test/cpp/end2end/client_lb_end2end_test.cc

   1 /*
   2  *
   3  * Copyright 2016 gRPC authors.
   4  *
   5  * Licensed under the Apache License, Version 2.0 (the "License");
   6  * you may not use this file except in compliance with the License.
   7  * You may obtain a copy of the License at
   8  *
   9  *     http://www.apache.org/licenses/LICENSE-2.0
  10  *
  11  * Unless required by applicable law or agreed to in writing, software
  12  * distributed under the License is distributed on an "AS IS" BASIS,
  13  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  14  * See the License for the specific language governing permissions and
  15  * limitations under the License.
  16  *
  17  */
  18
  19 #include <algorithm>
  20 #include <memory>
  21 #include <mutex>
  22 #include <random>
  23 #include <set>
  24 #include <thread>
  25
  26 #include <grpc/grpc.h>
  27 #include <grpc/support/alloc.h>
  28 #include <grpc/support/atm.h>
  29 #include <grpc/support/log.h>
  30 #include <grpc/support/string_util.h>
  31 #include <grpc/support/time.h>
  32 #include <grpcpp/channel.h>
  33 #include <grpcpp/client_context.h>
  34 #include <grpcpp/create_channel.h>
  35 #include <grpcpp/health_check_service_interface.h>
  36 #include <grpcpp/impl/codegen/sync.h>
  37 #include <grpcpp/server.h>
  38 #include <grpcpp/server_builder.h>
  39
  40 #include "src/core/ext/filters/client_channel/backup_poller.h"
  41 #include "src/core/ext/filters/client_channel/global_subchannel_pool.h"
  42 #include "src/core/ext/filters/client_channel/parse_address.h"
  43 #include "src/core/ext/filters/client_channel/resolver/fake/fake_resolver.h"
  44 #include "src/core/ext/filters/client_channel/server_address.h"
  45 #include "src/core/ext/filters/client_channel/service_config.h"
  46 #include "src/core/lib/backoff/backoff.h"
  47 #include "src/core/lib/channel/channel_args.h"
  48 #include "src/core/lib/gpr/env.h"
  49 #include "src/core/lib/gprpp/debug_location.h"
  50 #include "src/core/lib/gprpp/ref_counted_ptr.h"
  51 #include "src/core/lib/iomgr/tcp_client.h"
  52 #include "src/core/lib/security/credentials/fake/fake_credentials.h"
  53 #include "src/cpp/client/secure_credentials.h"
  54 #include "src/cpp/server/secure_server_credentials.h"
  55
  56 #include "src/proto/grpc/lb/v2/orca_load_report_for_test.pb.h"
  57 #include "src/proto/grpc/testing/echo.grpc.pb.h"
  58 #include "test/core/util/port.h"
  59 #include "test/core/util/test_config.h"
  60 #include "test/core/util/test_lb_policies.h"
  61 #include "test/cpp/end2end/test_service_impl.h"
  62
  63 #include <gmock/gmock.h>
  64 #include <gtest/gtest.h>
  65
  66 using grpc::testing::EchoRequest;
  67 using grpc::testing::EchoResponse;
  68 using std::chrono::system_clock;
  69
  70 // defined in tcp_client.cc
  71 extern grpc_tcp_client_vtable* grpc_tcp_client_impl;
  72
  73 static grpc_tcp_client_vtable* default_client_impl;
  74
  75 namespace grpc {
  76 namespace testing {
  77 namespace {
  78
  79 gpr_atm g_connection_delay_ms;
  80
  81 void tcp_client_connect_with_delay(grpc_closure* closure, grpc_endpoint** ep,
  82                                    grpc_pollset_set* interested_parties,
  83                                    const grpc_channel_args* channel_args,
  84                                    const grpc_resolved_address* addr,
  85                                    grpc_millis deadline) {
  86   const int delay_ms = gpr_atm_acq_load(&g_connection_delay_ms);
  87   if (delay_ms > 0) {
  88     gpr_sleep_until(grpc_timeout_milliseconds_to_deadline(delay_ms));
  89   }
  90   default_client_impl->connect(closure, ep, interested_parties, channel_args,
  91                                addr, deadline + delay_ms);
  92 }
  93
  94 grpc_tcp_client_vtable delayed_connect = {tcp_client_connect_with_delay};
  95
  96 // Subclass of TestServiceImpl that increments a request counter for
  97 // every call to the Echo RPC.
  98 class MyTestServiceImpl : public TestServiceImpl {
  99  public:
 100   Status Echo(ServerContext* context, const EchoRequest* request,
 101               EchoResponse* response) override {
 102     const udpa::data::orca::v1::OrcaLoadReport* load_report = nullptr;
 103     {
 104       grpc::internal::MutexLock lock(&mu_);
 105       ++request_count_;
 106       load_report = load_report_;
 107     }
 108     AddClient(context->peer());
 109     if (load_report != nullptr) {
 110       // TODO(roth): Once we provide a more standard server-side API for
 111       // populating this data, use that API here.
 112       context->AddTrailingMetadata("x-endpoint-load-metrics-bin",
 113                                    load_report->SerializeAsString());
 114     }
 115     return TestServiceImpl::Echo(context, request, response);
 116   }
 117
 118   int request_count() {
 119     grpc::internal::MutexLock lock(&mu_);
 120     return request_count_;
 121   }
 122
 123   void ResetCounters() {
 124     grpc::internal::MutexLock lock(&mu_);
 125     request_count_ = 0;
 126   }
 127
 128   std::set<grpc::string> clients() {
 129     grpc::internal::MutexLock lock(&clients_mu_);
 130     return clients_;
 131   }
 132
 133   void set_load_report(udpa::data::orca::v1::OrcaLoadReport* load_report) {
 134     grpc::internal::MutexLock lock(&mu_);
 135     load_report_ = load_report;
 136   }
 137
 138  private:
 139   void AddClient(const grpc::string& client) {
 140     grpc::internal::MutexLock lock(&clients_mu_);
 141     clients_.insert(client);
 142   }
 143
 144   grpc::internal::Mutex mu_;
 145   int request_count_ = 0;
 146   const udpa::data::orca::v1::OrcaLoadReport* load_report_ = nullptr;
 147   grpc::internal::Mutex clients_mu_;
 148   std::set<grpc::string> clients_;
 149 };
 150
 151 class FakeResolverResponseGeneratorWrapper {
 152  public:
 153   FakeResolverResponseGeneratorWrapper()
 154       : response_generator_(grpc_core::MakeRefCounted<
 155                             grpc_core::FakeResolverResponseGenerator>()) {}
 156
 157   FakeResolverResponseGeneratorWrapper(
 158       FakeResolverResponseGeneratorWrapper&& other) {
 159     response_generator_ = std::move(other.response_generator_);
 160   }
 161
 162   void SetNextResolution(const std::vector<int>& ports,
 163                          const char* service_config_json = nullptr) {
 164     grpc_core::ExecCtx exec_ctx;
 165     response_generator_->SetResponse(
 166         BuildFakeResults(ports, service_config_json));
 167   }
 168
 169   void SetNextResolutionUponError(const std::vector<int>& ports) {
 170     grpc_core::ExecCtx exec_ctx;
 171     response_generator_->SetReresolutionResponse(BuildFakeResults(ports));
 172   }
 173
 174   void SetFailureOnReresolution() {
 175     grpc_core::ExecCtx exec_ctx;
 176     response_generator_->SetFailureOnReresolution();
 177   }
 178
 179   grpc_core::FakeResolverResponseGenerator* Get() const {
 180     return response_generator_.get();
 181   }
 182
 183  private:
 184   static grpc_core::Resolver::Result BuildFakeResults(
 185       const std::vector<int>& ports,
 186       const char* service_config_json = nullptr) {
 187     grpc_core::Resolver::Result result;
 188     for (const int& port : ports) {
 189       char* lb_uri_str;
 190       gpr_asprintf(&lb_uri_str, "ipv4:127.0.0.1:%d", port);
 191       grpc_uri* lb_uri = grpc_uri_parse(lb_uri_str, true);
 192       GPR_ASSERT(lb_uri != nullptr);
 193       grpc_resolved_address address;
 194       GPR_ASSERT(grpc_parse_uri(lb_uri, &address));
 195       result.addresses.emplace_back(address.addr, address.len,
 196                                     nullptr /* args */);
 197       grpc_uri_destroy(lb_uri);
 198       gpr_free(lb_uri_str);
 199     }
 200     if (service_config_json != nullptr) {
 201       result.service_config = grpc_core::ServiceConfig::Create(
 202           service_config_json, &result.service_config_error);
 203       GPR_ASSERT(result.service_config != nullptr);
 204     }
 205     return result;
 206   }
 207
 208   grpc_core::RefCountedPtr<grpc_core::FakeResolverResponseGenerator>
 209       response_generator_;
 210 };
 211
 212 class ClientLbEnd2endTest : public ::testing::Test {
 213  protected:
 214   ClientLbEnd2endTest()
 215       : server_host_("localhost"),
 216         kRequestMessage_("Live long and prosper."),
 217         creds_(new SecureChannelCredentials(
 218             grpc_fake_transport_security_credentials_create())) {}
 219
 220   static void SetUpTestCase() {
 221     // Make the backup poller poll very frequently in order to pick up
 222     // updates from all the subchannels's FDs.
 223     GPR_GLOBAL_CONFIG_SET(grpc_client_channel_backup_poll_interval_ms, 1);
 224 #if TARGET_OS_IPHONE
 225     // Workaround Apple CFStream bug
 226     gpr_setenv("grpc_cfstream", "0");
 227 #endif
 228   }
 229
 230   void SetUp() override { grpc_init(); }
 231
 232   void TearDown() override {
 233     for (size_t i = 0; i < servers_.size(); ++i) {
 234       servers_[i]->Shutdown();
 235     }
 236     // Explicitly destroy all the members so that we can make sure grpc_shutdown
 237     // has finished by the end of this function, and thus all the registered
 238     // LB policy factories are removed.
 239     servers_.clear();
 240     creds_.reset();
 241     grpc_shutdown_blocking();
 242   }
 243
 244   void CreateServers(size_t num_servers,
 245                      std::vector<int> ports = std::vector<int>()) {
 246     servers_.clear();
 247     for (size_t i = 0; i < num_servers; ++i) {
 248       int port = 0;
 249       if (ports.size() == num_servers) port = ports[i];
 250       servers_.emplace_back(new ServerData(port));
 251     }
 252   }
 253
 254   void StartServer(size_t index) { servers_[index]->Start(server_host_); }
 255
 256   void StartServers(size_t num_servers,
 257                     std::vector<int> ports = std::vector<int>()) {
 258     CreateServers(num_servers, std::move(ports));
 259     for (size_t i = 0; i < num_servers; ++i) {
 260       StartServer(i);
 261     }
 262   }
 263
 264   std::vector<int> GetServersPorts(size_t start_index = 0) {
 265     std::vector<int> ports;
 266     for (size_t i = start_index; i < servers_.size(); ++i) {
 267       ports.push_back(servers_[i]->port_);
 268     }
 269     return ports;
 270   }
 271
 272   FakeResolverResponseGeneratorWrapper BuildResolverResponseGenerator() {
 273     return FakeResolverResponseGeneratorWrapper();
 274   }
 275
 276   std::unique_ptr<grpc::testing::EchoTestService::Stub> BuildStub(
 277       const std::shared_ptr<Channel>& channel) {
 278     return grpc::testing::EchoTestService::NewStub(channel);
 279   }
 280
 281   std::shared_ptr<Channel> BuildChannel(
 282       const grpc::string& lb_policy_name,
 283       const FakeResolverResponseGeneratorWrapper& response_generator,
 284       ChannelArguments args = ChannelArguments()) {
 285     if (lb_policy_name.size() > 0) {
 286       args.SetLoadBalancingPolicyName(lb_policy_name);
 287     }  // else, default to pick first
 288     args.SetPointer(GRPC_ARG_FAKE_RESOLVER_RESPONSE_GENERATOR,
 289                     response_generator.Get());
 290     return ::grpc::CreateCustomChannel("fake:///", creds_, args);
 291   }
 292
 293   bool SendRpc(
 294       const std::unique_ptr<grpc::testing::EchoTestService::Stub>& stub,
 295       EchoResponse* response = nullptr, int timeout_ms = 1000,
 296       Status* result = nullptr, bool wait_for_ready = false) {
 297     const bool local_response = (response == nullptr);
 298     if (local_response) response = new EchoResponse;
 299     EchoRequest request;
 300     request.set_message(kRequestMessage_);
 301     ClientContext context;
 302     context.set_deadline(grpc_timeout_milliseconds_to_deadline(timeout_ms));
 303     if (wait_for_ready) context.set_wait_for_ready(true);
 304     Status status = stub->Echo(&context, request, response);
 305     if (result != nullptr) *result = status;
 306     if (local_response) delete response;
 307     return status.ok();
 308   }
 309
 310   void CheckRpcSendOk(
 311       const std::unique_ptr<grpc::testing::EchoTestService::Stub>& stub,
 312       const grpc_core::DebugLocation& location, bool wait_for_ready = false) {
 313     EchoResponse response;
 314     Status status;
 315     const bool success =
 316         SendRpc(stub, &response, 2000, &status, wait_for_ready);
 317     ASSERT_TRUE(success) << "From " << location.file() << ":" << location.line()
 318                          << "\n"
 319                          << "Error: " << status.error_message() << " "
 320                          << status.error_details();
 321     ASSERT_EQ(response.message(), kRequestMessage_)
 322         << "From " << location.file() << ":" << location.line();
 323     if (!success) abort();
 324   }
 325
 326   void CheckRpcSendFailure(
 327       const std::unique_ptr<grpc::testing::EchoTestService::Stub>& stub) {
 328     const bool success = SendRpc(stub);
 329     EXPECT_FALSE(success);
 330   }
 331
 332   struct ServerData {
 333     int port_;
 334     std::unique_ptr<Server> server_;
 335     MyTestServiceImpl service_;
 336     std::unique_ptr<std::thread> thread_;
 337     bool server_ready_ = false;
 338     bool started_ = false;
 339
 340     explicit ServerData(int port = 0) {
 341       port_ = port > 0 ? port : grpc_pick_unused_port_or_die();
 342     }
 343
 344     void Start(const grpc::string& server_host) {
 345       gpr_log(GPR_INFO, "starting server on port %d", port_);
 346       started_ = true;
 347       grpc::internal::Mutex mu;
 348       grpc::internal::MutexLock lock(&mu);
 349       grpc::internal::CondVar cond;
 350       thread_.reset(new std::thread(
 351           std::bind(&ServerData::Serve, this, server_host, &mu, &cond)));
 352       cond.WaitUntil(&mu, [this] { return server_ready_; });
 353       server_ready_ = false;
 354       gpr_log(GPR_INFO, "server startup complete");
 355     }
 356
 357     void Serve(const grpc::string& server_host, grpc::internal::Mutex* mu,
 358                grpc::internal::CondVar* cond) {
 359       std::ostringstream server_address;
 360       server_address << server_host << ":" << port_;
 361       ServerBuilder builder;
 362       std::shared_ptr<ServerCredentials> creds(new SecureServerCredentials(
 363           grpc_fake_transport_security_server_credentials_create()));
 364       builder.AddListeningPort(server_address.str(), std::move(creds));
 365       builder.RegisterService(&service_);
 366       server_ = builder.BuildAndStart();
 367       grpc::internal::MutexLock lock(mu);
 368       server_ready_ = true;
 369       cond->Signal();
 370     }
 371
 372     void Shutdown() {
 373       if (!started_) return;
 374       server_->Shutdown(grpc_timeout_milliseconds_to_deadline(0));
 375       thread_->join();
 376       started_ = false;
 377     }
 378
 379     void SetServingStatus(const grpc::string& service, bool serving) {
 380       server_->GetHealthCheckService()->SetServingStatus(service, serving);
 381     }
 382   };
 383
 384   void ResetCounters() {
 385     for (const auto& server : servers_) server->service_.ResetCounters();
 386   }
 387
 388   void WaitForServer(
 389       const std::unique_ptr<grpc::testing::EchoTestService::Stub>& stub,
 390       size_t server_idx, const grpc_core::DebugLocation& location,
 391       bool ignore_failure = false) {
 392     do {
 393       if (ignore_failure) {
 394         SendRpc(stub);
 395       } else {
 396         CheckRpcSendOk(stub, location, true);
 397       }
 398     } while (servers_[server_idx]->service_.request_count() == 0);
 399     ResetCounters();
 400   }
 401
 402   bool WaitForChannelNotReady(Channel* channel, int timeout_seconds = 5) {
 403     const gpr_timespec deadline =
 404         grpc_timeout_seconds_to_deadline(timeout_seconds);
 405     grpc_connectivity_state state;
 406     while ((state = channel->GetState(false /* try_to_connect */)) ==
 407            GRPC_CHANNEL_READY) {
 408       if (!channel->WaitForStateChange(state, deadline)) return false;
 409     }
 410     return true;
 411   }
 412
 413   bool WaitForChannelReady(Channel* channel, int timeout_seconds = 5) {
 414     const gpr_timespec deadline =
 415         grpc_timeout_seconds_to_deadline(timeout_seconds);
 416     grpc_connectivity_state state;
 417     while ((state = channel->GetState(true /* try_to_connect */)) !=
 418            GRPC_CHANNEL_READY) {
 419       if (!channel->WaitForStateChange(state, deadline)) return false;
 420     }
 421     return true;
 422   }
 423
 424   bool SeenAllServers() {
 425     for (const auto& server : servers_) {
 426       if (server->service_.request_count() == 0) return false;
 427     }
 428     return true;
 429   }
 430
 431   // Updates \a connection_order by appending to it the index of the newly
 432   // connected server. Must be called after every single RPC.
 433   void UpdateConnectionOrder(
 434       const std::vector<std::unique_ptr<ServerData>>& servers,
 435       std::vector<int>* connection_order) {
 436     for (size_t i = 0; i < servers.size(); ++i) {
 437       if (servers[i]->service_.request_count() == 1) {
 438         // Was the server index known? If not, update connection_order.
 439         const auto it =
 440             std::find(connection_order->begin(), connection_order->end(), i);
 441         if (it == connection_order->end()) {
 442           connection_order->push_back(i);
 443           return;
 444         }
 445       }
 446     }
 447   }
 448
 449   const grpc::string server_host_;
 450   std::vector<std::unique_ptr<ServerData>> servers_;
 451   const grpc::string kRequestMessage_;
 452   std::shared_ptr<ChannelCredentials> creds_;
 453 };
 454
 455 TEST_F(ClientLbEnd2endTest, ChannelStateConnectingWhenResolving) {
 456   const int kNumServers = 3;
 457   StartServers(kNumServers);
 458   auto response_generator = BuildResolverResponseGenerator();
 459   auto channel = BuildChannel("", response_generator);
 460   auto stub = BuildStub(channel);
 461   // Initial state should be IDLE.
 462   EXPECT_EQ(channel->GetState(false /* try_to_connect */), GRPC_CHANNEL_IDLE);
 463   // Tell the channel to try to connect.
 464   // Note that this call also returns IDLE, since the state change has
 465   // not yet occurred; it just gets triggered by this call.
 466   EXPECT_EQ(channel->GetState(true /* try_to_connect */), GRPC_CHANNEL_IDLE);
 467   // Now that the channel is trying to connect, we should be in state
 468   // CONNECTING.
 469   EXPECT_EQ(channel->GetState(false /* try_to_connect */),
 470             GRPC_CHANNEL_CONNECTING);
 471   // Return a resolver result, which allows the connection attempt to proceed.
 472   response_generator.SetNextResolution(GetServersPorts());
 473   // We should eventually transition into state READY.
 474   EXPECT_TRUE(WaitForChannelReady(channel.get()));
 475 }
 476
 477 TEST_F(ClientLbEnd2endTest, PickFirst) {
 478   // Start servers and send one RPC per server.
 479   const int kNumServers = 3;
 480   StartServers(kNumServers);
 481   auto response_generator = BuildResolverResponseGenerator();
 482   auto channel = BuildChannel(
 483       "", response_generator);  // test that pick first is the default.
 484   auto stub = BuildStub(channel);
 485   response_generator.SetNextResolution(GetServersPorts());
 486   for (size_t i = 0; i < servers_.size(); ++i) {
 487     CheckRpcSendOk(stub, DEBUG_LOCATION);
 488   }
 489   // All requests should have gone to a single server.
 490   bool found = false;
 491   for (size_t i = 0; i < servers_.size(); ++i) {
 492     const int request_count = servers_[i]->service_.request_count();
 493     if (request_count == kNumServers) {
 494       found = true;
 495     } else {
 496       EXPECT_EQ(0, request_count);
 497     }
 498   }
 499   EXPECT_TRUE(found);
 500   // Check LB policy name for the channel.
 501   EXPECT_EQ("pick_first", channel->GetLoadBalancingPolicyName());
 502 }
 503
 504 TEST_F(ClientLbEnd2endTest, PickFirstProcessPending) {
 505   StartServers(1);  // Single server
 506   auto response_generator = BuildResolverResponseGenerator();
 507   auto channel = BuildChannel(
 508       "", response_generator);  // test that pick first is the default.
 509   auto stub = BuildStub(channel);
 510   response_generator.SetNextResolution({servers_[0]->port_});
 511   WaitForServer(stub, 0, DEBUG_LOCATION);
 512   // Create a new channel and its corresponding PF LB policy, which will pick
 513   // the subchannels in READY state from the previous RPC against the same
 514   // target (even if it happened over a different channel, because subchannels
 515   // are globally reused). Progress should happen without any transition from
 516   // this READY state.
 517   auto second_response_generator = BuildResolverResponseGenerator();
 518   auto second_channel = BuildChannel("", second_response_generator);
 519   auto second_stub = BuildStub(second_channel);
 520   second_response_generator.SetNextResolution({servers_[0]->port_});
 521   CheckRpcSendOk(second_stub, DEBUG_LOCATION);
 522 }
 523
 524 TEST_F(ClientLbEnd2endTest, PickFirstSelectsReadyAtStartup) {
 525   ChannelArguments args;
 526   constexpr int kInitialBackOffMs = 5000;
 527   args.SetInt(GRPC_ARG_INITIAL_RECONNECT_BACKOFF_MS, kInitialBackOffMs);
 528   // Create 2 servers, but start only the second one.
 529   std::vector<int> ports = {grpc_pick_unused_port_or_die(),
 530                             grpc_pick_unused_port_or_die()};
 531   CreateServers(2, ports);
 532   StartServer(1);
 533   auto response_generator1 = BuildResolverResponseGenerator();
 534   auto channel1 = BuildChannel("pick_first", response_generator1, args);
 535   auto stub1 = BuildStub(channel1);
 536   response_generator1.SetNextResolution(ports);
 537   // Wait for second server to be ready.
 538   WaitForServer(stub1, 1, DEBUG_LOCATION);
 539   // Create a second channel with the same addresses.  Its PF instance
 540   // should immediately pick the second subchannel, since it's already
 541   // in READY state.
 542   auto response_generator2 = BuildResolverResponseGenerator();
 543   auto channel2 = BuildChannel("pick_first", response_generator2, args);
 544   response_generator2.SetNextResolution(ports);
 545   // Check that the channel reports READY without waiting for the
 546   // initial backoff.
 547   EXPECT_TRUE(WaitForChannelReady(channel2.get(), 1 /* timeout_seconds */));
 548 }
 549
 550 TEST_F(ClientLbEnd2endTest, PickFirstBackOffInitialReconnect) {
 551   ChannelArguments args;
 552   constexpr int kInitialBackOffMs = 100;
 553   args.SetInt(GRPC_ARG_INITIAL_RECONNECT_BACKOFF_MS, kInitialBackOffMs);
 554   const std::vector<int> ports = {grpc_pick_unused_port_or_die()};
 555   const gpr_timespec t0 = gpr_now(GPR_CLOCK_MONOTONIC);
 556   auto response_generator = BuildResolverResponseGenerator();
 557   auto channel = BuildChannel("pick_first", response_generator, args);
 558   auto stub = BuildStub(channel);
 559   response_generator.SetNextResolution(ports);
 560   // The channel won't become connected (there's no server).
 561   ASSERT_FALSE(channel->WaitForConnected(
 562       grpc_timeout_milliseconds_to_deadline(kInitialBackOffMs * 2)));
 563   // Bring up a server on the chosen port.
 564   StartServers(1, ports);
 565   // Now it will.
 566   ASSERT_TRUE(channel->WaitForConnected(
 567       grpc_timeout_milliseconds_to_deadline(kInitialBackOffMs * 2)));
 568   const gpr_timespec t1 = gpr_now(GPR_CLOCK_MONOTONIC);
 569   const grpc_millis waited_ms = gpr_time_to_millis(gpr_time_sub(t1, t0));
 570   gpr_log(GPR_DEBUG, "Waited %" PRId64 " milliseconds", waited_ms);
 571   // We should have waited at least kInitialBackOffMs. We substract one to
 572   // account for test and precision accuracy drift.
 573   EXPECT_GE(waited_ms, kInitialBackOffMs - 1);
 574   // But not much more.
 575   EXPECT_GT(
 576       gpr_time_cmp(
 577           grpc_timeout_milliseconds_to_deadline(kInitialBackOffMs * 1.10), t1),
 578       0);
 579 }
 580
 581 TEST_F(ClientLbEnd2endTest, PickFirstBackOffMinReconnect) {
 582   ChannelArguments args;
 583   constexpr int kMinReconnectBackOffMs = 1000;
 584   args.SetInt(GRPC_ARG_MIN_RECONNECT_BACKOFF_MS, kMinReconnectBackOffMs);
 585   const std::vector<int> ports = {grpc_pick_unused_port_or_die()};
 586   auto response_generator = BuildResolverResponseGenerator();
 587   auto channel = BuildChannel("pick_first", response_generator, args);
 588   auto stub = BuildStub(channel);
 589   response_generator.SetNextResolution(ports);
 590   // Make connection delay a 10% longer than it's willing to in order to make
 591   // sure we are hitting the codepath that waits for the min reconnect backoff.
 592   gpr_atm_rel_store(&g_connection_delay_ms, kMinReconnectBackOffMs * 1.10);
 593   default_client_impl = grpc_tcp_client_impl;
 594   grpc_set_tcp_client_impl(&delayed_connect);
 595   const gpr_timespec t0 = gpr_now(GPR_CLOCK_MONOTONIC);
 596   channel->WaitForConnected(
 597       grpc_timeout_milliseconds_to_deadline(kMinReconnectBackOffMs * 2));
 598   const gpr_timespec t1 = gpr_now(GPR_CLOCK_MONOTONIC);
 599   const grpc_millis waited_ms = gpr_time_to_millis(gpr_time_sub(t1, t0));
 600   gpr_log(GPR_DEBUG, "Waited %" PRId64 " ms", waited_ms);
 601   // We should have waited at least kMinReconnectBackOffMs. We substract one to
 602   // account for test and precision accuracy drift.
 603   EXPECT_GE(waited_ms, kMinReconnectBackOffMs - 1);
 604   gpr_atm_rel_store(&g_connection_delay_ms, 0);
 605 }
 606
 607 TEST_F(ClientLbEnd2endTest, PickFirstResetConnectionBackoff) {
 608   ChannelArguments args;
 609   constexpr int kInitialBackOffMs = 1000;
 610   args.SetInt(GRPC_ARG_INITIAL_RECONNECT_BACKOFF_MS, kInitialBackOffMs);
 611   const std::vector<int> ports = {grpc_pick_unused_port_or_die()};
 612   auto response_generator = BuildResolverResponseGenerator();
 613   auto channel = BuildChannel("pick_first", response_generator, args);
 614   auto stub = BuildStub(channel);
 615   response_generator.SetNextResolution(ports);
 616   // The channel won't become connected (there's no server).
 617   EXPECT_FALSE(
 618       channel->WaitForConnected(grpc_timeout_milliseconds_to_deadline(10)));
 619   // Bring up a server on the chosen port.
 620   StartServers(1, ports);
 621   const gpr_timespec t0 = gpr_now(GPR_CLOCK_MONOTONIC);
 622   // Wait for connect, but not long enough.  This proves that we're
 623   // being throttled by initial backoff.
 624   EXPECT_FALSE(
 625       channel->WaitForConnected(grpc_timeout_milliseconds_to_deadline(10)));
 626   // Reset connection backoff.
 627   experimental::ChannelResetConnectionBackoff(channel.get());
 628   // Wait for connect.  Should happen ~immediately.
 629   EXPECT_TRUE(
 630       channel->WaitForConnected(grpc_timeout_milliseconds_to_deadline(10)));
 631   const gpr_timespec t1 = gpr_now(GPR_CLOCK_MONOTONIC);
 632   const grpc_millis waited_ms = gpr_time_to_millis(gpr_time_sub(t1, t0));
 633   gpr_log(GPR_DEBUG, "Waited %" PRId64 " milliseconds", waited_ms);
 634   // We should have waited less than kInitialBackOffMs.
 635   EXPECT_LT(waited_ms, kInitialBackOffMs);
 636 }
 637
 638 TEST_F(ClientLbEnd2endTest,
 639        PickFirstResetConnectionBackoffNextAttemptStartsImmediately) {
 640   ChannelArguments args;
 641   constexpr int kInitialBackOffMs = 1000;
 642   args.SetInt(GRPC_ARG_INITIAL_RECONNECT_BACKOFF_MS, kInitialBackOffMs);
 643   const std::vector<int> ports = {grpc_pick_unused_port_or_die()};
 644   auto response_generator = BuildResolverResponseGenerator();
 645   auto channel = BuildChannel("pick_first", response_generator, args);
 646   auto stub = BuildStub(channel);
 647   response_generator.SetNextResolution(ports);
 648   // Wait for connect, which should fail ~immediately, because the server
 649   // is not up.
 650   gpr_log(GPR_INFO, "=== INITIAL CONNECTION ATTEMPT");
 651   EXPECT_FALSE(
 652       channel->WaitForConnected(grpc_timeout_milliseconds_to_deadline(10)));
 653   // Reset connection backoff.
 654   // Note that the time at which the third attempt will be started is
 655   // actually computed at this point, so we record the start time here.
 656   gpr_log(GPR_INFO, "=== RESETTING BACKOFF");
 657   const gpr_timespec t0 = gpr_now(GPR_CLOCK_MONOTONIC);
 658   experimental::ChannelResetConnectionBackoff(channel.get());
 659   // Trigger a second connection attempt.  This should also fail
 660   // ~immediately, but the retry should be scheduled for
 661   // kInitialBackOffMs instead of applying the multiplier.
 662   gpr_log(GPR_INFO, "=== POLLING FOR SECOND CONNECTION ATTEMPT");
 663   EXPECT_FALSE(
 664       channel->WaitForConnected(grpc_timeout_milliseconds_to_deadline(10)));
 665   // Bring up a server on the chosen port.
 666   gpr_log(GPR_INFO, "=== STARTING BACKEND");
 667   StartServers(1, ports);
 668   // Wait for connect.  Should happen within kInitialBackOffMs.
 669   // Give an extra 100ms to account for the time spent in the second and
 670   // third connection attempts themselves (since what we really want to
 671   // measure is the time between the two).  As long as this is less than
 672   // the 1.6x increase we would see if the backoff state was not reset
 673   // properly, the test is still proving that the backoff was reset.
 674   constexpr int kWaitMs = kInitialBackOffMs + 100;
 675   gpr_log(GPR_INFO, "=== POLLING FOR THIRD CONNECTION ATTEMPT");
 676   EXPECT_TRUE(channel->WaitForConnected(
 677       grpc_timeout_milliseconds_to_deadline(kWaitMs)));
 678   const gpr_timespec t1 = gpr_now(GPR_CLOCK_MONOTONIC);
 679   const grpc_millis waited_ms = gpr_time_to_millis(gpr_time_sub(t1, t0));
 680   gpr_log(GPR_DEBUG, "Waited %" PRId64 " milliseconds", waited_ms);
 681   EXPECT_LT(waited_ms, kWaitMs);
 682 }
 683
 684 TEST_F(ClientLbEnd2endTest, PickFirstUpdates) {
 685   // Start servers and send one RPC per server.
 686   const int kNumServers = 3;
 687   StartServers(kNumServers);
 688   auto response_generator = BuildResolverResponseGenerator();
 689   auto channel = BuildChannel("pick_first", response_generator);
 690   auto stub = BuildStub(channel);
 691
 692   std::vector<int> ports;
 693
 694   // Perform one RPC against the first server.
 695   ports.emplace_back(servers_[0]->port_);
 696   response_generator.SetNextResolution(ports);
 697   gpr_log(GPR_INFO, "****** SET [0] *******");
 698   CheckRpcSendOk(stub, DEBUG_LOCATION);
 699   EXPECT_EQ(servers_[0]->service_.request_count(), 1);
 700
 701   // An empty update will result in the channel going into TRANSIENT_FAILURE.
 702   ports.clear();
 703   response_generator.SetNextResolution(ports);
 704   gpr_log(GPR_INFO, "****** SET none *******");
 705   grpc_connectivity_state channel_state;
 706   do {
 707     channel_state = channel->GetState(true /* try to connect */);
 708   } while (channel_state == GRPC_CHANNEL_READY);
 709   ASSERT_NE(channel_state, GRPC_CHANNEL_READY);
 710   servers_[0]->service_.ResetCounters();
 711
 712   // Next update introduces servers_[1], making the channel recover.
 713   ports.clear();
 714   ports.emplace_back(servers_[1]->port_);
 715   response_generator.SetNextResolution(ports);
 716   gpr_log(GPR_INFO, "****** SET [1] *******");
 717   WaitForServer(stub, 1, DEBUG_LOCATION);
 718   EXPECT_EQ(servers_[0]->service_.request_count(), 0);
 719
 720   // And again for servers_[2]
 721   ports.clear();
 722   ports.emplace_back(servers_[2]->port_);
 723   response_generator.SetNextResolution(ports);
 724   gpr_log(GPR_INFO, "****** SET [2] *******");
 725   WaitForServer(stub, 2, DEBUG_LOCATION);
 726   EXPECT_EQ(servers_[0]->service_.request_count(), 0);
 727   EXPECT_EQ(servers_[1]->service_.request_count(), 0);
 728
 729   // Check LB policy name for the channel.
 730   EXPECT_EQ("pick_first", channel->GetLoadBalancingPolicyName());
 731 }
 732
 733 TEST_F(ClientLbEnd2endTest, PickFirstUpdateSuperset) {
 734   // Start servers and send one RPC per server.
 735   const int kNumServers = 3;
 736   StartServers(kNumServers);
 737   auto response_generator = BuildResolverResponseGenerator();
 738   auto channel = BuildChannel("pick_first", response_generator);
 739   auto stub = BuildStub(channel);
 740
 741   std::vector<int> ports;
 742
 743   // Perform one RPC against the first server.
 744   ports.emplace_back(servers_[0]->port_);
 745   response_generator.SetNextResolution(ports);
 746   gpr_log(GPR_INFO, "****** SET [0] *******");
 747   CheckRpcSendOk(stub, DEBUG_LOCATION);
 748   EXPECT_EQ(servers_[0]->service_.request_count(), 1);
 749   servers_[0]->service_.ResetCounters();
 750
 751   // Send and superset update
 752   ports.clear();
 753   ports.emplace_back(servers_[1]->port_);
 754   ports.emplace_back(servers_[0]->port_);
 755   response_generator.SetNextResolution(ports);
 756   gpr_log(GPR_INFO, "****** SET superset *******");
 757   CheckRpcSendOk(stub, DEBUG_LOCATION);
 758   // We stick to the previously connected server.
 759   WaitForServer(stub, 0, DEBUG_LOCATION);
 760   EXPECT_EQ(0, servers_[1]->service_.request_count());
 761
 762   // Check LB policy name for the channel.
 763   EXPECT_EQ("pick_first", channel->GetLoadBalancingPolicyName());
 764 }
 765
 766 TEST_F(ClientLbEnd2endTest, PickFirstGlobalSubchannelPool) {
 767   // Start one server.
 768   const int kNumServers = 1;
 769   StartServers(kNumServers);
 770   std::vector<int> ports = GetServersPorts();
 771   // Create two channels that (by default) use the global subchannel pool.
 772   auto response_generator1 = BuildResolverResponseGenerator();
 773   auto channel1 = BuildChannel("pick_first", response_generator1);
 774   auto stub1 = BuildStub(channel1);
 775   response_generator1.SetNextResolution(ports);
 776   auto response_generator2 = BuildResolverResponseGenerator();
 777   auto channel2 = BuildChannel("pick_first", response_generator2);
 778   auto stub2 = BuildStub(channel2);
 779   response_generator2.SetNextResolution(ports);
 780   WaitForServer(stub1, 0, DEBUG_LOCATION);
 781   // Send one RPC on each channel.
 782   CheckRpcSendOk(stub1, DEBUG_LOCATION);
 783   CheckRpcSendOk(stub2, DEBUG_LOCATION);
 784   // The server receives two requests.
 785   EXPECT_EQ(2, servers_[0]->service_.request_count());
 786   // The two requests are from the same client port, because the two channels
 787   // share subchannels via the global subchannel pool.
 788   EXPECT_EQ(1UL, servers_[0]->service_.clients().size());
 789 }
 790
 791 TEST_F(ClientLbEnd2endTest, PickFirstLocalSubchannelPool) {
 792   // Start one server.
 793   const int kNumServers = 1;
 794   StartServers(kNumServers);
 795   std::vector<int> ports = GetServersPorts();
 796   // Create two channels that use local subchannel pool.
 797   ChannelArguments args;
 798   args.SetInt(GRPC_ARG_USE_LOCAL_SUBCHANNEL_POOL, 1);
 799   auto response_generator1 = BuildResolverResponseGenerator();
 800   auto channel1 = BuildChannel("pick_first", response_generator1, args);
 801   auto stub1 = BuildStub(channel1);
 802   response_generator1.SetNextResolution(ports);
 803   auto response_generator2 = BuildResolverResponseGenerator();
 804   auto channel2 = BuildChannel("pick_first", response_generator2, args);
 805   auto stub2 = BuildStub(channel2);
 806   response_generator2.SetNextResolution(ports);
 807   WaitForServer(stub1, 0, DEBUG_LOCATION);
 808   // Send one RPC on each channel.
 809   CheckRpcSendOk(stub1, DEBUG_LOCATION);
 810   CheckRpcSendOk(stub2, DEBUG_LOCATION);
 811   // The server receives two requests.
 812   EXPECT_EQ(2, servers_[0]->service_.request_count());
 813   // The two requests are from two client ports, because the two channels didn't
 814   // share subchannels with each other.
 815   EXPECT_EQ(2UL, servers_[0]->service_.clients().size());
 816 }
 817
 818 TEST_F(ClientLbEnd2endTest, PickFirstManyUpdates) {
 819   const int kNumUpdates = 1000;
 820   const int kNumServers = 3;
 821   StartServers(kNumServers);
 822   auto response_generator = BuildResolverResponseGenerator();
 823   auto channel = BuildChannel("pick_first", response_generator);
 824   auto stub = BuildStub(channel);
 825   std::vector<int> ports = GetServersPorts();
 826   for (size_t i = 0; i < kNumUpdates; ++i) {
 827     std::shuffle(ports.begin(), ports.end(),
 828                  std::mt19937(std::random_device()()));
 829     response_generator.SetNextResolution(ports);
 830     // We should re-enter core at the end of the loop to give the resolution
 831     // setting closure a chance to run.
 832     if ((i + 1) % 10 == 0) CheckRpcSendOk(stub, DEBUG_LOCATION);
 833   }
 834   // Check LB policy name for the channel.
 835   EXPECT_EQ("pick_first", channel->GetLoadBalancingPolicyName());
 836 }
 837
 838 TEST_F(ClientLbEnd2endTest, PickFirstReresolutionNoSelected) {
 839   // Prepare the ports for up servers and down servers.
 840   const int kNumServers = 3;
 841   const int kNumAliveServers = 1;
 842   StartServers(kNumAliveServers);
 843   std::vector<int> alive_ports, dead_ports;
 844   for (size_t i = 0; i < kNumServers; ++i) {
 845     if (i < kNumAliveServers) {
 846       alive_ports.emplace_back(servers_[i]->port_);
 847     } else {
 848       dead_ports.emplace_back(grpc_pick_unused_port_or_die());
 849     }
 850   }
 851   auto response_generator = BuildResolverResponseGenerator();
 852   auto channel = BuildChannel("pick_first", response_generator);
 853   auto stub = BuildStub(channel);
 854   // The initial resolution only contains dead ports. There won't be any
 855   // selected subchannel. Re-resolution will return the same result.
 856   response_generator.SetNextResolution(dead_ports);
 857   gpr_log(GPR_INFO, "****** INITIAL RESOLUTION SET *******");
 858   for (size_t i = 0; i < 10; ++i) CheckRpcSendFailure(stub);
 859   // Set a re-resolution result that contains reachable ports, so that the
 860   // pick_first LB policy can recover soon.
 861   response_generator.SetNextResolutionUponError(alive_ports);
 862   gpr_log(GPR_INFO, "****** RE-RESOLUTION SET *******");
 863   WaitForServer(stub, 0, DEBUG_LOCATION, true /* ignore_failure */);
 864   CheckRpcSendOk(stub, DEBUG_LOCATION);
 865   EXPECT_EQ(servers_[0]->service_.request_count(), 1);
 866   // Check LB policy name for the channel.
 867   EXPECT_EQ("pick_first", channel->GetLoadBalancingPolicyName());
 868 }
 869
 870 TEST_F(ClientLbEnd2endTest, PickFirstReconnectWithoutNewResolverResult) {
 871   std::vector<int> ports = {grpc_pick_unused_port_or_die()};
 872   StartServers(1, ports);
 873   auto response_generator = BuildResolverResponseGenerator();
 874   auto channel = BuildChannel("pick_first", response_generator);
 875   auto stub = BuildStub(channel);
 876   response_generator.SetNextResolution(ports);
 877   gpr_log(GPR_INFO, "****** INITIAL CONNECTION *******");
 878   WaitForServer(stub, 0, DEBUG_LOCATION);
 879   gpr_log(GPR_INFO, "****** STOPPING SERVER ******");
 880   servers_[0]->Shutdown();
 881   EXPECT_TRUE(WaitForChannelNotReady(channel.get()));
 882   gpr_log(GPR_INFO, "****** RESTARTING SERVER ******");
 883   StartServers(1, ports);
 884   WaitForServer(stub, 0, DEBUG_LOCATION);
 885 }
 886
 887 TEST_F(ClientLbEnd2endTest,
 888        PickFirstReconnectWithoutNewResolverResultStartsFromTopOfList) {
 889   std::vector<int> ports = {grpc_pick_unused_port_or_die(),
 890                             grpc_pick_unused_port_or_die()};
 891   CreateServers(2, ports);
 892   StartServer(1);
 893   auto response_generator = BuildResolverResponseGenerator();
 894   auto channel = BuildChannel("pick_first", response_generator);
 895   auto stub = BuildStub(channel);
 896   response_generator.SetNextResolution(ports);
 897   gpr_log(GPR_INFO, "****** INITIAL CONNECTION *******");
 898   WaitForServer(stub, 1, DEBUG_LOCATION);
 899   gpr_log(GPR_INFO, "****** STOPPING SERVER ******");
 900   servers_[1]->Shutdown();
 901   EXPECT_TRUE(WaitForChannelNotReady(channel.get()));
 902   gpr_log(GPR_INFO, "****** STARTING BOTH SERVERS ******");
 903   StartServers(2, ports);
 904   WaitForServer(stub, 0, DEBUG_LOCATION);
 905 }
 906
 907 TEST_F(ClientLbEnd2endTest, PickFirstCheckStateBeforeStartWatch) {
 908   std::vector<int> ports = {grpc_pick_unused_port_or_die()};
 909   StartServers(1, ports);
 910   auto response_generator = BuildResolverResponseGenerator();
 911   auto channel_1 = BuildChannel("pick_first", response_generator);
 912   auto stub_1 = BuildStub(channel_1);
 913   response_generator.SetNextResolution(ports);
 914   gpr_log(GPR_INFO, "****** RESOLUTION SET FOR CHANNEL 1 *******");
 915   WaitForServer(stub_1, 0, DEBUG_LOCATION);
 916   gpr_log(GPR_INFO, "****** CHANNEL 1 CONNECTED *******");
 917   servers_[0]->Shutdown();
 918   // Channel 1 will receive a re-resolution containing the same server. It will
 919   // create a new subchannel and hold a ref to it.
 920   StartServers(1, ports);
 921   gpr_log(GPR_INFO, "****** SERVER RESTARTED *******");
 922   auto response_generator_2 = BuildResolverResponseGenerator();
 923   auto channel_2 = BuildChannel("pick_first", response_generator_2);
 924   auto stub_2 = BuildStub(channel_2);
 925   response_generator_2.SetNextResolution(ports);
 926   gpr_log(GPR_INFO, "****** RESOLUTION SET FOR CHANNEL 2 *******");
 927   WaitForServer(stub_2, 0, DEBUG_LOCATION, true);
 928   gpr_log(GPR_INFO, "****** CHANNEL 2 CONNECTED *******");
 929   servers_[0]->Shutdown();
 930   // Wait until the disconnection has triggered the connectivity notification.
 931   // Otherwise, the subchannel may be picked for next call but will fail soon.
 932   EXPECT_TRUE(WaitForChannelNotReady(channel_2.get()));
 933   // Channel 2 will also receive a re-resolution containing the same server.
 934   // Both channels will ref the same subchannel that failed.
 935   StartServers(1, ports);
 936   gpr_log(GPR_INFO, "****** SERVER RESTARTED AGAIN *******");
 937   gpr_log(GPR_INFO, "****** CHANNEL 2 STARTING A CALL *******");
 938   // The first call after the server restart will succeed.
 939   CheckRpcSendOk(stub_2, DEBUG_LOCATION);
 940   gpr_log(GPR_INFO, "****** CHANNEL 2 FINISHED A CALL *******");
 941   // Check LB policy name for the channel.
 942   EXPECT_EQ("pick_first", channel_1->GetLoadBalancingPolicyName());
 943   // Check LB policy name for the channel.
 944   EXPECT_EQ("pick_first", channel_2->GetLoadBalancingPolicyName());
 945 }
 946
 947 TEST_F(ClientLbEnd2endTest, PickFirstIdleOnDisconnect) {
 948   // Start server, send RPC, and make sure channel is READY.
 949   const int kNumServers = 1;
 950   StartServers(kNumServers);
 951   auto response_generator = BuildResolverResponseGenerator();
 952   auto channel =
 953       BuildChannel("", response_generator);  // pick_first is the default.
 954   auto stub = BuildStub(channel);
 955   response_generator.SetNextResolution(GetServersPorts());
 956   CheckRpcSendOk(stub, DEBUG_LOCATION);
 957   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_READY);
 958   // Stop server.  Channel should go into state IDLE.
 959   response_generator.SetFailureOnReresolution();
 960   servers_[0]->Shutdown();
 961   EXPECT_TRUE(WaitForChannelNotReady(channel.get()));
 962   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_IDLE);
 963   servers_.clear();
 964 }
 965
 966 TEST_F(ClientLbEnd2endTest, PickFirstPendingUpdateAndSelectedSubchannelFails) {
 967   auto response_generator = BuildResolverResponseGenerator();
 968   auto channel =
 969       BuildChannel("", response_generator);  // pick_first is the default.
 970   auto stub = BuildStub(channel);
 971   // Create a number of servers, but only start 1 of them.
 972   CreateServers(10);
 973   StartServer(0);
 974   // Initially resolve to first server and make sure it connects.
 975   gpr_log(GPR_INFO, "Phase 1: Connect to first server.");
 976   response_generator.SetNextResolution({servers_[0]->port_});
 977   CheckRpcSendOk(stub, DEBUG_LOCATION, true /* wait_for_ready */);
 978   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_READY);
 979   // Send a resolution update with the remaining servers, none of which are
 980   // running yet, so the update will stay pending.  Note that it's important
 981   // to have multiple servers here, or else the test will be flaky; with only
 982   // one server, the pending subchannel list has already gone into
 983   // TRANSIENT_FAILURE due to hitting the end of the list by the time we
 984   // check the state.
 985   gpr_log(GPR_INFO,
 986           "Phase 2: Resolver update pointing to remaining "
 987           "(not started) servers.");
 988   response_generator.SetNextResolution(GetServersPorts(1 /* start_index */));
 989   // RPCs will continue to be sent to the first server.
 990   CheckRpcSendOk(stub, DEBUG_LOCATION);
 991   // Now stop the first server, so that the current subchannel list
 992   // fails.  This should cause us to immediately swap over to the
 993   // pending list, even though it's not yet connected.  The state should
 994   // be set to CONNECTING, since that's what the pending subchannel list
 995   // was doing when we swapped over.
 996   gpr_log(GPR_INFO, "Phase 3: Stopping first server.");
 997   servers_[0]->Shutdown();
 998   WaitForChannelNotReady(channel.get());
 999   // TODO(roth): This should always return CONNECTING, but it's flaky
1000   // between that and TRANSIENT_FAILURE.  I suspect that this problem
1001   // will go away once we move the backoff code out of the subchannel
1002   // and into the LB policies.
1003   EXPECT_THAT(channel->GetState(false),
1004               ::testing::AnyOf(GRPC_CHANNEL_CONNECTING,
1005                                GRPC_CHANNEL_TRANSIENT_FAILURE));
1006   // Now start the second server.
1007   gpr_log(GPR_INFO, "Phase 4: Starting second server.");
1008   StartServer(1);
1009   // The channel should go to READY state and RPCs should go to the
1010   // second server.
1011   WaitForChannelReady(channel.get());
1012   WaitForServer(stub, 1, DEBUG_LOCATION, true /* ignore_failure */);
1013 }
1014
1015 TEST_F(ClientLbEnd2endTest, PickFirstStaysIdleUponEmptyUpdate) {
1016   // Start server, send RPC, and make sure channel is READY.
1017   const int kNumServers = 1;
1018   StartServers(kNumServers);
1019   auto response_generator = BuildResolverResponseGenerator();
1020   auto channel =
1021       BuildChannel("", response_generator);  // pick_first is the default.
1022   auto stub = BuildStub(channel);
1023   response_generator.SetNextResolution(GetServersPorts());
1024   CheckRpcSendOk(stub, DEBUG_LOCATION);
1025   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_READY);
1026   // Stop server.  Channel should go into state IDLE.
1027   servers_[0]->Shutdown();
1028   EXPECT_TRUE(WaitForChannelNotReady(channel.get()));
1029   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_IDLE);
1030   // Now send resolver update that includes no addresses.  Channel
1031   // should stay in state IDLE.
1032   response_generator.SetNextResolution({});
1033   EXPECT_FALSE(channel->WaitForStateChange(
1034       GRPC_CHANNEL_IDLE, grpc_timeout_seconds_to_deadline(3)));
1035   // Now bring the backend back up and send a non-empty resolver update,
1036   // and then try to send an RPC.  Channel should go back into state READY.
1037   StartServer(0);
1038   response_generator.SetNextResolution(GetServersPorts());
1039   CheckRpcSendOk(stub, DEBUG_LOCATION);
1040   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_READY);
1041 }
1042
1043 TEST_F(ClientLbEnd2endTest, RoundRobin) {
1044   // Start servers and send one RPC per server.
1045   const int kNumServers = 3;
1046   StartServers(kNumServers);
1047   auto response_generator = BuildResolverResponseGenerator();
1048   auto channel = BuildChannel("round_robin", response_generator);
1049   auto stub = BuildStub(channel);
1050   response_generator.SetNextResolution(GetServersPorts());
1051   // Wait until all backends are ready.
1052   do {
1053     CheckRpcSendOk(stub, DEBUG_LOCATION);
1054   } while (!SeenAllServers());
1055   ResetCounters();
1056   // "Sync" to the end of the list. Next sequence of picks will start at the
1057   // first server (index 0).
1058   WaitForServer(stub, servers_.size() - 1, DEBUG_LOCATION);
1059   std::vector<int> connection_order;
1060   for (size_t i = 0; i < servers_.size(); ++i) {
1061     CheckRpcSendOk(stub, DEBUG_LOCATION);
1062     UpdateConnectionOrder(servers_, &connection_order);
1063   }
1064   // Backends should be iterated over in the order in which the addresses were
1065   // given.
1066   const auto expected = std::vector<int>{0, 1, 2};
1067   EXPECT_EQ(expected, connection_order);
1068   // Check LB policy name for the channel.
1069   EXPECT_EQ("round_robin", channel->GetLoadBalancingPolicyName());
1070 }
1071
1072 TEST_F(ClientLbEnd2endTest, RoundRobinProcessPending) {
1073   StartServers(1);  // Single server
1074   auto response_generator = BuildResolverResponseGenerator();
1075   auto channel = BuildChannel("round_robin", response_generator);
1076   auto stub = BuildStub(channel);
1077   response_generator.SetNextResolution({servers_[0]->port_});
1078   WaitForServer(stub, 0, DEBUG_LOCATION);
1079   // Create a new channel and its corresponding RR LB policy, which will pick
1080   // the subchannels in READY state from the previous RPC against the same
1081   // target (even if it happened over a different channel, because subchannels
1082   // are globally reused). Progress should happen without any transition from
1083   // this READY state.
1084   auto second_response_generator = BuildResolverResponseGenerator();
1085   auto second_channel = BuildChannel("round_robin", second_response_generator);
1086   auto second_stub = BuildStub(second_channel);
1087   second_response_generator.SetNextResolution({servers_[0]->port_});
1088   CheckRpcSendOk(second_stub, DEBUG_LOCATION);
1089 }
1090
1091 TEST_F(ClientLbEnd2endTest, RoundRobinUpdates) {
1092   // Start servers and send one RPC per server.
1093   const int kNumServers = 3;
1094   StartServers(kNumServers);
1095   auto response_generator = BuildResolverResponseGenerator();
1096   auto channel = BuildChannel("round_robin", response_generator);
1097   auto stub = BuildStub(channel);
1098   std::vector<int> ports;
1099   // Start with a single server.
1100   gpr_log(GPR_INFO, "*** FIRST BACKEND ***");
1101   ports.emplace_back(servers_[0]->port_);
1102   response_generator.SetNextResolution(ports);
1103   WaitForServer(stub, 0, DEBUG_LOCATION);
1104   // Send RPCs. They should all go servers_[0]
1105   for (size_t i = 0; i < 10; ++i) CheckRpcSendOk(stub, DEBUG_LOCATION);
1106   EXPECT_EQ(10, servers_[0]->service_.request_count());
1107   EXPECT_EQ(0, servers_[1]->service_.request_count());
1108   EXPECT_EQ(0, servers_[2]->service_.request_count());
1109   servers_[0]->service_.ResetCounters();
1110   // And now for the second server.
1111   gpr_log(GPR_INFO, "*** SECOND BACKEND ***");
1112   ports.clear();
1113   ports.emplace_back(servers_[1]->port_);
1114   response_generator.SetNextResolution(ports);
1115   // Wait until update has been processed, as signaled by the second backend
1116   // receiving a request.
1117   EXPECT_EQ(0, servers_[1]->service_.request_count());
1118   WaitForServer(stub, 1, DEBUG_LOCATION);
1119   for (size_t i = 0; i < 10; ++i) CheckRpcSendOk(stub, DEBUG_LOCATION);
1120   EXPECT_EQ(0, servers_[0]->service_.request_count());
1121   EXPECT_EQ(10, servers_[1]->service_.request_count());
1122   EXPECT_EQ(0, servers_[2]->service_.request_count());
1123   servers_[1]->service_.ResetCounters();
1124   // ... and for the last server.
1125   gpr_log(GPR_INFO, "*** THIRD BACKEND ***");
1126   ports.clear();
1127   ports.emplace_back(servers_[2]->port_);
1128   response_generator.SetNextResolution(ports);
1129   WaitForServer(stub, 2, DEBUG_LOCATION);
1130   for (size_t i = 0; i < 10; ++i) CheckRpcSendOk(stub, DEBUG_LOCATION);
1131   EXPECT_EQ(0, servers_[0]->service_.request_count());
1132   EXPECT_EQ(0, servers_[1]->service_.request_count());
1133   EXPECT_EQ(10, servers_[2]->service_.request_count());
1134   servers_[2]->service_.ResetCounters();
1135   // Back to all servers.
1136   gpr_log(GPR_INFO, "*** ALL BACKENDS ***");
1137   ports.clear();
1138   ports.emplace_back(servers_[0]->port_);
1139   ports.emplace_back(servers_[1]->port_);
1140   ports.emplace_back(servers_[2]->port_);
1141   response_generator.SetNextResolution(ports);
1142   WaitForServer(stub, 0, DEBUG_LOCATION);
1143   WaitForServer(stub, 1, DEBUG_LOCATION);
1144   WaitForServer(stub, 2, DEBUG_LOCATION);
1145   // Send three RPCs, one per server.
1146   for (size_t i = 0; i < 3; ++i) CheckRpcSendOk(stub, DEBUG_LOCATION);
1147   EXPECT_EQ(1, servers_[0]->service_.request_count());
1148   EXPECT_EQ(1, servers_[1]->service_.request_count());
1149   EXPECT_EQ(1, servers_[2]->service_.request_count());
1150   // An empty update will result in the channel going into TRANSIENT_FAILURE.
1151   gpr_log(GPR_INFO, "*** NO BACKENDS ***");
1152   ports.clear();
1153   response_generator.SetNextResolution(ports);
1154   grpc_connectivity_state channel_state;
1155   do {
1156     channel_state = channel->GetState(true /* try to connect */);
1157   } while (channel_state == GRPC_CHANNEL_READY);
1158   ASSERT_NE(channel_state, GRPC_CHANNEL_READY);
1159   servers_[0]->service_.ResetCounters();
1160   // Next update introduces servers_[1], making the channel recover.
1161   gpr_log(GPR_INFO, "*** BACK TO SECOND BACKEND ***");
1162   ports.clear();
1163   ports.emplace_back(servers_[1]->port_);
1164   response_generator.SetNextResolution(ports);
1165   WaitForServer(stub, 1, DEBUG_LOCATION);
1166   channel_state = channel->GetState(false /* try to connect */);
1167   ASSERT_EQ(channel_state, GRPC_CHANNEL_READY);
1168   // Check LB policy name for the channel.
1169   EXPECT_EQ("round_robin", channel->GetLoadBalancingPolicyName());
1170 }
1171
1172 TEST_F(ClientLbEnd2endTest, RoundRobinUpdateInError) {
1173   const int kNumServers = 3;
1174   StartServers(kNumServers);
1175   auto response_generator = BuildResolverResponseGenerator();
1176   auto channel = BuildChannel("round_robin", response_generator);
1177   auto stub = BuildStub(channel);
1178   std::vector<int> ports;
1179
1180   // Start with a single server.
1181   ports.emplace_back(servers_[0]->port_);
1182   response_generator.SetNextResolution(ports);
1183   WaitForServer(stub, 0, DEBUG_LOCATION);
1184   // Send RPCs. They should all go to servers_[0]
1185   for (size_t i = 0; i < 10; ++i) SendRpc(stub);
1186   EXPECT_EQ(10, servers_[0]->service_.request_count());
1187   EXPECT_EQ(0, servers_[1]->service_.request_count());
1188   EXPECT_EQ(0, servers_[2]->service_.request_count());
1189   servers_[0]->service_.ResetCounters();
1190
1191   // Shutdown one of the servers to be sent in the update.
1192   servers_[1]->Shutdown();
1193   ports.emplace_back(servers_[1]->port_);
1194   ports.emplace_back(servers_[2]->port_);
1195   response_generator.SetNextResolution(ports);
1196   WaitForServer(stub, 0, DEBUG_LOCATION);
1197   WaitForServer(stub, 2, DEBUG_LOCATION);
1198
1199   // Send three RPCs, one per server.
1200   for (size_t i = 0; i < kNumServers; ++i) SendRpc(stub);
1201   // The server in shutdown shouldn't receive any.
1202   EXPECT_EQ(0, servers_[1]->service_.request_count());
1203 }
1204
1205 TEST_F(ClientLbEnd2endTest, RoundRobinManyUpdates) {
1206   // Start servers and send one RPC per server.
1207   const int kNumServers = 3;
1208   StartServers(kNumServers);
1209   auto response_generator = BuildResolverResponseGenerator();
1210   auto channel = BuildChannel("round_robin", response_generator);
1211   auto stub = BuildStub(channel);
1212   std::vector<int> ports = GetServersPorts();
1213   for (size_t i = 0; i < 1000; ++i) {
1214     std::shuffle(ports.begin(), ports.end(),
1215                  std::mt19937(std::random_device()()));
1216     response_generator.SetNextResolution(ports);
1217     if (i % 10 == 0) CheckRpcSendOk(stub, DEBUG_LOCATION);
1218   }
1219   // Check LB policy name for the channel.
1220   EXPECT_EQ("round_robin", channel->GetLoadBalancingPolicyName());
1221 }
1222
1223 TEST_F(ClientLbEnd2endTest, RoundRobinConcurrentUpdates) {
1224   // TODO(dgq): replicate the way internal testing exercises the concurrent
1225   // update provisions of RR.
1226 }
1227
1228 TEST_F(ClientLbEnd2endTest, RoundRobinReresolve) {
1229   // Start servers and send one RPC per server.
1230   const int kNumServers = 3;
1231   std::vector<int> first_ports;
1232   std::vector<int> second_ports;
1233   first_ports.reserve(kNumServers);
1234   for (int i = 0; i < kNumServers; ++i) {
1235     first_ports.push_back(grpc_pick_unused_port_or_die());
1236   }
1237   second_ports.reserve(kNumServers);
1238   for (int i = 0; i < kNumServers; ++i) {
1239     second_ports.push_back(grpc_pick_unused_port_or_die());
1240   }
1241   StartServers(kNumServers, first_ports);
1242   auto response_generator = BuildResolverResponseGenerator();
1243   auto channel = BuildChannel("round_robin", response_generator);
1244   auto stub = BuildStub(channel);
1245   response_generator.SetNextResolution(first_ports);
1246   // Send a number of RPCs, which succeed.
1247   for (size_t i = 0; i < 100; ++i) {
1248     CheckRpcSendOk(stub, DEBUG_LOCATION);
1249   }
1250   // Kill all servers
1251   gpr_log(GPR_INFO, "****** ABOUT TO KILL SERVERS *******");
1252   for (size_t i = 0; i < servers_.size(); ++i) {
1253     servers_[i]->Shutdown();
1254   }
1255   gpr_log(GPR_INFO, "****** SERVERS KILLED *******");
1256   gpr_log(GPR_INFO, "****** SENDING DOOMED REQUESTS *******");
1257   // Client requests should fail. Send enough to tickle all subchannels.
1258   for (size_t i = 0; i < servers_.size(); ++i) CheckRpcSendFailure(stub);
1259   gpr_log(GPR_INFO, "****** DOOMED REQUESTS SENT *******");
1260   // Bring servers back up on a different set of ports. We need to do this to be
1261   // sure that the eventual success is *not* due to subchannel reconnection
1262   // attempts and that an actual re-resolution has happened as a result of the
1263   // RR policy going into transient failure when all its subchannels become
1264   // unavailable (in transient failure as well).
1265   gpr_log(GPR_INFO, "****** RESTARTING SERVERS *******");
1266   StartServers(kNumServers, second_ports);
1267   // Don't notify of the update. Wait for the LB policy's re-resolution to
1268   // "pull" the new ports.
1269   response_generator.SetNextResolutionUponError(second_ports);
1270   gpr_log(GPR_INFO, "****** SERVERS RESTARTED *******");
1271   gpr_log(GPR_INFO, "****** SENDING REQUEST TO SUCCEED *******");
1272   // Client request should eventually (but still fairly soon) succeed.
1273   const gpr_timespec deadline = grpc_timeout_seconds_to_deadline(5);
1274   gpr_timespec now = gpr_now(GPR_CLOCK_MONOTONIC);
1275   while (gpr_time_cmp(deadline, now) > 0) {
1276     if (SendRpc(stub)) break;
1277     now = gpr_now(GPR_CLOCK_MONOTONIC);
1278   }
1279   ASSERT_GT(gpr_time_cmp(deadline, now), 0);
1280 }
1281
1282 TEST_F(ClientLbEnd2endTest, RoundRobinSingleReconnect) {
1283   const int kNumServers = 3;
1284   StartServers(kNumServers);
1285   const auto ports = GetServersPorts();
1286   auto response_generator = BuildResolverResponseGenerator();
1287   auto channel = BuildChannel("round_robin", response_generator);
1288   auto stub = BuildStub(channel);
1289   response_generator.SetNextResolution(ports);
1290   for (size_t i = 0; i < kNumServers; ++i) {
1291     WaitForServer(stub, i, DEBUG_LOCATION);
1292   }
1293   for (size_t i = 0; i < servers_.size(); ++i) {
1294     CheckRpcSendOk(stub, DEBUG_LOCATION);
1295     EXPECT_EQ(1, servers_[i]->service_.request_count()) << "for backend #" << i;
1296   }
1297   // One request should have gone to each server.
1298   for (size_t i = 0; i < servers_.size(); ++i) {
1299     EXPECT_EQ(1, servers_[i]->service_.request_count());
1300   }
1301   const auto pre_death = servers_[0]->service_.request_count();
1302   // Kill the first server.
1303   servers_[0]->Shutdown();
1304   // Client request still succeed. May need retrying if RR had returned a pick
1305   // before noticing the change in the server's connectivity.
1306   while (!SendRpc(stub)) {
1307   }  // Retry until success.
1308   // Send a bunch of RPCs that should succeed.
1309   for (int i = 0; i < 10 * kNumServers; ++i) {
1310     CheckRpcSendOk(stub, DEBUG_LOCATION);
1311   }
1312   const auto post_death = servers_[0]->service_.request_count();
1313   // No requests have gone to the deceased server.
1314   EXPECT_EQ(pre_death, post_death);
1315   // Bring the first server back up.
1316   StartServer(0);
1317   // Requests should start arriving at the first server either right away (if
1318   // the server managed to start before the RR policy retried the subchannel) or
1319   // after the subchannel retry delay otherwise (RR's subchannel retried before
1320   // the server was fully back up).
1321   WaitForServer(stub, 0, DEBUG_LOCATION);
1322 }
1323
1324 // If health checking is required by client but health checking service
1325 // is not running on the server, the channel should be treated as healthy.
1326 TEST_F(ClientLbEnd2endTest,
1327        RoundRobinServersHealthCheckingUnimplementedTreatedAsHealthy) {
1328   StartServers(1);  // Single server
1329   ChannelArguments args;
1330   args.SetServiceConfigJSON(
1331       "{\"healthCheckConfig\": "
1332       "{\"serviceName\": \"health_check_service_name\"}}");
1333   auto response_generator = BuildResolverResponseGenerator();
1334   auto channel = BuildChannel("round_robin", response_generator, args);
1335   auto stub = BuildStub(channel);
1336   response_generator.SetNextResolution({servers_[0]->port_});
1337   EXPECT_TRUE(WaitForChannelReady(channel.get()));
1338   CheckRpcSendOk(stub, DEBUG_LOCATION);
1339 }
1340
1341 TEST_F(ClientLbEnd2endTest, RoundRobinWithHealthChecking) {
1342   EnableDefaultHealthCheckService(true);
1343   // Start servers.
1344   const int kNumServers = 3;
1345   StartServers(kNumServers);
1346   ChannelArguments args;
1347   args.SetServiceConfigJSON(
1348       "{\"healthCheckConfig\": "
1349       "{\"serviceName\": \"health_check_service_name\"}}");
1350   auto response_generator = BuildResolverResponseGenerator();
1351   auto channel = BuildChannel("round_robin", response_generator, args);
1352   auto stub = BuildStub(channel);
1353   response_generator.SetNextResolution(GetServersPorts());
1354   // Channel should not become READY, because health checks should be failing.
1355   gpr_log(GPR_INFO,
1356           "*** initial state: unknown health check service name for "
1357           "all servers");
1358   EXPECT_FALSE(WaitForChannelReady(channel.get(), 1));
1359   // Now set one of the servers to be healthy.
1360   // The channel should become healthy and all requests should go to
1361   // the healthy server.
1362   gpr_log(GPR_INFO, "*** server 0 healthy");
1363   servers_[0]->SetServingStatus("health_check_service_name", true);
1364   EXPECT_TRUE(WaitForChannelReady(channel.get()));
1365   for (int i = 0; i < 10; ++i) {
1366     CheckRpcSendOk(stub, DEBUG_LOCATION);
1367   }
1368   EXPECT_EQ(10, servers_[0]->service_.request_count());
1369   EXPECT_EQ(0, servers_[1]->service_.request_count());
1370   EXPECT_EQ(0, servers_[2]->service_.request_count());
1371   // Now set a second server to be healthy.
1372   gpr_log(GPR_INFO, "*** server 2 healthy");
1373   servers_[2]->SetServingStatus("health_check_service_name", true);
1374   WaitForServer(stub, 2, DEBUG_LOCATION);
1375   for (int i = 0; i < 10; ++i) {
1376     CheckRpcSendOk(stub, DEBUG_LOCATION);
1377   }
1378   EXPECT_EQ(5, servers_[0]->service_.request_count());
1379   EXPECT_EQ(0, servers_[1]->service_.request_count());
1380   EXPECT_EQ(5, servers_[2]->service_.request_count());
1381   // Now set the remaining server to be healthy.
1382   gpr_log(GPR_INFO, "*** server 1 healthy");
1383   servers_[1]->SetServingStatus("health_check_service_name", true);
1384   WaitForServer(stub, 1, DEBUG_LOCATION);
1385   for (int i = 0; i < 9; ++i) {
1386     CheckRpcSendOk(stub, DEBUG_LOCATION);
1387   }
1388   EXPECT_EQ(3, servers_[0]->service_.request_count());
1389   EXPECT_EQ(3, servers_[1]->service_.request_count());
1390   EXPECT_EQ(3, servers_[2]->service_.request_count());
1391   // Now set one server to be unhealthy again.  Then wait until the
1392   // unhealthiness has hit the client.  We know that the client will see
1393   // this when we send kNumServers requests and one of the remaining servers
1394   // sees two of the requests.
1395   gpr_log(GPR_INFO, "*** server 0 unhealthy");
1396   servers_[0]->SetServingStatus("health_check_service_name", false);
1397   do {
1398     ResetCounters();
1399     for (int i = 0; i < kNumServers; ++i) {
1400       CheckRpcSendOk(stub, DEBUG_LOCATION);
1401     }
1402   } while (servers_[1]->service_.request_count() != 2 &&
1403            servers_[2]->service_.request_count() != 2);
1404   // Now set the remaining two servers to be unhealthy.  Make sure the
1405   // channel leaves READY state and that RPCs fail.
1406   gpr_log(GPR_INFO, "*** all servers unhealthy");
1407   servers_[1]->SetServingStatus("health_check_service_name", false);
1408   servers_[2]->SetServingStatus("health_check_service_name", false);
1409   EXPECT_TRUE(WaitForChannelNotReady(channel.get()));
1410   CheckRpcSendFailure(stub);
1411   // Clean up.
1412   EnableDefaultHealthCheckService(false);
1413 }
1414
1415 TEST_F(ClientLbEnd2endTest, RoundRobinWithHealthCheckingInhibitPerChannel) {
1416   EnableDefaultHealthCheckService(true);
1417   // Start server.
1418   const int kNumServers = 1;
1419   StartServers(kNumServers);
1420   // Create a channel with health-checking enabled.
1421   ChannelArguments args;
1422   args.SetServiceConfigJSON(
1423       "{\"healthCheckConfig\": "
1424       "{\"serviceName\": \"health_check_service_name\"}}");
1425   auto response_generator1 = BuildResolverResponseGenerator();
1426   auto channel1 = BuildChannel("round_robin", response_generator1, args);
1427   auto stub1 = BuildStub(channel1);
1428   std::vector<int> ports = GetServersPorts();
1429   response_generator1.SetNextResolution(ports);
1430   // Create a channel with health checking enabled but inhibited.
1431   args.SetInt(GRPC_ARG_INHIBIT_HEALTH_CHECKING, 1);
1432   auto response_generator2 = BuildResolverResponseGenerator();
1433   auto channel2 = BuildChannel("round_robin", response_generator2, args);
1434   auto stub2 = BuildStub(channel2);
1435   response_generator2.SetNextResolution(ports);
1436   // First channel should not become READY, because health checks should be
1437   // failing.
1438   EXPECT_FALSE(WaitForChannelReady(channel1.get(), 1));
1439   CheckRpcSendFailure(stub1);
1440   // Second channel should be READY.
1441   EXPECT_TRUE(WaitForChannelReady(channel2.get(), 1));
1442   CheckRpcSendOk(stub2, DEBUG_LOCATION);
1443   // Enable health checks on the backend and wait for channel 1 to succeed.
1444   servers_[0]->SetServingStatus("health_check_service_name", true);
1445   CheckRpcSendOk(stub1, DEBUG_LOCATION, true /* wait_for_ready */);
1446   // Check that we created only one subchannel to the backend.
1447   EXPECT_EQ(1UL, servers_[0]->service_.clients().size());
1448   // Clean up.
1449   EnableDefaultHealthCheckService(false);
1450 }
1451
1452 TEST_F(ClientLbEnd2endTest, RoundRobinWithHealthCheckingServiceNamePerChannel) {
1453   EnableDefaultHealthCheckService(true);
1454   // Start server.
1455   const int kNumServers = 1;
1456   StartServers(kNumServers);
1457   // Create a channel with health-checking enabled.
1458   ChannelArguments args;
1459   args.SetServiceConfigJSON(
1460       "{\"healthCheckConfig\": "
1461       "{\"serviceName\": \"health_check_service_name\"}}");
1462   auto response_generator1 = BuildResolverResponseGenerator();
1463   auto channel1 = BuildChannel("round_robin", response_generator1, args);
1464   auto stub1 = BuildStub(channel1);
1465   std::vector<int> ports = GetServersPorts();
1466   response_generator1.SetNextResolution(ports);
1467   // Create a channel with health-checking enabled with a different
1468   // service name.
1469   ChannelArguments args2;
1470   args2.SetServiceConfigJSON(
1471       "{\"healthCheckConfig\": "
1472       "{\"serviceName\": \"health_check_service_name2\"}}");
1473   auto response_generator2 = BuildResolverResponseGenerator();
1474   auto channel2 = BuildChannel("round_robin", response_generator2, args2);
1475   auto stub2 = BuildStub(channel2);
1476   response_generator2.SetNextResolution(ports);
1477   // Allow health checks from channel 2 to succeed.
1478   servers_[0]->SetServingStatus("health_check_service_name2", true);
1479   // First channel should not become READY, because health checks should be
1480   // failing.
1481   EXPECT_FALSE(WaitForChannelReady(channel1.get(), 1));
1482   CheckRpcSendFailure(stub1);
1483   // Second channel should be READY.
1484   EXPECT_TRUE(WaitForChannelReady(channel2.get(), 1));
1485   CheckRpcSendOk(stub2, DEBUG_LOCATION);
1486   // Enable health checks for channel 1 and wait for it to succeed.
1487   servers_[0]->SetServingStatus("health_check_service_name", true);
1488   CheckRpcSendOk(stub1, DEBUG_LOCATION, true /* wait_for_ready */);
1489   // Check that we created only one subchannel to the backend.
1490   EXPECT_EQ(1UL, servers_[0]->service_.clients().size());
1491   // Clean up.
1492   EnableDefaultHealthCheckService(false);
1493 }
1494
1495 TEST_F(ClientLbEnd2endTest,
1496        RoundRobinWithHealthCheckingServiceNameChangesAfterSubchannelsCreated) {
1497   EnableDefaultHealthCheckService(true);
1498   // Start server.
1499   const int kNumServers = 1;
1500   StartServers(kNumServers);
1501   // Create a channel with health-checking enabled.
1502   const char* kServiceConfigJson =
1503       "{\"healthCheckConfig\": "
1504       "{\"serviceName\": \"health_check_service_name\"}}";
1505   auto response_generator = BuildResolverResponseGenerator();
1506   auto channel = BuildChannel("round_robin", response_generator);
1507   auto stub = BuildStub(channel);
1508   std::vector<int> ports = GetServersPorts();
1509   response_generator.SetNextResolution(ports, kServiceConfigJson);
1510   servers_[0]->SetServingStatus("health_check_service_name", true);
1511   EXPECT_TRUE(WaitForChannelReady(channel.get(), 1 /* timeout_seconds */));
1512   // Send an update on the channel to change it to use a health checking
1513   // service name that is not being reported as healthy.
1514   const char* kServiceConfigJson2 =
1515       "{\"healthCheckConfig\": "
1516       "{\"serviceName\": \"health_check_service_name2\"}}";
1517   response_generator.SetNextResolution(ports, kServiceConfigJson2);
1518   EXPECT_TRUE(WaitForChannelNotReady(channel.get()));
1519   // Clean up.
1520   EnableDefaultHealthCheckService(false);
1521 }
1522
1523 TEST_F(ClientLbEnd2endTest, ChannelIdleness) {
1524   // Start server.
1525   const int kNumServers = 1;
1526   StartServers(kNumServers);
1527   // Set max idle time and build the channel.
1528   ChannelArguments args;
1529   args.SetInt(GRPC_ARG_CLIENT_IDLE_TIMEOUT_MS, 1000);
1530   auto response_generator = BuildResolverResponseGenerator();
1531   auto channel = BuildChannel("", response_generator, args);
1532   auto stub = BuildStub(channel);
1533   // The initial channel state should be IDLE.
1534   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_IDLE);
1535   // After sending RPC, channel state should be READY.
1536   response_generator.SetNextResolution(GetServersPorts());
1537   CheckRpcSendOk(stub, DEBUG_LOCATION);
1538   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_READY);
1539   // After a period time not using the channel, the channel state should switch
1540   // to IDLE.
1541   gpr_sleep_until(grpc_timeout_milliseconds_to_deadline(1200));
1542   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_IDLE);
1543   // Sending a new RPC should awake the IDLE channel.
1544   response_generator.SetNextResolution(GetServersPorts());
1545   CheckRpcSendOk(stub, DEBUG_LOCATION);
1546   EXPECT_EQ(channel->GetState(false), GRPC_CHANNEL_READY);
1547 }
1548
1549 class ClientLbInterceptTrailingMetadataTest : public ClientLbEnd2endTest {
1550  protected:
1551   void SetUp() override {
1552     ClientLbEnd2endTest::SetUp();
1553     grpc_core::RegisterInterceptRecvTrailingMetadataLoadBalancingPolicy(
1554         ReportTrailerIntercepted, this);
1555   }
1556
1557   void TearDown() override { ClientLbEnd2endTest::TearDown(); }
1558
1559   int trailers_intercepted() {
1560     grpc::internal::MutexLock lock(&mu_);
1561     return trailers_intercepted_;
1562   }
1563
1564   const udpa::data::orca::v1::OrcaLoadReport* backend_load_report() {
1565     grpc::internal::MutexLock lock(&mu_);
1566     return load_report_.get();
1567   }
1568
1569  private:
1570   static void ReportTrailerIntercepted(
1571       void* arg, const grpc_core::LoadBalancingPolicy::BackendMetricData*
1572                      backend_metric_data) {
1573     ClientLbInterceptTrailingMetadataTest* self =
1574         static_cast<ClientLbInterceptTrailingMetadataTest*>(arg);
1575     grpc::internal::MutexLock lock(&self->mu_);
1576     self->trailers_intercepted_++;
1577     if (backend_metric_data != nullptr) {
1578       self->load_report_.reset(new udpa::data::orca::v1::OrcaLoadReport);
1579       self->load_report_->set_cpu_utilization(
1580           backend_metric_data->cpu_utilization);
1581       self->load_report_->set_mem_utilization(
1582           backend_metric_data->mem_utilization);
1583       self->load_report_->set_rps(backend_metric_data->requests_per_second);
1584       for (const auto& p : backend_metric_data->request_cost) {
1585         grpc_core::UniquePtr<char> name = p.first.dup();
1586         (*self->load_report_->mutable_request_cost())[name.get()] = p.second;
1587       }
1588       for (const auto& p : backend_metric_data->utilization) {
1589         grpc_core::UniquePtr<char> name = p.first.dup();
1590         (*self->load_report_->mutable_utilization())[name.get()] = p.second;
1591       }
1592     }
1593   }
1594
1595   grpc::internal::Mutex mu_;
1596   int trailers_intercepted_ = 0;
1597   std::unique_ptr<udpa::data::orca::v1::OrcaLoadReport> load_report_;
1598 };
1599
1600 TEST_F(ClientLbInterceptTrailingMetadataTest, InterceptsRetriesDisabled) {
1601   const int kNumServers = 1;
1602   const int kNumRpcs = 10;
1603   StartServers(kNumServers);
1604   auto response_generator = BuildResolverResponseGenerator();
1605   auto channel =
1606       BuildChannel("intercept_trailing_metadata_lb", response_generator);
1607   auto stub = BuildStub(channel);
1608   response_generator.SetNextResolution(GetServersPorts());
1609   for (size_t i = 0; i < kNumRpcs; ++i) {
1610     CheckRpcSendOk(stub, DEBUG_LOCATION);
1611   }
1612   // Check LB policy name for the channel.
1613   EXPECT_EQ("intercept_trailing_metadata_lb",
1614             channel->GetLoadBalancingPolicyName());
1615   EXPECT_EQ(kNumRpcs, trailers_intercepted());
1616   EXPECT_EQ(nullptr, backend_load_report());
1617 }
1618
1619 TEST_F(ClientLbInterceptTrailingMetadataTest, InterceptsRetriesEnabled) {
1620   const int kNumServers = 1;
1621   const int kNumRpcs = 10;
1622   StartServers(kNumServers);
1623   ChannelArguments args;
1624   args.SetServiceConfigJSON(
1625       "{\n"
1626       "  \"methodConfig\": [ {\n"
1627       "    \"name\": [\n"
1628       "      { \"service\": \"grpc.testing.EchoTestService\" }\n"
1629       "    ],\n"
1630       "    \"retryPolicy\": {\n"
1631       "      \"maxAttempts\": 3,\n"
1632       "      \"initialBackoff\": \"1s\",\n"
1633       "      \"maxBackoff\": \"120s\",\n"
1634       "      \"backoffMultiplier\": 1.6,\n"
1635       "      \"retryableStatusCodes\": [ \"ABORTED\" ]\n"
1636       "    }\n"
1637       "  } ]\n"
1638       "}");
1639   auto response_generator = BuildResolverResponseGenerator();
1640   auto channel =
1641       BuildChannel("intercept_trailing_metadata_lb", response_generator, args);
1642   auto stub = BuildStub(channel);
1643   response_generator.SetNextResolution(GetServersPorts());
1644   for (size_t i = 0; i < kNumRpcs; ++i) {
1645     CheckRpcSendOk(stub, DEBUG_LOCATION);
1646   }
1647   // Check LB policy name for the channel.
1648   EXPECT_EQ("intercept_trailing_metadata_lb",
1649             channel->GetLoadBalancingPolicyName());
1650   EXPECT_EQ(kNumRpcs, trailers_intercepted());
1651   EXPECT_EQ(nullptr, backend_load_report());
1652 }
1653
1654 TEST_F(ClientLbInterceptTrailingMetadataTest, BackendMetricData) {
1655   const int kNumServers = 1;
1656   const int kNumRpcs = 10;
1657   StartServers(kNumServers);
1658   udpa::data::orca::v1::OrcaLoadReport load_report;
1659   load_report.set_cpu_utilization(0.5);
1660   load_report.set_mem_utilization(0.75);
1661   load_report.set_rps(25);
1662   auto* request_cost = load_report.mutable_request_cost();
1663   (*request_cost)["foo"] = 0.8;
1664   (*request_cost)["bar"] = 1.4;
1665   auto* utilization = load_report.mutable_utilization();
1666   (*utilization)["baz"] = 1.1;
1667   (*utilization)["quux"] = 0.9;
1668   for (const auto& server : servers_) {
1669     server->service_.set_load_report(&load_report);
1670   }
1671   auto response_generator = BuildResolverResponseGenerator();
1672   auto channel =
1673       BuildChannel("intercept_trailing_metadata_lb", response_generator);
1674   auto stub = BuildStub(channel);
1675   response_generator.SetNextResolution(GetServersPorts());
1676   for (size_t i = 0; i < kNumRpcs; ++i) {
1677     CheckRpcSendOk(stub, DEBUG_LOCATION);
1678     auto* actual = backend_load_report();
1679     ASSERT_NE(actual, nullptr);
1680     // TODO(roth): Change this to use EqualsProto() once that becomes
1681     // available in OSS.
1682     EXPECT_EQ(actual->cpu_utilization(), load_report.cpu_utilization());
1683     EXPECT_EQ(actual->mem_utilization(), load_report.mem_utilization());
1684     EXPECT_EQ(actual->rps(), load_report.rps());
1685     EXPECT_EQ(actual->request_cost().size(), load_report.request_cost().size());
1686     for (const auto& p : actual->request_cost()) {
1687       auto it = load_report.request_cost().find(p.first);
1688       ASSERT_NE(it, load_report.request_cost().end());
1689       EXPECT_EQ(it->second, p.second);
1690     }
1691     EXPECT_EQ(actual->utilization().size(), load_report.utilization().size());
1692     for (const auto& p : actual->utilization()) {
1693       auto it = load_report.utilization().find(p.first);
1694       ASSERT_NE(it, load_report.utilization().end());
1695       EXPECT_EQ(it->second, p.second);
1696     }
1697   }
1698   // Check LB policy name for the channel.
1699   EXPECT_EQ("intercept_trailing_metadata_lb",
1700             channel->GetLoadBalancingPolicyName());
1701   EXPECT_EQ(kNumRpcs, trailers_intercepted());
1702 }
1703
1704 }  // namespace
1705 }  // namespace testing
1706 }  // namespace grpc
1707
1708 int main(int argc, char** argv) {
1709   ::testing::InitGoogleTest(&argc, argv);
1710   grpc::testing::TestEnvironment env(argc, argv);
1711   const auto result = RUN_ALL_TESTS();
1712   return result;
1713 }