pthreads生产者消费者死锁

我写了以下代码:

void *produce(void* arg)
{
 buffer* buff = (buffer *) arg;
 while (1)
 {
  pthread_mutex_lock(&mutex);
  if (elements_produced == JOB_SIZE)
  {
   pthread_mutex_unlock(&mutex);
   pthread_exit(NULL);
  }
  elements_produced++;

  while (buff->in_buff == CAPACITY)
  {
   pthread_cond_wait(&cond_empty, &mutex);
  }

  // produce
  buff->buffer[buff->tail] = rand();
  sum_produced += buff->buffer[buff->tail];
  printf(">produced %d\n", buff->buffer[buff->tail]);

  buff->tail = (buff->tail + 1) % CAPACITY;
  buff->in_buff++;
  pthread_cond_signal(&cond_empty);
  pthread_mutex_unlock(&mutex);
 }
 pthread_exit(NULL);
}

void *consume(void* arg)
{
 int rc;
 buffer* buff = (buffer *) arg;
 while (1)
 { 
  rc = pthread_mutex_lock(&mutex);

  if (elements_consumed == JOB_SIZE)
  {
   pthread_mutex_unlock(&mutex);
   pthread_exit(NULL);
   return 0;
  }
  elements_consumed++;

  while (buff->in_buff == 0)
  {   
   rc = pthread_cond_wait(&cond_empty, &mutex);
  }

  // consume  
  printf("<consumed %d\n", buff->buffer[buff->head]);
  sum_consumed += buff->buffer[buff->head];
  buff->head = (buff->head + 1) % CAPACITY;
  buff->in_buff--;
  pthread_cond_signal(&cond_full);
  pthread_mutex_unlock(&mutex);
 }
 pthread_exit(NULL);
 return 0;
}

所有变量都已正确初始化。任务是生成作业大小的元素并使用它们。它不时地被锁在死锁里。我对posix线程还很陌生,所以我可能遗漏了一些非常明显的东西(java/C#/python中的生产者/消费者做过很多次,但现在我真的陷入了困境)。我知道用信号量做要容易得多,但我需要用这种方法。

有什么建议吗?