C++11 并发指南六( atomic 类型详解二 std::atomic )
C++11 并發指南六(atomic 類型詳解一 atomic_flag 介紹)? 一文介紹了 C++11 中最簡單的原子類型 std::atomic_flag,但是 std::atomic_flag 過于簡單,只提供了 test_and_set 和 clear 兩個 API,不能滿足其他需求(如 store, load, exchange, compare_exchange 等),因此本文將介紹功能更加完善的 std::atomic 類。
std::atomic 基本介紹
std::atomic 是模板類,一個模板類型為 T 的原子對象中封裝了一個類型為 T 的值。
template <class T> struct atomic;原子類型對象的主要特點就是從不同線程訪問不會導致數據競爭(data race)。因此從不同線程訪問某個原子對象是良性 (well-defined) 行為,而通常對于非原子類型而言,并發訪問某個對象(如果不做任何同步操作)會導致未定義 (undifined) 行為發生。
C++11 標準中的基本 std::atomic 模板定義如下:
template < class T > struct atomic {bool is_lock_free() const volatile;bool is_lock_free() const;void store(T, memory_order = memory_order_seq_cst) volatile;void store(T, memory_order = memory_order_seq_cst);T load(memory_order = memory_order_seq_cst) const volatile;T load(memory_order = memory_order_seq_cst) const;operator T() const volatile;operator T() const;T exchange(T, memory_order = memory_order_seq_cst) volatile;T exchange(T, memory_order = memory_order_seq_cst);bool compare_exchange_weak(T &, T, memory_order, memory_order) volatile;bool compare_exchange_weak(T &, T, memory_order, memory_order);bool compare_exchange_strong(T &, T, memory_order, memory_order) volatile;bool compare_exchange_strong(T &, T, memory_order, memory_order);bool compare_exchange_weak(T &, T, memory_order = memory_order_seq_cst) volatile;bool compare_exchange_weak(T &, T, memory_order = memory_order_seq_cst);bool compare_exchange_strong(T &, T, memory_order = memory_order_seq_cst) volatile;bool compare_exchange_strong(T &, T, memory_order = memory_order_seq_cst);atomic() = default;constexpr atomic(T);atomic(const atomic &) = delete;atomic & operator=(const atomic &) = delete;atomic & operator=(const atomic &) volatile = delete;T operator=(T) volatile;T operator=(T); };另外,C++11 標準庫 std::atomic 提供了針對整形(integral)和指針類型的特化實現,分別定義如下:
針對整形(integal)的特化,其中 integal 代表了如下類型char, signed char, unsigned char, short, unsigned short, int, unsigned int, long, unsigned long, long long, unsigned long long, char16_t, char32_t, wchar_t:
?| template<> struct atomic<integral> { ????boolis_lock_free() constvolatile; ????boolis_lock_free() const; ????voidstore(integral, memory_order = memory_order_seq_cst) volatile; ????voidstore(integral, memory_order = memory_order_seq_cst); ????integral load(memory_order = memory_order_seq_cst)const volatile; ????integral load(memory_order = memory_order_seq_cst)const; ????operator integral()const volatile; ????operator integral()const; ????integral exchange(integral, memory_order = memory_order_seq_cst)volatile; ????integral exchange(integral, memory_order = memory_order_seq_cst); ????boolcompare_exchange_weak(integral&, integral, memory_order, memory_order)volatile; ????boolcompare_exchange_weak(integral&, integral, memory_order, memory_order); ????boolcompare_exchange_strong(integral&, integral, memory_order, memory_order)volatile; ????boolcompare_exchange_strong(integral&, integral, memory_order, memory_order); ????boolcompare_exchange_weak(integral&, integral, memory_order = memory_order_seq_cst)volatile; ????boolcompare_exchange_weak(integral&, integral, memory_order = memory_order_seq_cst); ????boolcompare_exchange_strong(integral&, integral, memory_order = memory_order_seq_cst)volatile; ????boolcompare_exchange_strong(integral&, integral, memory_order = memory_order_seq_cst); ????integral fetch_add(integral, memory_order = memory_order_seq_cst)volatile; ????integral fetch_add(integral, memory_order = memory_order_seq_cst); ????integral fetch_sub(integral, memory_order = memory_order_seq_cst)volatile; ????integral fetch_sub(integral, memory_order = memory_order_seq_cst); ????integral fetch_and(integral, memory_order = memory_order_seq_cst)volatile; ????integral fetch_and(integral, memory_order = memory_order_seq_cst); ????integral fetch_or(integral, memory_order = memory_order_seq_cst)volatile; ????integral fetch_or(integral, memory_order = memory_order_seq_cst); ????integral fetch_xor(integral, memory_order = memory_order_seq_cst)volatile; ????integral fetch_xor(integral, memory_order = memory_order_seq_cst); ????? ????atomic() =default; ????constexpratomic(integral); ????atomic(constatomic&) = delete; ????atomic& operator=(constatomic&) = delete; ????atomic& operator=(constatomic&) volatile= delete; ????? ????integral operator=(integral)volatile; ????integral operator=(integral); ????? ????integral operator++(int)volatile; ????integral operator++(int); ????integral operator--(int)volatile; ????integral operator--(int); ????integral operator++()volatile; ????integral operator++(); ????integral operator--()volatile; ????integral operator--(); ????integral operator+=(integral)volatile; ????integral operator+=(integral); ????integral operator-=(integral)volatile; ????integral operator-=(integral); ????integral operator&=(integral)volatile; ????integral operator&=(integral); ????integral operator|=(integral)volatile; ????integral operator|=(integral); ????integral operator^=(integral)volatile; ????integral operator^=(integral); }; |
針對指針的特化:
?| template<class T> struct atomic<T*> { ????boolis_lock_free() constvolatile; ????boolis_lock_free() const; ????voidstore(T*, memory_order = memory_order_seq_cst) volatile; ????voidstore(T*, memory_order = memory_order_seq_cst); ????T* load(memory_order = memory_order_seq_cst)const volatile; ????T* load(memory_order = memory_order_seq_cst)const; ????operator T*()const volatile; ????operator T*()const; ????T* exchange(T*, memory_order = memory_order_seq_cst)volatile; ????T* exchange(T*, memory_order = memory_order_seq_cst); ????boolcompare_exchange_weak(T*&, T*, memory_order, memory_order)volatile; ????boolcompare_exchange_weak(T*&, T*, memory_order, memory_order); ????boolcompare_exchange_strong(T*&, T*, memory_order, memory_order)volatile; ????boolcompare_exchange_strong(T*&, T*, memory_order, memory_order); ????boolcompare_exchange_weak(T*&, T*, memory_order = memory_order_seq_cst)volatile; ????boolcompare_exchange_weak(T*&, T*, memory_order = memory_order_seq_cst); ????boolcompare_exchange_strong(T*&, T*, memory_order = memory_order_seq_cst)volatile; ????boolcompare_exchange_strong(T*&, T*, memory_order = memory_order_seq_cst); ????T* fetch_add(ptrdiff_t, memory_order = memory_order_seq_cst)volatile; ????T* fetch_add(ptrdiff_t, memory_order = memory_order_seq_cst); ????T* fetch_sub(ptrdiff_t, memory_order = memory_order_seq_cst)volatile; ????T* fetch_sub(ptrdiff_t, memory_order = memory_order_seq_cst); ????atomic() =default; ????constexpratomic(T*); ????atomic(constatomic&) = delete; ????atomic& operator=(constatomic&) = delete; ????atomic& operator=(constatomic&) volatile= delete; ????T* operator=(T*)volatile; ????T* operator=(T*); ????T* operator++(int)volatile; ????T* operator++(int); ????T* operator--(int)volatile; ????T* operator--(int); ????T* operator++()volatile; ????T* operator++(); ????T* operator--()volatile; ????T* operator--(); ????T* operator+=(ptrdiff_t)volatile; ????T* operator+=(ptrdiff_t); ????T* operator-=(ptrdiff_t)volatile; ????T* operator-=(ptrdiff_t); }; |
std::atomic 成員函數
?好了,對 std::atomic 有了一個最基本認識之后我們來看 std::atomic 的成員函數吧。
std::atomic 構造函數
std::atomic 的構造函數如下:
| atomic() noexcept = default; |
| constexpr atomic (T val) noexcept; |
| atomic (const atomic&) = delete; |
請看下例:
?| #include <iostream>?????? // std::cout #include <atomic>???????? // std::atomic, std::atomic_flag, ATOMIC_FLAG_INIT #include <thread>???????? // std::thread, std::this_thread::yield #include <vector>???????? // std::vector // 由 false 初始化一個 std::atomic<bool> 類型的原子變量 std::atomic<bool> ready(false); std::atomic_flag winner = ATOMIC_FLAG_INIT; voiddo_count1m(intid) { ????while(!ready) { std::this_thread::yield(); } // 等待 ready 變為 true. ????for(volatile int i=0; i<1000000; ++i) {} // 計數 ????if(!winner.test_and_set()) { ??????std::cout <<"thread #" << id <<" won!\n"; ????} } int main () { ????std::vector<std::thread> threads; ????std::cout <<"spawning 10 threads that count to 1 million...\n"; ????for(int i=1; i<=10; ++i) threads.push_back(std::thread(count1m,i)); ????ready =true; ????for(auto& th : threads) th.join(); ????return0; } |
std::atomic::operator=() 函數
std::atomic 的賦值操作函數定義如下:
| T operator= (T val) noexcept; T operator= (T val) volatile noexcept; |
| atomic& operator= (const atomic&) = delete; atomic& operator= (const atomic&) volatile = delete; |
可以看出,普通的賦值拷貝操作已經被禁用。但是一個類型為 T 的變量可以賦值給相應的原子類型變量(相當與隱式轉換),該操作是原子的,內存序(Memory Order) 默認為順序一致性(std::memory_order_seq_cst),如果需要指定其他的內存序,需使用 std::atomic::store()。
?| #include <iostream>???????????? // std::cout #include <atomic>?????????????? // std::atomic #include <thread>?????????????? // std::thread, std::this_thread::yield std::atomic <int> foo = 0; void set_foo(int x) { ????foo = x;// 調用 std::atomic::operator=(). } voidprint_foo() { ????while(foo == 0) { // wait while foo == 0 ????????std::this_thread::yield(); ????} ????std::cout <<"foo: " << foo << '\n'; } int main() { ????std::threadfirst(print_foo); ????std::threadsecond(set_foo, 10); ????first.join(); ????second.join(); ????return0; } |
基本 std::atomic 類型操作
本節主要介紹基本 std::atomic 類型所具備的操作(即成員函數)。我們知道 std::atomic 是模板類,一個模板類型為 T 的原子對象中封裝了一個類型為 T 的值。本文<std::atomic 基本介紹>一節中也提到了 std::atomic 類模板除了基本類型以外,還針對整形和指針類型做了特化。 特化的 std::atomic 類型支持更多的操作,如 fetch_add, fetch_sub, fetch_and 等。本小節介紹基本 std::atomic 類型所具備的操作:
is_lock_free ?| bool is_lock_free() const volatile noexcept; bool is_lock_free() const noexcept; |
| void store (T val, memory_order sync = memory_order_seq_cst) volatilenoexcept; void store (T val, memory_order sync = memory_order_seq_cst) noexcept; |
| memory_order_relaxed | Relaxed |
| memory_order_release | Release |
| memory_order_seq_cst | Sequentially consistent |
| #include <iostream>?????? // std::cout #include <atomic>???????? // std::atomic, std::memory_order_relaxed #include <thread>???????? // std::thread std::atomic<int> foo(0);// 全局的原子對象 foo void set_foo(int x) { ????foo.store(x, std::memory_order_relaxed);// 設置(store) 原子對象 foo 的值 } voidprint_foo() { ????intx; ????do{ ????????x = foo.load(std::memory_order_relaxed);// 讀取(load) 原子對象 foo 的值 ????}while (x == 0); ????std::cout <<"foo: " << x << '\n'; } int main () { ????std::threadfirst(print_foo); // 線程 first 打印 foo 的值 ????std::threadsecond(set_foo, 10); // 線程 second 設置 foo 的值 ????first.join(); ????second.join(); ????return0; } |
| T load (memory_order sync = memory_order_seq_cst)const volatile noexcept; T load (memory_order sync = memory_order_seq_cst)const noexcept; |
| memory_order_relaxed | Relaxed |
| memory_order_consume | Consume |
| memory_order_acquire | Acquire |
| memory_order_seq_cst | Sequentially consistent |
| #include <iostream>?????? // std::cout #include <atomic>???????? // std::atomic, std::memory_order_relaxed #include <thread>???????? // std::thread std::atomic<int> foo(0);// 全局的原子對象 foo void set_foo(int x) { ????foo.store(x, std::memory_order_relaxed);// 設置(store) 原子對象 foo 的值 } voidprint_foo() { ????intx; ????do{ ????????x = foo.load(std::memory_order_relaxed);// 讀取(load) 原子對象 foo 的值 ????}while (x == 0); ????std::cout <<"foo: " << x << '\n'; } int main () { ????std::threadfirst(print_foo); // 線程 first 打印 foo 的值 ????std::threadsecond(set_foo, 10); // 線程 second 設置 foo 的值 ????first.join(); ????second.join(); ????return0; } |
| operator T() const volatilenoexcept; operator T() const noexcept; |
| #include <iostream>?????? // std::cout #include <atomic>???????? // std::atomic #include <thread>???????? // std::thread, std::this_thread::yield std::atomic<int> foo = 0; std::atomic<int> bar = 0; void set_foo(int x) { ????foo = x; } voidcopy_foo_to_bar() { ????// 如果 foo == 0,則該線程 yield, ????// 在 foo == 0 時, 實際也是隱含了類型轉換操作, ????// 因此也包含了 operator T() const 的調用. ????while(foo == 0) std::this_thread::yield(); ????// 實際調用了 operator T() const, 將foo 強制轉換成 int 類型, ????// 然后調用 operator=(). ????bar =static_cast<int>(foo); } voidprint_bar() { ????// 如果 bar == 0,則該線程 yield, ????// 在 bar == 0 時, 實際也是隱含了類型轉換操作, ????// 因此也包含了 operator T() const 的調用. ????while(bar == 0) std::this_thread::yield(); ????std::cout <<"bar: " << bar << '\n'; } int main () { ????std::threadfirst(print_bar); ????std::threadsecond(set_foo, 10); ????std::threadthird(copy_foo_to_bar); ????first.join(); ????second.join(); ????third.join(); ????return0; } |
?
exchange T exchange (T val, memory_order sync = memory_order_seq_cst) volatile noexcept; T exchange (T val, memory_order sync = memory_order_seq_cst) noexcept; 讀取并修改被封裝的值,exchange 會將 val 指定的值替換掉之前該原子對象封裝的值,并返回之前該原子對象封裝的值,整個過程是原子的(因此exchange 操作也稱為read-modify-write 操作)。sync參數指定內存序(Memory Order),可能的取值如下:| memory_order_relaxed | Relaxed |
| memory_order_consume | Consume |
| memory_order_acquire | Acquire |
| memory_order_release | Release |
| memory_order_acq_rel | Acquire/Release |
| memory_order_seq_cst | Sequentially consistent |
請看下面例子,各個線程計數至 1M,首先完成計數任務的線程打印自己的 ID,
?| #include <iostream>?????? // std::cout #include <atomic>???????? // std::atomic #include <thread>???????? // std::thread #include <vector>???????? // std::vector std::atomic<bool> ready(false); std::atomic<bool> winner(false); void count1m (int id) { ????while(!ready) {}????????????????? // wait for the ready signal ????for(int i = 0; i < 1000000; ++i) {}?? // go!, count to 1 million ????if(!winner.exchange(true)) { std::cout <<"thread #" << id <<" won!\n"; } }; int main () { ????std::vector<std::thread> threads; ????std::cout <<"spawning 10 threads that count to 1 million...\n"; ????for(int i = 1; i <= 10; ++i) threads.push_back(std::thread(count1m,i)); ????ready =true; ????for(auto& th : threads) th.join(); ????return0; } |
| bool compare_exchange_weak (T& expected, T val,memory_order sync = memory_order_seq_cst) volatile noexcept; bool compare_exchange_weak (T& expected, T val,memory_order sync = memory_order_seq_cst) noexcept; |
| bool compare_exchange_weak (T& expected, T val,memory_order success, memory_order failure) volatile noexcept; bool compare_exchange_weak (T& expected, T val,memory_order success, memory_order failure) noexcept; |
- 相等,則用 val 替換原子對象的舊值。
- 不相等,則用原子對象的舊值替換 expected ,因此調用該函數之后,如果被該原子對象封裝的值與參數?expected 所指定的值不相等,expected 中的內容就是原子對象的舊值。
在第(2)種情況下,內存序(Memory Order)的選擇取決于比較操作結果,如果比較結果為 true(即原子對象的值等于 expected),則選擇參數 success 指定的內存序,否則選擇參數 failure 所指定的內存序。
注意,該函數直接比較原子對象所封裝的值與參數 expected 的物理內容,所以某些情況下,對象的比較操作在使用 operator==() 判斷時相等,但 compare_exchange_weak 判斷時卻可能失敗,因為對象底層的物理內容中可能存在位對齊或其他邏輯表示相同但是物理表示不同的值(比如 true 和 2 或 3,它們在邏輯上都表示"真",但在物理上兩者的表示并不相同)。
與compare_exchange_strong 不同, weak 版本的 compare-and-exchange 操作允許(spuriously 地)返回 false(即原子對象所封裝的值與參數expected 的物理內容相同,但卻仍然返回 false),不過在某些需要循環操作的算法下這是可以接受的,并且在一些平臺下?compare_exchange_weak 的性能更好 。如果 compare_exchange_weak 的判斷確實發生了偽失敗(spurious failures)——即使原子對象所封裝的值與參數expected 的物理內容相同,但判斷操作的結果卻為 false,compare_exchange_weak函數返回 false,并且參數 expected 的值不會改變。
對于某些不需要采用循環操作的算法而言, 通常采用compare_exchange_strong 更好。另外,該函數的內存序由? sync 參數指定,可選條件如下:
| memory_order_relaxed | Relaxed |
| memory_order_consume | Consume |
| memory_order_acquire | Acquire |
| memory_order_release | Release |
| memory_order_acq_rel | Acquire/Release |
| memory_order_seq_cst | Sequentially consistent |
| #include <iostream>?????? // std::cout #include <atomic>???????? // std::atomic #include <thread>???????? // std::thread #include <vector>???????? // std::vector // a simple global linked list: structNode { int value; Node* next; }; std::atomic<Node*> list_head(nullptr); voidappend(int val) { ????// append an element to the list ????Node* newNode =new Node{val, list_head}; ????// next is the same as: list_head = newNode, but in a thread-safe way: ????while(!list_head.compare_exchange_weak(newNode->next,newNode)) {} ????// (with newNode->next updated accordingly if some other thread just appended another node) } int main () { ????// spawn 10 threads to fill the linked list: ????std::vector<std::thread> threads; ????for(int i = 0; i < 10; ++i) threads.push_back(std::thread(append, i)); ????for(auto& th : threads) th.join(); ????// print contents: ????for(Node* it = list_head; it!=nullptr; it=it->next) ????????std::cout <<' ' << it->value; ????std::cout <<'\n'; ????// cleanup: ????Node* it;while (it=list_head) {list_head=it->next;delete it;} ????return0; } |
| 9 8 7 6 5 4 3 2 1 0 |
?
compare_exchange_strong| bool compare_exchange_strong (T& expected, T val,memory_order sync = memory_order_seq_cst) volatile noexcept; bool compare_exchange_strong (T& expected, T val,memory_order sync = memory_order_seq_cst) noexcept; |
| bool compare_exchange_strong (T& expected, T val,memory_order success, memory_order failure) volatile noexcept; bool compare_exchange_strong (T& expected, T val,memory_order success, memory_order failure) noexcept; |
- 相等,則用 val 替換原子對象的舊值。
- 不相等,則用原子對象的舊值替換 expected ,因此調用該函數之后,如果被該原子對象封裝的值與參數?expected 所指定的值不相等,expected 中的內容就是原子對象的舊值。
在第(2)種情況下,內存序(Memory Order)的選擇取決于比較操作結果,如果比較結果為 true(即原子對象的值等于 expected),則選擇參數 success 指定的內存序,否則選擇參數 failure 所指定的內存序。
注意,該函數直接比較原子對象所封裝的值與參數 expected 的物理內容,所以某些情況下,對象的比較操作在使用 operator==() 判斷時相等,但 compare_exchange_weak 判斷時卻可能失敗,因為對象底層的物理內容中可能存在位對齊或其他邏輯表示相同但是物理表示不同的值(比如 true 和 2 或 3,它們在邏輯上都表示"真",但在物理上兩者的表示并不相同)。 與compare_exchange_weak 不同, strong版本的 compare-and-exchange 操作不允許(spuriously 地)返回 false,即原子對象所封裝的值與參數expected 的物理內容相同,比較操作一定會為 true。不過在某些平臺下,如果算法本身需要循環操作來做檢查, compare_exchange_weak 的性能會更好。
因此對于某些不需要采用循環操作的算法而言, 通常采用compare_exchange_strong 更好。另外,該函數的內存序由? sync 參數指定,可選條件如下:
| memory_order_relaxed | Relaxed |
| memory_order_consume | Consume |
| memory_order_acquire | Acquire |
| memory_order_release | Release |
| memory_order_acq_rel | Acquire/Release |
| memory_order_seq_cst | Sequentially consistent |
| #include <iostream>?????? // std::cout #include <atomic>???????? // std::atomic #include <thread>???????? // std::thread #include <vector>???????? // std::vector // a simple global linked list: structNode { int value; Node* next; }; std::atomic<Node*> list_head(nullptr); voidappend(int val) { ????// append an element to the list ????Node* newNode =new Node{val, list_head}; ????// next is the same as: list_head = newNode, but in a thread-safe way: ????while(!(list_head.compare_exchange_strong(newNode->next, newNode))); ????// (with newNode->next updated accordingly if some other thread just appended another node) } int main () { ????// spawn 10 threads to fill the linked list: ????std::vector<std::thread> threads; ????for(int i = 0; i < 10; ++i) threads.push_back(std::thread(append, i)); ????for(auto& th : threads) th.join(); ????// print contents: ????for(Node* it = list_head; it!=nullptr; it=it->next) ????????std::cout <<' ' << it->value; ????std::cout <<'\n'; ????// cleanup: ????Node* it;while (it=list_head) {list_head=it->next;delete it;} ????return0; } |
好了,本文花了大量的篇幅介紹 std::atomic 基本類型,下一篇博客我會給大家介紹 C++11 的標準庫中std::atomic 針對整形(integral)和指針類型的特化版本做了哪些改進。
總結
以上是生活随笔為你收集整理的C++11 并发指南六( atomic 类型详解二 std::atomic )的全部內容,希望文章能夠幫你解決所遇到的問題。
- 上一篇: 1.5 Python基础知识 - whi
- 下一篇: LVS NAT/DR