RE: [RFC v2] KVM: x86/vmx: Suppress posted interrupt notification when CPU is in host

From: "Tian, Kevin" <kevin.tian@intel.com>
To: "Gao, Chao" <chao.gao@intel.com>,
	"kvm@vger.kernel.org" <kvm@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"Kohler, Jon" <jon@nutanix.com>
Cc: "Christopherson,, Sean" <seanjc@google.com>,
	Paolo Bonzini <pbonzini@redhat.com>,
	Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
	Dave Hansen <dave.hansen@linux.intel.com>,
	"x86@kernel.org" <x86@kernel.org>,
	"H. Peter Anvin" <hpa@zytor.com>
Subject: RE: [RFC v2] KVM: x86/vmx: Suppress posted interrupt notification when CPU is in host
Date: Wed, 28 Sep 2022 08:29:48 +0000	[thread overview]
Message-ID: <BL1PR11MB52712631CBBED273F5E4E1828C549@BL1PR11MB5271.namprd11.prod.outlook.com> (raw)
In-Reply-To: <20220923085806.384344-1-chao.gao@intel.com>

> From: Gao, Chao <chao.gao@intel.com>
> Sent: Friday, September 23, 2022 4:58 PM
> 
> PIN (Posted interrupt notification) is useless to host as KVM always syncs
> pending guest interrupts in PID to guest's vAPIC before each VM entry. In
> fact, Sending PINs to a CPU running in host will lead to additional
> overhead due to interrupt handling.
> 
> Currently, software path, vmx_deliver_posted_interrupt(), is optimized to
> issue PINs only if target vCPU is in IN_GUEST_MODE. But hardware paths
> (VT-d and Intel IPI virtualization) aren't optimized.
> 
> Set PID.SN right after VM exits and clear it before VM entry to minimize
> the chance of hardware issuing PINs to a CPU when it's in host.
> 
> Also honour PID.SN bit in vmx_deliver_posted_interrupt().
> 
> Opportunistically clean up vmx_vcpu_pi_put(); when a vCPU is preempted,
> it is pointless to update PID.NV to wakeup vector since notification is
> anyway suppressed. And since PID.SN should be already set for running
> vCPUs, so, don't set it again for preempted vCPUs.
> 
> When IPI virtualization is enabled, this patch increases "perf bench" [*]
> by 6.56%, and PIN count in 1 second drops from tens of thousands to
> hundreds. But cpuid loop test shows this patch causes 1.58% overhead in
> VM-exit round-trip latency.
> 
> [*] test cmd: perf bench sched pipe -T. Note that we change the source
> code to pin two threads to two different vCPUs so that it can reproduce
> stable results.
> 
> Signed-off-by: Chao Gao <chao.gao@intel.com>
> ---
> RFC: I am not sure whether the benefits outweighs the extra VM-exit cost.
> 
> Changes in v2 (addressed comments from Kevin):
> - measure/estimate the impact to non-IPC-intensive cases
> - don't tie PID.SN to vcpu->mode. Instead, clear PID.SN
>   right before VM-entry and set it after VM-exit.

One correction here. My comment in v1 [1] was actually close to Sean's
suggestion, i.e. opposite to above description:

--
  I wonder whether it makes more sense to have 'sn' closely sync-ed
  with vcpu->mode, e.g. having a kvm_x86_set_vcpu_mode() ops
  to translate vcpu->mode into vmx/svm specific hardware bits like
  'sn' here. Then call it in common place when vcpu->mode is changed.
--

[1] https://lore.kernel.org/lkml/BN9PR11MB52766B74ADFBAEC0AA205E298CB39@BN9PR11MB5276.namprd11.prod.outlook.com/