多線程訪問unordered_map的運行時錯誤

Question

我有這個多線程的代碼，它試圖使用stl unordered_map創建一個線程本地單例對象。

這是code。我在這裏複製的代碼逐字：

#include <iostream> 
#include <unordered_map> 
#include <vector> 
#include <thread> 
#include <algorithm> 
#include <mutex> 
using namespace std; 

class single 
{ 
public: 
    // Every thread needs to call this to get its individual instance 
    static single* getInstance(unsigned int threadId); 

    static void print(unsigned int threadId) 
    { 
     std::cout << "threadId:" << threadId << ", Singleton: " << _instances[threadId] << std::endl; 
    } 

protected: 
    // Made protected not private so that singleton can be subclassed 
    single();           // Clients cant construct objects directly 
    ~single();           // cannot be destroyed by clients 

    single(const single &) = delete;     // non-copyable 
    single& operator=(const single &) = delete;   // can't be copy assigned 
    single(single &&) = delete;       // non-move constructible 
    single & operator=(single &&) = delete;   // non-move assignable 

private: 
    static std::unordered_map<unsigned,single*> _instances; 
    static std::mutex _lock; 

}; 

std::mutex single::_lock; 
std::unordered_map<unsigned,single*> single::_instances; 

single::single(){} 

single::~single(){} 

single* single::getInstance(unsigned int threadId) 
{ 
    if(_instances.count(threadId) == 0) 
    { 
     std::lock_guard<std::mutex> lg(_lock); 
     if(_instances.count(threadId) == 0) 
     { 
      _instances[threadId] = new single; 
      std::cout <<"Created By ThreadId: " << threadId <<std::endl; 
     } 
    } 

    return _instances[threadId]; 
} 

void Run(unsigned int threadId) 
{ 
    single::getInstance(threadId)->print(threadId); 
} 

int main() 
{ 
    std::vector<std::thread> workers; 
    const unsigned threadCount = 16; 

    for(auto i = 0; i != threadCount; ++i) 
    { 
     workers.push_back(std::thread(Run, i)); 
    } 

    for_each(workers.begin(), workers.end(), std::mem_fn(&thread::join)); 

    return 0; 
} 

我用unordered_map的count()常數函數（線程安全的？），以檢查是否該線程的實例被創建。如果計數爲0，則創建實例並將其作爲密鑰的值存儲爲threadId。我還添加了一個lock_guard以防止多線程併發插入到無序的映射靜態對象，但它似乎是一些競爭條件，我無法發現並且此代碼有時會出錯。

任何人都可以向我解釋在這段代碼的哪部分競爭條件正在發生，如果這個解決方案可以使線程本地單身人士工作？

Answer 1

誰能給我解釋這其中的部分代碼的比賽條件正在發生的事情...

不，這不是線程安全的。一旦至少有一位作家，所有訪問都需要通過鎖來保護。所以第一個_instances.count()是無人防守的，可能導致數據競爭。

...如果這個解決方案可以使線程本地單例工作？

你可以看看使用thread_local控制本地線程存儲時限。

您也可以將防盜鎖移動到unordered_map的任何訪問之前。

single* single::getInstance(unsigned int threadId) 
{ 
    std::lock_guard<std::mutex> lg(_lock); 
    if(_instances.count(threadId) == 0) 
    { 
     _instances[threadId] = new single; 
     std::cout <<"Created By ThreadId: " << threadId <<std::endl; 
    } 
    return _instances[threadId]; 
} 

的線程數

Demo code here.

---

側面說明，根據問題正在得到解決，你可以線程數限制爲std::thread::hardware_concurrency限制。