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

[Tian, Kevin]

> From: Chao Gao <chao.gao@xxxxxxxxx>
> Sent: Friday, June 17, 2022 7:47 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().
> 
> When IPI virtualization is enabled, this patch increases "perf bench" [*]
> by 4% from 8.12 us/ops to 7.80 us/ops.
> 
> [*] 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@xxxxxxxxx>
> ---
>  arch/x86/kvm/vmx/posted_intr.c | 28 ++--------------------------
>  arch/x86/kvm/vmx/vmx.c         | 24 +++++++++++++++++++++++-
>  2 files changed, 25 insertions(+), 27 deletions(-)
> 
> diff --git a/arch/x86/kvm/vmx/posted_intr.c
> b/arch/x86/kvm/vmx/posted_intr.c
> index 237a1f40f939..a0458f72df99 100644
> --- a/arch/x86/kvm/vmx/posted_intr.c
> +++ b/arch/x86/kvm/vmx/posted_intr.c
> @@ -70,12 +70,6 @@ void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int
> cpu)
>  	 * needs to be changed.
>  	 */
>  	if (pi_desc->nv != POSTED_INTR_WAKEUP_VECTOR && vcpu->cpu ==
> cpu) {
> -		/*
> -		 * Clear SN if it was set due to being preempted.  Again, do
> -		 * this even if there is no assigned device for simplicity.
> -		 */
> -		if (pi_test_and_clear_sn(pi_desc))
> -			goto after_clear_sn;
>  		return;
>  	}
> 
> @@ -99,12 +93,8 @@ void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int
> cpu)
>  	do {
>  		old.control = new.control = READ_ONCE(pi_desc->control);
> 
> -		/*
> -		 * Clear SN (as above) and refresh the destination APIC ID to
> -		 * handle task migration (@cpu != vcpu->cpu).
> -		 */
>  		new.ndst = dest;
> -		new.sn = 0;
> +		new.sn = 1;

A comment is appreciated.

> 
>  		/*
>  		 * Restore the notification vector; in the blocking case, the
> @@ -114,19 +104,6 @@ void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int
> cpu)
>  	} while (pi_try_set_control(pi_desc, old.control, new.control));
> 
>  	local_irq_restore(flags);
> -
> -after_clear_sn:
> -
> -	/*
> -	 * Clear SN before reading the bitmap.  The VT-d firmware
> -	 * writes the bitmap and reads SN atomically (5.2.3 in the
> -	 * spec), so it doesn't really have a memory barrier that
> -	 * pairs with this, but we cannot do that and we need one.
> -	 */
> -	smp_mb__after_atomic();
> -
> -	if (!pi_is_pir_empty(pi_desc))
> -		pi_set_on(pi_desc);
>  }
> 
>  static bool vmx_can_use_vtd_pi(struct kvm *kvm)
> @@ -154,13 +131,12 @@ static void pi_enable_wakeup_handler(struct
> kvm_vcpu *vcpu)
>  		      &per_cpu(wakeup_vcpus_on_cpu, vcpu->cpu));
>  	raw_spin_unlock(&per_cpu(wakeup_vcpus_on_cpu_lock, vcpu-
> >cpu));
> 
> -	WARN(pi_desc->sn, "PI descriptor SN field set before blocking");
> -
>  	do {
>  		old.control = new.control = READ_ONCE(pi_desc->control);
> 
>  		/* set 'NV' to 'wakeup vector' */
>  		new.nv = POSTED_INTR_WAKEUP_VECTOR;
> +		new.sn = 0;
>  	} while (pi_try_set_control(pi_desc, old.control, new.control));
> 

there is a problem a few lines downwards:

	/*
	 * Send a wakeup IPI to this CPU if an interrupt may have been posted
	 * before the notification vector was updated, in which case the IRQ
	 * will arrive on the non-wakeup vector.  An IPI is needed as calling
	 * try_to_wake_up() from ->sched_out() isn't allowed (IRQs are not
	 * enabled until it is safe to call try_to_wake_up() on the task being
	 * scheduled out).
	 */
	if (pi_test_on(&new))
		apic->send_IPI_self(POSTED_INTR_WAKEUP_VECTOR);

'on' is not set when SN is set. This is different from original assumption
which has SN cleared in above window. In this case pi_test_on()
should be replaced with pi_is_pir_empty().

There is another simplification possible in vmx_vcpu_pi_put():

	if (kvm_vcpu_is_blocking(vcpu) && !vmx_interrupt_blocked(vcpu))
		pi_enable_wakeup_handler(vcpu);

	/*
	 * Set SN when the vCPU is preempted.  Note, the vCPU can both be seen
	 * as blocking and preempted, e.g. if it's preempted between setting
	 * its wait state and manually scheduling out.
	 */
	if (vcpu->preempted)
		pi_set_sn(pi_desc);

With this patch 'sn' is already set when a runnable vcpu is preempted
hence above is required only for a blocking vcpu. And in the
blocking case if the notification is anyway suppressed it's pointless to
further change the notification vector. Then it could be simplified as:

	if (!vcpu->preempted && kvm_vcpu_is_blocking(vcpu) &&
		!vmx_interrupt_blocked(vcpu))
		pi_enable_wakeup_handler(vcpu);

>  	/*
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index a3c5504601a8..fa915b1680eb 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -4036,6 +4036,9 @@ static int vmx_deliver_posted_interrupt(struct
> kvm_vcpu *vcpu, int vector)
>  	if (pi_test_and_set_pir(vector, &vmx->pi_desc))
>  		return 0;
> 
> +	if (pi_test_sn(&vmx->pi_desc))
> +		return 0;
> +
>  	/* If a previous notification has sent the IPI, nothing to do.  */
>  	if (pi_test_and_set_on(&vmx->pi_desc))
>  		return 0;
> @@ -6520,8 +6523,17 @@ static int vmx_sync_pir_to_irr(struct kvm_vcpu
> *vcpu)
>  	if (KVM_BUG_ON(!enable_apicv, vcpu->kvm))
>  		return -EIO;
> 
> -	if (pi_test_on(&vmx->pi_desc)) {
> +	if (pi_test_on(&vmx->pi_desc) || pi_test_sn(&vmx->pi_desc)) {

this has potential side-effect in vmexit/vmentry path where pi_desc is
always scanned now. While reducing interrupts help the IPC test case,
do you observe any impact on other scenarios where interrupts are not
the main cause of vmexits?

>  		pi_clear_on(&vmx->pi_desc);
> +
> +		/*
> +		 * IN_GUEST_MODE means we are about to enter vCPU.
> Allow
> +		 * PIN (posted interrupt notification) to deliver is key
> +		 * to interrupt posting. Clear PID.SN.
> +		 */
> +		if (vcpu->mode == IN_GUEST_MODE)
> +			pi_clear_sn(&vmx->pi_desc);

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. 

> +
>  		/*
>  		 * IOMMU can write to PID.ON, so the barrier matters even
> on UP.
>  		 * But on x86 this is just a compiler barrier anyway.
> @@ -6976,6 +6988,16 @@ static fastpath_t vmx_vcpu_run(struct kvm_vcpu
> *vcpu)
>  	/* The actual VMENTER/EXIT is in the .noinstr.text section. */
>  	vmx_vcpu_enter_exit(vcpu, vmx);
> 
> +	/*
> +	 * Suppress notification right after VM exits to minimize the
> +	 * window where VT-d or remote CPU may send a useless notification
> +	 * when posting interrupts to a VM. Note that the notification is
> +	 * useless because KVM syncs pending interrupts in PID.IRR to vAPIC
> +	 * IRR before VM entry.
> +	 */
> +	if (kvm_vcpu_apicv_active(vcpu))
> +		pi_set_sn(&vmx->pi_desc);
> +
>  	/*
>  	 * We do not use IBRS in the kernel. If this vCPU has used the
>  	 * SPEC_CTRL MSR it may have left it on; save the value and
> --
> 2.25.1