diff --git a/felix/bpf-gpl/tc.c b/felix/bpf-gpl/tc.c
index 98b1c13ddec..ecd532af2a6 100644
--- a/felix/bpf-gpl/tc.c
+++ b/felix/bpf-gpl/tc.c
@@ -55,6 +55,12 @@ int calico_tc_main(struct __sk_buff *skb)
 	 */
 	skb->mark = SKB_MARK;
 #endif
+
+	if (CALI_F_LO && CALI_F_TO_HOST) {
+		/* Do nothing, it is a packet that just looped around. */
+		return TC_ACT_UNSPEC;
+	}
+
 	/* Optimisation: if another BPF program has already pre-approved the packet,
 	 * skip all processing. */
 	if (CALI_F_FROM_HOST && skb->mark == CALI_SKB_MARK_BYPASS) {
@@ -263,6 +269,16 @@ static CALI_BPF_INLINE void calico_tc_process_ct_lookup(struct cali_tc_ctx *ctx)
 		  skb_mark_equals(ctx->skb, CALI_SKB_MARK_BYPASS_MASK, CALI_SKB_MARK_SKIP_FIB)));
 
 	if (HAS_HOST_CONFLICT_PROG &&
+			/* Do not do conflict resolution for host-self loop. Unlike with
+			 * traffic to another backend, we are not able to tell traffic to
+			 * self via service from straight to self.
+			 */
+			!CALI_F_LO &&
+			/* Do conflict resolution on other device if it clashes with
+			 * traffic looped via the NAT_IF but it hasn't been seen yet and
+			 * is not looped via the NAT_IF, that is, it is from host, but not
+			 * to a service.
+			 */
 			(ctx->state->ct_result.flags & CALI_CT_FLAG_VIA_NAT_IF) &&
 			!(ctx->skb->mark & (CALI_SKB_MARK_FROM_NAT_IFACE_OUT | CALI_SKB_MARK_SEEN))) {
 		CALI_DEBUG("Host source SNAT conflict\n");
diff --git a/felix/fv/bpf_test.go b/felix/fv/bpf_test.go
index 90dc9cb4b8f..d0aa7af415a 100644
--- a/felix/fv/bpf_test.go
+++ b/felix/fv/bpf_test.go
@@ -4257,6 +4257,35 @@ func describeBPFTests(opts ...bpfTestOpt) bool {
 				})
 			})
 
+			It("should have connectivity from host-networked pods via service to host-networked backend", func() {
+				By("Setting up the service")
+				hostW[0].ConfigureInInfra(infra)
+				testSvc := k8sService("host-svc", clusterIP, hostW[0], 80, 8055, 0, testOpts.protocol)
+				testSvcNamespace := testSvc.ObjectMeta.Namespace
+				k8sClient := infra.(*infrastructure.K8sDatastoreInfra).K8sClient
+				_, err := k8sClient.CoreV1().Services(testSvcNamespace).Create(context.Background(), testSvc, metav1.CreateOptions{})
+				Expect(err).NotTo(HaveOccurred())
+				Eventually(k8sGetEpsForServiceFunc(k8sClient, testSvc), "10s").Should(HaveLen(1),
+					"Service endpoints didn't get created? Is controller-manager happy?")
+
+				By("Testing connectivity")
+				port := uint16(testSvc.Spec.Ports[0].Port)
+
+				hostW0SrcIP := ExpectWithSrcIPs(felixIP(0))
+				hostW1SrcIP := ExpectWithSrcIPs(felixIP(1))
+				if !testOpts.connTimeEnabled {
+					switch testOpts.tunnel {
+					case "ipip":
+						hostW0SrcIP = ExpectWithSrcIPs(tc.Felixes[0].ExpectedIPIPTunnelAddr)
+						hostW1SrcIP = ExpectWithSrcIPs(tc.Felixes[1].ExpectedIPIPTunnelAddr)
+					}
+				}
+
+				cc.Expect(Some, hostW[0], TargetIP(clusterIP), ExpectWithPorts(port), hostW0SrcIP)
+				cc.Expect(Some, hostW[1], TargetIP(clusterIP), ExpectWithPorts(port), hostW1SrcIP)
+				cc.CheckConnectivity()
+			})
+
 		})
 
 		Describe("with BPF disabled to begin with", func() {