概要
此线程池拥有一个被所有工作线程共享的任务队列。线程池用户提交的任务,被线程池保存在任务队列中,工作线程从任务队列中获取任务并执行。
任务是可拥有返回值的、无参数的可调用(callable)对象,或者是经 std::bind 绑定了可调用对象及其参数后的调用包装器。具体而言可以是
- 自由函数(也称为全局函数)
- lambda
- 函数对象(也称为函数符)
- 类成员函数
- 包装了上述类型的 std::function
- bind 调用包装器
该线程池异步地执行任务。当任务被提交进线程池后,用户不必等待任务执行和返回结果。
实现
以下代码给出了此线程池的实现。
class Thread_Pool {
private:
struct Task_Wrapper { ...
};
atomic<bool> _done_; // #2
Lockwise_Queue<Task_Wrapper> _queue_; // #3
unsigned _workersize_;
thread* _workers_; // #4
void work() {
while (!_done_.load(memory_order_acquire)) {
Task_Wrapper task;
if (_queue_.pop(task))
task();
else
std::this_thread::yield();
}
}
public:
Thread_Pool() : _done_(false) { // #1
try {
_workersize_ = thread::hardware_concurrency(); // #5
_workers_ = new thread[_workersize_];
for (unsigned i = 0; i < _workersize_; ++i) {
_workers_[i] = thread(&Thread_Pool::work, this); // #6
}
} catch (...) { // #7
_done_.store(true, memory_order_release);
for (unsigned i = 0; i < _workersize_; ++i) {
if (_workers_[i].joinable())
_workers_[i].join();
}
delete[] _workers_;
throw;
}
}
~Thread_Pool() {
_done_.store(true, memory_order_release);
for (unsigned i = 0; i < _workersize_; ++i) {
if (_workers_[i].joinable())
_workers_[i].join();
}
delete[] _workers_;
}
template<class Callable>
future<typename std::result_of<Callable()>::type> submit(Callable c) { // #8
typedef typename std::result_of<Callable()>::type R;
packaged_task<R()> task(c);
future<R> r = task.get_future();
_queue_.push(std::move(task)); // #9
return r; // #10
}
};
我们从构造 Thread_Pool 对象(#1)开始了解这个线程池。atomic<bool> 数据成员用于标志线程池是否结束,并强制同步内存顺序(#2);Task_Wrapper 具体化了线程安全的任务队列 Lockwise_Queue<>(#3);thread* 用于引用所有的工作线程对象(#4)。Task_Wrapper 和 Lockwise_Queue<> 稍后再做说明。
线程池通过 thread::hardware_concurrency() 获取当前硬件支持的并发线程数量(#5),并依据此数量创建出工作线程。Thread_Pool 对象的成员函数 work() 作为所有工作线程的初始函数(#6),这使得线程池中的任务队列能被所有工作线程共享。创建 thread 对象和 new 操作可能失败并引发异常,因此用 try-catch 捕获潜在的异常。处理异常过程中,需要标志线程池结束,保证任何创建的线程都能正常的停止,并回收内存资源(#7)。线程池对象析构时的工作与此一致。
Thread_Pool 对象构建完成后,任务通过 Thread_Pool::submit<>() 被提交进入线程池(#8)。为了支持任务的异步执行,任务先被封装在 std::packaged_task<> 中,再被放入线程安全的任务队列(#9)。任务执行结果被封装在返回的 std::future<> 对象中(#10),允许用户在未来需要结果时,等待任务结束并获取结果。
因为每一个任务都是一个特定类型的 std::packaged_task<> 对象,为了实现任务队列的泛型化,需要设计一个通用的数据结构 Task_Wrapper,用于封装特定类型的 std::packaged_task<> 对象。
struct Task_Wrapper {
struct Task_Base {
virtual ~Task_Base() {}
virtual void call() = 0;
};
template<class T>
struct Task : Task_Base { // #5
T _t_;
Task(T&& t) : _t_(std::move(t)) {} // #6
void call() { _t_(); } // #9
};
Task_Base* _ptr_; // #7
Task_Wrapper() : _ptr_(nullptr) {};
template<class T>
Task_Wrapper(T&& t) : _ptr_(new Task<T>(std::move(t))) {} // #1
// support move
Task_Wrapper(Task_Wrapper&& other) { // #2
_ptr_ = other._ptr_;
other._ptr_ = nullptr;
}
Task_Wrapper& operator=(Task_Wrapper&& other) { // #3
_ptr_ = other._ptr_;
other._ptr_ = nullptr;
return *this;
}
// no copy
Task_Wrapper(Task_Wrapper&) = delete;
Task_Wrapper& operator=(Task_Wrapper&) = delete;
~Task_Wrapper() {
if (_ptr_) delete _ptr_;
}
void operator()() const { // #4
_ptr_->call(); // #8
}
};
std::packaged_task<> 的实例只是可移动的,而不可复制。Task_Wrapper 必须能移动封装 std::packaged_task<R()> 对象(#1)。为了保持一致性,Task_Wrapper 也实现了移动构造(#2)和移动赋值(#3),同时实现了 operator()(#4)。ABC 的继承结构(#5)用于支持泛型化地封装和调用 std::packaged_task<> 对象。std::packaged_task<> 封装在派生类 Task<> 中(#6),由指向非泛型的抽象基类 Task_Base 的指针引用派生类对象(#7)。对 Task_Wrapper 对象的调用由虚调用(#8)委托给派生类并执行实际的任务(#9)。
另一个关键的数据结构是线程安全的任务队列 Lockwise_Queue<>。
template<class T>
class Lockwise_Queue {
private:
struct Spinlock_Mutex { // #3
atomic_flag _af_;
Spinlock_Mutex() : _af_(false) {}
void lock() {
while (_af_.test_and_set(memory_order_acquire));
}
void unlock() {
_af_.clear(memory_order_release);
}
} mutable _m_; // #2
condition_variable _cv_;
queue<T> _q_; // #1
public:
Lockwise_Queue() {}
void push(const T& element) {
lock_guard<Spinlock_Mutex> lk(_m_);
_q_.push(std::move(element));
_cv_.notify_one();
}
void push(T&& element) { // #4
lock_guard<Spinlock_Mutex> lk(_m_);
_q_.push(std::move(element));
_cv_.notify_one();
}
bool pop(T& element) { // #5
lock_guard<Spinlock_Mutex> lk(_m_);
if (_q_.empty())
return false;
element = std::move(_q_.front());
_q_.pop();
return true;
}
bool empty() const {
lock_guard<Spinlock_Mutex> lk(_m_);
return _q_.empty();
}
};
所有 Task_Wrapper 对象保存在 std::queue<> 中(#1)。互斥元和条件变量控制工作线程对任务队列的并发访问(#2)。为了提高并发程度,采用非阻塞自旋锁作为互斥元(#3)。任务的入队和出队操作,分别由支持移动语义的 push 函数(#4) 和 pop 函数(#5)完成。
验证
为了验证此线程池满足概要中描述的能力,设计了如下的各类可调用对象。
void shoot() {
std::printf("\n\t[Free Function] Let an arrow fly...\n");
}
bool shoot(long n) {
std::printf("\n\t[Free Function] Let %ld arrows fly...\n", n);
return false;
}
auto shootAnarrow = [] {
std::printf("\n\t[Lambda] Let an arrow fly...\n");
};
auto shootNarrows = [](long n) -> bool {
std::printf("\n\t[Lambda] Let %ld arrows fly...\n", n);
return true;
};
class Archer {
public:
void operator()() {
std::printf("\n\t[Functor] Let an arrow fly...\n");
}
bool operator()(long n) {
std::printf("\n\t[Functor] Let %ld arrows fly...\n", n);
return false;
}
void shoot() {
std::printf("\n\t[Member Function] Let an arrow fly...\n");
}
bool shoot(long n) {
std::printf("\n\t[Member Function] Let %ld arrows fly...\n", n);
return true;
}
};
对这些函数做好必要的参数封装,将其提交给线程池,
atomic<bool> go(false);
time_point<steady_clock> start = steady_clock::now();
minutes PERIOD(1);
Thread_Pool pool;
thread t1([&go, &pool, &PERIOD, start] { // test free function of void()
while (!go.load(memory_order_acquire))
std::this_thread::yield();
void (*task)() = shoot;
for (long x = 0; steady_clock::now() - start <= PERIOD; ++x) {
pool.submit(task);
//pool.submit(std::bind<void(*)()>(shoot));
std::this_thread::yield();
}
});
thread t2([&go, &pool, &PERIOD, start] { // test free function of bool(long)
while (!go.load(memory_order_acquire))
std::this_thread::yield();
bool (*task)(long) = shoot;
for (long x = 2; steady_clock::now() - start <= PERIOD; ++x) {
future<bool> r = pool.submit(std::bind(task, x));
//future<bool> r = pool.submit(std::bind<bool(*)(long)>(shoot, x));
std::this_thread::yield();
}
});
thread t3([&go, &pool, &PERIOD, start] { // test lambda of void()
while (!go.load(memory_order_acquire))
std::this_thread::yield();
for (long x = 0; steady_clock::now() - start <= PERIOD; ++x) {
pool.submit(shootAnarrow);
std::this_thread::yield();
}
});
thread t4([&go, &pool, &PERIOD, start] { // test lambda of bool(long)
while (!go.load(memory_order_acquire))
std::this_thread::yield();
for (long x = 2; steady_clock::now() - start <= PERIOD; ++x) {
future<bool> r = pool.submit(std::bind(shootNarrows, x));
std::this_thread::yield();
}
});
thread t5([&go, &pool, &PERIOD, start] { // test functor of void()
while (!go.load(memory_order_acquire))
std::this_thread::yield();
Archer hoyt;
for (long x = 0; steady_clock::now() - start <= PERIOD; ++x) {
pool.submit(hoyt);
std::this_thread::yield();
}
});
thread t6([&go, &pool, &PERIOD, start] { // test functor of bool(long)
while (!go.load(memory_order_acquire))
std::this_thread::yield();
Archer hoyt;
for (long x = 2; steady_clock::now() - start <= PERIOD; ++x) {
future<bool> r = pool.submit(std::bind(hoyt, x));
std::this_thread::yield();
}
});
thread t7([&go, &pool, &PERIOD, start] { // test member function of void()
while (!go.load(memory_order_acquire))
std::this_thread::yield();
Archer hoyt;
for (long x = 0; steady_clock::now() - start <= PERIOD; ++x) {
pool.submit(std::bind<void(Archer::*)()>(&Archer::shoot, &hoyt));
//pool.submit(std::bind(static_cast<void(Archer::*)()>(&Archer::shoot), &hoyt));
std::this_thread::yield();
}
});
thread t8([&go, &pool, &PERIOD, start] { // test member function of bool(long)
while (!go.load(memory_order_acquire))
std::this_thread::yield();
Archer hoyt;
for (long x = 2; steady_clock::now() - start <= PERIOD; ++x) {
future<bool> r = pool.submit(std::bind<bool(Archer::*)(long)>(&Archer::shoot, &hoyt, x));
//future<bool> r = pool.submit(std::bind(static_cast<bool(Archer::*)(long)>(&Archer::shoot), &hoyt, x));
std::this_thread::yield();
}
});
thread t9([&go, &pool, &PERIOD, start] { // test std::function<> of void()
while (!go.load(memory_order_acquire))
std::this_thread::yield();
std::function<void()> task = static_cast<void(*)()>(shoot);
for (long x = 0; steady_clock::now() - start <= PERIOD; ++x) {
pool.submit(task);
std::this_thread::yield();
}
});
thread t10([&go, &pool, &PERIOD, start] { // test std::function<> of bool(long)
while (!go.load(memory_order_acquire))
std::this_thread::yield();
std::function<bool(long)> task = static_cast<bool(*)(long)>(shoot);
for (long x = 2; steady_clock::now() - start <= PERIOD; ++x) {
future<bool> r = pool.submit(std::bind(task, x));
std::this_thread::yield();
}
});
编译代码 g++ -std=c++11 a_simple_thread_pool.cpp
成功后执行 ./a.out
。以下是执行过程中的部分输出,
...
[Functor] Let an arrow fly...
[Free Function] Let 9224 arrows fly...
[Free Function] Let 9445 arrows fly...
[Member Function] Let 9375 arrows fly...
[Lambda] Let 9449 arrows fly...
[Free Function] Let an arrow fly...
[Lambda] Let an arrow fly...
[Member Function] Let an arrow fly...
[Functor] Let 9469 arrows fly...
...
最后
完整示例请参考 [github] a_simple_thread_pool 。
作者参考了 C++并发编程实战 / (美)威廉姆斯 (Williams, A.) 著; 周全等译. - 北京: 人民邮电出版社, 2015.6 (2016.4重印) 一书中的部分设计思路。借此机会对 Anthony Williams 及周全等译者表示感谢。