Singleton multithreading programs (1)

Static variables with block scope

Static variables with block scope will be created exactly once. This characteristic is the base of the so called Meyers Singleton, named after Scott Meyers.  This is by far the most elegant implementation of the singleton pattern.

#include <thread>

class MySingleton{
public:
  static MySingleton& getInstance(){
    static MySingleton instance;
    return instance;
  }
private:
  MySingleton();
  ~MySingleton();
  MySingleton(const MySingleton&)= delete;
  MySingleton& operator=(const MySingleton&)= delete;

};

MySingleton::MySingleton()= default;
MySingleton::~MySingleton()= default;

int main(){

  MySingleton::getInstance();

}

 

By using the keyword default, you can request special methods from the compiler. They are Special because only compiler can create them. With delete, the result is, that the automatically generated methods (constructor, for example) from the compiler will not be created and, therefore, can not be called. If you try to use them you'll get an compile time error. What's the point of the Meyers Singleton in multithreading programs? The Meyers Singleton is thread safe.

A side note: Double-checked locking pattern

Wrong beliefe exists, that an additional way for the thread safe initialization of a singleton in a multithreading environment is the double-checked locking pattern. The double-checked locking pattern is - in general -  an unsafe way to initialize a singleton. It assumes guarantees in the classical implementation, which aren't given by the Java, C# or C++ memory model. The assumption is, that the access of the singleton is atomic.

But, what is the double-checked locking pattern? The first idea to implement the singleton pattern in a thread safe way, is  to protected the initialization of the singleton by a lock.

mutex myMutex;

class MySingleton{
public:  
  static MySingleton& getInstance(){    
    lock_guard<mutex> myLock(myMutex);      
    if( !instance ) instance= new MySingleton();    
    return *instance;  
  }
private:  
  MySingleton();  
  ~MySingleton();  
  MySingleton(const MySingleton&)= delete;  
  MySingleton& operator=(const MySingleton&)= delete;
  static MySingleton* instance;
};
MySingleton::MySingleton()= default;
MySingleton::~MySingleton()= default;
MySingleton* MySingleton::instance= nullptr;

 

 

Any issues? Yes and no. The implementation is thread safe. But there is a great performance penalty. Each access of the singleton in line 6 is protected by an expansive lock. That applies also for the reading access. Most time it's not necessary. Here comes the double-checked locking pattern to our rescue.

 

static MySingleton& getInstance(){    
  if ( !instance ){      
    lock_guard<mutex> myLock(myMutex);      
    if( !instance ) instance= new MySingleton();    
  }
  return *instance; 
}

 

 

I use inexpensive pointer comparison  in the line 2 instead of an expensive lock a. Only if I get a null pointer, I apply the expensive lock on the singleton (line 3). Because there is the possibility that another thread will initialize the singleton between the pointer comparison (line 2) and the lock (line3), I have to perform an additional pointer comparison the on line 4. So the name is clear. Two times a check and one time a lock.

Smart? Yes. Thread safe? No.

What is the problem? The call instance= new MySingleton() in line 4 consists of at least three steps.

1. Allocate memory for MySingleton
2. Create the MySingleton object in the memory
3. Let instance refer to the MySingleton object

The problem: there is no guarantee about  the sequence of these steps. For example, out of optimization reasons, the processor can reorder the steps to the sequence 1,3 and 2. So, in the first step the memory will be allocated and in the second step, instance refers to an incomplete singleton. If at that time another thread tries to access the singleton, it compares the pointer and gets the answer true. So, the other thread has the illusion that it's dealing with a complete singleton.

The consequence is simple: program behaviour is undefined.

 

최적화에 의해서 명령어 순서가 1,3 ,2 순서로 뒤 바뀔 수 있기 때문에 다른 스레드에서 잘못된 instance 포인터를 얻어가게 되어 문제가 발생 할 수 있게 된다

핵심은 Mutex 가  reorder 를 막아주진 못한다는 것!!!

 

 

What's next?

At first, I thought, I should continue in the next post with the singleton pattern. But to write about the singleton pattern, you should have a basic knowledge of the memory model. So I continue in the sequence of my German blog. The next post will be about-thread local storage. In case we are done with the high end API of multithreading in C++, I'll go further with the low end API. (Proofreader Alexey Elymanov)

 

 

ref : https://www.modernescpp.com/index.php/thread-safe-initialization-of-data