c++ - pthread广播然后等待？

Question

我正在尝试设置多个线程处于等待状态，直到它们收到 pthread_cond_broadcast()。

完成一项工作后，我希望线程回到它们的等待状态。

我还希望调用 pthread_cond_broadcast() 的进程在继续之前等待所有线程返回到它们的等待状态。在这种情况下，调用广播的是 main() 函数。我正在尝试让 main(0 在调用广播后执行 pthread_cond_wait()。

void* Work::job(void* id)
{
    int idx = (long)id;

    while(1)
    {
        pthread_mutex_lock(&job_lock);

        while(!jobs_complete)
        {
            // wait for main to broadcast
            pthread_cond_wait(&can_work, &job_lock);
            pthread_mutex_unlock(&job_lock);

            // work here

            pthread_mutex_lock(&job_lock);
            ++jobs_completed;

            if(jobs_completed == NUM_THREADS)
            {
                jobs_complete = true;
                pthread_cond_signal(&jobs_done);
                pthread_mutex_unlock(&job_lock);
            }
            pthread_mutex_unlock(&job_lock);
        }

        pthread_mutex_unlock(&job_lock);
    }

    return NULL;
}

NUM_THREADS 是 4，job_lock 是一个 pthread_mutex_t，can_work 和 jobs_done是 pthread_cond_t，jobs_completed 是一个 bool 而 jobs_complete 是一个 int。

// work

jobs_completed = false;
jobs_complete = 0;
pthread_mutex_lock(&job_lock);
pthread_cond_broadcast(&can_work);
pthread_cond_wait(&jobs_complete);
pthread_mutex_unlock(&job_lock);

// work that depends on jobs_complete

现在，我正在通过调用 pthread_cond_broadcast() 然后在它之后调用 pthread_cond_wait() 来执行此操作，但这似乎陷入了僵局。

任何人都可以解释我应该如何做或哪里出错了吗？如果有任何帮助，我将不胜感激。

谢谢!

Answer 1

我只是发布这个(这几乎是所有 C 代码，但 pthreads 也是如此，所以请稍微放松一下)来演示一种做我认为你正在尝试的事情的方法去完成。显然，您可能希望将其中的大部分内容正确地封装在适当的类等中。希望向您展示的是条件变量、互斥量以及它们与谓词管理和通知的关系是如何工作的。

希望您觉得它有用。祝你有美好的一天。

#include <iostream>
#include <unistd.h>
#include <pthread.h>
using namespace std;

// our global condition variable and mutex
pthread_cond_t cv = PTHREAD_COND_INITIALIZER;
pthread_mutex_t mtx = PTHREAD_MUTEX_INITIALIZER;

// our predicate values.
bool finished = false;
int jobs_waiting = 0;
int jobs_completed = 0;

// our thread proc
static void *worker_proc(void* p)
{
    intptr_t id = (intptr_t)p;  // our id
    size_t n_completed = 0;     // our job completion count

    // always latch prior to eval'ing predicate vars.
    pthread_mutex_lock(&mtx);
    while (!finished)
    {
        // wait for finish or work-waiting predicate
        while (!finished && jobs_waiting == 0)
            pthread_cond_wait(&cv, &mtx);

        // we own the mutex, so we're free to look at, modify
        //  etc. the values(s) that we're using for our predicate
        if (finished)
            break;

        // must be a job_waiting, reduce that number by one, then
        //  unlock the mutex and start our work. Note that we're
        //  changing the predicate (jobs_waiting is part of it) and
        //  we therefore need to let anyone that is monitoring know.
        --jobs_waiting;
        pthread_cond_broadcast(&cv);
        pthread_mutex_unlock(&mtx);

        // DO WORK HERE (this just runs a lame summation)
        for (int i=0,x=0;i<1048576; x += ++i);
        ++n_completed;

        // finished work latch mutex and setup changes
        pthread_mutex_lock(&mtx);
        ++jobs_completed;
        pthread_cond_broadcast(&cv);
    }

    // final report
    cout << id << ": jobs completed = " << n_completed << endl;

    // we always exit owning the mutex, so unlock it now. but
    //  let anyone else know they should be quitting as well.
    pthread_cond_broadcast(&cv);
    pthread_mutex_unlock(&mtx);
    return p;
}

// sets up a batch of work and waits for it to finish.
void run_batch(int num)
{
    pthread_mutex_lock(&mtx);
    jobs_waiting = num;
    jobs_completed = 0;
    pthread_cond_broadcast(&cv);

    // wait or all jobs to complete.
    while (jobs_completed != num)
        pthread_cond_wait(&cv, &mtx);

    // we own this coming out, so let it go.
    pthread_mutex_unlock(&mtx);
}

// main entry point.
int main()
{
    // number of threads in our crew
    static const size_t N = 7;
    pthread_t thrds[N] = {0};

    // startup thread crew.
    intptr_t id = 0;
    for (size_t i=0; i<N; ++i)
        pthread_create(thrds + i, NULL, worker_proc, (void*)(++id));

    // run through batches. each batch is one larger
    //  than the prior batch. this should result in some
    //  interesting job-counts per-thread.
    for (int i=0; i<64; ++i)
        run_batch(i);

    // flag for shutdown state.
    pthread_mutex_lock(&mtx);
    finished = true;
    pthread_cond_broadcast(&cv);
    pthread_mutex_unlock(&mtx);
    for (size_t i=0; i<N; pthread_join(thrds[i++], NULL));

    return 0;
}

示例输出 #1

3: jobs completed = 256
6: jobs completed = 282
5: jobs completed = 292
2: jobs completed = 242
1: jobs completed = 339
4: jobs completed = 260
7: jobs completed = 409

示例输出 #2

6: jobs completed = 882
1: jobs completed = 210
4: jobs completed = 179
5: jobs completed = 178
2: jobs completed = 187
7: jobs completed = 186
3: jobs completed = 194

示例输出 #3

1: jobs completed = 268
6: jobs completed = 559
3: jobs completed = 279
5: jobs completed = 270
2: jobs completed = 164
4: jobs completed = 317
7: jobs completed = 159

---

固定批量大小

相同的代码，但改变了这个:

for (int i=0; i<64; ++i)
    run_batch(i);

为此:

for (int i=0; i<64; ++i)
    run_batch(N);

给出以下内容，这可能更接近您真正要找的内容。

示例输出 #1

4: jobs completed = 65
2: jobs completed = 63
5: jobs completed = 66
3: jobs completed = 63
1: jobs completed = 64
7: jobs completed = 63
6: jobs completed = 64

示例输出 #2

3: jobs completed = 65
5: jobs completed = 62
1: jobs completed = 67
7: jobs completed = 63
2: jobs completed = 65
6: jobs completed = 61
4: jobs completed = 65

示例输出 #3

2: jobs completed = 58
4: jobs completed = 61
5: jobs completed = 69
7: jobs completed = 68
3: jobs completed = 61
1: jobs completed = 64
6: jobs completed = 67