xds: mesh gateway CDS requests are now allowed to receive an empty CDS reply (#6787)

This is the rest of the fix for #6543 that was incompletely fixed in #6576.

agent/proxycfg/snapshot.go

@@ -83,7 +83,7 @@ func (s *ConfigSnapshot) Valid() bool {
 	case structs.ServiceKindConnectProxy:
 		return s.Roots != nil && s.ConnectProxy.Leaf != nil
 	case structs.ServiceKindMeshGateway:
-		return s.Roots != nil && (s.MeshGateway.WatchedServicesSet || len(s.MeshGateway.WatchedServices) > 0)
+		return s.Roots != nil && (s.MeshGateway.WatchedServicesSet || len(s.MeshGateway.ServiceGroups) > 0)
 	default:
 		return false
 	}

agent/xds/server.go

@@ -215,6 +215,10 @@ func (s *Server) process(stream ADSStream, reqCh <-chan *envoy.DiscoveryRequest)
 			typeURL:   ClusterType,
 			resources: s.clustersFromSnapshot,
 			stream:    stream,
+			allowEmptyFn: func(cfgSnap *proxycfg.ConfigSnapshot) bool {
+				// Mesh gateways are allowed to inform CDS of no clusters.
+				return cfgSnap.Kind == structs.ServiceKindMeshGateway
+			},
 		},
 		RouteType: &xDSType{
 			typeURL:   RouteType,
@@ -370,8 +374,9 @@ type xDSType struct {
 	// previous request we already responded to and 2) if the proxy rejected the
 	// last version we sent with a Nack then req.VersionInfo will be the older
 	// version it's hanging on to.
-	lastVersion uint64
+	lastVersion  uint64
-	resources   func(cfgSnap *proxycfg.ConfigSnapshot, token string) ([]proto.Message, error)
+	resources    func(cfgSnap *proxycfg.ConfigSnapshot, token string) ([]proto.Message, error)
+	allowEmptyFn func(cfgSnap *proxycfg.ConfigSnapshot) bool
 }
 
 func (t *xDSType) Recv(req *envoy.DiscoveryRequest) {
@@ -392,13 +397,16 @@ func (t *xDSType) SendIfNew(cfgSnap *proxycfg.ConfigSnapshot, version uint64, no
 	if err != nil {
 		return err
 	}
+
+	allowEmpty := t.allowEmptyFn != nil && t.allowEmptyFn(cfgSnap)
+
 	// Zero length resource responses should be ignored and are the result of no
 	// data yet. Notice that this caused a bug originally where we had zero
 	// healthy endpoints for an upstream that would cause Envoy to hang waiting
 	// for the EDS response. This is fixed though by ensuring we send an explicit
 	// empty LoadAssignment resource for the cluster rather than allowing junky
 	// empty resources.
-	if len(resources) == 0 {
+	if len(resources) == 0 && !allowEmpty {
 		// Nothing to send yet
 		return nil
 	}

test/integration/connect/envoy/case-gateway-without-services/bind.hcl

@@ -0,0 +1,2 @@
+bind_addr = "0.0.0.0"
+advertise_addr = "{{ GetInterfaceIP \"eth0\" }}"

test/integration/connect/envoy/case-gateway-without-services/gateway.hcl

@@ -0,0 +1,5 @@
+services {
+  name = "mesh-gateway"
+  kind = "mesh-gateway"
+  port = 4431
+}

test/integration/connect/envoy/case-gateway-without-services/s1.hcl

@@ -0,0 +1 @@
+# We don't want an s1 service

test/integration/connect/envoy/case-gateway-without-services/s2.hcl

@@ -0,0 +1 @@
+# We don't want an s2 service

test/integration/connect/envoy/case-gateway-without-services/setup.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+
+set -eEuo pipefail
+
+gen_envoy_bootstrap mesh-gateway 19000 primary true

test/integration/connect/envoy/case-gateway-without-services/vars.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+
+export REQUIRED_SERVICES="gateway-primary"

test/integration/connect/envoy/case-gateway-without-services/verify.bats

@@ -0,0 +1,11 @@
+#!/usr/bin/env bats
+
+load helpers
+
+@test "gateway-primary proxy admin is up on :19000" {
+  retry_default curl -f -s localhost:19000/stats -o /dev/null
+}
+
+@test "gateway-primary listener is up on :4431" {
+  retry_default nc -z localhost:4431
+}