diff --git a/include/linux/irqchip/arm-gic-v3.h b/include/linux/irqchip/arm-gic-v3.h
index c00c4c33e432..b26eccc78fb1 100644
--- a/include/linux/irqchip/arm-gic-v3.h
+++ b/include/linux/irqchip/arm-gic-v3.h
@@ -503,6 +503,7 @@
 
 #define ICH_HCR_EN			(1 << 0)
 #define ICH_HCR_UIE			(1 << 1)
+#define ICH_HCR_NPIE			(1 << 3)
 #define ICH_HCR_TC			(1 << 10)
 #define ICH_HCR_TALL0			(1 << 11)
 #define ICH_HCR_TALL1			(1 << 12)
diff --git a/include/linux/irqchip/arm-gic.h b/include/linux/irqchip/arm-gic.h
index d3453ee072fc..68d8b1f73682 100644
--- a/include/linux/irqchip/arm-gic.h
+++ b/include/linux/irqchip/arm-gic.h
@@ -84,6 +84,7 @@
 
 #define GICH_HCR_EN			(1 << 0)
 #define GICH_HCR_UIE			(1 << 1)
+#define GICH_HCR_NPIE			(1 << 3)
 
 #define GICH_LR_VIRTUALID		(0x3ff << 0)
 #define GICH_LR_PHYSID_CPUID_SHIFT	(10)
diff --git a/virt/kvm/arm/vgic/vgic-v2.c b/virt/kvm/arm/vgic/vgic-v2.c
index e9d840a75e7b..29556f71b691 100644
--- a/virt/kvm/arm/vgic/vgic-v2.c
+++ b/virt/kvm/arm/vgic/vgic-v2.c
@@ -37,6 +37,13 @@ void vgic_v2_init_lrs(void)
 		vgic_v2_write_lr(i, 0);
 }
 
+void vgic_v2_set_npie(struct kvm_vcpu *vcpu)
+{
+	struct vgic_v2_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v2;
+
+	cpuif->vgic_hcr |= GICH_HCR_NPIE;
+}
+
 void vgic_v2_set_underflow(struct kvm_vcpu *vcpu)
 {
 	struct vgic_v2_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v2;
@@ -64,7 +71,7 @@ void vgic_v2_fold_lr_state(struct kvm_vcpu *vcpu)
 	int lr;
 	unsigned long flags;
 
-	cpuif->vgic_hcr &= ~GICH_HCR_UIE;
+	cpuif->vgic_hcr &= ~(GICH_HCR_UIE | GICH_HCR_NPIE);
 
 	for (lr = 0; lr < vgic_cpu->used_lrs; lr++) {
 		u32 val = cpuif->vgic_lr[lr];
diff --git a/virt/kvm/arm/vgic/vgic-v3.c b/virt/kvm/arm/vgic/vgic-v3.c
index 6b329414e57a..0ff2006f3781 100644
--- a/virt/kvm/arm/vgic/vgic-v3.c
+++ b/virt/kvm/arm/vgic/vgic-v3.c
@@ -26,6 +26,13 @@ static bool group1_trap;
 static bool common_trap;
 static bool gicv4_enable;
 
+void vgic_v3_set_npie(struct kvm_vcpu *vcpu)
+{
+	struct vgic_v3_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v3;
+
+	cpuif->vgic_hcr |= ICH_HCR_NPIE;
+}
+
 void vgic_v3_set_underflow(struct kvm_vcpu *vcpu)
 {
 	struct vgic_v3_cpu_if *cpuif = &vcpu->arch.vgic_cpu.vgic_v3;
@@ -47,7 +54,7 @@ void vgic_v3_fold_lr_state(struct kvm_vcpu *vcpu)
 	int lr;
 	unsigned long flags;
 
-	cpuif->vgic_hcr &= ~ICH_HCR_UIE;
+	cpuif->vgic_hcr &= ~(ICH_HCR_UIE | ICH_HCR_NPIE);
 
 	for (lr = 0; lr < vgic_cpu->used_lrs; lr++) {
 		u64 val = cpuif->vgic_lr[lr];
diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
index 0001858a2c23..8201899126f6 100644
--- a/virt/kvm/arm/vgic/vgic.c
+++ b/virt/kvm/arm/vgic/vgic.c
@@ -710,22 +710,37 @@ static inline void vgic_set_underflow(struct kvm_vcpu *vcpu)
 		vgic_v3_set_underflow(vcpu);
 }
 
+static inline void vgic_set_npie(struct kvm_vcpu *vcpu)
+{
+	if (kvm_vgic_global_state.type == VGIC_V2)
+		vgic_v2_set_npie(vcpu);
+	else
+		vgic_v3_set_npie(vcpu);
+}
+
 /* Requires the ap_list_lock to be held. */
-static int compute_ap_list_depth(struct kvm_vcpu *vcpu)
+static int compute_ap_list_depth(struct kvm_vcpu *vcpu,
+				 bool *multi_sgi)
 {
 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 	struct vgic_irq *irq;
 	int count = 0;
 
+	*multi_sgi = false;
+
 	DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 
 	list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 		spin_lock(&irq->irq_lock);
 		/* GICv2 SGIs can count for more than one... */
-		if (vgic_irq_is_sgi(irq->intid) && irq->source)
-			count += hweight8(irq->source);
-		else
+		if (vgic_irq_is_sgi(irq->intid) && irq->source) {
+			int w = hweight8(irq->source);
+
+			count += w;
+			*multi_sgi |= (w > 1);
+		} else {
 			count++;
+		}
 		spin_unlock(&irq->irq_lock);
 	}
 	return count;
@@ -736,28 +751,43 @@ static void vgic_flush_lr_state(struct kvm_vcpu *vcpu)
 {
 	struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
 	struct vgic_irq *irq;
-	int count = 0;
+	int count;
+	bool npie = false;
+	bool multi_sgi;
+	u8 prio = 0xff;
 
 	DEBUG_SPINLOCK_BUG_ON(!spin_is_locked(&vgic_cpu->ap_list_lock));
 
-	if (compute_ap_list_depth(vcpu) > kvm_vgic_global_state.nr_lr)
+	count = compute_ap_list_depth(vcpu, &multi_sgi);
+	if (count > kvm_vgic_global_state.nr_lr || multi_sgi)
 		vgic_sort_ap_list(vcpu);
 
+	count = 0;
+
 	list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
 		spin_lock(&irq->irq_lock);
 
-		if (unlikely(vgic_target_oracle(irq) != vcpu))
-			goto next;
-
 		/*
-		 * If we get an SGI with multiple sources, try to get
-		 * them in all at once.
+		 * If we have multi-SGIs in the pipeline, we need to
+		 * guarantee that they are all seen before any IRQ of
+		 * lower priority. In that case, we need to filter out
+		 * these interrupts by exiting early. This is easy as
+		 * the AP list has been sorted already.
 		 */
-		do {
-			vgic_populate_lr(vcpu, irq, count++);
-		} while (irq->source && count < kvm_vgic_global_state.nr_lr);
+		if (multi_sgi && irq->priority > prio) {
+			spin_unlock(&irq->irq_lock);
+			break;
+		}
+
+		if (likely(vgic_target_oracle(irq) == vcpu)) {
+			vgic_populate_lr(vcpu, irq, count++);
+
+			if (irq->source) {
+				npie = true;
+				prio = irq->priority;
+			}
+		}
 
-next:
 		spin_unlock(&irq->irq_lock);
 
 		if (count == kvm_vgic_global_state.nr_lr) {
@@ -768,6 +798,9 @@ next:
 		}
 	}
 
+	if (npie)
+		vgic_set_npie(vcpu);
+
 	vcpu->arch.vgic_cpu.used_lrs = count;
 
 	/* Nuke remaining LRs */
diff --git a/virt/kvm/arm/vgic/vgic.h b/virt/kvm/arm/vgic/vgic.h
index 5b11859a1a1e..f5b8519e5546 100644
--- a/virt/kvm/arm/vgic/vgic.h
+++ b/virt/kvm/arm/vgic/vgic.h
@@ -160,6 +160,7 @@ void vgic_v2_fold_lr_state(struct kvm_vcpu *vcpu);
 void vgic_v2_populate_lr(struct kvm_vcpu *vcpu, struct vgic_irq *irq, int lr);
 void vgic_v2_clear_lr(struct kvm_vcpu *vcpu, int lr);
 void vgic_v2_set_underflow(struct kvm_vcpu *vcpu);
+void vgic_v2_set_npie(struct kvm_vcpu *vcpu);
 int vgic_v2_has_attr_regs(struct kvm_device *dev, struct kvm_device_attr *attr);
 int vgic_v2_dist_uaccess(struct kvm_vcpu *vcpu, bool is_write,
 			 int offset, u32 *val);
@@ -189,6 +190,7 @@ void vgic_v3_fold_lr_state(struct kvm_vcpu *vcpu);
 void vgic_v3_populate_lr(struct kvm_vcpu *vcpu, struct vgic_irq *irq, int lr);
 void vgic_v3_clear_lr(struct kvm_vcpu *vcpu, int lr);
 void vgic_v3_set_underflow(struct kvm_vcpu *vcpu);
+void vgic_v3_set_npie(struct kvm_vcpu *vcpu);
 void vgic_v3_set_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcr);
 void vgic_v3_get_vmcr(struct kvm_vcpu *vcpu, struct vgic_vmcr *vmcr);
 void vgic_v3_enable(struct kvm_vcpu *vcpu);