brintos

brintos / llvm-project-archived public Read only

0
0
Text · 16.6 KiB · 7f72747 Raw
535 lines · cpp
1//========- unittests/Support/ThreadPools.cpp - ThreadPools.h tests --========//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-exception6//7//===----------------------------------------------------------------------===//8 9#include "llvm/Support/ThreadPool.h"10#include "llvm/ADT/STLExtras.h"11#include "llvm/ADT/ScopeExit.h"12#include "llvm/ADT/SetVector.h"13#include "llvm/ADT/SmallVector.h"14#include "llvm/Config/llvm-config.h" // for LLVM_ENABLE_THREADS15#include "llvm/Support/CommandLine.h"16#include "llvm/Support/Program.h"17#include "llvm/Support/TargetSelect.h"18#include "llvm/Support/Threading.h"19#include "llvm/TargetParser/Host.h"20#include "llvm/TargetParser/Triple.h"21 22#ifdef _WIN3223#include "llvm/Support/Windows/WindowsSupport.h"24#endif25 26#include <chrono>27#include <thread>28 29#include "gtest/gtest.h"30 31namespace testing {32namespace internal {33// Specialize gtest construct to provide friendlier name in the output.34#if LLVM_ENABLE_THREADS35template <> std::string GetTypeName<llvm::StdThreadPool>() {36  return "llvm::StdThreadPool";37}38#endif39template <> std::string GetTypeName<llvm::SingleThreadExecutor>() {40  return "llvm::SingleThreadExecutor";41}42} // namespace internal43} // namespace testing44 45using namespace llvm;46 47// Fixture for the unittests, allowing to *temporarily* disable the unittests48// on a particular platform49template <typename ThreadPoolImpl> class ThreadPoolTest : public testing::Test {50  Triple Host;51  SmallVector<Triple::ArchType, 4> UnsupportedArchs;52  SmallVector<Triple::OSType, 4> UnsupportedOSs;53  SmallVector<Triple::EnvironmentType, 1> UnsupportedEnvironments;54 55protected:56  // This is intended for platform as a temporary "XFAIL"57  bool isUnsupportedOSOrEnvironment() {58    Triple Host(Triple::normalize(sys::getProcessTriple()));59 60    if (find(UnsupportedEnvironments, Host.getEnvironment()) !=61        UnsupportedEnvironments.end())62      return true;63 64    if (is_contained(UnsupportedOSs, Host.getOS()))65      return true;66 67    if (is_contained(UnsupportedArchs, Host.getArch()))68      return true;69 70    return false;71  }72 73  ThreadPoolTest() {74    // Add unsupported configuration here, example:75    //   UnsupportedArchs.push_back(Triple::x86_64);76 77    // See https://llvm.org/bugs/show_bug.cgi?id=2582978    UnsupportedArchs.push_back(Triple::ppc64le);79    UnsupportedArchs.push_back(Triple::ppc64);80  }81 82  /// Make sure this thread not progress faster than the main thread.83  void waitForMainThread() { waitForPhase(1); }84 85  /// Set the readiness of the main thread.86  void setMainThreadReady() { setPhase(1); }87 88  /// Wait until given phase is set using setPhase(); first "main" phase is 1.89  /// See also PhaseResetHelper below.90  void waitForPhase(int Phase) {91    std::unique_lock<std::mutex> LockGuard(CurrentPhaseMutex);92    CurrentPhaseCondition.wait(93        LockGuard, [&] { return CurrentPhase == Phase || CurrentPhase < 0; });94  }95  /// If a thread waits on another phase, the test could bail out on a failed96  /// assertion and ThreadPool destructor would wait() on all threads, which97  /// would deadlock on the task waiting. Create this helper to automatically98  /// reset the phase and unblock such threads.99  struct PhaseResetHelper {100    PhaseResetHelper(ThreadPoolTest *test) : test(test) {}101    ~PhaseResetHelper() { test->setPhase(-1); }102    ThreadPoolTest *test;103  };104 105  /// Advance to the given phase.106  void setPhase(int Phase) {107    {108      std::unique_lock<std::mutex> LockGuard(CurrentPhaseMutex);109      assert(Phase == CurrentPhase + 1 || Phase < 0);110      CurrentPhase = Phase;111    }112    CurrentPhaseCondition.notify_all();113  }114 115  void SetUp() override { CurrentPhase = 0; }116 117  SmallVector<llvm::BitVector, 0> RunOnAllSockets(ThreadPoolStrategy S);118 119  std::condition_variable CurrentPhaseCondition;120  std::mutex CurrentPhaseMutex;121  int CurrentPhase; // -1 = error, 0 = setup, 1 = ready, 2+ = custom122};123 124using ThreadPoolImpls = ::testing::Types<125#if LLVM_ENABLE_THREADS126    StdThreadPool,127#endif128    SingleThreadExecutor>;129 130TYPED_TEST_SUITE(ThreadPoolTest, ThreadPoolImpls, );131 132#define CHECK_UNSUPPORTED()                                                    \133  do {                                                                         \134    if (this->isUnsupportedOSOrEnvironment())                                  \135      GTEST_SKIP();                                                            \136  } while (0);137 138TYPED_TEST(ThreadPoolTest, AsyncBarrier) {139  CHECK_UNSUPPORTED();140  // test that async & barrier work together properly.141 142  std::atomic_int checked_in{0};143 144  TypeParam Pool;145  for (size_t i = 0; i < 5; ++i) {146    Pool.async([this, &checked_in] {147      this->waitForMainThread();148      ++checked_in;149    });150  }151  ASSERT_EQ(0, checked_in);152  this->setMainThreadReady();153  Pool.wait();154  ASSERT_EQ(5, checked_in);155}156 157static void TestFunc(std::atomic_int &checked_in, int i) { checked_in += i; }158 159TYPED_TEST(ThreadPoolTest, AsyncBarrierArgs) {160  CHECK_UNSUPPORTED();161  // Test that async works with a function requiring multiple parameters.162  std::atomic_int checked_in{0};163 164  DefaultThreadPool Pool;165  for (size_t i = 0; i < 5; ++i) {166    Pool.async(TestFunc, std::ref(checked_in), i);167  }168  Pool.wait();169  ASSERT_EQ(10, checked_in);170}171 172TYPED_TEST(ThreadPoolTest, Async) {173  CHECK_UNSUPPORTED();174  DefaultThreadPool Pool;175  std::atomic_int i{0};176  Pool.async([this, &i] {177    this->waitForMainThread();178    ++i;179  });180  Pool.async([&i] { ++i; });181  ASSERT_NE(2, i.load());182  this->setMainThreadReady();183  Pool.wait();184  ASSERT_EQ(2, i.load());185}186 187TYPED_TEST(ThreadPoolTest, AsyncMoveOnly) {188  CHECK_UNSUPPORTED();189  DefaultThreadPool Pool;190  std::promise<int> p;191  std::future<int> f = p.get_future();192  Pool.async([this, p = std::move(p)]() mutable {193    this->waitForMainThread();194    p.set_value(42);195  });196  this->setMainThreadReady();197  Pool.wait();198  ASSERT_EQ(42, f.get());199}200 201TYPED_TEST(ThreadPoolTest, AsyncRAIICaptures) {202  CHECK_UNSUPPORTED();203  DefaultThreadPool Pool(hardware_concurrency(2));204 205  // We use a task group and a non-atomic value to stress test that the chaining206  // of tasks via a captured RAII object in fact chains and synchronizes within207  // a group.208  ThreadPoolTaskGroup Group(Pool);209  int value = 0;210 211  // Create an RAII object that when destroyed schedules more work. This makes212  // it easy to check that the RAII is resolved at the same point as a task runs213  // on the thread pool.214  auto schedule_next = llvm::make_scope_exit([&Group, &value] {215    // We sleep before scheduling the final task to make it much more likely216    // that an incorrect implementation actually exbitits a bug. Without the217    // sleep, we may get "lucky" and have the second task finish before the218    // assertion below fails even with an incorrect implementaiton. The219    // sleep is making _failures_ more reliable, it is not needed for220    // correctness and this test should only flakily _pass_, never flakily221    // fail.222    std::this_thread::sleep_for(std::chrono::milliseconds(10));223    Group.async([&value] { value = 42; });224  });225 226  // Now schedule the initial task, moving the RAII object to schedule the final227  // task into its captures.228  Group.async([schedule_next = std::move(schedule_next)]() {229    // Nothing to do here, the captured RAII object does the work.230  });231 232  // Both tasks should complete here, synchronizing with the read of value.233  Group.wait();234  ASSERT_EQ(42, value);235}236 237TYPED_TEST(ThreadPoolTest, GetFuture) {238  CHECK_UNSUPPORTED();239  DefaultThreadPool Pool(hardware_concurrency(2));240  std::atomic_int i{0};241  Pool.async([this, &i] {242    this->waitForMainThread();243    ++i;244  });245  // Force the future using get()246  Pool.async([&i] { ++i; }).get();247  ASSERT_NE(2, i.load());248  this->setMainThreadReady();249  Pool.wait();250  ASSERT_EQ(2, i.load());251}252 253TYPED_TEST(ThreadPoolTest, GetFutureWithResult) {254  CHECK_UNSUPPORTED();255  DefaultThreadPool Pool(hardware_concurrency(2));256  auto F1 = Pool.async([] { return 1; });257  auto F2 = Pool.async([] { return 2; });258 259  this->setMainThreadReady();260  Pool.wait();261  ASSERT_EQ(1, F1.get());262  ASSERT_EQ(2, F2.get());263}264 265TYPED_TEST(ThreadPoolTest, GetFutureWithResultAndArgs) {266  CHECK_UNSUPPORTED();267  DefaultThreadPool Pool(hardware_concurrency(2));268  auto Fn = [](int x) { return x; };269  auto F1 = Pool.async(Fn, 1);270  auto F2 = Pool.async(Fn, 2);271 272  this->setMainThreadReady();273  Pool.wait();274  ASSERT_EQ(1, F1.get());275  ASSERT_EQ(2, F2.get());276}277 278TYPED_TEST(ThreadPoolTest, PoolDestruction) {279  CHECK_UNSUPPORTED();280  // Test that we are waiting on destruction281  std::atomic_int checked_in{0};282  {283    DefaultThreadPool Pool;284    for (size_t i = 0; i < 5; ++i) {285      Pool.async([this, &checked_in] {286        this->waitForMainThread();287        ++checked_in;288      });289    }290    ASSERT_EQ(0, checked_in);291    this->setMainThreadReady();292  }293  ASSERT_EQ(5, checked_in);294}295 296// Check running tasks in different groups.297TYPED_TEST(ThreadPoolTest, Groups) {298  CHECK_UNSUPPORTED();299  // Need at least two threads, as the task in group2300  // might block a thread until all tasks in group1 finish.301  ThreadPoolStrategy S = hardware_concurrency(2);302  if (S.compute_thread_count() < 2)303    GTEST_SKIP();304  DefaultThreadPool Pool(S);305  typename TestFixture::PhaseResetHelper Helper(this);306  ThreadPoolTaskGroup Group1(Pool);307  ThreadPoolTaskGroup Group2(Pool);308 309  // Check that waiting for an empty group is a no-op.310  Group1.wait();311 312  std::atomic_int checked_in1{0};313  std::atomic_int checked_in2{0};314 315  for (size_t i = 0; i < 5; ++i) {316    Group1.async([this, &checked_in1] {317      this->waitForMainThread();318      ++checked_in1;319    });320  }321  Group2.async([this, &checked_in2] {322    this->waitForPhase(2);323    ++checked_in2;324  });325  ASSERT_EQ(0, checked_in1);326  ASSERT_EQ(0, checked_in2);327  // Start first group and wait for it.328  this->setMainThreadReady();329  Group1.wait();330  ASSERT_EQ(5, checked_in1);331  // Second group has not yet finished, start it and wait for it.332  ASSERT_EQ(0, checked_in2);333  this->setPhase(2);334  Group2.wait();335  ASSERT_EQ(5, checked_in1);336  ASSERT_EQ(1, checked_in2);337}338 339// Check recursive tasks.340TYPED_TEST(ThreadPoolTest, RecursiveGroups) {341  CHECK_UNSUPPORTED();342  DefaultThreadPool Pool;343  ThreadPoolTaskGroup Group(Pool);344 345  std::atomic_int checked_in1{0};346 347  for (size_t i = 0; i < 5; ++i) {348    Group.async([this, &Pool, &checked_in1] {349      this->waitForMainThread();350 351      ThreadPoolTaskGroup LocalGroup(Pool);352 353      // Check that waiting for an empty group is a no-op.354      LocalGroup.wait();355 356      std::atomic_int checked_in2{0};357      for (size_t i = 0; i < 5; ++i) {358        LocalGroup.async([&checked_in2] { ++checked_in2; });359      }360      LocalGroup.wait();361      ASSERT_EQ(5, checked_in2);362 363      ++checked_in1;364    });365  }366  ASSERT_EQ(0, checked_in1);367  this->setMainThreadReady();368  Group.wait();369  ASSERT_EQ(5, checked_in1);370}371 372TYPED_TEST(ThreadPoolTest, RecursiveWaitDeadlock) {373  CHECK_UNSUPPORTED();374  ThreadPoolStrategy S = hardware_concurrency(2);375  if (S.compute_thread_count() < 2)376    GTEST_SKIP();377  DefaultThreadPool Pool(S);378  typename TestFixture::PhaseResetHelper Helper(this);379  ThreadPoolTaskGroup Group(Pool);380 381  // Test that a thread calling wait() for a group and is waiting for more tasks382  // returns when the last task finishes in a different thread while the waiting383  // thread was waiting for more tasks to process while waiting.384 385  // Task A runs in the first thread. It finishes and leaves386  // the background thread waiting for more tasks.387  Group.async([this] {388    this->waitForMainThread();389    this->setPhase(2);390  });391  // Task B is run in a second thread, it launches yet another392  // task C in a different group, which will be handled by the waiting393  // thread started above.394  Group.async([this, &Pool] {395    this->waitForPhase(2);396    ThreadPoolTaskGroup LocalGroup(Pool);397    LocalGroup.async([this] {398      this->waitForPhase(3);399      // Give the other thread enough time to check that there's no task400      // to process and suspend waiting for a notification. This is indeed racy,401      // but probably the best that can be done.402      std::this_thread::sleep_for(std::chrono::milliseconds(10));403    });404    // And task B only now will wait for the tasks in the group (=task C)405    // to finish. This test checks that it does not deadlock. If the406    // `NotifyGroup` handling in ThreadPool::processTasks() didn't take place,407    // this task B would be stuck waiting for tasks to arrive.408    this->setPhase(3);409    LocalGroup.wait();410  });411  this->setMainThreadReady();412  Group.wait();413}414 415#if LLVM_ENABLE_THREADS == 1416 417// FIXME: Skip some tests below on non-Windows because multi-socket systems418// were not fully tested on Unix yet, and llvm::get_thread_affinity_mask()419// isn't implemented for Unix (need AffinityMask in Support/Unix/Program.inc).420#ifdef _WIN32421 422template <typename ThreadPoolImpl>423SmallVector<llvm::BitVector, 0>424ThreadPoolTest<ThreadPoolImpl>::RunOnAllSockets(ThreadPoolStrategy S) {425  llvm::SetVector<llvm::BitVector> ThreadsUsed;426  std::mutex Lock;427  {428    std::condition_variable AllThreads;429    std::mutex AllThreadsLock;430    unsigned Active = 0;431 432    DefaultThreadPool Pool(S);433    for (size_t I = 0; I < S.compute_thread_count(); ++I) {434      Pool.async([&] {435        {436          std::lock_guard<std::mutex> Guard(AllThreadsLock);437          ++Active;438          AllThreads.notify_one();439        }440        this->waitForMainThread();441        std::lock_guard<std::mutex> Guard(Lock);442        auto Mask = llvm::get_thread_affinity_mask();443        ThreadsUsed.insert(Mask);444      });445    }446    EXPECT_EQ(true, ThreadsUsed.empty());447    {448      std::unique_lock<std::mutex> Guard(AllThreadsLock);449      AllThreads.wait(Guard,450                      [&]() { return Active == S.compute_thread_count(); });451    }452    this->setMainThreadReady();453  }454  return ThreadsUsed.takeVector();455}456 457TYPED_TEST(ThreadPoolTest, AllThreads_UseAllRessources) {458  CHECK_UNSUPPORTED();459  // After Windows 11, the OS is free to deploy the threads on any CPU socket.460  // We cannot relibly ensure that all thread affinity mask are covered,461  // therefore this test should not run.462  if (llvm::RunningWindows11OrGreater())463    GTEST_SKIP();464  auto ThreadsUsed = this->RunOnAllSockets({});465  ASSERT_EQ(llvm::get_cpus(), ThreadsUsed.size());466}467 468TYPED_TEST(ThreadPoolTest, AllThreads_OneThreadPerCore) {469  CHECK_UNSUPPORTED();470  // After Windows 11, the OS is free to deploy the threads on any CPU socket.471  // We cannot relibly ensure that all thread affinity mask are covered,472  // therefore this test should not run.473  if (llvm::RunningWindows11OrGreater())474    GTEST_SKIP();475  auto ThreadsUsed =476      this->RunOnAllSockets(llvm::heavyweight_hardware_concurrency());477  ASSERT_EQ(llvm::get_cpus(), ThreadsUsed.size());478}479 480// From TestMain.cpp.481extern const char *TestMainArgv0;482 483// Just a reachable symbol to ease resolving of the executable's path.484static cl::opt<std::string> ThreadPoolTestStringArg1("thread-pool-string-arg1");485 486#ifdef _WIN32487#define setenv(name, var, ignore) _putenv_s(name, var)488#endif489 490TYPED_TEST(ThreadPoolTest, AffinityMask) {491  CHECK_UNSUPPORTED();492 493  // Skip this test if less than 4 threads are available.494  if (llvm::hardware_concurrency().compute_thread_count() < 4)495    GTEST_SKIP();496 497  using namespace llvm::sys;498  if (getenv("LLVM_THREADPOOL_AFFINITYMASK")) {499    auto ThreadsUsed = this->RunOnAllSockets({});500    // Ensure the threads only ran on CPUs 0-3.501    // NOTE: Don't use ASSERT* here because this runs in a subprocess,502    // and will show up as un-executed in the parent.503    assert(llvm::all_of(ThreadsUsed,504                        [](auto &T) { return T.getData().front() < 16UL; }) &&505           "Threads ran on more CPUs than expected! The affinity mask does not "506           "seem to work.");507    return;508  }509  std::string Executable =510      sys::fs::getMainExecutable(TestMainArgv0, &ThreadPoolTestStringArg1);511  const auto *TestInfo = testing::UnitTest::GetInstance()->current_test_info();512  std::string Arg = (Twine("--gtest_filter=") + TestInfo->test_suite_name() +513                     "." + TestInfo->name())514                        .str();515  StringRef argv[] = {Executable, Arg};516 517  // Add environment variable to the environment of the child process.518  int Res = setenv("LLVM_THREADPOOL_AFFINITYMASK", "1", false);519  ASSERT_EQ(0, Res);520 521  std::string Error;522  bool ExecutionFailed;523  BitVector Affinity;524  Affinity.resize(4);525  Affinity.set(0, 4); // Use CPUs 0,1,2,3.526  int Ret = sys::ExecuteAndWait(Executable, argv, {}, {}, 0, 0, &Error,527                                &ExecutionFailed, nullptr, &Affinity);528  Res = setenv("LLVM_THREADPOOL_AFFINITYMASK", "", false);529  ASSERT_EQ(0, Res);530  ASSERT_EQ(0, Ret);531}532 533#endif // #ifdef _WIN32534#endif // #if LLVM_ENABLE_THREADS == 1535