KVM: arm64: vgic-v3: Implement MMIO-based LPI invalidation

Since GICv4.1, it has become legal for an implementation to advertise
GICR_{INVLPIR,INVALLR,SYNCR} while having an ITS, allowing for a more
efficient invalidation scheme (no guest command queue contention when
multiple CPUs are generating invalidations).

Provide the invalidation registers as a primitive to their ITS
counterpart. Note that we don't advertise them to the guest yet
(the architecture allows an implementation to do this).

Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Oliver Upton <oupton@google.com>
Link: https://lore.kernel.org/r/20220405182327.205520-4-maz@kernel.org

arch/arm64/kvm/vgic/vgic-its.c

@@ -1272,6 +1272,11 @@ static int vgic_its_cmd_handle_clear(struct kvm *kvm, struct vgic_its *its,
 	return 0;
 }
 
+int vgic_its_inv_lpi(struct kvm *kvm, struct vgic_irq *irq)
+{
+	return update_lpi_config(kvm, irq, NULL, true);
+}
+
 /*
  * The INV command syncs the configuration bits from the memory table.
  * Must be called with the its_lock mutex held.
@@ -1288,7 +1293,41 @@ static int vgic_its_cmd_handle_inv(struct kvm *kvm, struct vgic_its *its,
 	if (!ite)
 		return E_ITS_INV_UNMAPPED_INTERRUPT;
 
-	return update_lpi_config(kvm, ite->irq, NULL, true);
+	return vgic_its_inv_lpi(kvm, ite->irq);
+}
+
+/**
+ * vgic_its_invall - invalidate all LPIs targetting a given vcpu
+ * @vcpu: the vcpu for which the RD is targetted by an invalidation
+ *
+ * Contrary to the INVALL command, this targets a RD instead of a
+ * collection, and we don't need to hold the its_lock, since no ITS is
+ * involved here.
+ */
+int vgic_its_invall(struct kvm_vcpu *vcpu)
+{
+	struct kvm *kvm = vcpu->kvm;
+	int irq_count, i = 0;
+	u32 *intids;
+
+	irq_count = vgic_copy_lpi_list(kvm, vcpu, &intids);
+	if (irq_count < 0)
+		return irq_count;
+
+	for (i = 0; i < irq_count; i++) {
+		struct vgic_irq *irq = vgic_get_irq(kvm, NULL, intids[i]);
+		if (!irq)
+			continue;
+		update_lpi_config(kvm, irq, vcpu, false);
+		vgic_put_irq(kvm, irq);
+	}
+
+	kfree(intids);
+
+	if (vcpu->arch.vgic_cpu.vgic_v3.its_vpe.its_vm)
+		its_invall_vpe(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe);
+
+	return 0;
 }
 
 /*
@@ -1305,32 +1344,13 @@ static int vgic_its_cmd_handle_invall(struct kvm *kvm, struct vgic_its *its,
 	u32 coll_id = its_cmd_get_collection(its_cmd);
 	struct its_collection *collection;
 	struct kvm_vcpu *vcpu;
-	struct vgic_irq *irq;
-	u32 *intids;
-	int irq_count, i;
 
 	collection = find_collection(its, coll_id);
 	if (!its_is_collection_mapped(collection))
 		return E_ITS_INVALL_UNMAPPED_COLLECTION;
 
 	vcpu = kvm_get_vcpu(kvm, collection->target_addr);
-
-	irq_count = vgic_copy_lpi_list(kvm, vcpu, &intids);
-	if (irq_count < 0)
-		return irq_count;
-
-	for (i = 0; i < irq_count; i++) {
-		irq = vgic_get_irq(kvm, NULL, intids[i]);
-		if (!irq)
-			continue;
-		update_lpi_config(kvm, irq, vcpu, false);
-		vgic_put_irq(kvm, irq);
-	}
-
-	kfree(intids);
-
-	if (vcpu->arch.vgic_cpu.vgic_v3.its_vpe.its_vm)
-		its_invall_vpe(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe);
+	vgic_its_invall(vcpu);
 
 	return 0;
 }

arch/arm64/kvm/vgic/vgic-mmio-v3.c

@@ -543,6 +543,63 @@ static void vgic_mmio_write_pendbase(struct kvm_vcpu *vcpu,
 			   pendbaser) != old_pendbaser);
 }
 
+static unsigned long vgic_mmio_read_sync(struct kvm_vcpu *vcpu,
+					 gpa_t addr, unsigned int len)
+{
+	return !!atomic_read(&vcpu->arch.vgic_cpu.syncr_busy);
+}
+
+static void vgic_set_rdist_busy(struct kvm_vcpu *vcpu, bool busy)
+{
+	if (busy) {
+		atomic_inc(&vcpu->arch.vgic_cpu.syncr_busy);
+		smp_mb__after_atomic();
+	} else {
+		smp_mb__before_atomic();
+		atomic_dec(&vcpu->arch.vgic_cpu.syncr_busy);
+	}
+}
+
+static void vgic_mmio_write_invlpi(struct kvm_vcpu *vcpu,
+				   gpa_t addr, unsigned int len,
+				   unsigned long val)
+{
+	struct vgic_irq *irq;
+
+	/*
+	 * If the guest wrote only to the upper 32bit part of the
+	 * register, drop the write on the floor, as it is only for
+	 * vPEs (which we don't support for obvious reasons).
+	 *
+	 * Also discard the access if LPIs are not enabled.
+	 */
+	if ((addr & 4) || !vgic_lpis_enabled(vcpu))
+		return;
+
+	vgic_set_rdist_busy(vcpu, true);
+
+	irq = vgic_get_irq(vcpu->kvm, NULL, lower_32_bits(val));
+	if (irq) {
+		vgic_its_inv_lpi(vcpu->kvm, irq);
+		vgic_put_irq(vcpu->kvm, irq);
+	}
+
+	vgic_set_rdist_busy(vcpu, false);
+}
+
+static void vgic_mmio_write_invall(struct kvm_vcpu *vcpu,
+				   gpa_t addr, unsigned int len,
+				   unsigned long val)
+{
+	/* See vgic_mmio_write_invlpi() for the early return rationale */
+	if ((addr & 4) || !vgic_lpis_enabled(vcpu))
+		return;
+
+	vgic_set_rdist_busy(vcpu, true);
+	vgic_its_invall(vcpu);
+	vgic_set_rdist_busy(vcpu, false);
+}
+
 /*
  * The GICv3 per-IRQ registers are split to control PPIs and SGIs in the
  * redistributors, while SPIs are covered by registers in the distributor
@@ -648,6 +705,15 @@ static const struct vgic_register_region vgic_v3_rd_registers[] = {
 	REGISTER_DESC_WITH_LENGTH(GICR_PENDBASER,
 		vgic_mmio_read_pendbase, vgic_mmio_write_pendbase, 8,
 		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
+	REGISTER_DESC_WITH_LENGTH(GICR_INVLPIR,
+		vgic_mmio_read_raz, vgic_mmio_write_invlpi, 8,
+		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
+	REGISTER_DESC_WITH_LENGTH(GICR_INVALLR,
+		vgic_mmio_read_raz, vgic_mmio_write_invall, 8,
+		VGIC_ACCESS_64bit | VGIC_ACCESS_32bit),
+	REGISTER_DESC_WITH_LENGTH(GICR_SYNCR,
+		vgic_mmio_read_sync, vgic_mmio_write_wi, 4,
+		VGIC_ACCESS_32bit),
 	REGISTER_DESC_WITH_LENGTH(GICR_IDREGS,
 		vgic_mmio_read_v3_idregs, vgic_mmio_write_wi, 48,
 		VGIC_ACCESS_32bit),

arch/arm64/kvm/vgic/vgic.h

@@ -318,6 +318,10 @@ void vgic_lpi_translation_cache_init(struct kvm *kvm);
 void vgic_lpi_translation_cache_destroy(struct kvm *kvm);
 void vgic_its_invalidate_cache(struct kvm *kvm);
 
+/* GICv4.1 MMIO interface */
+int vgic_its_inv_lpi(struct kvm *kvm, struct vgic_irq *irq);
+int vgic_its_invall(struct kvm_vcpu *vcpu);
+
 bool vgic_supports_direct_msis(struct kvm *kvm);
 int vgic_v4_init(struct kvm *kvm);
 void vgic_v4_teardown(struct kvm *kvm);

include/kvm/arm_vgic.h

@@ -344,6 +344,7 @@ struct vgic_cpu {
 	struct vgic_io_device	rd_iodev;
 	struct vgic_redist_region *rdreg;
 	u32 rdreg_index;
+	atomic_t syncr_busy;
 
 	/* Contains the attributes and gpa of the LPI pending tables. */
 	u64 pendbaser;