| 1 | //===----------------------------------------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | // UNSUPPORTED: c++03 |
| 10 | // UNSUPPORTED: no-threads |
| 11 | // UNSUPPORTED: no-exceptions |
| 12 | |
| 13 | #define TESTING_CXA_GUARD |
| 14 | #include "../src/cxa_guard_impl.h" |
| 15 | #include <unordered_map> |
| 16 | #include <thread> |
| 17 | #include <atomic> |
| 18 | #include <array> |
| 19 | #include <cassert> |
| 20 | #include <memory> |
| 21 | #include <vector> |
| 22 | |
| 23 | #include "make_test_thread.h" |
| 24 | #include "test_macros.h" |
| 25 | |
| 26 | |
| 27 | using namespace __cxxabiv1; |
| 28 | |
| 29 | // Misc test configuration. It's used to tune the flakyness of the test. |
| 30 | // ThreadsPerTest - The number of threads used |
| 31 | constexpr int ThreadsPerTest = 10; |
| 32 | // The number of instances of a test to run concurrently. |
| 33 | constexpr int ConcurrentRunsPerTest = 10; |
| 34 | // The number of times to rerun each test. |
| 35 | constexpr int TestSamples = 50; |
| 36 | |
| 37 | |
| 38 | |
| 39 | void BusyWait() { |
| 40 | std::this_thread::yield(); |
| 41 | } |
| 42 | |
| 43 | void YieldAfterBarrier() { |
| 44 | std::this_thread::sleep_for(rtime: std::chrono::nanoseconds(10)); |
| 45 | std::this_thread::yield(); |
| 46 | } |
| 47 | |
| 48 | struct Barrier { |
| 49 | explicit Barrier(int n) : m_threads(n), m_remaining(n) { } |
| 50 | Barrier(Barrier const&) = delete; |
| 51 | Barrier& operator=(Barrier const&) = delete; |
| 52 | |
| 53 | void arrive_and_wait() const { |
| 54 | --m_remaining; |
| 55 | while (m_remaining.load()) { |
| 56 | BusyWait(); |
| 57 | } |
| 58 | } |
| 59 | |
| 60 | void arrive_and_drop() const { |
| 61 | --m_remaining; |
| 62 | } |
| 63 | |
| 64 | void wait_for_threads(int n) const { |
| 65 | while ((m_threads - m_remaining.load()) < n) { |
| 66 | std::this_thread::yield(); |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | private: |
| 71 | const int m_threads; |
| 72 | mutable std::atomic<int> m_remaining; |
| 73 | }; |
| 74 | |
| 75 | |
| 76 | enum class InitResult { |
| 77 | COMPLETE, |
| 78 | PERFORMED, |
| 79 | WAITED, |
| 80 | ABORTED |
| 81 | }; |
| 82 | constexpr InitResult COMPLETE = InitResult::COMPLETE; |
| 83 | constexpr InitResult PERFORMED = InitResult::PERFORMED; |
| 84 | constexpr InitResult WAITED = InitResult::WAITED; |
| 85 | constexpr InitResult ABORTED = InitResult::ABORTED; |
| 86 | |
| 87 | |
| 88 | template <class Impl, class GuardType, class Init> |
| 89 | InitResult check_guard(GuardType *g, Init init) { |
| 90 | uint8_t *first_byte = reinterpret_cast<uint8_t*>(g); |
| 91 | if (std::__libcpp_atomic_load(first_byte, std::_AO_Acquire) == 0) { |
| 92 | Impl impl(g); |
| 93 | if (impl.cxa_guard_acquire() == INIT_IS_PENDING) { |
| 94 | #ifndef TEST_HAS_NO_EXCEPTIONS |
| 95 | try { |
| 96 | #endif |
| 97 | init(); |
| 98 | impl.cxa_guard_release(); |
| 99 | return PERFORMED; |
| 100 | #ifndef TEST_HAS_NO_EXCEPTIONS |
| 101 | } catch (...) { |
| 102 | impl.cxa_guard_abort(); |
| 103 | return ABORTED; |
| 104 | } |
| 105 | #endif |
| 106 | } |
| 107 | return WAITED; |
| 108 | } |
| 109 | return COMPLETE; |
| 110 | } |
| 111 | |
| 112 | |
| 113 | template <class GuardType, class Impl> |
| 114 | struct FunctionLocalStatic { |
| 115 | FunctionLocalStatic() {} |
| 116 | FunctionLocalStatic(FunctionLocalStatic const&) = delete; |
| 117 | |
| 118 | template <class InitFunc> |
| 119 | InitResult access(InitFunc&& init) { |
| 120 | auto res = check_guard<Impl>(&guard_object, init); |
| 121 | ++result_counts[static_cast<int>(res)]; |
| 122 | return res; |
| 123 | } |
| 124 | |
| 125 | template <class InitFn> |
| 126 | struct AccessCallback { |
| 127 | void operator()() const { this_obj->access(init); } |
| 128 | |
| 129 | FunctionLocalStatic *this_obj; |
| 130 | InitFn init; |
| 131 | }; |
| 132 | |
| 133 | template <class InitFn, class Callback = AccessCallback< InitFn > > |
| 134 | Callback access_callback(InitFn init) { |
| 135 | return Callback{this, init}; |
| 136 | } |
| 137 | |
| 138 | int get_count(InitResult I) const { |
| 139 | return result_counts[static_cast<int>(I)].load(); |
| 140 | } |
| 141 | |
| 142 | int num_completed() const { |
| 143 | return get_count(I: COMPLETE) + get_count(I: PERFORMED) + get_count(I: WAITED); |
| 144 | } |
| 145 | |
| 146 | int num_waiting() const { |
| 147 | return waiting_threads.load(); |
| 148 | } |
| 149 | |
| 150 | private: |
| 151 | GuardType guard_object = {}; |
| 152 | std::atomic<int> waiting_threads{0}; |
| 153 | std::array<std::atomic<int>, 4> result_counts{}; |
| 154 | static_assert(static_cast<int>(ABORTED) == 3, "only 4 result kinds expected" ); |
| 155 | }; |
| 156 | |
| 157 | struct ThreadGroup { |
| 158 | ThreadGroup() = default; |
| 159 | ThreadGroup(ThreadGroup const&) = delete; |
| 160 | |
| 161 | template <class ...Args> |
| 162 | void Create(Args&& ...args) { |
| 163 | threads.emplace_back(std::forward<Args>(args)...); |
| 164 | } |
| 165 | |
| 166 | template <class Callback> |
| 167 | void CreateThreadsWithBarrier(int N, Callback cb) { |
| 168 | auto start = std::make_shared<Barrier>(args: N + 1); |
| 169 | for (int I=0; I < N; ++I) { |
| 170 | Create([start, cb]() { |
| 171 | start->arrive_and_wait(); |
| 172 | cb(); |
| 173 | }); |
| 174 | } |
| 175 | start->arrive_and_wait(); |
| 176 | } |
| 177 | |
| 178 | void JoinAll() { |
| 179 | for (auto& t : threads) { |
| 180 | t.join(); |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | private: |
| 185 | std::vector<std::thread> threads; |
| 186 | }; |
| 187 | |
| 188 | |
| 189 | template <class GuardType, class Impl> |
| 190 | void test_free_for_all(int num_waiters) { |
| 191 | FunctionLocalStatic<GuardType, Impl> test_obj; |
| 192 | |
| 193 | ThreadGroup threads; |
| 194 | |
| 195 | bool already_init = false; |
| 196 | threads.CreateThreadsWithBarrier(num_waiters, |
| 197 | test_obj.access_callback([&]() { |
| 198 | assert(!already_init); |
| 199 | already_init = true; |
| 200 | }) |
| 201 | ); |
| 202 | |
| 203 | // wait for the other threads to finish initialization. |
| 204 | threads.JoinAll(); |
| 205 | |
| 206 | assert(test_obj.get_count(PERFORMED) == 1); |
| 207 | assert(test_obj.get_count(COMPLETE) + test_obj.get_count(WAITED) == num_waiters - 1); |
| 208 | } |
| 209 | |
| 210 | template <class GuardType, class Impl> |
| 211 | void test_waiting_for_init(int num_waiters) { |
| 212 | FunctionLocalStatic<GuardType, Impl> test_obj; |
| 213 | |
| 214 | ThreadGroup threads; |
| 215 | |
| 216 | Barrier start_init(2); |
| 217 | threads.Create(test_obj.access_callback( |
| 218 | [&]() { |
| 219 | start_init.arrive_and_wait(); |
| 220 | // Take our sweet time completing the initialization... |
| 221 | // |
| 222 | // There's a race condition between the other threads reaching the |
| 223 | // start_init barrier, and them actually hitting the cxa guard. |
| 224 | // But we're trying to test the waiting logic, we want as many |
| 225 | // threads to enter the waiting loop as possible. |
| 226 | YieldAfterBarrier(); |
| 227 | } |
| 228 | )); |
| 229 | start_init.wait_for_threads(n: 1); |
| 230 | |
| 231 | threads.CreateThreadsWithBarrier(num_waiters, |
| 232 | test_obj.access_callback([]() { assert(false); }) |
| 233 | ); |
| 234 | // unblock the initializing thread |
| 235 | start_init.arrive_and_drop(); |
| 236 | |
| 237 | // wait for the other threads to finish initialization. |
| 238 | threads.JoinAll(); |
| 239 | |
| 240 | assert(test_obj.get_count(PERFORMED) == 1); |
| 241 | assert(test_obj.get_count(ABORTED) == 0); |
| 242 | assert(test_obj.get_count(COMPLETE) + test_obj.get_count(WAITED) == num_waiters); |
| 243 | } |
| 244 | |
| 245 | |
| 246 | template <class GuardType, class Impl> |
| 247 | void test_aborted_init(int num_waiters) { |
| 248 | FunctionLocalStatic<GuardType, Impl> test_obj; |
| 249 | |
| 250 | Barrier start_init(2); |
| 251 | ThreadGroup threads; |
| 252 | threads.Create(test_obj.access_callback( |
| 253 | [&]() { |
| 254 | start_init.arrive_and_wait(); |
| 255 | YieldAfterBarrier(); |
| 256 | throw 42; |
| 257 | }) |
| 258 | ); |
| 259 | start_init.wait_for_threads(n: 1); |
| 260 | |
| 261 | bool already_init = false; |
| 262 | threads.CreateThreadsWithBarrier(num_waiters, |
| 263 | test_obj.access_callback([&]() { |
| 264 | assert(!already_init); |
| 265 | already_init = true; |
| 266 | }) |
| 267 | ); |
| 268 | // unblock the initializing thread |
| 269 | start_init.arrive_and_drop(); |
| 270 | |
| 271 | // wait for the other threads to finish initialization. |
| 272 | threads.JoinAll(); |
| 273 | |
| 274 | assert(test_obj.get_count(ABORTED) == 1); |
| 275 | assert(test_obj.get_count(PERFORMED) == 1); |
| 276 | assert(test_obj.get_count(WAITED) + test_obj.get_count(COMPLETE) == num_waiters - 1); |
| 277 | } |
| 278 | |
| 279 | |
| 280 | template <class GuardType, class Impl> |
| 281 | void test_completed_init(int num_waiters) { |
| 282 | |
| 283 | FunctionLocalStatic<GuardType, Impl> test_obj; |
| 284 | |
| 285 | test_obj.access([]() {}); // initialize the object |
| 286 | assert(test_obj.num_waiting() == 0); |
| 287 | assert(test_obj.num_completed() == 1); |
| 288 | assert(test_obj.get_count(PERFORMED) == 1); |
| 289 | |
| 290 | ThreadGroup threads; |
| 291 | threads.CreateThreadsWithBarrier(num_waiters, |
| 292 | test_obj.access_callback([]() { assert(false); }) |
| 293 | ); |
| 294 | // wait for the other threads to finish initialization. |
| 295 | threads.JoinAll(); |
| 296 | |
| 297 | assert(test_obj.get_count(ABORTED) == 0); |
| 298 | assert(test_obj.get_count(PERFORMED) == 1); |
| 299 | assert(test_obj.get_count(WAITED) == 0); |
| 300 | assert(test_obj.get_count(COMPLETE) == num_waiters); |
| 301 | } |
| 302 | |
| 303 | template <class Impl> |
| 304 | void test_impl() { |
| 305 | using TestFn = void(*)(int); |
| 306 | TestFn TestList[] = { |
| 307 | test_free_for_all<uint32_t, Impl>, |
| 308 | test_free_for_all<uint32_t, Impl>, |
| 309 | test_waiting_for_init<uint32_t, Impl>, |
| 310 | test_waiting_for_init<uint64_t, Impl>, |
| 311 | test_aborted_init<uint32_t, Impl>, |
| 312 | test_aborted_init<uint64_t, Impl>, |
| 313 | test_completed_init<uint32_t, Impl>, |
| 314 | test_completed_init<uint64_t, Impl> |
| 315 | }; |
| 316 | |
| 317 | for (auto test_func : TestList) { |
| 318 | ThreadGroup test_threads; |
| 319 | test_threads.CreateThreadsWithBarrier(ConcurrentRunsPerTest, [=]() { |
| 320 | for (int I = 0; I < TestSamples; ++I) { |
| 321 | test_func(ThreadsPerTest); |
| 322 | } |
| 323 | }); |
| 324 | test_threads.JoinAll(); |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | void test_all_impls() { |
| 329 | using MutexImpl = SelectImplementation<Implementation::GlobalMutex>::type; |
| 330 | |
| 331 | // Attempt to test the Futex based implementation if it's supported on the |
| 332 | // target platform. |
| 333 | using RealFutexImpl = SelectImplementation<Implementation::Futex>::type; |
| 334 | using FutexImpl = typename std::conditional< |
| 335 | PlatformSupportsFutex(), |
| 336 | RealFutexImpl, |
| 337 | MutexImpl |
| 338 | >::type; |
| 339 | |
| 340 | test_impl<MutexImpl>(); |
| 341 | if (PlatformSupportsFutex()) |
| 342 | test_impl<FutexImpl>(); |
| 343 | } |
| 344 | |
| 345 | // A dummy |
| 346 | template <bool Dummy = true> |
| 347 | void test_futex_syscall() { |
| 348 | if (!PlatformSupportsFutex()) |
| 349 | return; |
| 350 | int lock1 = 0; |
| 351 | int lock2 = 0; |
| 352 | int lock3 = 0; |
| 353 | std::thread waiter1 = support::make_test_thread([&]() { |
| 354 | int expect = 0; |
| 355 | PlatformFutexWait(&lock1, expect); |
| 356 | assert(lock1 == 1); |
| 357 | }); |
| 358 | std::thread waiter2 = support::make_test_thread([&]() { |
| 359 | int expect = 0; |
| 360 | PlatformFutexWait(&lock2, expect); |
| 361 | assert(lock2 == 2); |
| 362 | }); |
| 363 | std::thread waiter3 = support::make_test_thread([&]() { |
| 364 | int expect = 42; // not the value |
| 365 | PlatformFutexWait(&lock3, expect); // doesn't block |
| 366 | }); |
| 367 | std::thread waker = support::make_test_thread([&]() { |
| 368 | lock1 = 1; |
| 369 | PlatformFutexWake(&lock1); |
| 370 | lock2 = 2; |
| 371 | PlatformFutexWake(&lock2); |
| 372 | }); |
| 373 | waiter1.join(); |
| 374 | waiter2.join(); |
| 375 | waiter3.join(); |
| 376 | waker.join(); |
| 377 | } |
| 378 | |
| 379 | int main(int, char**) { |
| 380 | // Test each multi-threaded implementation with real threads. |
| 381 | test_all_impls(); |
| 382 | // Test the basic sanity of the futex syscall wrappers. |
| 383 | test_futex_syscall(); |
| 384 | |
| 385 | return 0; |
| 386 | } |
| 387 | |