Anyway, I don't object to this patch which also proves altruism is a good
strategy in the world.

Thanks
Lai

>
> Paolo
>
> > Cc: Waiman Long <longman@redhat.com>
> > Cc: Peter Zijlstra <peterz@infradead.org>
> > Cc: Paolo Bonzini <pbonzini@redhat.com>
> > Cc: Radim Kr=C4=8Dm=C3=A1=C5=99 <rkrcmar@redhat.com>
> > Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
> > ---
> >  arch/x86/kvm/x86.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 01e18ca..c6d951c 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -7206,7 +7206,7 @@ static void kvm_sched_yield(struct kvm *kvm, unsi=
gned long dest_id)
> >
> >       rcu_read_unlock();
> >
> > -     if (target)
> > +     if (target && READ_ONCE(target->ready))
> >               kvm_vcpu_yield_to(target);
> >  }
> >
> > @@ -7246,6 +7246,7 @@ int kvm_emulate_hypercall(struct kvm_vcpu *vcpu)
> >               break;
> >       case KVM_HC_KICK_CPU:
> >               kvm_pv_kick_cpu_op(vcpu->kvm, a0, a1);
> > +             kvm_sched_yield(vcpu->kvm, a1);
> >               ret =3D 0;
> >               break;
> >  #ifdef CONFIG_X86_64
> >
>