std::thread

• 每个程序有一个执行 main() 函数的主线程，将函数添加为 std::thread 的参数即可启动另一个线程，两个线程会同时运行

#include <iostream>
#include <thread>

void f() { std::cout << "hello world"; }

int main() {
  std::thread t{f};
  t.join();  // 等待新起的线程退出
}

• std::thread 的参数也可以是函数对象或者 lambda

#include <iostream>
#include <thread>

struct A {
  void operator()() const { std::cout << 1; }
};

int main() {
  A a;
  std::thread t1(a);  // 会调用 A 的拷贝构造函数
  std::thread t2(A());  // most vexing parse，声明名为 t2 参数类型为 A 的函数
  std::thread t3{A()};
  std::thread t4((A()));
  std::thread t5{[] { std::cout << 1; }};
  t1.join();
  t3.join();
  t4.join();
  t5.join();
}

• 在线程销毁前要对其调用 join 等待线程退出或 detach 将线程分离，否则 std::thread 的析构函数会调用 std::terminate 终止程序，注意分离线程可能出现空悬引用的隐患

#include <iostream>
#include <thread>

class A {
 public:
  A(int& x) : x_(x) {}

  void operator()() const {
    for (int i = 0; i < 1000000; ++i) {
      call(x_);  // 存在对象析构后引用空悬的隐患
    }
  }

 private:
  void call(int& x) {}

 private:
  int& x_;
};

void f() {
  int x = 0;
  A a{x};
  std::thread t{a};
  t.detach();  // 不等待 t 结束
}  // 函数结束后 t 可能还在运行，而 x 已经销毁，a.x_ 为空悬引用

int main() {
  std::thread t{f};  // 导致空悬引用
  t.join();
}

• join 会在线程结束后清理 std::thread，使其与完成的线程不再关联，因此对一个线程只能进行一次 join

#include <thread>

int main() {
  std::thread t([] {});
  t.join();
  t.join();  // 错误
}

• 如果线程运行过程中发生异常，之后的 join 会被忽略，为此需要捕获异常，并在抛出异常前 join

#include <thread>

int main() {
  std::thread t([] {});
  try {
    throw 0;
  } catch (int x) {
    t.join();  // 处理异常前先 join()
    throw x;   // 再将异常抛出
  }
  t.join();  // 之前抛异常，不会执行到此处
}

• C++20 提供了 std::jthread，它会在析构函数中对线程 join

#include <thread>

int main() {
  std::jthread t([] {});
}

• detach 分离线程会让线程在后台运行，一般将这种在后台运行的线程称为守护线程，守护线程与主线程无法直接交互，也不能被 join

std::thread t([] {});
t.detach();
assert(!t.joinable());

• 创建守护线程一般是为了长时间运行，比如有一个文档处理应用，为了同时编辑多个文档，每次新开一个文档，就可以开一个对应的守护线程

void edit_document(const std::string& filename) {
  open_document_and_display_gui(filename);
  while (!done_editing()) {
    user_command cmd = get_user_input();
    if (cmd.type == open_new_document) {
      const std::string new_name = get_filename_from_user();
      std::thread t(edit_document, new_name);
      t.detach();
    } else {
      process_user_input(cmd);
    }
  }
}

为带参数的函数创建线程

• 有参数的函数也能传给 std::thread，参数的默认实参会被忽略

#include <thread>

void f(int i = 1) {}

int main() {
  std::thread t{f, 42};  // std::thread t{f} 则会出错，因为默认实参会被忽略
  t.join();
}

• 参数的引用类型也会被忽略，为此要使用 std::ref

#include <cassert>
#include <thread>

void f(int& i) { ++i; }

int main() {
  int i = 1;
  std::thread t{f, std::ref(i)};
  t.join();
  assert(i == 2);
}

• 如果对一个实例的 non-static 成员函数创建线程，第一个参数类型为成员函数指针，第二个参数类型为实例指针，后续参数为函数参数

#include <iostream>
#include <thread>

class A {
 public:
  void f(int i) { std::cout << i; }
};

int main() {
  A a;
  std::thread t1{&A::f, &a, 42};  // 调用 a->f(42)
  std::thread t2{&A::f, a, 42};   // 拷贝构造 tmp_a，再调用 tmp_a.f(42)
  t1.join();
  t2.join();
}

• 如果要为参数是 move-only 类型的函数创建线程，则需要使用 std::move 传入参数

#include <iostream>
#include <thread>
#include <utility>

void f(std::unique_ptr<int> p) { std::cout << *p; }

int main() {
  std::unique_ptr<int> p(new int(42));
  std::thread t{f, std::move(p)};
  t.join();
}

转移线程所有权

• std::thread 是 move-only 类型，不能拷贝，只能通过移动转移所有权，但不能转移所有权到 joinable 的线程

#include <thread>
#include <utility>

void f() {}
void g() {}

int main() {
  std::thread a{f};
  std::thread b = std::move(a);
  assert(!a.joinable());
  assert(b.joinable());
  a = std::thread{g};
  assert(a.joinable());
  assert(b.joinable());
  // a = std::move(b);  // 错误，不能转移所有权到 joinable 的线程
  a.join();
  a = std::move(b);
  assert(a.joinable());
  assert(!b.joinable());
  a.join();
}

• 移动操作同样适用于支持移动的容器

#include <algorithm>
#include <thread>
#include <vector>

int main() {
  std::vector<std::thread> v;
  for (int i = 0; i < 10; ++i) {
    v.emplace_back([] {});
  }
  std::for_each(std::begin(v), std::end(v), std::mem_fn(&std::thread::join));
}

• std::thread 可以作为函数返回值

#include <thread>

std::thread f() {
  return std::thread{[] {}};
}

int main() {
  std::thread t{f()};
  t.join();
}

• std::thread 也可以作为函数参数

#include <thread>
#include <utility>

void f(std::thread t) { t.join(); }

int main() {
  f(std::thread([] {}));
  std::thread t([] {});
  f(std::move(t));
}

• 实现一个可以直接用 std::thread 构造的自动清理线程的类

#include <stdexcept>
#include <thread>
#include <utility>

class scoped_thread {
 public:
  explicit scoped_thread(std::thread x) : t_(std::move(x)) {
    if (!t_.joinable()) {
      throw std::logic_error("no thread");
    }
  }
  ~scoped_thread() { t_.join(); }
  scoped_thread(const scoped_thread&) = delete;
  scoped_thread& operator=(const scoped_thread&) = delete;

 private:
  std::thread t_;
};

int main() {
  scoped_thread t{std::thread{[] {}}};
}

• 类似 std::jthread 的类

#include <thread>

class Jthread {
 public:
  Jthread() noexcept = default;

  template <typename T, typename... Ts>
  explicit Jthread(T&& f, Ts&&... args)
      : t_(std::forward<T>(f), std::forward<Ts>(args)...) {}

  explicit Jthread(std::thread x) noexcept : t_(std::move(x)) {}
  Jthread(Jthread&& rhs) noexcept : t_(std::move(rhs.t_)) {}

  Jthread& operator=(Jthread&& rhs) noexcept {
    if (joinable()) {
      join();
    }
    t_ = std::move(rhs.t_);
    return *this;
  }

  Jthread& operator=(std::thread t) noexcept {
    if (joinable()) {
      join();
    }
    t_ = std::move(t);
    return *this;
  }

  ~Jthread() noexcept {
    if (joinable()) {
      join();
    }
  }

  void swap(Jthread&& rhs) noexcept { t_.swap(rhs.t_); }
  std::thread::id get_id() const noexcept { return t_.get_id(); }
  bool joinable() const noexcept { return t_.joinable(); }
  void join() { t_.join(); }
  void detach() { t_.detach(); }
  std::thread& as_thread() noexcept { return t_; }
  const std::thread& as_thread() const noexcept { return t_; }

 private:
  std::thread t_;
};

int main() {
  Jthread t{[] {}};
}

查看硬件支持的线程数量

• hardware_concurrency 会返回硬件支持的并发线程数

#include <iostream>
#include <thread>

int main() {
  unsigned int n = std::thread::hardware_concurrency();
  std::cout << n << " concurrent threads are supported.\n";
}

• 并行版本的 std::accumulate

#include <algorithm>
#include <cassert>
#include <functional>
#include <iterator>
#include <numeric>
#include <thread>
#include <vector>

template <typename Iterator, typename T>
struct accumulate_block {
  void operator()(Iterator first, Iterator last, T& res) {
    res = std::accumulate(first, last, res);
  }
};

template <typename Iterator, typename T>
T parallel_accumulate(Iterator first, Iterator last, T init) {
  long len = std::distance(first, last);
  if (!len) {
    return init;
  }
  long min_per_thread = 25;
  long max_threads = (len + min_per_thread - 1) / min_per_thread;
  long hardware_threads = std::thread::hardware_concurrency();
  long num_threads =  // 线程数量
      std::min(hardware_threads == 0 ? 2 : hardware_threads, max_threads);
  long block_size = len / num_threads;  // 每个线程中的数据量
  std::vector<T> res(num_threads);
  std::vector<std::thread> threads(num_threads - 1);  // 已有主线程故少一个线程
  Iterator block_start = first;
  for (long i = 0; i < num_threads - 1; ++i) {
    Iterator block_end = block_start;
    std::advance(block_end, block_size);  // block_end 指向当前块尾部
    threads[i] = std::thread{accumulate_block<Iterator, T>{}, block_start,
                             block_end, std::ref(res[i])};
    block_start = block_end;
  }
  accumulate_block<Iterator, T>()(block_start, last, res[num_threads - 1]);
  std::for_each(threads.begin(), threads.end(),
                std::mem_fn(&std::thread::join));
  return std::accumulate(res.begin(), res.end(), init);
}

int main() {
  std::vector<long> v(1000000);
  std::iota(std::begin(v), std::end(v), 0);
  long res = std::accumulate(std::begin(v), std::end(v), 0);
  assert(parallel_accumulate(std::begin(v), std::end(v), 0) == res);
}

线程号

• 可以通过对线程实例调用成员函数 get_id 或在当前线程中调用 std::this_thread::get_id 获取 线程号，其本质是一个无符号整型的封装，允许拷贝和比较，因此可以将其作为容器的键值，如果两个线程的线程号相等，则两者是同一线程或都是空线程（一般空线程的线程号为 0）

#include <iostream>
#include <thread>

#ifdef _WIN32
#include <Windows.h>
#elif defined __GNUC__
#include <syscall.h>
#include <unistd.h>

#endif

void print_current_thread_id() {
#ifdef _WIN32
  std::cout << std::this_thread::get_id() << std::endl;       // 19840
  std::cout << GetCurrentThreadId() << std::endl;             // 19840
  std::cout << GetThreadId(GetCurrentThread()) << std::endl;  // 19840
#elif defined __GNUC__
  std::cout << std::this_thread::get_id() << std::endl;  // 1
  std::cout << pthread_self() << std::endl;              // 140250646003520
  std::cout << getpid() << std::endl;  // 1502109，ps aux 显示此 pid
  std::cout << syscall(SYS_gettid) << std::endl;  // 1502109
#endif
}

std::thread::id master_thread_id = std::this_thread::get_id();

void f() {
  if (std::this_thread::get_id() == master_thread_id) {
    // do_master_thread_work();
  }
  // do_common_work();
}

int main() {
  print_current_thread_id();
  f();
  std::thread t{f};
  t.join();
}

CPU 亲和性（affinity）

• 将线程绑定到一个指定的 CPU core 上运行，避免多核 CPU 上下文切换和 cache miss 的开销

#ifdef _WIN32
#include <Windows.h>
#elif defined __GNUC__
#include <pthread.h>
#include <sched.h>
#include <string.h>
#endif

#include <iostream>
#include <thread>

void affinity_cpu(std::thread::native_handle_type t, int cpu_id) {
#ifdef _WIN32
  if (!SetThreadAffinityMask(t, 1ll << cpu_id)) {
    std::cerr << "fail to affinity" << GetLastError() << std::endl;
  }
#elif defined __GNUC__
  cpu_set_t cpu_set;
  CPU_ZERO(&cpu_set);
  CPU_SET(cpu_id, &cpu_set);
  int res = pthread_setaffinity_np(t, sizeof(cpu_set), &cpu_set);
  if (res != 0) {
    errno = res;
    std::cerr << "fail to affinity" << strerror(errno) << std::endl;
  }
#endif
}

void affinity_cpu_on_current_thread(int cpu_id) {
#ifdef _WIN32
  if (!SetThreadAffinityMask(GetCurrentThread(), 1ll << cpu_id)) {
    std::cerr << "fail to affinity" << GetLastError() << std::endl;
  }
#elif defined __GNUC__
  cpu_set_t cpu_set;
  CPU_ZERO(&cpu_set);
  CPU_SET(cpu_id, &cpu_set);
  int res = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set), &cpu_set);
  if (res != 0) {
    errno = res;
    std::cerr << "fail to affinity" << strerror(errno) << std::endl;
  }
#endif
}

void f() { affinity_cpu_on_current_thread(0); }

int main() {
  std::thread t1{[] {}};
  affinity_cpu(t1.native_handle(), 1);
  std::thread t2{f};
  t1.join();
  t2.join();
}