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:   Tue, 27 Jul 2021 01:25:37 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     "Li,Rongqing" <lirongqing@baidu.com>
Cc:     Wanpeng Li <kernellwp@gmail.com>,
        Paolo Bonzini <pbonzini@redhat.com>,
        Ingo Molnar <mingo@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        kvm <kvm@vger.kernel.org>, LKML <linux-kernel@vger.kernel.org>
Subject: Re: =?utf-8?B?562U5aSN?= =?utf-8?Q?=3A?= [PATCH] KVM: Consider SMT
 idle status when halt polling
Message-ID: <YP9gkSk+CHdKLP/Q@google.com>
References: <20210722035807.36937-1-lirongqing@baidu.com>
 <CANRm+Cx-5Yyxx5A4+qkYa01MG4BCdwXPd++bmxzOid+XL267cQ@mail.gmail.com>
 <4efe4fdb91b747da93d7980c10d016c9@baidu.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <4efe4fdb91b747da93d7980c10d016c9@baidu.com>
Precedence: bulk

On Thu, Jul 22, 2021, Li,Rongqing wrote:
> > > > SMT siblings share caches and other hardware, halt polling will
> > > > degrade its sibling performance if its sibling is busy
> > >
> > > Do you have any real scenario benefits? As the polling nature, some
> > > cloud providers will configure to their preferred balance of cpu usage
> > > and performance, and other cloud providers for their NFV scenarios
> > > which are more sensitive to latency are vCPU and pCPU 1:1 pin，you
> > > destroy these setups.
> > >
> > >     Wanpeng
> > 
> 
> 
> Run a copy (single thread) Unixbench, with or without a busy poll program in
> its SMT sibling,  and Unixbench score can lower 1/3 with SMT busy polling
> program

Rather than disallowing halt-polling entirely, on x86 it should be sufficient to
simply have the hardware thread yield to its sibling(s) via PAUSE.  It probably
won't get back all performance, but I would expect it to be close. 

This compiles on all KVM architectures, and AFAICT the intended usage of
cpu_relax() is identical for all architectures.

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 6980dabe9df5..a07ecb3c67fb 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -3111,6 +3111,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
                                goto out;
                        }
                        poll_end = cur = ktime_get();
+                       cpu_relax();
                } while (kvm_vcpu_can_poll(cur, stop));
        }