2

I am trying to implement thread pool in C++ using pthread. I want to encapsulate logic related to threads management in one object which is taking ownership of these threads. That means whenever this object is destroyed, threads must be stopped and cleaned up.

What is the best way to stop and destroy threads? Is detaching at start and cancellation at stop a good solution? Or maybe it is better to cancel and join threads? See my code, I'll appreciate any relevant comments.

WorkerThreadManager.h:

#include "WorkerThreadManagerInterface.h"
#include "utils/mutex.h"
#include <queue>
#include <semaphore.h>

#include <iostream>

class WorkerThreadManager : public WorkerThreadManagerInterface
{
    public:
        WorkerThreadManager(unsigned threadsNumber = 5);
        virtual ~WorkerThreadManager();

        virtual void    PushTask(thread_function_t A_threadFun, result_function_t A_resultFun);
        void    SignalResults();

    private:
        static void*    WorkerThread(void* A_data);

        void    PushResult(int A_result, result_function_t A_resultFun);

        typedef boost::function<void ()> signal_function_t;

        struct worker_thread_data_t
        {
            worker_thread_data_t(thread_function_t A_threadFun, result_function_t A_resultFun) :
                threadFun(A_threadFun), resultFun(A_resultFun) {}
            worker_thread_data_t() {}

            thread_function_t       threadFun;
            result_function_t       resultFun;
        };


        const unsigned                      m_threadsNumber;
        pthread_t*                          m_pthreads;

        utils::Mutex                        m_tasksMutex;
        sem_t                               m_tasksSem;
        std::queue<worker_thread_data_t>    m_tasks;

        utils::Mutex                        m_resultsMutex;
        std::queue<signal_function_t>       m_results;
};

WorkerThreadManager.cpp:

#include "WorkerThreadManager.h"
#include "gateway_log.h"
#include <pthread.h>

/**
 * @brief Creates semaphore and starts threads.
 */
WorkerThreadManager::WorkerThreadManager(unsigned threadsNumber) : m_threadsNumber(threadsNumber)
{
    if ( sem_init(&m_tasksSem, 0, 0) )
    {
        std::stringstream ss;
        ss << "Semaphore could not be initialized: " << errno << " - " << strerror(errno);
        LOG_FATAL(ss);
        throw std::runtime_error(ss.str());
    }

    m_pthreads = new pthread_t[m_threadsNumber];
    for (unsigned i = 0; i < m_threadsNumber; ++i)
    {
        int rc = pthread_create(&m_pthreads[i], NULL, WorkerThreadManager::WorkerThread, (void*) this );
        if(rc)
        {
            std::stringstream ss;
            ss << "Pthread could not be started: " << errno << " - " << strerror(errno);
            LOG_FATAL(ss.str());

            if ( sem_destroy(&m_tasksSem) )
                LOG_ERROR("Semaphore could not be destroyed: " << errno << " - " << strerror(errno));

            delete [] m_pthreads;

            throw std::runtime_error(ss.str());
        }
        else
        {
            LOG_DEBUG("Worker thread started " << m_pthreads[i]);

            if(pthread_detach(m_pthreads[i]))
                LOG_WARN("Failed to detach worker thread");
        }
    }
}

/**
 * @brief Cancels all threads, destroys semaphore
 */
WorkerThreadManager::~WorkerThreadManager()
{
    LOG_DEBUG("~WorkerThreadManager()");

    for(unsigned i = 0; i < m_threadsNumber; ++i)
    {
        if ( pthread_cancel(m_pthreads[i]) )
            LOG_ERROR("Worker thread cancellation failed");
    }

    if ( sem_destroy(&m_tasksSem) )
        LOG_ERROR("Semaphore could not be destroyed: " << errno << " - " << strerror(errno));

    delete [] m_pthreads;
}

/**
 * @brief Adds new task to queue, so worker threads can
 * @param A_threadFun function which will be executed by thread
 * @param A_resultFun function which will be enqueued for calling with return value of A_threadFun as parameter
 *          after worker thread executes A_threadFun.
 */
void WorkerThreadManager::PushTask(thread_function_t A_threadFun, result_function_t A_resultFun)
{
    utils::ScopedLock mutex(m_tasksMutex);

    worker_thread_data_t    data(A_threadFun, A_resultFun);
    m_tasks.push( data );
    sem_post(&m_tasksSem);
    LOG_DEBUG("Task for worker threads has been added to queue");
}

/**
 * @brief   Executes result functions (if there are any) to give feedback 
 *  to classes which requested task execution in worker thread.
 */
void WorkerThreadManager::SignalResults()
{
    while(true)
    {
        signal_function_t signal;
        {
            utils::ScopedLock mutex(m_resultsMutex);
            if(m_results.size())
            {
                signal = m_results.front();
                m_results.pop();
            }
            else
                return;
        }

        signal();
    }
}

/**
 * @brief Enqueues result of function executed in worker thread.
 * @param A_result return value of function executed in worker thread
 * @param A_resultFun function which will be enqueued for calling with A_result as a parameter.
 */
void WorkerThreadManager::PushResult(int A_result, result_function_t A_resultFun)
{
    utils::ScopedLock mutex(m_resultsMutex);

    signal_function_t signal = boost::bind(A_resultFun, A_result);
    m_results.push( signal );
}


/**
 * @brief   worker thread body
 * @param A_data pointer to WorkerThreadManager instance
 */
void* WorkerThreadManager::WorkerThread(void* A_data)
{
    WorkerThreadManager* manager = reinterpret_cast<WorkerThreadManager*>(A_data);
    LOG_DEBUG("Starting worker thread loop");
    while (1)
    {
        if ( -1 == sem_wait(&manager->m_tasksSem) && errno == EINTR )
        {
            LOG_DEBUG("sem_wait interrupted with signal");
            continue;
        }
        LOG_DEBUG("WorkerThread:::::: about to call lock mutex");

        worker_thread_data_t data;
        {
            utils::ScopedLock mutex(manager->m_tasksMutex);
            data = manager->m_tasks.front();
            manager->m_results.pop();
        }

        LOG_DEBUG("WorkerThread:::::: about to call resultFun");
        int result  = data.threadFun();
        LOG_DEBUG("WorkerThread:::::: after call resultFun");
        pthread_testcancel();

        manager->PushResult(result, data.resultFun);
    }

    return NULL;
}

main.cpp:

#include "gateway_log.h"
#include "WorkerThreadManager.h"
#include <memory>

class A {
public:
    int Fun() { LOG_DEBUG("Fun before sleep"); sleep(8); LOG_DEBUG("Fun after sleep");return 0; }
    void Result(int a) { LOG_DEBUG("Result: " << a); }
};


int main()
{
    sd::auto_ptr<WorkerThreadManager> workerThreadManager = new WorkerThreadManager;
    A a;
    workerThreadManager->PushTask(boost::bind(&A::Fun, &a), boost::bind(&A::Result, &a, _1));
    sleep(3);
    LOG_DEBUG("deleting workerThreadManager");
    workerThreadManager.reset();                    // <<<--- CRASH
    LOG_DEBUG("deleted workerThreadManager");
    sleep(10);
    LOG_DEBUG("after sleep");    

    return 0;
}

Please note that there is a problem with this code described here.

Community
  • 1
  • 1
Marcin
  • 279
  • 1
  • 4
  • 13
  • As a point of opinion: `pthread_cancel` is the devil :) It ends up with harder to control termination and might or might not play nicely with C++ (depending on the implementation). http://stackoverflow.com/questions/4760687/cancelling-a-thread-using-pthread-cancel-good-practice-or-bad – Dave S Sep 17 '12 at 19:57
  • 1
    The best way to shut down a thread that is part of a thread pool is to queue a "job of death". When a thread pulls a job from the queue, it checks if it's a "job of death" and if so, terminates itself. – David Schwartz Sep 17 '12 at 19:58
  • Oh yes, in auto_ptr initialisation I miss "new WorkerThreadManager", my apologise. While I was trying to prepare this post I copied and pasted code from my application. I also cleaned the example from some test code, so line numbers you see in backtrace may be not exactly the same as in the pasted code. – Marcin Sep 17 '12 at 20:07
  • On a point of StackOverflow etiquette, I think it's better to ask two questions as two separate questions, then people can be clear which question they are answering and you can accept one answer for each question. – Bryan Sep 17 '12 at 20:16
  • @Bryan I'm separating questions, so this question relates to thread pool cleanup and a [new question](http://stackoverflow.com/questions/12466613/crash-related-to-boostfunction-usage-in-thread-pool) relates to crash – Marcin Sep 17 '12 at 20:30
  • Thanks. I agree with Dave S that there are good answers to this question at http://stackoverflow.com/questions/4760687/cancelling-a-thread-using-pthread-cancel-good-practice-or-bad – Bryan Sep 17 '12 at 20:47

1 Answers1

2

Regarding safe stop: I prefer pthread_join. I do not use pthread_cancel - I am using special stop message, but I always have event-driven thread (mean thread with some queue of messages). When thread get exit-message it stops its loop and then join returns to my main code.

Regarding your code - I would recommend to create class Thread encapsulating single thread. Pool shall have Thread objects created on heap - like now you have array of pthread_t. IF you need synchronization between pool and threads - then you cannot exit pool destructor without being sure Thread objects are destroyed.

Pete Becker
  • 74,985
  • 8
  • 76
  • 165
PiotrNycz
  • 23,099
  • 7
  • 66
  • 112
  • How should I wake up my worker threads when I want to join them? The easiest solution but probably not the nicest would be to call `post_sem` once for each worker thread. – Marcin Sep 17 '12 at 20:46
  • You can add some flag `exitThread` - set it to `true` then post semaphore as many times as there threads to stop. But this is just good advice after a very quick look, don't take it too serious ;) – PiotrNycz Sep 17 '12 at 20:56