簡單的執行緒池

cnblogs 發表於 2021-11-25

概要

此執行緒池擁有一個被所有工作執行緒共享的任務佇列。執行緒池使用者提交的任務,被執行緒池儲存在任務佇列中,工作執行緒從任務佇列中獲取任務並執行。

gist

任務是可擁有返回值的、無引數的可呼叫(callable)物件,或者是經 std::bind 繫結了可呼叫物件及其引數後的呼叫包裝器。具體而言可以是

  • 自由函式(也稱為全域性函式)
  • lambda
  • 函式物件(也稱為函式符)
  • 類成員函式
  • 包裝了上述型別的 std::function
  • bind 呼叫包裝器

該執行緒池非同步地執行任務。當任務被提交進執行緒池後,使用者不必等待任務執行和返回結果。

實現

以下程式碼給出了此執行緒池的實現。

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
    }

};

我們從構造 Thread_Pool 物件(#1)開始瞭解這個執行緒池。atomic<bool> 資料成員用於標誌執行緒池是否結束,並強制同步記憶體順序(#2);Task_Wrapper 具體化了執行緒安全的任務佇列 Lockwise_Queue<>(#3);thread* 用於引用所有的工作執行緒物件(#4)。Task_Wrapper 和 Lockwise_Queue<> 稍後再做說明。

執行緒池通過 thread::hardware_concurrency() 獲取當前硬體支援的併發執行緒數量(#5),並依據此數量建立出工作執行緒。Thread_Pool 物件的成員函式 work() 作為所有工作執行緒的初始函式(#6),這使得執行緒池中的任務佇列能被所有工作執行緒共享。建立 thread 物件和 new 操作可能失敗並引發異常,因此用 try-catch 捕獲潛在的異常。處理異常過程中,需要標誌執行緒池結束,保證任何建立的執行緒都能正常的停止,並回收記憶體資源(#7)。執行緒池物件析構時的工作與此一致。

Thread_Pool 物件構建完成後,任務通過 Thread_Pool::submit<>() 被提交進入執行緒池(#8)。為了支援任務的非同步執行,任務先被封裝在 std::packaged_task<> 中,再被放入執行緒安全的任務佇列(#9)。任務執行結果被封裝在返回的 std::future<> 物件中(#10),允許使用者在未來需要結果時,等待任務結束並獲取結果。


因為每一個任務都是一個特定型別的 std::packaged_task<> 物件,為了實現任務佇列的泛型化,需要設計一個通用的資料結構 Task_Wrapper,用於封裝特定型別的 std::packaged_task<> 物件。

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<> 的例項只是可移動的,而不可複製。Task_Wrapper 必須能移動封裝 std::packaged_task<R()> 物件(#1)。為了保持一致性,Task_Wrapper 也實現了移動構造(#2)和移動賦值(#3),同時實現了 operator()(#4)。ABC 的繼承結構(#5)用於支援泛型化地封裝和呼叫 std::packaged_task<> 物件。std::packaged_task<> 封裝在派生類 Task<> 中(#6),由指向非泛型的抽象基類 Task_Base 的指標引用派生類物件(#7)。對 Task_Wrapper 物件的呼叫由虛呼叫(#8)委託給派生類並執行實際的任務(#9)。


另一個關鍵的資料結構是執行緒安全的任務佇列 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 物件儲存在 std::queue<> 中(#1)。互斥元和條件變數控制工作執行緒對任務佇列的併發訪問(#2)。為了提高併發程度,採用非阻塞自旋鎖作為互斥元(#3)。任務的入隊和出隊操作,分別由支援移動語義的 push 函式(#4) 和 pop 函式(#5)完成。

驗證

為了驗證此執行緒池滿足概要中描述的能力,設計瞭如下的各類可呼叫物件。

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;
    }

};

對這些函式做好必要的引數封裝,將其提交給執行緒池,

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();
    }
});

編譯程式碼 g++ -std=c++11 a_simple_thread_pool.cpp 成功後執行 ./a.out。以下是執行過程中的部分輸出,

...

[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...

...

最後

完整示例請參考 [github] a_simple_thread_pool

作者參考了 C++併發程式設計實戰 / (美)威廉姆斯 (Williams, A.) 著; 周全等譯. - 北京: 人民郵電出版社, 2015.6 (2016.4重印) 一書中的部分設計思路。藉此機會對 Anthony Williams 及周全等譯者表示感謝。