[C++]boost提供的几种lock-free方案以及std::atomic实现无锁队列

boost方案

boost提供了三种无锁方案

boost::lockfree::queue：

支持多个生产者和多个消费者线程的无锁队列。

boost::lockfree::stack：

支持多个生产者和多个消费者线程的无锁栈。

boost::lockfree::spsc_queue：

仅支持单个生产者和单个消费者线程的无锁队列。相比boost::lockfree::queue，其效率更高。

 

注：这些API内部是通过轻量级原子锁实现的lock-free，不是真正意义的无锁。我看到的资料中，貌似只有linux kernel中fifo实现了真正意义上的无锁，但是仅用于与单个消费者单个生产者的环境。

 

boost官方文档：

http://www.boost.org/doc/libs/1_60_0/doc/html/lockfree.html

 

queue容量和自增长的问题

可以设置初始容量，添加新元素时如果容量不够，则总容量可能自动增长：queue在当前操作系统上如果支持lock-free，则不会自动增长，如果不支持lock-free，才会自动增长。不同的操作系统其内存分配机制不同，这样会导致在某些操作系统上的queue不支持lockfree

boost::lockfree::spsc_queue> q;
printf("boost::lockfree:queue is lock free:%s", q.is_lock_free() ? "true" : "false");	//true

//push的返回值：1，push成功；0，push失败。
size_t s1 = q.push(9);	//1
size_t s2 = q.push(9);	//1
size_t s3 = q.push(9);	//0

boost::lockfree::queue, boost::lockfree::capacity<2>> q2;
size_t s2_1 = q2.push(9);	//1
size_t s2_2 = q2.push(9);	//1
size_t s2_3 = q2.push(9);	//0

boost::lockfree::queue, boost::lockfree::capacity<2>> q3;
size_t s3_1 = q3.push(9);	//1
size_t s3_2 = q3.push(9);	//1
size_t s3_3 = q3.push(9);	//0
size_t s3_4 = q3.push(9);	//0

 

 

如果不需要考虑多线程或者自己实现同步，还有一种方案：boost::circular_buffer

http://www.boost.org/doc/libs/1_60_0/doc/html/circular_buffer.html

 

 

C++11 std::atomic方案

网上有人借用std::atomic实现的一套无锁队列，其内部实现参考了boost::lockfree::queue的设计思路，可以适用于多个消费者多个生产者线程。

A High Performance Lock Free Ring Queue

http://www.codeproject.com/Tips/754393/A-High-Performance-Lock-Free-Ring-Queue

下面代码我在原文基础上做了修改：最新的编译器已不支持std::atomic_flag在构造函数中初始化。

 lfringqueue.h

#ifndef INCLUDED_UTILS_LFRINGQUEUE
#define INCLUDED_UTILS_LFRINGQUEUE

#define _ENABLE_ATOMIC_ALIGNMENT_FIX
#define ATOMIC_FLAG_INIT 0


#pragma once


#include 
#include 
#include 
#include 
#include 
#include 
#include 

// Lock free ring queue 

template < typename _TyData, long _uiCount = 100000 >
class lfringqueue
{
public:
    lfringqueue( long uiCount = _uiCount ) : m_lTailIterator(0), m_lHeadIterator(0), m_uiCount( uiCount )
    {
        m_queue = new _TyData*[m_uiCount];
        memset( m_queue, 0, sizeof(_TyData*) * m_uiCount );
    }

    ~lfringqueue()
    {
        if ( m_queue )
            delete [] m_queue;
    }

 
    bool enqueue( _TyData *pdata, unsigned int uiRetries = 1000 )
    {
        if ( NULL == pdata )
        {
            // Null enqueues are not allowed
            return false;
        }

        unsigned int uiCurrRetries = 0;
        while ( uiCurrRetries < uiRetries )
        {
            // Release fence in order to prevent memory reordering 
            // of any read or write with following write
            std::atomic_thread_fence(std::memory_order_release);
            
            long lHeadIterator = m_lHeadIterator;

            if ( NULL == m_queue[lHeadIterator] )
            {
                long lHeadIteratorOrig = lHeadIterator;

                ++lHeadIterator;
                if ( lHeadIterator >= m_uiCount )
                        lHeadIterator = 0;

                // Don't worry if this CAS fails.  It only means some thread else has
                // already inserted an item and set it.
                if ( std::atomic_compare_exchange_strong( &m_lHeadIterator, &lHeadIteratorOrig, lHeadIterator ) )             
                {
                    // void* are always atomic (you wont set a partial pointer).
                    m_queue[lHeadIteratorOrig] = pdata;
                  
                    if ( m_lEventSet.test_and_set( ))
                    {
                        m_bHasItem.test_and_set();
                    }
                    return true;
                }
            }
            else
            {
                // The queue is full.  Spin a few times to check to see if an item is popped off.
                ++uiCurrRetries;
            }
        }
        return false;
    }

    bool dequeue( _TyData **ppdata )
    {
        if ( !ppdata )
        {
            // Null dequeues are not allowed!
            return false;
        }

        bool bDone = false;
        bool bCheckQueue = true;

        while ( !bDone )
        {
            // Acquire fence in order to prevent memory reordering 
            // of any read or write with following read
            std::atomic_thread_fence(std::memory_order_acquire);
            //MemoryBarrier();
            long lTailIterator = m_lTailIterator;
            _TyData *pdata = m_queue[lTailIterator];
            //volatile _TyData *pdata = m_queue[lTailIterator];            
            if ( NULL != pdata )
            {
                bCheckQueue = true;
                long lTailIteratorOrig = lTailIterator;

                ++lTailIterator;
                if ( lTailIterator >= m_uiCount )
                        lTailIterator = 0;

                //if ( lTailIteratorOrig == atomic_cas( (volatile long*)&m_lTailIterator, lTailIterator, lTailIteratorOrig ))
                if ( std::atomic_compare_exchange_strong( &m_lTailIterator, &lTailIteratorOrig, lTailIterator ))
                {
                        // Sets of sizeof(void*) are always atomic (you wont set a partial pointer).
                        m_queue[lTailIteratorOrig] = NULL;

                        // Gets of sizeof(void*) are always atomic (you wont get a partial pointer).
                        *ppdata = (_TyData*)pdata;

                        return true;
                }
            }
            else
            {
                bDone = true;
                m_lEventSet.clear();
            }
        }
        *ppdata = NULL;
        return false;
    }
	
	
    long countguess() const
    {
        long lCount = trycount();

        if ( 0 != lCount )
                return lCount;

        // If the queue is full then the item right before the tail item will be valid.  If it
        // is empty then the item should be set to NULL.
        long lLastInsert = m_lTailIterator - 1;
        if ( lLastInsert < 0 )
                lLastInsert = m_uiCount - 1;

        _TyData *pdata = m_queue[lLastInsert];
        if ( pdata != NULL ) 
                return m_uiCount;

        return 0;
    }

    long getmaxsize() const
    {
        return m_uiCount;
    }

    bool HasItem()
    {
        return m_bHasItem.test_and_set();
    }

    void SetItemFlagBack()
    {
        m_bHasItem.clear();
    }
	
private:
    long trycount() const
    {
        long lHeadIterator = m_lHeadIterator;
        long lTailIterator = m_lTailIterator;

        if ( lTailIterator > lHeadIterator )
                return m_uiCount - lTailIterator + lHeadIterator;

        // This has a bug where it returns 0 if the queue is full.
        return lHeadIterator - lTailIterator;
    }
	
private:    
    std::atomic m_lHeadIterator;  // enqueue index
    std::atomic m_lTailIterator;  // dequeue index
    _TyData **m_queue;                  // array of pointers to the data
    long m_uiCount;                     // size of the array
    std::atomic_flag m_lEventSet = ATOMIC_FLAG_INIT;       // a flag to use whether we should change the item flag
    std::atomic_flag m_bHasItem = ATOMIC_FLAG_INIT;        // a flag to indicate whether there is an item enqueued
};

#endif //INCLUDED_UTILS_LFRINGQUEUE

 

 

测试：

/* 
 * File:   main.cpp
 * Author: Peng
 *
 * Created on February 22, 2014, 9:55 PM
 */

#include 
#include "lfringqueue.h"
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include 

#include 
#include 
#include 

#include 

const long NUM_DATA = 10;
const int NUM_ENQUEUE_THREAD = 1;
const int NUM_DEQUEUE_THREAD = 1;
const long NUM_ITEM = 1000000;
        
using namespace std;
class Data
{
public:   
    Data( int i = 0 ) : m_iData(i)
    {
        stringstream ss;
        ss << i;
        m_szDataString = ss.str();
        //sprintf( m_szDataString, "%l-d", i);    
    }
        
    bool operator< ( const Data & aData) const
    {
        if ( m_iData < aData.m_iData)
            return true;
        else
            return false;
    }
    
    int& GetData()
    {
        return m_iData;
    }
private:   
    int m_iData;
    string m_szDataString;
    //char m_szDataString[MAX_DATA_SIZE];
};

Data DataArray[NUM_DATA];

constexpr long size = 0.5 * NUM_DATA;
lfringqueue < Data, 1000> LockFreeQueue; 
boost::lockfree::queue BoostQueue(1000);

// Since there is a chance that the searched number cannot be found, so the function should return boolean
bool BinarySearchNumberInSortedArray( Data datas[], int iStart, int iEnd, int SearchedNum, int &iFound )
{
    if ( iEnd - iStart <= 1 )
    {
        if ( datas[iStart].GetData() == SearchedNum )
        {
            iFound = iStart;
            return true;
        }
        else if ( datas[iEnd].GetData() == SearchedNum )
        {
            iFound = iEnd;
            return true;
        }
        else
            return false;
    }
    
    int mid = 0.5 * ( iStart + iEnd );
    
    if ( datas[mid].GetData() == SearchedNum )
    {
        iFound = mid;
        return true;
    }
    
    if ( datas[mid].GetData() > SearchedNum )
    {
        if ( mid - 1 >= 0)
            return BinarySearchNumberInSortedArray ( datas, iStart, mid - 1, SearchedNum, iFound);
        else
            return false;
    }
    else
    {
        if ( mid + 1 <= iEnd )
            return BinarySearchNumberInSortedArray ( datas, mid + 1, iEnd, SearchedNum, iFound);
        else
            return false;
    }
}
bool GenerateRandomNumber_FindPointerToTheNumber_EnQueue()
{
    std::uniform_int_distribution dis(1, NUM_DATA);
    default_random_engine engine{};
       
    for ( long i = 0; i < NUM_ITEM; i++ )
    {
        int x = dis ( engine );
        
        int iFoundIndex;
        if ( BinarySearchNumberInSortedArray(DataArray, 0, NUM_DATA - 1, x, iFoundIndex ) )
        {
            Data* pData = &DataArray[iFoundIndex];
            LockFreeQueue.enqueue( pData );
            //BoostQueue.push( pData );
        }
    }
}
bool Dequeue()
{
    Data *pData;

    for ( long i = 0; i < NUM_ITEM; i ++)
    {
        while (  LockFreeQueue.dequeue( &pData ) );       
        //while (  BoostQueue.pop( pData ) ) ;      
    }    
}

int main(int argc, char** argv) 
{
    for ( int i = 1; i < NUM_DATA + 1; i++ )
    {
        Data data(i);
        DataArray[i-1] = data;
    }
     
    std::thread PublishThread[NUM_ENQUEUE_THREAD]; 
    std::thread ConsumerThread[NUM_DEQUEUE_THREAD];
    std::chrono::duration elapsed_seconds;
  
    for ( int i = 0; i < NUM_ENQUEUE_THREAD;  i++ )
    {
        PublishThread[i] = std::thread( GenerateRandomNumber_FindPointerToTheNumber_EnQueue ); 
    }
    
    auto start = std::chrono::high_resolution_clock::now();
    for ( int i = 0; i < NUM_DEQUEUE_THREAD; i++ )
    {
        ConsumerThread[i] = std::thread{ Dequeue};
    }
    
    for ( int i = 0; i < NUM_DEQUEUE_THREAD; i++ )
    {
        ConsumerThread[i].join();
    }   

    auto end = std::chrono::high_resolution_clock::now();
    elapsed_seconds = end - start;
    std::cout << "Enqueue and Dequeue 1 million item in:" << elapsed_seconds.count() << std::endl;
   

    for ( int i = 0; i < NUM_ENQUEUE_THREAD; i++ )
    {
        PublishThread[i].join();
    }
             
    return 0;
}