OLD | NEW |
| (Empty) |
1 /* | |
2 * Copyright 2014 The WebRTC Project Authors. All rights reserved. | |
3 * | |
4 * Use of this source code is governed by a BSD-style license | |
5 * that can be found in the LICENSE file in the root of the source | |
6 * tree. An additional intellectual property rights grant can be found | |
7 * in the file PATENTS. All contributing project authors may | |
8 * be found in the AUTHORS file in the root of the source tree. | |
9 */ | |
10 | |
11 #include <memory> | |
12 #include <set> | |
13 #include <vector> | |
14 | |
15 #include "webrtc/base/arraysize.h" | |
16 #include "webrtc/base/checks.h" | |
17 #include "webrtc/base/criticalsection.h" | |
18 #include "webrtc/base/event.h" | |
19 #include "webrtc/base/gunit.h" | |
20 #include "webrtc/base/platform_thread.h" | |
21 #include "webrtc/base/thread.h" | |
22 | |
23 namespace rtc { | |
24 | |
25 namespace { | |
26 | |
27 const int kLongTime = 10000; // 10 seconds | |
28 const int kNumThreads = 16; | |
29 const int kOperationsToRun = 1000; | |
30 | |
31 class UniqueValueVerifier { | |
32 public: | |
33 void Verify(const std::vector<int>& values) { | |
34 for (size_t i = 0; i < values.size(); ++i) { | |
35 std::pair<std::set<int>::iterator, bool> result = | |
36 all_values_.insert(values[i]); | |
37 // Each value should only be taken by one thread, so if this value | |
38 // has already been added, something went wrong. | |
39 EXPECT_TRUE(result.second) | |
40 << " Thread=" << Thread::Current() << " value=" << values[i]; | |
41 } | |
42 } | |
43 | |
44 void Finalize() {} | |
45 | |
46 private: | |
47 std::set<int> all_values_; | |
48 }; | |
49 | |
50 class CompareAndSwapVerifier { | |
51 public: | |
52 CompareAndSwapVerifier() : zero_count_(0) {} | |
53 | |
54 void Verify(const std::vector<int>& values) { | |
55 for (auto v : values) { | |
56 if (v == 0) { | |
57 EXPECT_EQ(0, zero_count_) << "Thread=" << Thread::Current(); | |
58 ++zero_count_; | |
59 } else { | |
60 EXPECT_EQ(1, v) << " Thread=" << Thread::Current(); | |
61 } | |
62 } | |
63 } | |
64 | |
65 void Finalize() { | |
66 EXPECT_EQ(1, zero_count_); | |
67 } | |
68 private: | |
69 int zero_count_; | |
70 }; | |
71 | |
72 class RunnerBase : public MessageHandler { | |
73 public: | |
74 explicit RunnerBase(int value) | |
75 : threads_active_(0), | |
76 start_event_(true, false), | |
77 done_event_(true, false), | |
78 shared_value_(value) {} | |
79 | |
80 bool Run() { | |
81 // Signal all threads to start. | |
82 start_event_.Set(); | |
83 | |
84 // Wait for all threads to finish. | |
85 return done_event_.Wait(kLongTime); | |
86 } | |
87 | |
88 void SetExpectedThreadCount(int count) { | |
89 threads_active_ = count; | |
90 } | |
91 | |
92 int shared_value() const { return shared_value_; } | |
93 | |
94 protected: | |
95 // Derived classes must override OnMessage, and call BeforeStart and AfterEnd | |
96 // at the beginning and the end of OnMessage respectively. | |
97 void BeforeStart() { | |
98 ASSERT_TRUE(start_event_.Wait(kLongTime)); | |
99 } | |
100 | |
101 // Returns true if all threads have finished. | |
102 bool AfterEnd() { | |
103 if (AtomicOps::Decrement(&threads_active_) == 0) { | |
104 done_event_.Set(); | |
105 return true; | |
106 } | |
107 return false; | |
108 } | |
109 | |
110 int threads_active_; | |
111 Event start_event_; | |
112 Event done_event_; | |
113 int shared_value_; | |
114 }; | |
115 | |
116 class LOCKABLE CriticalSectionLock { | |
117 public: | |
118 void Lock() EXCLUSIVE_LOCK_FUNCTION() { | |
119 cs_.Enter(); | |
120 } | |
121 void Unlock() UNLOCK_FUNCTION() { | |
122 cs_.Leave(); | |
123 } | |
124 | |
125 private: | |
126 CriticalSection cs_; | |
127 }; | |
128 | |
129 template <class Lock> | |
130 class LockRunner : public RunnerBase { | |
131 public: | |
132 LockRunner() : RunnerBase(0) {} | |
133 | |
134 void OnMessage(Message* msg) override { | |
135 BeforeStart(); | |
136 | |
137 lock_.Lock(); | |
138 | |
139 EXPECT_EQ(0, shared_value_); | |
140 int old = shared_value_; | |
141 | |
142 // Use a loop to increase the chance of race. | |
143 for (int i = 0; i < kOperationsToRun; ++i) { | |
144 ++shared_value_; | |
145 } | |
146 EXPECT_EQ(old + kOperationsToRun, shared_value_); | |
147 shared_value_ = 0; | |
148 | |
149 lock_.Unlock(); | |
150 | |
151 AfterEnd(); | |
152 } | |
153 | |
154 private: | |
155 Lock lock_; | |
156 }; | |
157 | |
158 template <class Op, class Verifier> | |
159 class AtomicOpRunner : public RunnerBase { | |
160 public: | |
161 explicit AtomicOpRunner(int initial_value) : RunnerBase(initial_value) {} | |
162 | |
163 void OnMessage(Message* msg) override { | |
164 BeforeStart(); | |
165 | |
166 std::vector<int> values; | |
167 values.reserve(kOperationsToRun); | |
168 | |
169 // Generate a bunch of values by updating shared_value_ atomically. | |
170 for (int i = 0; i < kOperationsToRun; ++i) { | |
171 values.push_back(Op::AtomicOp(&shared_value_)); | |
172 } | |
173 | |
174 { // Add them all to the set. | |
175 CritScope cs(&all_values_crit_); | |
176 verifier_.Verify(values); | |
177 } | |
178 | |
179 if (AfterEnd()) { | |
180 verifier_.Finalize(); | |
181 } | |
182 } | |
183 | |
184 private: | |
185 CriticalSection all_values_crit_; | |
186 Verifier verifier_; | |
187 }; | |
188 | |
189 struct IncrementOp { | |
190 static int AtomicOp(int* i) { return AtomicOps::Increment(i); } | |
191 }; | |
192 | |
193 struct DecrementOp { | |
194 static int AtomicOp(int* i) { return AtomicOps::Decrement(i); } | |
195 }; | |
196 | |
197 struct CompareAndSwapOp { | |
198 static int AtomicOp(int* i) { return AtomicOps::CompareAndSwap(i, 0, 1); } | |
199 }; | |
200 | |
201 void StartThreads(std::vector<std::unique_ptr<Thread>>* threads, | |
202 MessageHandler* handler) { | |
203 for (int i = 0; i < kNumThreads; ++i) { | |
204 std::unique_ptr<Thread> thread(new Thread()); | |
205 thread->Start(); | |
206 thread->Post(RTC_FROM_HERE, handler); | |
207 threads->push_back(std::move(thread)); | |
208 } | |
209 } | |
210 | |
211 } // namespace | |
212 | |
213 TEST(AtomicOpsTest, Simple) { | |
214 int value = 0; | |
215 EXPECT_EQ(1, AtomicOps::Increment(&value)); | |
216 EXPECT_EQ(1, value); | |
217 EXPECT_EQ(2, AtomicOps::Increment(&value)); | |
218 EXPECT_EQ(2, value); | |
219 EXPECT_EQ(1, AtomicOps::Decrement(&value)); | |
220 EXPECT_EQ(1, value); | |
221 EXPECT_EQ(0, AtomicOps::Decrement(&value)); | |
222 EXPECT_EQ(0, value); | |
223 } | |
224 | |
225 TEST(AtomicOpsTest, SimplePtr) { | |
226 class Foo {}; | |
227 Foo* volatile foo = nullptr; | |
228 std::unique_ptr<Foo> a(new Foo()); | |
229 std::unique_ptr<Foo> b(new Foo()); | |
230 // Reading the initial value should work as expected. | |
231 EXPECT_TRUE(rtc::AtomicOps::AcquireLoadPtr(&foo) == nullptr); | |
232 // Setting using compare and swap should work. | |
233 EXPECT_TRUE(rtc::AtomicOps::CompareAndSwapPtr( | |
234 &foo, static_cast<Foo*>(nullptr), a.get()) == nullptr); | |
235 EXPECT_TRUE(rtc::AtomicOps::AcquireLoadPtr(&foo) == a.get()); | |
236 // Setting another value but with the wrong previous pointer should fail | |
237 // (remain a). | |
238 EXPECT_TRUE(rtc::AtomicOps::CompareAndSwapPtr( | |
239 &foo, static_cast<Foo*>(nullptr), b.get()) == a.get()); | |
240 EXPECT_TRUE(rtc::AtomicOps::AcquireLoadPtr(&foo) == a.get()); | |
241 // Replacing a with b should work. | |
242 EXPECT_TRUE(rtc::AtomicOps::CompareAndSwapPtr(&foo, a.get(), b.get()) == | |
243 a.get()); | |
244 EXPECT_TRUE(rtc::AtomicOps::AcquireLoadPtr(&foo) == b.get()); | |
245 } | |
246 | |
247 TEST(AtomicOpsTest, Increment) { | |
248 // Create and start lots of threads. | |
249 AtomicOpRunner<IncrementOp, UniqueValueVerifier> runner(0); | |
250 std::vector<std::unique_ptr<Thread>> threads; | |
251 StartThreads(&threads, &runner); | |
252 runner.SetExpectedThreadCount(kNumThreads); | |
253 | |
254 // Release the hounds! | |
255 EXPECT_TRUE(runner.Run()); | |
256 EXPECT_EQ(kOperationsToRun * kNumThreads, runner.shared_value()); | |
257 } | |
258 | |
259 TEST(AtomicOpsTest, Decrement) { | |
260 // Create and start lots of threads. | |
261 AtomicOpRunner<DecrementOp, UniqueValueVerifier> runner( | |
262 kOperationsToRun * kNumThreads); | |
263 std::vector<std::unique_ptr<Thread>> threads; | |
264 StartThreads(&threads, &runner); | |
265 runner.SetExpectedThreadCount(kNumThreads); | |
266 | |
267 // Release the hounds! | |
268 EXPECT_TRUE(runner.Run()); | |
269 EXPECT_EQ(0, runner.shared_value()); | |
270 } | |
271 | |
272 TEST(AtomicOpsTest, CompareAndSwap) { | |
273 // Create and start lots of threads. | |
274 AtomicOpRunner<CompareAndSwapOp, CompareAndSwapVerifier> runner(0); | |
275 std::vector<std::unique_ptr<Thread>> threads; | |
276 StartThreads(&threads, &runner); | |
277 runner.SetExpectedThreadCount(kNumThreads); | |
278 | |
279 // Release the hounds! | |
280 EXPECT_TRUE(runner.Run()); | |
281 EXPECT_EQ(1, runner.shared_value()); | |
282 } | |
283 | |
284 TEST(GlobalLockTest, Basic) { | |
285 // Create and start lots of threads. | |
286 LockRunner<GlobalLock> runner; | |
287 std::vector<std::unique_ptr<Thread>> threads; | |
288 StartThreads(&threads, &runner); | |
289 runner.SetExpectedThreadCount(kNumThreads); | |
290 | |
291 // Release the hounds! | |
292 EXPECT_TRUE(runner.Run()); | |
293 EXPECT_EQ(0, runner.shared_value()); | |
294 } | |
295 | |
296 TEST(CriticalSectionTest, Basic) { | |
297 // Create and start lots of threads. | |
298 LockRunner<CriticalSectionLock> runner; | |
299 std::vector<std::unique_ptr<Thread>> threads; | |
300 StartThreads(&threads, &runner); | |
301 runner.SetExpectedThreadCount(kNumThreads); | |
302 | |
303 // Release the hounds! | |
304 EXPECT_TRUE(runner.Run()); | |
305 EXPECT_EQ(0, runner.shared_value()); | |
306 } | |
307 | |
308 class PerfTestData { | |
309 public: | |
310 PerfTestData(int expected_count, Event* event) | |
311 : cache_line_barrier_1_(), cache_line_barrier_2_(), | |
312 expected_count_(expected_count), event_(event) { | |
313 cache_line_barrier_1_[0]++; // Avoid 'is not used'. | |
314 cache_line_barrier_2_[0]++; // Avoid 'is not used'. | |
315 } | |
316 ~PerfTestData() {} | |
317 | |
318 void AddToCounter(int add) { | |
319 rtc::CritScope cs(&lock_); | |
320 my_counter_ += add; | |
321 if (my_counter_ == expected_count_) | |
322 event_->Set(); | |
323 } | |
324 | |
325 int64_t total() const { | |
326 // Assume that only one thread is running now. | |
327 return my_counter_; | |
328 } | |
329 | |
330 private: | |
331 uint8_t cache_line_barrier_1_[64]; | |
332 CriticalSection lock_; | |
333 uint8_t cache_line_barrier_2_[64]; | |
334 int64_t my_counter_ = 0; | |
335 const int expected_count_; | |
336 Event* const event_; | |
337 }; | |
338 | |
339 class PerfTestThread { | |
340 public: | |
341 PerfTestThread() : thread_(&ThreadFunc, this, "CsPerf") {} | |
342 | |
343 void Start(PerfTestData* data, int repeats, int id) { | |
344 RTC_DCHECK(!thread_.IsRunning()); | |
345 RTC_DCHECK(!data_); | |
346 data_ = data; | |
347 repeats_ = repeats; | |
348 my_id_ = id; | |
349 thread_.Start(); | |
350 } | |
351 | |
352 void Stop() { | |
353 RTC_DCHECK(thread_.IsRunning()); | |
354 RTC_DCHECK(data_); | |
355 thread_.Stop(); | |
356 repeats_ = 0; | |
357 data_ = nullptr; | |
358 my_id_ = 0; | |
359 } | |
360 | |
361 private: | |
362 static bool ThreadFunc(void* param) { | |
363 PerfTestThread* me = static_cast<PerfTestThread*>(param); | |
364 for (int i = 0; i < me->repeats_; ++i) | |
365 me->data_->AddToCounter(me->my_id_); | |
366 return false; | |
367 } | |
368 | |
369 PlatformThread thread_; | |
370 PerfTestData* data_ = nullptr; | |
371 int repeats_ = 0; | |
372 int my_id_ = 0; | |
373 }; | |
374 | |
375 // Comparison of output of this test as tested on a MacBook Pro Retina, 15-inch, | |
376 // Mid 2014, 2,8 GHz Intel Core i7, 16 GB 1600 MHz DDR3, | |
377 // running OS X El Capitan, 10.11.2. | |
378 // | |
379 // Native mutex implementation: | |
380 // Approximate CPU usage: | |
381 // System: ~16% | |
382 // User mode: ~1.3% | |
383 // Idle: ~82% | |
384 // Unit test output: | |
385 // [ OK ] CriticalSectionTest.Performance (234545 ms) | |
386 // | |
387 // Special partially spin lock based implementation: | |
388 // Approximate CPU usage: | |
389 // System: ~75% | |
390 // User mode: ~16% | |
391 // Idle: ~8% | |
392 // Unit test output: | |
393 // [ OK ] CriticalSectionTest.Performance (2107 ms) | |
394 // | |
395 // The test is disabled by default to avoid unecessarily loading the bots. | |
396 TEST(CriticalSectionTest, DISABLED_Performance) { | |
397 PerfTestThread threads[8]; | |
398 Event event(false, false); | |
399 | |
400 static const int kThreadRepeats = 10000000; | |
401 static const int kExpectedCount = kThreadRepeats * arraysize(threads); | |
402 PerfTestData test_data(kExpectedCount, &event); | |
403 | |
404 for (auto& t : threads) | |
405 t.Start(&test_data, kThreadRepeats, 1); | |
406 | |
407 event.Wait(Event::kForever); | |
408 | |
409 for (auto& t : threads) | |
410 t.Stop(); | |
411 } | |
412 | |
413 } // namespace rtc | |
OLD | NEW |