1 //
2 // Copyright 2017 gRPC authors.
3 //
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
7 //
8 // http://www.apache.org/licenses/LICENSE-2.0
9 //
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
15 //
16
17 #include <deque>
18 #include <memory>
19 #include <mutex>
20 #include <set>
21 #include <sstream>
22 #include <string>
23 #include <thread>
24
25 #include <gmock/gmock.h>
26 #include <gtest/gtest.h>
27
28 #include "absl/memory/memory.h"
29 #include "absl/strings/str_cat.h"
30 #include "absl/strings/str_format.h"
31
32 #include <grpc/grpc.h>
33 #include <grpc/support/alloc.h>
34 #include <grpc/support/log.h>
35 #include <grpc/support/time.h>
36 #include <grpcpp/channel.h>
37 #include <grpcpp/client_context.h>
38 #include <grpcpp/create_channel.h>
39 #include <grpcpp/impl/codegen/sync.h>
40 #include <grpcpp/server.h>
41 #include <grpcpp/server_builder.h>
42
43 #include "src/core/ext/filters/client_channel/backup_poller.h"
44 #include "src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.h"
45 #include "src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb_balancer_addresses.h"
46 #include "src/core/ext/filters/client_channel/resolver/fake/fake_resolver.h"
47 #include "src/core/ext/filters/client_channel/server_address.h"
48 #include "src/core/ext/service_config/service_config.h"
49 #include "src/core/lib/address_utils/parse_address.h"
50 #include "src/core/lib/gpr/env.h"
51 #include "src/core/lib/gprpp/ref_counted_ptr.h"
52 #include "src/core/lib/iomgr/sockaddr.h"
53 #include "src/core/lib/security/credentials/fake/fake_credentials.h"
54 #include "src/cpp/client/secure_credentials.h"
55 #include "src/cpp/server/secure_server_credentials.h"
56 #include "src/proto/grpc/lb/v1/load_balancer.grpc.pb.h"
57 #include "src/proto/grpc/testing/echo.grpc.pb.h"
58 #include "test/core/util/port.h"
59 #include "test/core/util/resolve_localhost_ip46.h"
60 #include "test/core/util/test_config.h"
61 #include "test/cpp/end2end/counted_service.h"
62 #include "test/cpp/end2end/test_service_impl.h"
63 #include "test/cpp/util/test_config.h"
64
65 // TODO(dgq): Other scenarios in need of testing:
66 // - Send a serverlist with faulty ip:port addresses (port > 2^16, etc).
67 // - Test reception of invalid serverlist
68 // - Test against a non-LB server.
69 // - Random LB server closing the stream unexpectedly.
70 //
71 // Findings from end to end testing to be covered here:
72 // - Handling of LB servers restart, including reconnection after backing-off
73 // retries.
74 // - Destruction of load balanced channel (and therefore of grpclb instance)
75 // while:
76 // 1) the internal LB call is still active. This should work by virtue
77 // of the weak reference the LB call holds. The call should be terminated as
78 // part of the grpclb shutdown process.
79 // 2) the retry timer is active. Again, the weak reference it holds should
80 // prevent a premature call to \a glb_destroy.
81
82 using std::chrono::system_clock;
83
84 using grpc::lb::v1::LoadBalancer;
85 using grpc::lb::v1::LoadBalanceRequest;
86 using grpc::lb::v1::LoadBalanceResponse;
87
88 namespace grpc {
89 namespace testing {
90 namespace {
91
92 constexpr char kDefaultServiceConfig[] =
93 "{\n"
94 " \"loadBalancingConfig\":[\n"
95 " { \"grpclb\":{} }\n"
96 " ]\n"
97 "}";
98
99 using BackendService = CountedService<TestServiceImpl>;
100 using BalancerService = CountedService<LoadBalancer::Service>;
101
102 const char g_kCallCredsMdKey[] = "Balancer should not ...";
103 const char g_kCallCredsMdValue[] = "... receive me";
104
105 class BackendServiceImpl : public BackendService {
106 public:
BackendServiceImpl()107 BackendServiceImpl() {}
108
Echo(ServerContext * context,const EchoRequest * request,EchoResponse * response)109 Status Echo(ServerContext* context, const EchoRequest* request,
110 EchoResponse* response) override {
111 // Backend should receive the call credentials metadata.
112 auto call_credentials_entry =
113 context->client_metadata().find(g_kCallCredsMdKey);
114 EXPECT_NE(call_credentials_entry, context->client_metadata().end());
115 if (call_credentials_entry != context->client_metadata().end()) {
116 EXPECT_EQ(call_credentials_entry->second, g_kCallCredsMdValue);
117 }
118 IncreaseRequestCount();
119 const auto status = TestServiceImpl::Echo(context, request, response);
120 IncreaseResponseCount();
121 AddClient(context->peer());
122 return status;
123 }
124
Start()125 void Start() {}
126
Shutdown()127 void Shutdown() {}
128
clients()129 std::set<std::string> clients() {
130 grpc::internal::MutexLock lock(&clients_mu_);
131 return clients_;
132 }
133
134 private:
AddClient(const std::string & client)135 void AddClient(const std::string& client) {
136 grpc::internal::MutexLock lock(&clients_mu_);
137 clients_.insert(client);
138 }
139
140 grpc::internal::Mutex clients_mu_;
141 std::set<std::string> clients_ ABSL_GUARDED_BY(&clients_mu_);
142 };
143
Ip4ToPackedString(const char * ip_str)144 std::string Ip4ToPackedString(const char* ip_str) {
145 struct in_addr ip4;
146 GPR_ASSERT(inet_pton(AF_INET, ip_str, &ip4) == 1);
147 return std::string(reinterpret_cast<const char*>(&ip4), sizeof(ip4));
148 }
149
Ip6ToPackedString(const char * ip_str)150 std::string Ip6ToPackedString(const char* ip_str) {
151 struct in6_addr ip6;
152 GPR_ASSERT(inet_pton(AF_INET6, ip_str, &ip6) == 1);
153 return std::string(reinterpret_cast<const char*>(&ip6), sizeof(ip6));
154 }
155
156 struct ClientStats {
157 size_t num_calls_started = 0;
158 size_t num_calls_finished = 0;
159 size_t num_calls_finished_with_client_failed_to_send = 0;
160 size_t num_calls_finished_known_received = 0;
161 std::map<std::string, size_t> drop_token_counts;
162
operator +=grpc::testing::__anone7fed90b0111::ClientStats163 ClientStats& operator+=(const ClientStats& other) {
164 num_calls_started += other.num_calls_started;
165 num_calls_finished += other.num_calls_finished;
166 num_calls_finished_with_client_failed_to_send +=
167 other.num_calls_finished_with_client_failed_to_send;
168 num_calls_finished_known_received +=
169 other.num_calls_finished_known_received;
170 for (const auto& p : other.drop_token_counts) {
171 drop_token_counts[p.first] += p.second;
172 }
173 return *this;
174 }
175
Resetgrpc::testing::__anone7fed90b0111::ClientStats176 void Reset() {
177 num_calls_started = 0;
178 num_calls_finished = 0;
179 num_calls_finished_with_client_failed_to_send = 0;
180 num_calls_finished_known_received = 0;
181 drop_token_counts.clear();
182 }
183 };
184
185 class BalancerServiceImpl : public BalancerService {
186 public:
187 using Stream = ServerReaderWriter<LoadBalanceResponse, LoadBalanceRequest>;
188 using ResponseDelayPair = std::pair<LoadBalanceResponse, int>;
189
BalancerServiceImpl(int client_load_reporting_interval_seconds)190 explicit BalancerServiceImpl(int client_load_reporting_interval_seconds)
191 : client_load_reporting_interval_seconds_(
192 client_load_reporting_interval_seconds) {}
193
BalanceLoad(ServerContext * context,Stream * stream)194 Status BalanceLoad(ServerContext* context, Stream* stream) override {
195 gpr_log(GPR_INFO, "LB[%p]: BalanceLoad", this);
196 {
197 grpc::internal::MutexLock lock(&mu_);
198 if (serverlist_done_) goto done;
199 }
200 {
201 // Balancer shouldn't receive the call credentials metadata.
202 EXPECT_EQ(context->client_metadata().find(g_kCallCredsMdKey),
203 context->client_metadata().end());
204 LoadBalanceRequest request;
205 std::vector<ResponseDelayPair> responses_and_delays;
206
207 if (!stream->Read(&request)) {
208 goto done;
209 } else {
210 if (request.has_initial_request()) {
211 grpc::internal::MutexLock lock(&mu_);
212 service_names_.push_back(request.initial_request().name());
213 }
214 }
215 IncreaseRequestCount();
216 gpr_log(GPR_INFO, "LB[%p]: received initial message '%s'", this,
217 request.DebugString().c_str());
218
219 // TODO(juanlishen): Initial response should always be the first response.
220 if (client_load_reporting_interval_seconds_ > 0) {
221 LoadBalanceResponse initial_response;
222 initial_response.mutable_initial_response()
223 ->mutable_client_stats_report_interval()
224 ->set_seconds(client_load_reporting_interval_seconds_);
225 stream->Write(initial_response);
226 }
227
228 {
229 grpc::internal::MutexLock lock(&mu_);
230 responses_and_delays = responses_and_delays_;
231 }
232 for (const auto& response_and_delay : responses_and_delays) {
233 SendResponse(stream, response_and_delay.first,
234 response_and_delay.second);
235 }
236 {
237 grpc::internal::MutexLock lock(&mu_);
238 while (!serverlist_done_) {
239 serverlist_cond_.Wait(&mu_);
240 }
241 }
242
243 if (client_load_reporting_interval_seconds_ > 0) {
244 request.Clear();
245 while (stream->Read(&request)) {
246 gpr_log(GPR_INFO, "LB[%p]: received client load report message '%s'",
247 this, request.DebugString().c_str());
248 GPR_ASSERT(request.has_client_stats());
249 ClientStats load_report;
250 load_report.num_calls_started =
251 request.client_stats().num_calls_started();
252 load_report.num_calls_finished =
253 request.client_stats().num_calls_finished();
254 load_report.num_calls_finished_with_client_failed_to_send =
255 request.client_stats()
256 .num_calls_finished_with_client_failed_to_send();
257 load_report.num_calls_finished_known_received =
258 request.client_stats().num_calls_finished_known_received();
259 for (const auto& drop_token_count :
260 request.client_stats().calls_finished_with_drop()) {
261 load_report
262 .drop_token_counts[drop_token_count.load_balance_token()] =
263 drop_token_count.num_calls();
264 }
265 // We need to acquire the lock here in order to prevent the notify_one
266 // below from firing before its corresponding wait is executed.
267 grpc::internal::MutexLock lock(&mu_);
268 load_report_queue_.emplace_back(std::move(load_report));
269 load_report_cond_.Signal();
270 }
271 }
272 }
273 done:
274 gpr_log(GPR_INFO, "LB[%p]: done", this);
275 return Status::OK;
276 }
277
add_response(const LoadBalanceResponse & response,int send_after_ms)278 void add_response(const LoadBalanceResponse& response, int send_after_ms) {
279 grpc::internal::MutexLock lock(&mu_);
280 responses_and_delays_.push_back(std::make_pair(response, send_after_ms));
281 }
282
Start()283 void Start() {
284 grpc::internal::MutexLock lock(&mu_);
285 serverlist_done_ = false;
286 responses_and_delays_.clear();
287 load_report_queue_.clear();
288 }
289
Shutdown()290 void Shutdown() {
291 NotifyDoneWithServerlists();
292 gpr_log(GPR_INFO, "LB[%p]: shut down", this);
293 }
294
WaitForLoadReport()295 ClientStats WaitForLoadReport() {
296 grpc::internal::MutexLock lock(&mu_);
297 if (load_report_queue_.empty()) {
298 while (load_report_queue_.empty()) {
299 load_report_cond_.Wait(&mu_);
300 }
301 }
302 ClientStats load_report = std::move(load_report_queue_.front());
303 load_report_queue_.pop_front();
304 return load_report;
305 }
306
NotifyDoneWithServerlists()307 void NotifyDoneWithServerlists() {
308 grpc::internal::MutexLock lock(&mu_);
309 if (!serverlist_done_) {
310 serverlist_done_ = true;
311 serverlist_cond_.SignalAll();
312 }
313 }
314
service_names()315 std::vector<std::string> service_names() {
316 grpc::internal::MutexLock lock(&mu_);
317 return service_names_;
318 }
319
320 private:
SendResponse(Stream * stream,const LoadBalanceResponse & response,int delay_ms)321 void SendResponse(Stream* stream, const LoadBalanceResponse& response,
322 int delay_ms) {
323 gpr_log(GPR_INFO, "LB[%p]: sleeping for %d ms...", this, delay_ms);
324 if (delay_ms > 0) {
325 gpr_sleep_until(grpc_timeout_milliseconds_to_deadline(delay_ms));
326 }
327 gpr_log(GPR_INFO, "LB[%p]: Woke up! Sending response '%s'", this,
328 response.DebugString().c_str());
329 IncreaseResponseCount();
330 stream->Write(response);
331 }
332
333 const int client_load_reporting_interval_seconds_;
334 std::vector<ResponseDelayPair> responses_and_delays_;
335 std::vector<std::string> service_names_;
336
337 grpc::internal::Mutex mu_;
338 grpc::internal::CondVar serverlist_cond_;
339 bool serverlist_done_ ABSL_GUARDED_BY(mu_) = false;
340 grpc::internal::CondVar load_report_cond_;
341 std::deque<ClientStats> load_report_queue_ ABSL_GUARDED_BY(mu_);
342 };
343
344 class GrpclbEnd2endTest : public ::testing::Test {
345 protected:
GrpclbEnd2endTest(size_t num_backends,size_t num_balancers,int client_load_reporting_interval_seconds)346 GrpclbEnd2endTest(size_t num_backends, size_t num_balancers,
347 int client_load_reporting_interval_seconds)
348 : server_host_("localhost"),
349 num_backends_(num_backends),
350 num_balancers_(num_balancers),
351 client_load_reporting_interval_seconds_(
352 client_load_reporting_interval_seconds) {}
353
SetUpTestCase()354 static void SetUpTestCase() {
355 // Make the backup poller poll very frequently in order to pick up
356 // updates from all the subchannels's FDs.
357 GPR_GLOBAL_CONFIG_SET(grpc_client_channel_backup_poll_interval_ms, 1);
358 #if TARGET_OS_IPHONE
359 // Workaround Apple CFStream bug
360 gpr_setenv("grpc_cfstream", "0");
361 #endif
362 grpc_init();
363 }
364
TearDownTestCase()365 static void TearDownTestCase() { grpc_shutdown(); }
366
SetUp()367 void SetUp() override {
368 bool localhost_resolves_to_ipv4 = false;
369 bool localhost_resolves_to_ipv6 = false;
370 grpc_core::LocalhostResolves(&localhost_resolves_to_ipv4,
371 &localhost_resolves_to_ipv6);
372 ipv6_only_ = !localhost_resolves_to_ipv4 && localhost_resolves_to_ipv6;
373 response_generator_ =
374 grpc_core::MakeRefCounted<grpc_core::FakeResolverResponseGenerator>();
375 // Start the backends.
376 for (size_t i = 0; i < num_backends_; ++i) {
377 backends_.emplace_back(new ServerThread<BackendServiceImpl>("backend"));
378 backends_.back()->Start(server_host_);
379 }
380 // Start the load balancers.
381 for (size_t i = 0; i < num_balancers_; ++i) {
382 balancers_.emplace_back(new ServerThread<BalancerServiceImpl>(
383 "balancer", client_load_reporting_interval_seconds_));
384 balancers_.back()->Start(server_host_);
385 }
386 ResetStub();
387 }
388
TearDown()389 void TearDown() override {
390 ShutdownAllBackends();
391 for (auto& balancer : balancers_) balancer->Shutdown();
392 }
393
StartAllBackends()394 void StartAllBackends() {
395 for (auto& backend : backends_) backend->Start(server_host_);
396 }
397
StartBackend(size_t index)398 void StartBackend(size_t index) { backends_[index]->Start(server_host_); }
399
ShutdownAllBackends()400 void ShutdownAllBackends() {
401 for (auto& backend : backends_) backend->Shutdown();
402 }
403
ShutdownBackend(size_t index)404 void ShutdownBackend(size_t index) { backends_[index]->Shutdown(); }
405
ResetStub(int fallback_timeout=0,const std::string & expected_targets="",int subchannel_cache_delay_ms=0)406 void ResetStub(int fallback_timeout = 0,
407 const std::string& expected_targets = "",
408 int subchannel_cache_delay_ms = 0) {
409 ChannelArguments args;
410 if (fallback_timeout > 0) args.SetGrpclbFallbackTimeout(fallback_timeout);
411 args.SetPointer(GRPC_ARG_FAKE_RESOLVER_RESPONSE_GENERATOR,
412 response_generator_.get());
413 if (!expected_targets.empty()) {
414 args.SetString(GRPC_ARG_FAKE_SECURITY_EXPECTED_TARGETS, expected_targets);
415 }
416 if (subchannel_cache_delay_ms > 0) {
417 args.SetInt(GRPC_ARG_GRPCLB_SUBCHANNEL_CACHE_INTERVAL_MS,
418 subchannel_cache_delay_ms);
419 }
420 std::ostringstream uri;
421 uri << "fake:///" << kApplicationTargetName_;
422 // TODO(dgq): templatize tests to run everything using both secure and
423 // insecure channel credentials.
424 grpc_channel_credentials* channel_creds =
425 grpc_fake_transport_security_credentials_create();
426 grpc_call_credentials* call_creds = grpc_md_only_test_credentials_create(
427 g_kCallCredsMdKey, g_kCallCredsMdValue, false);
428 std::shared_ptr<ChannelCredentials> creds(
429 new SecureChannelCredentials(grpc_composite_channel_credentials_create(
430 channel_creds, call_creds, nullptr)));
431 call_creds->Unref();
432 channel_creds->Unref();
433 channel_ = ::grpc::CreateCustomChannel(uri.str(), creds, args);
434 stub_ = grpc::testing::EchoTestService::NewStub(channel_);
435 }
436
ResetBackendCounters()437 void ResetBackendCounters() {
438 for (auto& backend : backends_) backend->service_.ResetCounters();
439 }
440
WaitForLoadReports()441 ClientStats WaitForLoadReports() {
442 ClientStats client_stats;
443 for (auto& balancer : balancers_) {
444 client_stats += balancer->service_.WaitForLoadReport();
445 }
446 return client_stats;
447 }
448
SeenAllBackends(size_t start_index=0,size_t stop_index=0)449 bool SeenAllBackends(size_t start_index = 0, size_t stop_index = 0) {
450 if (stop_index == 0) stop_index = backends_.size();
451 for (size_t i = start_index; i < stop_index; ++i) {
452 if (backends_[i]->service_.request_count() == 0) return false;
453 }
454 return true;
455 }
456
SendRpcAndCount(int * num_total,int * num_ok,int * num_failure,int * num_drops)457 void SendRpcAndCount(int* num_total, int* num_ok, int* num_failure,
458 int* num_drops) {
459 const Status status = SendRpc();
460 if (status.ok()) {
461 ++*num_ok;
462 } else {
463 if (status.error_message() == "drop directed by grpclb balancer") {
464 ++*num_drops;
465 } else {
466 ++*num_failure;
467 }
468 }
469 ++*num_total;
470 }
471
WaitForAllBackends(int num_requests_multiple_of=1,size_t start_index=0,size_t stop_index=0)472 std::tuple<int, int, int> WaitForAllBackends(int num_requests_multiple_of = 1,
473 size_t start_index = 0,
474 size_t stop_index = 0) {
475 int num_ok = 0;
476 int num_failure = 0;
477 int num_drops = 0;
478 int num_total = 0;
479 while (!SeenAllBackends(start_index, stop_index)) {
480 SendRpcAndCount(&num_total, &num_ok, &num_failure, &num_drops);
481 }
482 while (num_total % num_requests_multiple_of != 0) {
483 SendRpcAndCount(&num_total, &num_ok, &num_failure, &num_drops);
484 }
485 ResetBackendCounters();
486 gpr_log(GPR_INFO,
487 "Performed %d warm up requests (a multiple of %d) against the "
488 "backends. %d succeeded, %d failed, %d dropped.",
489 num_total, num_requests_multiple_of, num_ok, num_failure,
490 num_drops);
491 return std::make_tuple(num_ok, num_failure, num_drops);
492 }
493
WaitForBackend(size_t backend_idx)494 void WaitForBackend(size_t backend_idx) {
495 do {
496 (void)SendRpc();
497 } while (backends_[backend_idx]->service_.request_count() == 0);
498 ResetBackendCounters();
499 }
500
501 struct AddressData {
502 int port;
503 std::string balancer_name;
504 };
505
CreateLbAddressesFromAddressDataList(const std::vector<AddressData> & address_data)506 grpc_core::ServerAddressList CreateLbAddressesFromAddressDataList(
507 const std::vector<AddressData>& address_data) {
508 grpc_core::ServerAddressList addresses;
509 for (const auto& addr : address_data) {
510 absl::StatusOr<grpc_core::URI> lb_uri =
511 grpc_core::URI::Parse(absl::StrCat(
512 ipv6_only_ ? "ipv6:[::1]:" : "ipv4:127.0.0.1:", addr.port));
513 GPR_ASSERT(lb_uri.ok());
514 grpc_resolved_address address;
515 GPR_ASSERT(grpc_parse_uri(*lb_uri, &address));
516 grpc_arg arg = grpc_channel_arg_string_create(
517 const_cast<char*>(GRPC_ARG_DEFAULT_AUTHORITY),
518 const_cast<char*>(addr.balancer_name.c_str()));
519 grpc_channel_args* args =
520 grpc_channel_args_copy_and_add(nullptr, &arg, 1);
521 addresses.emplace_back(address.addr, address.len, args);
522 }
523 return addresses;
524 }
525
MakeResolverResult(const std::vector<AddressData> & balancer_address_data,const std::vector<AddressData> & backend_address_data={},const char * service_config_json=kDefaultServiceConfig)526 grpc_core::Resolver::Result MakeResolverResult(
527 const std::vector<AddressData>& balancer_address_data,
528 const std::vector<AddressData>& backend_address_data = {},
529 const char* service_config_json = kDefaultServiceConfig) {
530 grpc_core::Resolver::Result result;
531 result.addresses =
532 CreateLbAddressesFromAddressDataList(backend_address_data);
533 grpc_error_handle error = GRPC_ERROR_NONE;
534 result.service_config =
535 grpc_core::ServiceConfig::Create(nullptr, service_config_json, &error);
536 GPR_ASSERT(error == GRPC_ERROR_NONE);
537 grpc_core::ServerAddressList balancer_addresses =
538 CreateLbAddressesFromAddressDataList(balancer_address_data);
539 grpc_arg arg = CreateGrpclbBalancerAddressesArg(&balancer_addresses);
540 result.args = grpc_channel_args_copy_and_add(nullptr, &arg, 1);
541 return result;
542 }
543
SetNextResolutionAllBalancers(const char * service_config_json=kDefaultServiceConfig)544 void SetNextResolutionAllBalancers(
545 const char* service_config_json = kDefaultServiceConfig) {
546 std::vector<AddressData> addresses;
547 for (size_t i = 0; i < balancers_.size(); ++i) {
548 addresses.emplace_back(AddressData{balancers_[i]->port_, ""});
549 }
550 SetNextResolution(addresses, {}, service_config_json);
551 }
552
SetNextResolution(const std::vector<AddressData> & balancer_address_data,const std::vector<AddressData> & backend_address_data={},const char * service_config_json=kDefaultServiceConfig)553 void SetNextResolution(
554 const std::vector<AddressData>& balancer_address_data,
555 const std::vector<AddressData>& backend_address_data = {},
556 const char* service_config_json = kDefaultServiceConfig) {
557 grpc_core::ExecCtx exec_ctx;
558 grpc_core::Resolver::Result result = MakeResolverResult(
559 balancer_address_data, backend_address_data, service_config_json);
560 response_generator_->SetResponse(std::move(result));
561 }
562
SetNextReresolutionResponse(const std::vector<AddressData> & balancer_address_data,const std::vector<AddressData> & backend_address_data={},const char * service_config_json=kDefaultServiceConfig)563 void SetNextReresolutionResponse(
564 const std::vector<AddressData>& balancer_address_data,
565 const std::vector<AddressData>& backend_address_data = {},
566 const char* service_config_json = kDefaultServiceConfig) {
567 grpc_core::ExecCtx exec_ctx;
568 grpc_core::Resolver::Result result = MakeResolverResult(
569 balancer_address_data, backend_address_data, service_config_json);
570 response_generator_->SetReresolutionResponse(std::move(result));
571 }
572
GetBackendPorts(size_t start_index=0,size_t stop_index=0) const573 std::vector<int> GetBackendPorts(size_t start_index = 0,
574 size_t stop_index = 0) const {
575 if (stop_index == 0) stop_index = backends_.size();
576 std::vector<int> backend_ports;
577 for (size_t i = start_index; i < stop_index; ++i) {
578 backend_ports.push_back(backends_[i]->port_);
579 }
580 return backend_ports;
581 }
582
ScheduleResponseForBalancer(size_t i,const LoadBalanceResponse & response,int delay_ms)583 void ScheduleResponseForBalancer(size_t i,
584 const LoadBalanceResponse& response,
585 int delay_ms) {
586 balancers_[i]->service_.add_response(response, delay_ms);
587 }
588
BuildResponseForBackends(const std::vector<int> & backend_ports,const std::map<std::string,size_t> & drop_token_counts)589 LoadBalanceResponse BuildResponseForBackends(
590 const std::vector<int>& backend_ports,
591 const std::map<std::string, size_t>& drop_token_counts) {
592 LoadBalanceResponse response;
593 for (const auto& drop_token_count : drop_token_counts) {
594 for (size_t i = 0; i < drop_token_count.second; ++i) {
595 auto* server = response.mutable_server_list()->add_servers();
596 server->set_drop(true);
597 server->set_load_balance_token(drop_token_count.first);
598 }
599 }
600 for (const int& backend_port : backend_ports) {
601 auto* server = response.mutable_server_list()->add_servers();
602 server->set_ip_address(ipv6_only_ ? Ip6ToPackedString("::1")
603 : Ip4ToPackedString("127.0.0.1"));
604 server->set_port(backend_port);
605 static int token_count = 0;
606 server->set_load_balance_token(
607 absl::StrFormat("token%03d", ++token_count));
608 }
609 return response;
610 }
611
SendRpc(EchoResponse * response=nullptr,int timeout_ms=1000,bool wait_for_ready=false,const Status & expected_status=Status::OK)612 Status SendRpc(EchoResponse* response = nullptr, int timeout_ms = 1000,
613 bool wait_for_ready = false,
614 const Status& expected_status = Status::OK) {
615 const bool local_response = (response == nullptr);
616 if (local_response) response = new EchoResponse;
617 EchoRequest request;
618 request.set_message(kRequestMessage_);
619 if (!expected_status.ok()) {
620 auto* error = request.mutable_param()->mutable_expected_error();
621 error->set_code(expected_status.error_code());
622 error->set_error_message(expected_status.error_message());
623 }
624 ClientContext context;
625 context.set_deadline(grpc_timeout_milliseconds_to_deadline(timeout_ms));
626 if (wait_for_ready) context.set_wait_for_ready(true);
627 Status status = stub_->Echo(&context, request, response);
628 if (local_response) delete response;
629 return status;
630 }
631
CheckRpcSendOk(const size_t times=1,const int timeout_ms=1000,bool wait_for_ready=false)632 void CheckRpcSendOk(const size_t times = 1, const int timeout_ms = 1000,
633 bool wait_for_ready = false) {
634 for (size_t i = 0; i < times; ++i) {
635 EchoResponse response;
636 const Status status = SendRpc(&response, timeout_ms, wait_for_ready);
637 EXPECT_TRUE(status.ok()) << "code=" << status.error_code()
638 << " message=" << status.error_message();
639 EXPECT_EQ(response.message(), kRequestMessage_);
640 }
641 }
642
CheckRpcSendFailure()643 void CheckRpcSendFailure() {
644 const Status status = SendRpc();
645 EXPECT_FALSE(status.ok());
646 }
647
648 template <typename T>
649 struct ServerThread {
650 template <typename... Args>
ServerThreadgrpc::testing::__anone7fed90b0111::GrpclbEnd2endTest::ServerThread651 explicit ServerThread(const std::string& type, Args&&... args)
652 : port_(grpc_pick_unused_port_or_die()),
653 type_(type),
654 service_(std::forward<Args>(args)...) {}
655
Startgrpc::testing::__anone7fed90b0111::GrpclbEnd2endTest::ServerThread656 void Start(const std::string& server_host) {
657 gpr_log(GPR_INFO, "starting %s server on port %d", type_.c_str(), port_);
658 GPR_ASSERT(!running_);
659 running_ = true;
660 service_.Start();
661 grpc::internal::Mutex mu;
662 // We need to acquire the lock here in order to prevent the notify_one
663 // by ServerThread::Serve from firing before the wait below is hit.
664 grpc::internal::MutexLock lock(&mu);
665 grpc::internal::CondVar cond;
666 thread_ = absl::make_unique<std::thread>(
667 std::bind(&ServerThread::Serve, this, server_host, &mu, &cond));
668 cond.Wait(&mu);
669 gpr_log(GPR_INFO, "%s server startup complete", type_.c_str());
670 }
671
Servegrpc::testing::__anone7fed90b0111::GrpclbEnd2endTest::ServerThread672 void Serve(const std::string& server_host, grpc::internal::Mutex* mu,
673 grpc::internal::CondVar* cond) {
674 // We need to acquire the lock here in order to prevent the notify_one
675 // below from firing before its corresponding wait is executed.
676 grpc::internal::MutexLock lock(mu);
677 std::ostringstream server_address;
678 server_address << server_host << ":" << port_;
679 ServerBuilder builder;
680 std::shared_ptr<ServerCredentials> creds(new SecureServerCredentials(
681 grpc_fake_transport_security_server_credentials_create()));
682 builder.AddListeningPort(server_address.str(), creds);
683 builder.RegisterService(&service_);
684 server_ = builder.BuildAndStart();
685 cond->Signal();
686 }
687
Shutdowngrpc::testing::__anone7fed90b0111::GrpclbEnd2endTest::ServerThread688 void Shutdown() {
689 if (!running_) return;
690 gpr_log(GPR_INFO, "%s about to shutdown", type_.c_str());
691 service_.Shutdown();
692 server_->Shutdown(grpc_timeout_milliseconds_to_deadline(0));
693 thread_->join();
694 gpr_log(GPR_INFO, "%s shutdown completed", type_.c_str());
695 running_ = false;
696 }
697
698 const int port_;
699 std::string type_;
700 T service_;
701 std::unique_ptr<Server> server_;
702 std::unique_ptr<std::thread> thread_;
703 bool running_ = false;
704 };
705
706 const std::string server_host_;
707 const size_t num_backends_;
708 const size_t num_balancers_;
709 const int client_load_reporting_interval_seconds_;
710 bool ipv6_only_ = false;
711 std::shared_ptr<Channel> channel_;
712 std::unique_ptr<grpc::testing::EchoTestService::Stub> stub_;
713 std::vector<std::unique_ptr<ServerThread<BackendServiceImpl>>> backends_;
714 std::vector<std::unique_ptr<ServerThread<BalancerServiceImpl>>> balancers_;
715 grpc_core::RefCountedPtr<grpc_core::FakeResolverResponseGenerator>
716 response_generator_;
717 const std::string kRequestMessage_ = "Live long and prosper.";
718 const std::string kApplicationTargetName_ = "application_target_name";
719 };
720
721 class SingleBalancerTest : public GrpclbEnd2endTest {
722 public:
SingleBalancerTest()723 SingleBalancerTest() : GrpclbEnd2endTest(4, 1, 0) {}
724 };
725
TEST_F(SingleBalancerTest,Vanilla)726 TEST_F(SingleBalancerTest, Vanilla) {
727 SetNextResolutionAllBalancers();
728 const size_t kNumRpcsPerAddress = 100;
729 ScheduleResponseForBalancer(
730 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
731 // Make sure that trying to connect works without a call.
732 channel_->GetState(true /* try_to_connect */);
733 // We need to wait for all backends to come online.
734 WaitForAllBackends();
735 // Send kNumRpcsPerAddress RPCs per server.
736 CheckRpcSendOk(kNumRpcsPerAddress * num_backends_);
737
738 // Each backend should have gotten 100 requests.
739 for (size_t i = 0; i < backends_.size(); ++i) {
740 EXPECT_EQ(kNumRpcsPerAddress, backends_[i]->service_.request_count());
741 }
742 balancers_[0]->service_.NotifyDoneWithServerlists();
743 // The balancer got a single request.
744 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
745 // and sent a single response.
746 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
747
748 // Check LB policy name for the channel.
749 EXPECT_EQ("grpclb", channel_->GetLoadBalancingPolicyName());
750 }
751
TEST_F(SingleBalancerTest,SubchannelCaching)752 TEST_F(SingleBalancerTest, SubchannelCaching) {
753 ResetStub(/*fallback_timeout=*/0, /*expected_targets=*/"",
754 /*subchannel_cache_delay_ms=*/1500);
755 SetNextResolutionAllBalancers();
756 // Initially send all backends.
757 ScheduleResponseForBalancer(
758 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
759 // Then remove backends 0 and 1.
760 ScheduleResponseForBalancer(
761 0, BuildResponseForBackends(GetBackendPorts(2), {}), 1000);
762 // Now re-add backend 1.
763 ScheduleResponseForBalancer(
764 0, BuildResponseForBackends(GetBackendPorts(1), {}), 1000);
765 // Wait for all backends to come online.
766 WaitForAllBackends();
767 // Send RPCs for long enough to get all responses.
768 gpr_timespec deadline = grpc_timeout_milliseconds_to_deadline(3000);
769 do {
770 CheckRpcSendOk();
771 } while (gpr_time_cmp(gpr_now(GPR_CLOCK_MONOTONIC), deadline) < 0);
772 // Backend 0 should have received less traffic than the others.
773 // Backend 1 would have received less traffic than 2 and 3.
774 gpr_log(GPR_INFO, "BACKEND 0: %" PRIuPTR " requests",
775 backends_[0]->service_.request_count());
776 EXPECT_GT(backends_[0]->service_.request_count(), 0);
777 for (size_t i = 1; i < backends_.size(); ++i) {
778 gpr_log(GPR_INFO, "BACKEND %" PRIuPTR ": %" PRIuPTR " requests", i,
779 backends_[i]->service_.request_count());
780 EXPECT_GT(backends_[i]->service_.request_count(),
781 backends_[0]->service_.request_count())
782 << "backend " << i;
783 if (i >= 2) {
784 EXPECT_GT(backends_[i]->service_.request_count(),
785 backends_[1]->service_.request_count())
786 << "backend " << i;
787 }
788 }
789 // Backend 1 should never have lost its connection from the client.
790 EXPECT_EQ(1UL, backends_[1]->service_.clients().size());
791 balancers_[0]->service_.NotifyDoneWithServerlists();
792 // The balancer got a single request.
793 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
794 // And sent 3 responses.
795 EXPECT_EQ(3U, balancers_[0]->service_.response_count());
796 }
797
TEST_F(SingleBalancerTest,ReturnServerStatus)798 TEST_F(SingleBalancerTest, ReturnServerStatus) {
799 SetNextResolutionAllBalancers();
800 ScheduleResponseForBalancer(
801 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
802 // We need to wait for all backends to come online.
803 WaitForAllBackends();
804 // Send a request that the backend will fail, and make sure we get
805 // back the right status.
806 Status expected(StatusCode::INVALID_ARGUMENT, "He's dead, Jim!");
807 Status actual = SendRpc(/*response=*/nullptr, /*timeout_ms=*/1000,
808 /*wait_for_ready=*/false, expected);
809 EXPECT_EQ(actual.error_code(), expected.error_code());
810 EXPECT_EQ(actual.error_message(), expected.error_message());
811 }
812
TEST_F(SingleBalancerTest,SelectGrpclbWithMigrationServiceConfig)813 TEST_F(SingleBalancerTest, SelectGrpclbWithMigrationServiceConfig) {
814 SetNextResolutionAllBalancers(
815 "{\n"
816 " \"loadBalancingConfig\":[\n"
817 " { \"does_not_exist\":{} },\n"
818 " { \"grpclb\":{} }\n"
819 " ]\n"
820 "}");
821 ScheduleResponseForBalancer(
822 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
823 CheckRpcSendOk(1, 1000 /* timeout_ms */, true /* wait_for_ready */);
824 balancers_[0]->service_.NotifyDoneWithServerlists();
825 // The balancer got a single request.
826 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
827 // and sent a single response.
828 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
829 // Check LB policy name for the channel.
830 EXPECT_EQ("grpclb", channel_->GetLoadBalancingPolicyName());
831 }
832
TEST_F(SingleBalancerTest,SelectGrpclbWithMigrationServiceConfigAndNoAddresses)833 TEST_F(SingleBalancerTest,
834 SelectGrpclbWithMigrationServiceConfigAndNoAddresses) {
835 const int kFallbackTimeoutMs = 200 * grpc_test_slowdown_factor();
836 ResetStub(kFallbackTimeoutMs);
837 SetNextResolution({}, {},
838 "{\n"
839 " \"loadBalancingConfig\":[\n"
840 " { \"does_not_exist\":{} },\n"
841 " { \"grpclb\":{} }\n"
842 " ]\n"
843 "}");
844 // Try to connect.
845 EXPECT_EQ(GRPC_CHANNEL_IDLE, channel_->GetState(true));
846 // Should go into state TRANSIENT_FAILURE when we enter fallback mode.
847 const gpr_timespec deadline = grpc_timeout_seconds_to_deadline(1);
848 grpc_connectivity_state state;
849 while ((state = channel_->GetState(false)) !=
850 GRPC_CHANNEL_TRANSIENT_FAILURE) {
851 ASSERT_TRUE(channel_->WaitForStateChange(state, deadline));
852 }
853 // Check LB policy name for the channel.
854 EXPECT_EQ("grpclb", channel_->GetLoadBalancingPolicyName());
855 }
856
TEST_F(SingleBalancerTest,UsePickFirstChildPolicy)857 TEST_F(SingleBalancerTest, UsePickFirstChildPolicy) {
858 SetNextResolutionAllBalancers(
859 "{\n"
860 " \"loadBalancingConfig\":[\n"
861 " { \"grpclb\":{\n"
862 " \"childPolicy\":[\n"
863 " { \"pick_first\":{} }\n"
864 " ]\n"
865 " } }\n"
866 " ]\n"
867 "}");
868 ScheduleResponseForBalancer(
869 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
870 const size_t kNumRpcs = num_backends_ * 2;
871 CheckRpcSendOk(kNumRpcs, 1000 /* timeout_ms */, true /* wait_for_ready */);
872 balancers_[0]->service_.NotifyDoneWithServerlists();
873 // Check that all requests went to the first backend. This verifies
874 // that we used pick_first instead of round_robin as the child policy.
875 EXPECT_EQ(backends_[0]->service_.request_count(), kNumRpcs);
876 for (size_t i = 1; i < backends_.size(); ++i) {
877 EXPECT_EQ(backends_[i]->service_.request_count(), 0UL);
878 }
879 // The balancer got a single request.
880 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
881 // and sent a single response.
882 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
883 // Check LB policy name for the channel.
884 EXPECT_EQ("grpclb", channel_->GetLoadBalancingPolicyName());
885 }
886
TEST_F(SingleBalancerTest,SwapChildPolicy)887 TEST_F(SingleBalancerTest, SwapChildPolicy) {
888 SetNextResolutionAllBalancers(
889 "{\n"
890 " \"loadBalancingConfig\":[\n"
891 " { \"grpclb\":{\n"
892 " \"childPolicy\":[\n"
893 " { \"pick_first\":{} }\n"
894 " ]\n"
895 " } }\n"
896 " ]\n"
897 "}");
898 ScheduleResponseForBalancer(
899 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
900 const size_t kNumRpcs = num_backends_ * 2;
901 CheckRpcSendOk(kNumRpcs, 1000 /* timeout_ms */, true /* wait_for_ready */);
902 // Check that all requests went to the first backend. This verifies
903 // that we used pick_first instead of round_robin as the child policy.
904 EXPECT_EQ(backends_[0]->service_.request_count(), kNumRpcs);
905 for (size_t i = 1; i < backends_.size(); ++i) {
906 EXPECT_EQ(backends_[i]->service_.request_count(), 0UL);
907 }
908 // Send new resolution that removes child policy from service config.
909 SetNextResolutionAllBalancers();
910 WaitForAllBackends();
911 CheckRpcSendOk(kNumRpcs, 1000 /* timeout_ms */, true /* wait_for_ready */);
912 // Check that every backend saw the same number of requests. This verifies
913 // that we used round_robin.
914 for (size_t i = 0; i < backends_.size(); ++i) {
915 EXPECT_EQ(backends_[i]->service_.request_count(), 2UL);
916 }
917 // Done.
918 balancers_[0]->service_.NotifyDoneWithServerlists();
919 // The balancer got a single request.
920 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
921 // and sent a single response.
922 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
923 // Check LB policy name for the channel.
924 EXPECT_EQ("grpclb", channel_->GetLoadBalancingPolicyName());
925 }
926
TEST_F(SingleBalancerTest,SameBackendListedMultipleTimes)927 TEST_F(SingleBalancerTest, SameBackendListedMultipleTimes) {
928 SetNextResolutionAllBalancers();
929 // Same backend listed twice.
930 std::vector<int> ports;
931 ports.push_back(backends_[0]->port_);
932 ports.push_back(backends_[0]->port_);
933 const size_t kNumRpcsPerAddress = 10;
934 ScheduleResponseForBalancer(0, BuildResponseForBackends(ports, {}), 0);
935 // We need to wait for the backend to come online.
936 WaitForBackend(0);
937 // Send kNumRpcsPerAddress RPCs per server.
938 CheckRpcSendOk(kNumRpcsPerAddress * ports.size());
939 // Backend should have gotten 20 requests.
940 EXPECT_EQ(kNumRpcsPerAddress * 2, backends_[0]->service_.request_count());
941 // And they should have come from a single client port, because of
942 // subchannel sharing.
943 EXPECT_EQ(1UL, backends_[0]->service_.clients().size());
944 balancers_[0]->service_.NotifyDoneWithServerlists();
945 }
946
TEST_F(SingleBalancerTest,SecureNaming)947 TEST_F(SingleBalancerTest, SecureNaming) {
948 ResetStub(0, kApplicationTargetName_ + ";lb");
949 SetNextResolution({AddressData{balancers_[0]->port_, "lb"}});
950 const size_t kNumRpcsPerAddress = 100;
951 ScheduleResponseForBalancer(
952 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
953 // Make sure that trying to connect works without a call.
954 channel_->GetState(true /* try_to_connect */);
955 // We need to wait for all backends to come online.
956 WaitForAllBackends();
957 // Send kNumRpcsPerAddress RPCs per server.
958 CheckRpcSendOk(kNumRpcsPerAddress * num_backends_);
959
960 // Each backend should have gotten 100 requests.
961 for (size_t i = 0; i < backends_.size(); ++i) {
962 EXPECT_EQ(kNumRpcsPerAddress, backends_[i]->service_.request_count());
963 }
964 balancers_[0]->service_.NotifyDoneWithServerlists();
965 // The balancer got a single request.
966 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
967 // and sent a single response.
968 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
969 // Check LB policy name for the channel.
970 EXPECT_EQ("grpclb", channel_->GetLoadBalancingPolicyName());
971 }
972
TEST_F(SingleBalancerTest,SecureNamingDeathTest)973 TEST_F(SingleBalancerTest, SecureNamingDeathTest) {
974 GRPC_GTEST_FLAG_SET_DEATH_TEST_STYLE("threadsafe");
975 // Make sure that we blow up (via abort() from the security connector) when
976 // the name from the balancer doesn't match expectations.
977 ASSERT_DEATH_IF_SUPPORTED(
978 {
979 ResetStub(0, kApplicationTargetName_ + ";lb");
980 SetNextResolution({AddressData{balancers_[0]->port_, "woops"}});
981 channel_->WaitForConnected(grpc_timeout_seconds_to_deadline(1));
982 },
983 "");
984 }
985
TEST_F(SingleBalancerTest,InitiallyEmptyServerlist)986 TEST_F(SingleBalancerTest, InitiallyEmptyServerlist) {
987 SetNextResolutionAllBalancers();
988 const int kServerlistDelayMs = 500 * grpc_test_slowdown_factor();
989 const int kCallDeadlineMs = kServerlistDelayMs * 2;
990 // First response is an empty serverlist, sent right away.
991 ScheduleResponseForBalancer(0, LoadBalanceResponse(), 0);
992 // Send non-empty serverlist only after kServerlistDelayMs
993 ScheduleResponseForBalancer(
994 0, BuildResponseForBackends(GetBackendPorts(), {}), kServerlistDelayMs);
995 const auto t0 = system_clock::now();
996 // Client will block: LB will initially send empty serverlist.
997 CheckRpcSendOk(1, kCallDeadlineMs, true /* wait_for_ready */);
998 const auto ellapsed_ms =
999 std::chrono::duration_cast<std::chrono::milliseconds>(
1000 system_clock::now() - t0);
1001 // but eventually, the LB sends a serverlist update that allows the call to
1002 // proceed. The call delay must be larger than the delay in sending the
1003 // populated serverlist but under the call's deadline (which is enforced by
1004 // the call's deadline).
1005 EXPECT_GT(ellapsed_ms.count(), kServerlistDelayMs);
1006 balancers_[0]->service_.NotifyDoneWithServerlists();
1007 // The balancer got a single request.
1008 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1009 // and sent two responses.
1010 EXPECT_EQ(2U, balancers_[0]->service_.response_count());
1011 }
1012
TEST_F(SingleBalancerTest,AllServersUnreachableFailFast)1013 TEST_F(SingleBalancerTest, AllServersUnreachableFailFast) {
1014 SetNextResolutionAllBalancers();
1015 const size_t kNumUnreachableServers = 5;
1016 std::vector<int> ports;
1017 for (size_t i = 0; i < kNumUnreachableServers; ++i) {
1018 ports.push_back(grpc_pick_unused_port_or_die());
1019 }
1020 ScheduleResponseForBalancer(0, BuildResponseForBackends(ports, {}), 0);
1021 const Status status = SendRpc();
1022 // The error shouldn't be DEADLINE_EXCEEDED.
1023 EXPECT_EQ(StatusCode::UNAVAILABLE, status.error_code());
1024 balancers_[0]->service_.NotifyDoneWithServerlists();
1025 // The balancer got a single request.
1026 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1027 // and sent a single response.
1028 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1029 }
1030
TEST_F(SingleBalancerTest,Fallback)1031 TEST_F(SingleBalancerTest, Fallback) {
1032 SetNextResolutionAllBalancers();
1033 const int kFallbackTimeoutMs = 200 * grpc_test_slowdown_factor();
1034 const int kServerlistDelayMs = 500 * grpc_test_slowdown_factor();
1035 const size_t kNumBackendsInResolution = backends_.size() / 2;
1036
1037 ResetStub(kFallbackTimeoutMs);
1038 std::vector<AddressData> balancer_addresses;
1039 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1040 std::vector<AddressData> backend_addresses;
1041 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1042 backend_addresses.emplace_back(AddressData{backends_[i]->port_, ""});
1043 }
1044 SetNextResolution(balancer_addresses, backend_addresses);
1045
1046 // Send non-empty serverlist only after kServerlistDelayMs.
1047 ScheduleResponseForBalancer(
1048 0,
1049 BuildResponseForBackends(
1050 GetBackendPorts(kNumBackendsInResolution /* start_index */), {}),
1051 kServerlistDelayMs);
1052
1053 // Wait until all the fallback backends are reachable.
1054 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1055 WaitForBackend(i);
1056 }
1057
1058 // The first request.
1059 gpr_log(GPR_INFO, "========= BEFORE FIRST BATCH ==========");
1060 CheckRpcSendOk(kNumBackendsInResolution);
1061 gpr_log(GPR_INFO, "========= DONE WITH FIRST BATCH ==========");
1062
1063 // Fallback is used: each backend returned by the resolver should have
1064 // gotten one request.
1065 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1066 EXPECT_EQ(1U, backends_[i]->service_.request_count());
1067 }
1068 for (size_t i = kNumBackendsInResolution; i < backends_.size(); ++i) {
1069 EXPECT_EQ(0U, backends_[i]->service_.request_count());
1070 }
1071
1072 // Wait until the serverlist reception has been processed and all backends
1073 // in the serverlist are reachable.
1074 for (size_t i = kNumBackendsInResolution; i < backends_.size(); ++i) {
1075 WaitForBackend(i);
1076 }
1077
1078 // Send out the second request.
1079 gpr_log(GPR_INFO, "========= BEFORE SECOND BATCH ==========");
1080 CheckRpcSendOk(backends_.size() - kNumBackendsInResolution);
1081 gpr_log(GPR_INFO, "========= DONE WITH SECOND BATCH ==========");
1082
1083 // Serverlist is used: each backend returned by the balancer should
1084 // have gotten one request.
1085 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1086 EXPECT_EQ(0U, backends_[i]->service_.request_count());
1087 }
1088 for (size_t i = kNumBackendsInResolution; i < backends_.size(); ++i) {
1089 EXPECT_EQ(1U, backends_[i]->service_.request_count());
1090 }
1091
1092 balancers_[0]->service_.NotifyDoneWithServerlists();
1093 // The balancer got a single request.
1094 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1095 // and sent a single response.
1096 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1097 }
1098
TEST_F(SingleBalancerTest,FallbackUpdate)1099 TEST_F(SingleBalancerTest, FallbackUpdate) {
1100 SetNextResolutionAllBalancers();
1101 const int kFallbackTimeoutMs = 200 * grpc_test_slowdown_factor();
1102 const int kServerlistDelayMs = 500 * grpc_test_slowdown_factor();
1103 const size_t kNumBackendsInResolution = backends_.size() / 3;
1104 const size_t kNumBackendsInResolutionUpdate = backends_.size() / 3;
1105
1106 ResetStub(kFallbackTimeoutMs);
1107 std::vector<AddressData> balancer_addresses;
1108 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1109 std::vector<AddressData> backend_addresses;
1110 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1111 backend_addresses.emplace_back(AddressData{backends_[i]->port_, ""});
1112 }
1113 SetNextResolution(balancer_addresses, backend_addresses);
1114
1115 // Send non-empty serverlist only after kServerlistDelayMs.
1116 ScheduleResponseForBalancer(
1117 0,
1118 BuildResponseForBackends(
1119 GetBackendPorts(kNumBackendsInResolution +
1120 kNumBackendsInResolutionUpdate /* start_index */),
1121 {}),
1122 kServerlistDelayMs);
1123
1124 // Wait until all the fallback backends are reachable.
1125 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1126 WaitForBackend(i);
1127 }
1128
1129 // The first request.
1130 gpr_log(GPR_INFO, "========= BEFORE FIRST BATCH ==========");
1131 CheckRpcSendOk(kNumBackendsInResolution);
1132 gpr_log(GPR_INFO, "========= DONE WITH FIRST BATCH ==========");
1133
1134 // Fallback is used: each backend returned by the resolver should have
1135 // gotten one request.
1136 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1137 EXPECT_EQ(1U, backends_[i]->service_.request_count());
1138 }
1139 for (size_t i = kNumBackendsInResolution; i < backends_.size(); ++i) {
1140 EXPECT_EQ(0U, backends_[i]->service_.request_count());
1141 }
1142
1143 balancer_addresses.clear();
1144 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1145 backend_addresses.clear();
1146 for (size_t i = kNumBackendsInResolution;
1147 i < kNumBackendsInResolution + kNumBackendsInResolutionUpdate; ++i) {
1148 backend_addresses.emplace_back(AddressData{backends_[i]->port_, ""});
1149 }
1150 SetNextResolution(balancer_addresses, backend_addresses);
1151
1152 // Wait until the resolution update has been processed and all the new
1153 // fallback backends are reachable.
1154 for (size_t i = kNumBackendsInResolution;
1155 i < kNumBackendsInResolution + kNumBackendsInResolutionUpdate; ++i) {
1156 WaitForBackend(i);
1157 }
1158
1159 // Send out the second request.
1160 gpr_log(GPR_INFO, "========= BEFORE SECOND BATCH ==========");
1161 CheckRpcSendOk(kNumBackendsInResolutionUpdate);
1162 gpr_log(GPR_INFO, "========= DONE WITH SECOND BATCH ==========");
1163
1164 // The resolution update is used: each backend in the resolution update should
1165 // have gotten one request.
1166 for (size_t i = 0; i < kNumBackendsInResolution; ++i) {
1167 EXPECT_EQ(0U, backends_[i]->service_.request_count());
1168 }
1169 for (size_t i = kNumBackendsInResolution;
1170 i < kNumBackendsInResolution + kNumBackendsInResolutionUpdate; ++i) {
1171 EXPECT_EQ(1U, backends_[i]->service_.request_count());
1172 }
1173 for (size_t i = kNumBackendsInResolution + kNumBackendsInResolutionUpdate;
1174 i < backends_.size(); ++i) {
1175 EXPECT_EQ(0U, backends_[i]->service_.request_count());
1176 }
1177
1178 // Wait until the serverlist reception has been processed and all backends
1179 // in the serverlist are reachable.
1180 for (size_t i = kNumBackendsInResolution + kNumBackendsInResolutionUpdate;
1181 i < backends_.size(); ++i) {
1182 WaitForBackend(i);
1183 }
1184
1185 // Send out the third request.
1186 gpr_log(GPR_INFO, "========= BEFORE THIRD BATCH ==========");
1187 CheckRpcSendOk(backends_.size() - kNumBackendsInResolution -
1188 kNumBackendsInResolutionUpdate);
1189 gpr_log(GPR_INFO, "========= DONE WITH THIRD BATCH ==========");
1190
1191 // Serverlist is used: each backend returned by the balancer should
1192 // have gotten one request.
1193 for (size_t i = 0;
1194 i < kNumBackendsInResolution + kNumBackendsInResolutionUpdate; ++i) {
1195 EXPECT_EQ(0U, backends_[i]->service_.request_count());
1196 }
1197 for (size_t i = kNumBackendsInResolution + kNumBackendsInResolutionUpdate;
1198 i < backends_.size(); ++i) {
1199 EXPECT_EQ(1U, backends_[i]->service_.request_count());
1200 }
1201
1202 balancers_[0]->service_.NotifyDoneWithServerlists();
1203 // The balancer got a single request.
1204 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1205 // and sent a single response.
1206 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1207 }
1208
TEST_F(SingleBalancerTest,FallbackAfterStartup_LoseContactWithBalancerThenBackends)1209 TEST_F(SingleBalancerTest,
1210 FallbackAfterStartup_LoseContactWithBalancerThenBackends) {
1211 // First two backends are fallback, last two are pointed to by balancer.
1212 const size_t kNumFallbackBackends = 2;
1213 const size_t kNumBalancerBackends = backends_.size() - kNumFallbackBackends;
1214 std::vector<AddressData> backend_addresses;
1215 for (size_t i = 0; i < kNumFallbackBackends; ++i) {
1216 backend_addresses.emplace_back(AddressData{backends_[i]->port_, ""});
1217 }
1218 std::vector<AddressData> balancer_addresses;
1219 for (size_t i = 0; i < balancers_.size(); ++i) {
1220 balancer_addresses.emplace_back(AddressData{balancers_[i]->port_, ""});
1221 }
1222 SetNextResolution(balancer_addresses, backend_addresses);
1223 ScheduleResponseForBalancer(
1224 0, BuildResponseForBackends(GetBackendPorts(kNumFallbackBackends), {}),
1225 0);
1226 // Try to connect.
1227 channel_->GetState(true /* try_to_connect */);
1228 WaitForAllBackends(1 /* num_requests_multiple_of */,
1229 kNumFallbackBackends /* start_index */);
1230 // Stop balancer. RPCs should continue going to backends from balancer.
1231 balancers_[0]->Shutdown();
1232 CheckRpcSendOk(100 * kNumBalancerBackends);
1233 for (size_t i = kNumFallbackBackends; i < backends_.size(); ++i) {
1234 EXPECT_EQ(100UL, backends_[i]->service_.request_count());
1235 }
1236 // Stop backends from balancer. This should put us in fallback mode.
1237 for (size_t i = kNumFallbackBackends; i < backends_.size(); ++i) {
1238 ShutdownBackend(i);
1239 }
1240 WaitForAllBackends(1 /* num_requests_multiple_of */, 0 /* start_index */,
1241 kNumFallbackBackends /* stop_index */);
1242 // Restart the backends from the balancer. We should *not* start
1243 // sending traffic back to them at this point (although the behavior
1244 // in xds may be different).
1245 for (size_t i = kNumFallbackBackends; i < backends_.size(); ++i) {
1246 StartBackend(i);
1247 }
1248 CheckRpcSendOk(100 * kNumBalancerBackends);
1249 for (size_t i = 0; i < kNumFallbackBackends; ++i) {
1250 EXPECT_EQ(100UL, backends_[i]->service_.request_count());
1251 }
1252 // Now start the balancer again. This should cause us to exit
1253 // fallback mode.
1254 balancers_[0]->Start(server_host_);
1255 ScheduleResponseForBalancer(
1256 0, BuildResponseForBackends(GetBackendPorts(kNumFallbackBackends), {}),
1257 0);
1258 WaitForAllBackends(1 /* num_requests_multiple_of */,
1259 kNumFallbackBackends /* start_index */);
1260 }
1261
TEST_F(SingleBalancerTest,FallbackAfterStartup_LoseContactWithBackendsThenBalancer)1262 TEST_F(SingleBalancerTest,
1263 FallbackAfterStartup_LoseContactWithBackendsThenBalancer) {
1264 // First two backends are fallback, last two are pointed to by balancer.
1265 const size_t kNumFallbackBackends = 2;
1266 const size_t kNumBalancerBackends = backends_.size() - kNumFallbackBackends;
1267 std::vector<AddressData> backend_addresses;
1268 for (size_t i = 0; i < kNumFallbackBackends; ++i) {
1269 backend_addresses.emplace_back(AddressData{backends_[i]->port_, ""});
1270 }
1271 std::vector<AddressData> balancer_addresses;
1272 for (size_t i = 0; i < balancers_.size(); ++i) {
1273 balancer_addresses.emplace_back(AddressData{balancers_[i]->port_, ""});
1274 }
1275 SetNextResolution(balancer_addresses, backend_addresses);
1276 ScheduleResponseForBalancer(
1277 0, BuildResponseForBackends(GetBackendPorts(kNumFallbackBackends), {}),
1278 0);
1279 // Try to connect.
1280 channel_->GetState(true /* try_to_connect */);
1281 WaitForAllBackends(1 /* num_requests_multiple_of */,
1282 kNumFallbackBackends /* start_index */);
1283 // Stop backends from balancer. Since we are still in contact with
1284 // the balancer at this point, RPCs should be failing.
1285 for (size_t i = kNumFallbackBackends; i < backends_.size(); ++i) {
1286 ShutdownBackend(i);
1287 }
1288 CheckRpcSendFailure();
1289 // Stop balancer. This should put us in fallback mode.
1290 balancers_[0]->Shutdown();
1291 WaitForAllBackends(1 /* num_requests_multiple_of */, 0 /* start_index */,
1292 kNumFallbackBackends /* stop_index */);
1293 // Restart the backends from the balancer. We should *not* start
1294 // sending traffic back to them at this point (although the behavior
1295 // in xds may be different).
1296 for (size_t i = kNumFallbackBackends; i < backends_.size(); ++i) {
1297 StartBackend(i);
1298 }
1299 CheckRpcSendOk(100 * kNumBalancerBackends);
1300 for (size_t i = 0; i < kNumFallbackBackends; ++i) {
1301 EXPECT_EQ(100UL, backends_[i]->service_.request_count());
1302 }
1303 // Now start the balancer again. This should cause us to exit
1304 // fallback mode.
1305 balancers_[0]->Start(server_host_);
1306 ScheduleResponseForBalancer(
1307 0, BuildResponseForBackends(GetBackendPorts(kNumFallbackBackends), {}),
1308 0);
1309 WaitForAllBackends(1 /* num_requests_multiple_of */,
1310 kNumFallbackBackends /* start_index */);
1311 }
1312
TEST_F(SingleBalancerTest,FallbackEarlyWhenBalancerChannelFails)1313 TEST_F(SingleBalancerTest, FallbackEarlyWhenBalancerChannelFails) {
1314 const int kFallbackTimeoutMs = 10000 * grpc_test_slowdown_factor();
1315 ResetStub(kFallbackTimeoutMs);
1316 // Return an unreachable balancer and one fallback backend.
1317 std::vector<AddressData> balancer_addresses;
1318 balancer_addresses.emplace_back(
1319 AddressData{grpc_pick_unused_port_or_die(), ""});
1320 std::vector<AddressData> backend_addresses;
1321 backend_addresses.emplace_back(AddressData{backends_[0]->port_, ""});
1322 SetNextResolution(balancer_addresses, backend_addresses);
1323 // Send RPC with deadline less than the fallback timeout and make sure it
1324 // succeeds.
1325 CheckRpcSendOk(/* times */ 1, /* timeout_ms */ 1000,
1326 /* wait_for_ready */ false);
1327 }
1328
TEST_F(SingleBalancerTest,FallbackEarlyWhenBalancerCallFails)1329 TEST_F(SingleBalancerTest, FallbackEarlyWhenBalancerCallFails) {
1330 const int kFallbackTimeoutMs = 10000 * grpc_test_slowdown_factor();
1331 ResetStub(kFallbackTimeoutMs);
1332 // Return one balancer and one fallback backend.
1333 std::vector<AddressData> balancer_addresses;
1334 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1335 std::vector<AddressData> backend_addresses;
1336 backend_addresses.emplace_back(AddressData{backends_[0]->port_, ""});
1337 SetNextResolution(balancer_addresses, backend_addresses);
1338 // Balancer drops call without sending a serverlist.
1339 balancers_[0]->service_.NotifyDoneWithServerlists();
1340 // Send RPC with deadline less than the fallback timeout and make sure it
1341 // succeeds.
1342 CheckRpcSendOk(/* times */ 1, /* timeout_ms */ 1000,
1343 /* wait_for_ready */ false);
1344 }
1345
TEST_F(SingleBalancerTest,FallbackControlledByBalancer_BeforeFirstServerlist)1346 TEST_F(SingleBalancerTest, FallbackControlledByBalancer_BeforeFirstServerlist) {
1347 const int kFallbackTimeoutMs = 10000 * grpc_test_slowdown_factor();
1348 ResetStub(kFallbackTimeoutMs);
1349 // Return one balancer and one fallback backend.
1350 std::vector<AddressData> balancer_addresses;
1351 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1352 std::vector<AddressData> backend_addresses;
1353 backend_addresses.emplace_back(AddressData{backends_[0]->port_, ""});
1354 SetNextResolution(balancer_addresses, backend_addresses);
1355 // Balancer explicitly tells client to fallback.
1356 LoadBalanceResponse resp;
1357 resp.mutable_fallback_response();
1358 ScheduleResponseForBalancer(0, resp, 0);
1359 // Send RPC with deadline less than the fallback timeout and make sure it
1360 // succeeds.
1361 CheckRpcSendOk(/* times */ 1, /* timeout_ms */ 1000,
1362 /* wait_for_ready */ false);
1363 }
1364
TEST_F(SingleBalancerTest,FallbackControlledByBalancer_AfterFirstServerlist)1365 TEST_F(SingleBalancerTest, FallbackControlledByBalancer_AfterFirstServerlist) {
1366 // Return one balancer and one fallback backend (backend 0).
1367 std::vector<AddressData> balancer_addresses;
1368 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1369 std::vector<AddressData> backend_addresses;
1370 backend_addresses.emplace_back(AddressData{backends_[0]->port_, ""});
1371 SetNextResolution(balancer_addresses, backend_addresses);
1372 // Balancer initially sends serverlist, then tells client to fall back,
1373 // then sends the serverlist again.
1374 // The serverlist points to backend 1.
1375 LoadBalanceResponse serverlist_resp =
1376 BuildResponseForBackends({backends_[1]->port_}, {});
1377 LoadBalanceResponse fallback_resp;
1378 fallback_resp.mutable_fallback_response();
1379 ScheduleResponseForBalancer(0, serverlist_resp, 0);
1380 ScheduleResponseForBalancer(0, fallback_resp, 100);
1381 ScheduleResponseForBalancer(0, serverlist_resp, 100);
1382 // Requests initially go to backend 1, then go to backend 0 in
1383 // fallback mode, then go back to backend 1 when we exit fallback.
1384 WaitForBackend(1);
1385 WaitForBackend(0);
1386 WaitForBackend(1);
1387 }
1388
TEST_F(SingleBalancerTest,BackendsRestart)1389 TEST_F(SingleBalancerTest, BackendsRestart) {
1390 SetNextResolutionAllBalancers();
1391 const size_t kNumRpcsPerAddress = 100;
1392 ScheduleResponseForBalancer(
1393 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
1394 // Make sure that trying to connect works without a call.
1395 channel_->GetState(true /* try_to_connect */);
1396 // Send kNumRpcsPerAddress RPCs per server.
1397 CheckRpcSendOk(kNumRpcsPerAddress * num_backends_);
1398 // Stop backends. RPCs should fail.
1399 ShutdownAllBackends();
1400 CheckRpcSendFailure();
1401 // Restart backends. RPCs should start succeeding again.
1402 StartAllBackends();
1403 CheckRpcSendOk(1 /* times */, 2000 /* timeout_ms */,
1404 true /* wait_for_ready */);
1405 // The balancer got a single request.
1406 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1407 // and sent a single response.
1408 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1409 }
1410
TEST_F(SingleBalancerTest,ServiceNameFromLbPolicyConfig)1411 TEST_F(SingleBalancerTest, ServiceNameFromLbPolicyConfig) {
1412 constexpr char kServiceConfigWithTarget[] =
1413 "{\n"
1414 " \"loadBalancingConfig\":[\n"
1415 " { \"grpclb\":{\n"
1416 " \"serviceName\":\"test_service\"\n"
1417 " }}\n"
1418 " ]\n"
1419 "}";
1420
1421 SetNextResolutionAllBalancers(kServiceConfigWithTarget);
1422 ScheduleResponseForBalancer(
1423 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
1424 // Make sure that trying to connect works without a call.
1425 channel_->GetState(true /* try_to_connect */);
1426 // We need to wait for all backends to come online.
1427 WaitForAllBackends();
1428 EXPECT_EQ(balancers_[0]->service_.service_names().back(), "test_service");
1429 }
1430
1431 class UpdatesTest : public GrpclbEnd2endTest {
1432 public:
UpdatesTest()1433 UpdatesTest() : GrpclbEnd2endTest(4, 3, 0) {}
1434 };
1435
TEST_F(UpdatesTest,UpdateBalancersButKeepUsingOriginalBalancer)1436 TEST_F(UpdatesTest, UpdateBalancersButKeepUsingOriginalBalancer) {
1437 SetNextResolutionAllBalancers();
1438 const std::vector<int> first_backend{GetBackendPorts()[0]};
1439 const std::vector<int> second_backend{GetBackendPorts()[1]};
1440 ScheduleResponseForBalancer(0, BuildResponseForBackends(first_backend, {}),
1441 0);
1442 ScheduleResponseForBalancer(1, BuildResponseForBackends(second_backend, {}),
1443 0);
1444
1445 // Wait until the first backend is ready.
1446 WaitForBackend(0);
1447
1448 // Send 10 requests.
1449 gpr_log(GPR_INFO, "========= BEFORE FIRST BATCH ==========");
1450 CheckRpcSendOk(10);
1451 gpr_log(GPR_INFO, "========= DONE WITH FIRST BATCH ==========");
1452
1453 // All 10 requests should have gone to the first backend.
1454 EXPECT_EQ(10U, backends_[0]->service_.request_count());
1455
1456 // Balancer 0 got a single request.
1457 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1458 // and sent a single response.
1459 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1460 EXPECT_EQ(0U, balancers_[1]->service_.request_count());
1461 EXPECT_EQ(0U, balancers_[1]->service_.response_count());
1462 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1463 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1464
1465 std::vector<AddressData> addresses;
1466 addresses.emplace_back(AddressData{balancers_[1]->port_, ""});
1467 gpr_log(GPR_INFO, "========= ABOUT TO UPDATE 1 ==========");
1468 SetNextResolution(addresses);
1469 gpr_log(GPR_INFO, "========= UPDATE 1 DONE ==========");
1470
1471 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1472 gpr_timespec deadline = gpr_time_add(
1473 gpr_now(GPR_CLOCK_REALTIME), gpr_time_from_millis(10000, GPR_TIMESPAN));
1474 // Send 10 seconds worth of RPCs
1475 do {
1476 CheckRpcSendOk();
1477 } while (gpr_time_cmp(gpr_now(GPR_CLOCK_REALTIME), deadline) < 0);
1478 // The current LB call is still working, so grpclb continued using it to the
1479 // first balancer, which doesn't assign the second backend.
1480 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1481
1482 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1483 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1484 EXPECT_EQ(0U, balancers_[1]->service_.request_count());
1485 EXPECT_EQ(0U, balancers_[1]->service_.response_count());
1486 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1487 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1488 }
1489
1490 // Send an update with the same set of LBs as the one in SetUp() in order to
1491 // verify that the LB channel inside grpclb keeps the initial connection (which
1492 // by definition is also present in the update).
TEST_F(UpdatesTest,UpdateBalancersRepeated)1493 TEST_F(UpdatesTest, UpdateBalancersRepeated) {
1494 SetNextResolutionAllBalancers();
1495 const std::vector<int> first_backend{GetBackendPorts()[0]};
1496 const std::vector<int> second_backend{GetBackendPorts()[0]};
1497
1498 ScheduleResponseForBalancer(0, BuildResponseForBackends(first_backend, {}),
1499 0);
1500 ScheduleResponseForBalancer(1, BuildResponseForBackends(second_backend, {}),
1501 0);
1502
1503 // Wait until the first backend is ready.
1504 WaitForBackend(0);
1505
1506 // Send 10 requests.
1507 gpr_log(GPR_INFO, "========= BEFORE FIRST BATCH ==========");
1508 CheckRpcSendOk(10);
1509 gpr_log(GPR_INFO, "========= DONE WITH FIRST BATCH ==========");
1510
1511 // All 10 requests should have gone to the first backend.
1512 EXPECT_EQ(10U, backends_[0]->service_.request_count());
1513
1514 balancers_[0]->service_.NotifyDoneWithServerlists();
1515 // Balancer 0 got a single request.
1516 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1517 // and sent a single response.
1518 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1519 EXPECT_EQ(0U, balancers_[1]->service_.request_count());
1520 EXPECT_EQ(0U, balancers_[1]->service_.response_count());
1521 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1522 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1523
1524 std::vector<AddressData> addresses;
1525 addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1526 addresses.emplace_back(AddressData{balancers_[1]->port_, ""});
1527 addresses.emplace_back(AddressData{balancers_[2]->port_, ""});
1528 gpr_log(GPR_INFO, "========= ABOUT TO UPDATE 1 ==========");
1529 SetNextResolution(addresses);
1530 gpr_log(GPR_INFO, "========= UPDATE 1 DONE ==========");
1531
1532 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1533 gpr_timespec deadline = gpr_time_add(
1534 gpr_now(GPR_CLOCK_REALTIME), gpr_time_from_millis(10000, GPR_TIMESPAN));
1535 // Send 10 seconds worth of RPCs
1536 do {
1537 CheckRpcSendOk();
1538 } while (gpr_time_cmp(gpr_now(GPR_CLOCK_REALTIME), deadline) < 0);
1539 // grpclb continued using the original LB call to the first balancer, which
1540 // doesn't assign the second backend.
1541 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1542 balancers_[0]->service_.NotifyDoneWithServerlists();
1543
1544 addresses.clear();
1545 addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1546 addresses.emplace_back(AddressData{balancers_[1]->port_, ""});
1547 gpr_log(GPR_INFO, "========= ABOUT TO UPDATE 2 ==========");
1548 SetNextResolution(addresses);
1549 gpr_log(GPR_INFO, "========= UPDATE 2 DONE ==========");
1550
1551 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1552 deadline = gpr_time_add(gpr_now(GPR_CLOCK_REALTIME),
1553 gpr_time_from_millis(10000, GPR_TIMESPAN));
1554 // Send 10 seconds worth of RPCs
1555 do {
1556 CheckRpcSendOk();
1557 } while (gpr_time_cmp(gpr_now(GPR_CLOCK_REALTIME), deadline) < 0);
1558 // grpclb continued using the original LB call to the first balancer, which
1559 // doesn't assign the second backend.
1560 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1561 balancers_[0]->service_.NotifyDoneWithServerlists();
1562 }
1563
TEST_F(UpdatesTest,UpdateBalancersDeadUpdate)1564 TEST_F(UpdatesTest, UpdateBalancersDeadUpdate) {
1565 std::vector<AddressData> addresses;
1566 addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1567 SetNextResolution(addresses);
1568 const std::vector<int> first_backend{GetBackendPorts()[0]};
1569 const std::vector<int> second_backend{GetBackendPorts()[1]};
1570
1571 ScheduleResponseForBalancer(0, BuildResponseForBackends(first_backend, {}),
1572 0);
1573 ScheduleResponseForBalancer(1, BuildResponseForBackends(second_backend, {}),
1574 0);
1575
1576 // Start servers and send 10 RPCs per server.
1577 gpr_log(GPR_INFO, "========= BEFORE FIRST BATCH ==========");
1578 CheckRpcSendOk(10);
1579 gpr_log(GPR_INFO, "========= DONE WITH FIRST BATCH ==========");
1580 // All 10 requests should have gone to the first backend.
1581 EXPECT_EQ(10U, backends_[0]->service_.request_count());
1582
1583 // Kill balancer 0
1584 gpr_log(GPR_INFO, "********** ABOUT TO KILL BALANCER 0 *************");
1585 balancers_[0]->Shutdown();
1586 gpr_log(GPR_INFO, "********** KILLED BALANCER 0 *************");
1587
1588 // This is serviced by the existing RR policy
1589 gpr_log(GPR_INFO, "========= BEFORE SECOND BATCH ==========");
1590 CheckRpcSendOk(10);
1591 gpr_log(GPR_INFO, "========= DONE WITH SECOND BATCH ==========");
1592 // All 10 requests should again have gone to the first backend.
1593 EXPECT_EQ(20U, backends_[0]->service_.request_count());
1594 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1595
1596 // Balancer 0 got a single request.
1597 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1598 // and sent a single response.
1599 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1600 EXPECT_EQ(0U, balancers_[1]->service_.request_count());
1601 EXPECT_EQ(0U, balancers_[1]->service_.response_count());
1602 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1603 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1604
1605 addresses.clear();
1606 addresses.emplace_back(AddressData{balancers_[1]->port_, ""});
1607 gpr_log(GPR_INFO, "========= ABOUT TO UPDATE 1 ==========");
1608 SetNextResolution(addresses);
1609 gpr_log(GPR_INFO, "========= UPDATE 1 DONE ==========");
1610
1611 // Wait until update has been processed, as signaled by the second backend
1612 // receiving a request. In the meantime, the client continues to be serviced
1613 // (by the first backend) without interruption.
1614 EXPECT_EQ(0U, backends_[1]->service_.request_count());
1615 WaitForBackend(1);
1616
1617 // This is serviced by the updated RR policy
1618 backends_[1]->service_.ResetCounters();
1619 gpr_log(GPR_INFO, "========= BEFORE THIRD BATCH ==========");
1620 CheckRpcSendOk(10);
1621 gpr_log(GPR_INFO, "========= DONE WITH THIRD BATCH ==========");
1622 // All 10 requests should have gone to the second backend.
1623 EXPECT_EQ(10U, backends_[1]->service_.request_count());
1624
1625 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1626 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1627 // The second balancer, published as part of the first update, may end up
1628 // getting two requests (that is, 1 <= #req <= 2) if the LB call retry timer
1629 // firing races with the arrival of the update containing the second
1630 // balancer.
1631 EXPECT_GE(balancers_[1]->service_.request_count(), 1U);
1632 EXPECT_GE(balancers_[1]->service_.response_count(), 1U);
1633 EXPECT_LE(balancers_[1]->service_.request_count(), 2U);
1634 EXPECT_LE(balancers_[1]->service_.response_count(), 2U);
1635 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1636 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1637 }
1638
TEST_F(UpdatesTest,ReresolveDeadBackend)1639 TEST_F(UpdatesTest, ReresolveDeadBackend) {
1640 ResetStub(500);
1641 // The first resolution contains the addresses of a balancer that never
1642 // responds, and a fallback backend.
1643 std::vector<AddressData> balancer_addresses;
1644 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1645 std::vector<AddressData> backend_addresses;
1646 backend_addresses.emplace_back(AddressData{backends_[0]->port_, ""});
1647 SetNextResolution(balancer_addresses, backend_addresses);
1648 // Ask channel to connect to trigger resolver creation.
1649 channel_->GetState(true);
1650 // The re-resolution result will contain the addresses of the same balancer
1651 // and a new fallback backend.
1652 balancer_addresses.clear();
1653 balancer_addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1654 backend_addresses.clear();
1655 backend_addresses.emplace_back(AddressData{backends_[1]->port_, ""});
1656 SetNextReresolutionResponse(balancer_addresses, backend_addresses);
1657
1658 // Start servers and send 10 RPCs per server.
1659 gpr_log(GPR_INFO, "========= BEFORE FIRST BATCH ==========");
1660 CheckRpcSendOk(10);
1661 gpr_log(GPR_INFO, "========= DONE WITH FIRST BATCH ==========");
1662 // All 10 requests should have gone to the fallback backend.
1663 EXPECT_EQ(10U, backends_[0]->service_.request_count());
1664
1665 // Kill backend 0.
1666 gpr_log(GPR_INFO, "********** ABOUT TO KILL BACKEND 0 *************");
1667 backends_[0]->Shutdown();
1668 gpr_log(GPR_INFO, "********** KILLED BACKEND 0 *************");
1669
1670 // Wait until re-resolution has finished, as signaled by the second backend
1671 // receiving a request.
1672 WaitForBackend(1);
1673
1674 gpr_log(GPR_INFO, "========= BEFORE SECOND BATCH ==========");
1675 CheckRpcSendOk(10);
1676 gpr_log(GPR_INFO, "========= DONE WITH SECOND BATCH ==========");
1677 // All 10 requests should have gone to the second backend.
1678 EXPECT_EQ(10U, backends_[1]->service_.request_count());
1679
1680 balancers_[0]->service_.NotifyDoneWithServerlists();
1681 balancers_[1]->service_.NotifyDoneWithServerlists();
1682 balancers_[2]->service_.NotifyDoneWithServerlists();
1683 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1684 EXPECT_EQ(0U, balancers_[0]->service_.response_count());
1685 EXPECT_EQ(0U, balancers_[1]->service_.request_count());
1686 EXPECT_EQ(0U, balancers_[1]->service_.response_count());
1687 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1688 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1689 }
1690
1691 // TODO(juanlishen): Should be removed when the first response is always the
1692 // initial response. Currently, if client load reporting is not enabled, the
1693 // balancer doesn't send initial response. When the backend shuts down, an
1694 // unexpected re-resolution will happen. This test configuration is a workaround
1695 // for test ReresolveDeadBalancer.
1696 class UpdatesWithClientLoadReportingTest : public GrpclbEnd2endTest {
1697 public:
UpdatesWithClientLoadReportingTest()1698 UpdatesWithClientLoadReportingTest() : GrpclbEnd2endTest(4, 3, 2) {}
1699 };
1700
TEST_F(UpdatesWithClientLoadReportingTest,ReresolveDeadBalancer)1701 TEST_F(UpdatesWithClientLoadReportingTest, ReresolveDeadBalancer) {
1702 const std::vector<int> first_backend{GetBackendPorts()[0]};
1703 const std::vector<int> second_backend{GetBackendPorts()[1]};
1704 ScheduleResponseForBalancer(0, BuildResponseForBackends(first_backend, {}),
1705 0);
1706 ScheduleResponseForBalancer(1, BuildResponseForBackends(second_backend, {}),
1707 0);
1708
1709 // Ask channel to connect to trigger resolver creation.
1710 channel_->GetState(true);
1711 std::vector<AddressData> addresses;
1712 addresses.emplace_back(AddressData{balancers_[0]->port_, ""});
1713 SetNextResolution(addresses);
1714 addresses.clear();
1715 addresses.emplace_back(AddressData{balancers_[1]->port_, ""});
1716 SetNextReresolutionResponse(addresses);
1717
1718 // Start servers and send 10 RPCs per server.
1719 gpr_log(GPR_INFO, "========= BEFORE FIRST BATCH ==========");
1720 CheckRpcSendOk(10);
1721 gpr_log(GPR_INFO, "========= DONE WITH FIRST BATCH ==========");
1722 // All 10 requests should have gone to the first backend.
1723 EXPECT_EQ(10U, backends_[0]->service_.request_count());
1724
1725 // Kill backend 0.
1726 gpr_log(GPR_INFO, "********** ABOUT TO KILL BACKEND 0 *************");
1727 backends_[0]->Shutdown();
1728 gpr_log(GPR_INFO, "********** KILLED BACKEND 0 *************");
1729
1730 CheckRpcSendFailure();
1731
1732 // Balancer 0 got a single request.
1733 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1734 // and sent a single response.
1735 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1736 EXPECT_EQ(0U, balancers_[1]->service_.request_count());
1737 EXPECT_EQ(0U, balancers_[1]->service_.response_count());
1738 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1739 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1740
1741 // Kill balancer 0.
1742 gpr_log(GPR_INFO, "********** ABOUT TO KILL BALANCER 0 *************");
1743 balancers_[0]->Shutdown();
1744 gpr_log(GPR_INFO, "********** KILLED BALANCER 0 *************");
1745
1746 // Wait until re-resolution has finished, as signaled by the second backend
1747 // receiving a request.
1748 WaitForBackend(1);
1749
1750 // This is serviced by the new serverlist.
1751 gpr_log(GPR_INFO, "========= BEFORE SECOND BATCH ==========");
1752 CheckRpcSendOk(10);
1753 gpr_log(GPR_INFO, "========= DONE WITH SECOND BATCH ==========");
1754 // All 10 requests should have gone to the second backend.
1755 EXPECT_EQ(10U, backends_[1]->service_.request_count());
1756
1757 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1758 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1759 // After balancer 0 is killed, we restart an LB call immediately (because we
1760 // disconnect to a previously connected balancer). Although we will cancel
1761 // this call when the re-resolution update is done and another LB call restart
1762 // is needed, this old call may still succeed reaching the LB server if
1763 // re-resolution is slow. So balancer 1 may have received 2 requests and sent
1764 // 2 responses.
1765 EXPECT_GE(balancers_[1]->service_.request_count(), 1U);
1766 EXPECT_GE(balancers_[1]->service_.response_count(), 1U);
1767 EXPECT_LE(balancers_[1]->service_.request_count(), 2U);
1768 EXPECT_LE(balancers_[1]->service_.response_count(), 2U);
1769 EXPECT_EQ(0U, balancers_[2]->service_.request_count());
1770 EXPECT_EQ(0U, balancers_[2]->service_.response_count());
1771 }
1772
TEST_F(SingleBalancerTest,Drop)1773 TEST_F(SingleBalancerTest, Drop) {
1774 SetNextResolutionAllBalancers();
1775 const size_t kNumRpcsPerAddress = 100;
1776 const int num_of_drop_by_rate_limiting_addresses = 1;
1777 const int num_of_drop_by_load_balancing_addresses = 2;
1778 const int num_of_drop_addresses = num_of_drop_by_rate_limiting_addresses +
1779 num_of_drop_by_load_balancing_addresses;
1780 const int num_total_addresses = num_backends_ + num_of_drop_addresses;
1781 ScheduleResponseForBalancer(
1782 0,
1783 BuildResponseForBackends(
1784 GetBackendPorts(),
1785 {{"rate_limiting", num_of_drop_by_rate_limiting_addresses},
1786 {"load_balancing", num_of_drop_by_load_balancing_addresses}}),
1787 0);
1788 // Wait until all backends are ready.
1789 WaitForAllBackends();
1790 // Send kNumRpcsPerAddress RPCs for each server and drop address.
1791 size_t num_drops = 0;
1792 for (size_t i = 0; i < kNumRpcsPerAddress * num_total_addresses; ++i) {
1793 EchoResponse response;
1794 const Status status = SendRpc(&response);
1795 if (!status.ok() &&
1796 status.error_message() == "drop directed by grpclb balancer") {
1797 ++num_drops;
1798 } else {
1799 EXPECT_TRUE(status.ok()) << "code=" << status.error_code()
1800 << " message=" << status.error_message();
1801 EXPECT_EQ(response.message(), kRequestMessage_);
1802 }
1803 }
1804 EXPECT_EQ(kNumRpcsPerAddress * num_of_drop_addresses, num_drops);
1805 // Each backend should have gotten 100 requests.
1806 for (size_t i = 0; i < backends_.size(); ++i) {
1807 EXPECT_EQ(kNumRpcsPerAddress, backends_[i]->service_.request_count());
1808 }
1809 // The balancer got a single request.
1810 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1811 // and sent a single response.
1812 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1813 }
1814
TEST_F(SingleBalancerTest,DropAllFirst)1815 TEST_F(SingleBalancerTest, DropAllFirst) {
1816 SetNextResolutionAllBalancers();
1817 // All registered addresses are marked as "drop".
1818 const int num_of_drop_by_rate_limiting_addresses = 1;
1819 const int num_of_drop_by_load_balancing_addresses = 1;
1820 ScheduleResponseForBalancer(
1821 0,
1822 BuildResponseForBackends(
1823 {}, {{"rate_limiting", num_of_drop_by_rate_limiting_addresses},
1824 {"load_balancing", num_of_drop_by_load_balancing_addresses}}),
1825 0);
1826 const Status status = SendRpc(nullptr, 1000, true);
1827 EXPECT_FALSE(status.ok());
1828 EXPECT_EQ(status.error_message(), "drop directed by grpclb balancer");
1829 }
1830
TEST_F(SingleBalancerTest,DropAll)1831 TEST_F(SingleBalancerTest, DropAll) {
1832 SetNextResolutionAllBalancers();
1833 ScheduleResponseForBalancer(
1834 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
1835 const int num_of_drop_by_rate_limiting_addresses = 1;
1836 const int num_of_drop_by_load_balancing_addresses = 1;
1837 ScheduleResponseForBalancer(
1838 0,
1839 BuildResponseForBackends(
1840 {}, {{"rate_limiting", num_of_drop_by_rate_limiting_addresses},
1841 {"load_balancing", num_of_drop_by_load_balancing_addresses}}),
1842 1000);
1843
1844 // First call succeeds.
1845 CheckRpcSendOk();
1846 // But eventually, the update with only dropped servers is processed and calls
1847 // fail.
1848 Status status;
1849 do {
1850 status = SendRpc(nullptr, 1000, true);
1851 } while (status.ok());
1852 EXPECT_FALSE(status.ok());
1853 EXPECT_EQ(status.error_message(), "drop directed by grpclb balancer");
1854 }
1855
1856 class SingleBalancerWithClientLoadReportingTest : public GrpclbEnd2endTest {
1857 public:
SingleBalancerWithClientLoadReportingTest()1858 SingleBalancerWithClientLoadReportingTest() : GrpclbEnd2endTest(4, 1, 3) {}
1859 };
1860
TEST_F(SingleBalancerWithClientLoadReportingTest,Vanilla)1861 TEST_F(SingleBalancerWithClientLoadReportingTest, Vanilla) {
1862 SetNextResolutionAllBalancers();
1863 const size_t kNumRpcsPerAddress = 100;
1864 ScheduleResponseForBalancer(
1865 0, BuildResponseForBackends(GetBackendPorts(), {}), 0);
1866 // Wait until all backends are ready.
1867 int num_ok = 0;
1868 int num_failure = 0;
1869 int num_drops = 0;
1870 std::tie(num_ok, num_failure, num_drops) = WaitForAllBackends();
1871 // Send kNumRpcsPerAddress RPCs per server.
1872 CheckRpcSendOk(kNumRpcsPerAddress * num_backends_);
1873 // Each backend should have gotten 100 requests.
1874 for (size_t i = 0; i < backends_.size(); ++i) {
1875 EXPECT_EQ(kNumRpcsPerAddress, backends_[i]->service_.request_count());
1876 }
1877 balancers_[0]->service_.NotifyDoneWithServerlists();
1878 // The balancer got a single request.
1879 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
1880 // and sent a single response.
1881 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
1882
1883 ClientStats client_stats;
1884 do {
1885 client_stats += WaitForLoadReports();
1886 } while (client_stats.num_calls_finished !=
1887 kNumRpcsPerAddress * num_backends_ + num_ok);
1888 EXPECT_EQ(kNumRpcsPerAddress * num_backends_ + num_ok,
1889 client_stats.num_calls_started);
1890 EXPECT_EQ(kNumRpcsPerAddress * num_backends_ + num_ok,
1891 client_stats.num_calls_finished);
1892 EXPECT_EQ(0U, client_stats.num_calls_finished_with_client_failed_to_send);
1893 EXPECT_EQ(kNumRpcsPerAddress * num_backends_ + (num_ok + num_drops),
1894 client_stats.num_calls_finished_known_received);
1895 EXPECT_THAT(client_stats.drop_token_counts, ::testing::ElementsAre());
1896 }
1897
TEST_F(SingleBalancerWithClientLoadReportingTest,BalancerRestart)1898 TEST_F(SingleBalancerWithClientLoadReportingTest, BalancerRestart) {
1899 SetNextResolutionAllBalancers();
1900 const size_t kNumBackendsFirstPass = 2;
1901 const size_t kNumBackendsSecondPass =
1902 backends_.size() - kNumBackendsFirstPass;
1903 // Balancer returns backends starting at index 1.
1904 ScheduleResponseForBalancer(
1905 0,
1906 BuildResponseForBackends(GetBackendPorts(0, kNumBackendsFirstPass), {}),
1907 0);
1908 // Wait until all backends returned by the balancer are ready.
1909 int num_ok = 0;
1910 int num_failure = 0;
1911 int num_drops = 0;
1912 std::tie(num_ok, num_failure, num_drops) =
1913 WaitForAllBackends(/* num_requests_multiple_of */ 1, /* start_index */ 0,
1914 /* stop_index */ kNumBackendsFirstPass);
1915 balancers_[0]->service_.NotifyDoneWithServerlists();
1916 ClientStats client_stats = WaitForLoadReports();
1917 EXPECT_EQ(static_cast<size_t>(num_ok), client_stats.num_calls_started);
1918 EXPECT_EQ(static_cast<size_t>(num_ok), client_stats.num_calls_finished);
1919 EXPECT_EQ(0U, client_stats.num_calls_finished_with_client_failed_to_send);
1920 EXPECT_EQ(static_cast<size_t>(num_ok),
1921 client_stats.num_calls_finished_known_received);
1922 EXPECT_THAT(client_stats.drop_token_counts, ::testing::ElementsAre());
1923 // Shut down the balancer.
1924 balancers_[0]->Shutdown();
1925 // Send 10 more requests per backend. This will continue using the
1926 // last serverlist we received from the balancer before it was shut down.
1927 ResetBackendCounters();
1928 CheckRpcSendOk(kNumBackendsFirstPass);
1929 // Each backend should have gotten 1 request.
1930 for (size_t i = 0; i < kNumBackendsFirstPass; ++i) {
1931 EXPECT_EQ(1UL, backends_[i]->service_.request_count());
1932 }
1933 // Now restart the balancer, this time pointing to all backends.
1934 balancers_[0]->Start(server_host_);
1935 ScheduleResponseForBalancer(
1936 0, BuildResponseForBackends(GetBackendPorts(kNumBackendsFirstPass), {}),
1937 0);
1938 // Wait for queries to start going to one of the new backends.
1939 // This tells us that we're now using the new serverlist.
1940 do {
1941 CheckRpcSendOk();
1942 } while (backends_[2]->service_.request_count() == 0 &&
1943 backends_[3]->service_.request_count() == 0);
1944 // Send one RPC per backend.
1945 CheckRpcSendOk(kNumBackendsSecondPass);
1946 balancers_[0]->service_.NotifyDoneWithServerlists();
1947 // Check client stats.
1948 client_stats = WaitForLoadReports();
1949 EXPECT_EQ(kNumBackendsSecondPass + 1, client_stats.num_calls_started);
1950 EXPECT_EQ(kNumBackendsSecondPass + 1, client_stats.num_calls_finished);
1951 EXPECT_EQ(0U, client_stats.num_calls_finished_with_client_failed_to_send);
1952 EXPECT_EQ(kNumBackendsSecondPass + 1,
1953 client_stats.num_calls_finished_known_received);
1954 EXPECT_THAT(client_stats.drop_token_counts, ::testing::ElementsAre());
1955 }
1956
TEST_F(SingleBalancerWithClientLoadReportingTest,Drop)1957 TEST_F(SingleBalancerWithClientLoadReportingTest, Drop) {
1958 SetNextResolutionAllBalancers();
1959 const size_t kNumRpcsPerAddress = 3;
1960 const int num_of_drop_by_rate_limiting_addresses = 2;
1961 const int num_of_drop_by_load_balancing_addresses = 1;
1962 const int num_of_drop_addresses = num_of_drop_by_rate_limiting_addresses +
1963 num_of_drop_by_load_balancing_addresses;
1964 const int num_total_addresses = num_backends_ + num_of_drop_addresses;
1965 ScheduleResponseForBalancer(
1966 0,
1967 BuildResponseForBackends(
1968 GetBackendPorts(),
1969 {{"rate_limiting", num_of_drop_by_rate_limiting_addresses},
1970 {"load_balancing", num_of_drop_by_load_balancing_addresses}}),
1971 0);
1972 // Wait until all backends are ready.
1973 int num_warmup_ok = 0;
1974 int num_warmup_failure = 0;
1975 int num_warmup_drops = 0;
1976 std::tie(num_warmup_ok, num_warmup_failure, num_warmup_drops) =
1977 WaitForAllBackends(num_total_addresses /* num_requests_multiple_of */);
1978 const int num_total_warmup_requests =
1979 num_warmup_ok + num_warmup_failure + num_warmup_drops;
1980 size_t num_drops = 0;
1981 for (size_t i = 0; i < kNumRpcsPerAddress * num_total_addresses; ++i) {
1982 EchoResponse response;
1983 const Status status = SendRpc(&response);
1984 if (!status.ok() &&
1985 status.error_message() == "drop directed by grpclb balancer") {
1986 ++num_drops;
1987 } else {
1988 EXPECT_TRUE(status.ok()) << "code=" << status.error_code()
1989 << " message=" << status.error_message();
1990 EXPECT_EQ(response.message(), kRequestMessage_);
1991 }
1992 }
1993 EXPECT_EQ(kNumRpcsPerAddress * num_of_drop_addresses, num_drops);
1994 // Each backend should have gotten 100 requests.
1995 for (size_t i = 0; i < backends_.size(); ++i) {
1996 EXPECT_EQ(kNumRpcsPerAddress, backends_[i]->service_.request_count());
1997 }
1998 balancers_[0]->service_.NotifyDoneWithServerlists();
1999 // The balancer got a single request.
2000 EXPECT_EQ(1U, balancers_[0]->service_.request_count());
2001 // and sent a single response.
2002 EXPECT_EQ(1U, balancers_[0]->service_.response_count());
2003
2004 const ClientStats client_stats = WaitForLoadReports();
2005 EXPECT_EQ(
2006 kNumRpcsPerAddress * num_total_addresses + num_total_warmup_requests,
2007 client_stats.num_calls_started);
2008 EXPECT_EQ(
2009 kNumRpcsPerAddress * num_total_addresses + num_total_warmup_requests,
2010 client_stats.num_calls_finished);
2011 EXPECT_EQ(0U, client_stats.num_calls_finished_with_client_failed_to_send);
2012 EXPECT_EQ(kNumRpcsPerAddress * num_backends_ + num_warmup_ok,
2013 client_stats.num_calls_finished_known_received);
2014 // The number of warmup request is a multiple of the number of addresses.
2015 // Therefore, all addresses in the scheduled balancer response are hit the
2016 // same number of times.
2017 const int num_times_drop_addresses_hit =
2018 num_warmup_drops / num_of_drop_addresses;
2019 EXPECT_THAT(
2020 client_stats.drop_token_counts,
2021 ::testing::ElementsAre(
2022 ::testing::Pair("load_balancing",
2023 (kNumRpcsPerAddress + num_times_drop_addresses_hit)),
2024 ::testing::Pair(
2025 "rate_limiting",
2026 (kNumRpcsPerAddress + num_times_drop_addresses_hit) * 2)));
2027 }
2028
2029 } // namespace
2030 } // namespace testing
2031 } // namespace grpc
2032
main(int argc,char ** argv)2033 int main(int argc, char** argv) {
2034 grpc::testing::TestEnvironment env(argc, argv);
2035 ::testing::InitGoogleTest(&argc, argv);
2036 const auto result = RUN_ALL_TESTS();
2037 return result;
2038 }
2039