global static boolean pointer causes segmentation fault using pthread

Question

New to pthread programming, and stuck on this error when working on a C++&C mixed code.

What I have done is to call the c code in the thread created by the c++ code. There is a static boolean pointer is_center used in the thread and should got free when the thread finishes.

However I noticed that every time when the program processed into the c function, the value of the boolean pointer would be changed and the segmentation fault then happened due to the free(). And the problem only happens when the c code is used. Remove the c code and the multi-thread c++ part works well.

Detail code is as follows:

static bool *is_center;

// omit other codes in between ...

void streamCluster( PStream* stream)
{
    // some code here ...
    while(1){
        // some code here ...
        is_center = (bool*)calloc(points.num,sizeof(bool));

        // start the parallel thread here.
        // the c code is invoked in this function.
        localSearch(&points,kmin, kmax,&kfinal); // parallel

        free(is_center);
    }

And the function using parallel is as follows (my c code is invoked in each thread):

void localSearch( Points* points, long kmin, long kmax, long* kfinal ) {
    pthread_barrier_t barrier;
    pthread_t* threads = new pthread_t[nproc];
    pkmedian_arg_t* arg = new pkmedian_arg_t[nproc];

    pthread_barrier_init(&barrier,NULL,nproc);

    for( int i = 0; i < nproc; i++ ) {
            arg[i].points = points;
            arg[i].kmin = kmin;
            arg[i].kmax = kmax;
            arg[i].pid = i;
            arg[i].kfinal = kfinal;
            arg[i].barrier = &barrier;

            pthread_create(threads+i,NULL,localSearchSub,(void*)&arg[i]);
    }

    for ( int i = 0; i < nproc; i++) {
        pthread_join(threads[i],NULL);
    }

    delete[] threads;
    delete[] arg;
    pthread_barrier_destroy(&barrier);
}

Finally the function calling my c code:

void* localSearchSub(void* arg_) {                                                                                                                                                        

    int eventSet = PAPI_NULL;                                                                                                                                                                                                                                                                                                                                     
    begin_papi_thread(&eventSet);                                                                                                                                                         

    pkmedian_arg_t* arg= (pkmedian_arg_t*)arg_;                                                                                                                                             
    pkmedian(arg->points,arg->kmin,arg->kmax,arg->kfinal,arg->pid,arg->barrier);                                                                                                            

    end_papi_thread(&eventSet);                                                                                                                                                                                                                                                                                                                                                 

    return NULL;                                                                                                                                                                            
}   

And from gdb, what I have got for the is_center is:

Breakpoint 2, localSearchSub (arg_=0x600000000000bc40) at streamcluster.cpp:1711
1711      end_papi_thread(&eventSet);
(gdb) s

Hardware watchpoint 1: is_center

Old value = (bool *) 0x600000000000bba0
New value = (bool *) 0xa93f3
0x400000000000d8d1 in localSearchSub (arg_=0x600000000000bc40) at streamcluster.cpp:1711
1711      end_papi_thread(&eventSet);

Any suggestions? Thanks in advance!

Some new information about the code: for the c code, I am using the PAPI package. I write my own papi wrapper to initialize and read system counters. The code is as follows:

void begin_papi_thread(int* eventSet)                                                                                                                                                     
{                                                                                                                                                                                         
    int thread_id = pthread_self();                                                                                                                                                       
    // Events                                                                                                                                                                             
    if (PAPI_create_eventset(eventSet)) {                                                                                                                                                 
        PAPI_perror(return_value, error_string, PAPI_MAX_STR_LEN);                                                                                                                        
        printf("*** ERROR *** Failed to create event set for thread %d: %s\n.", thread_id, error_string);                                                                                 
    }                                                                                                                                                                                     
    if((return_value = PAPI_add_events(*eventSet, event_code, event_num)) != PAPI_OK)                                                                                                     
    {                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        
        printf("*** ERROR *** Failed to add event for thread %d: %d.\n", thread_id, return_value);                                                                                                                                                                                                                                                                         
    }                                                                                                                                                                                     
    // Start counting                                                                                                                                                                     
    if ((return_value = PAPI_start(*eventSet)) != PAPI_OK) {                                                                                                                                                                                                                                                                                                              
        PAPI_perror(return_value, error_string, PAPI_MAX_STR_LEN);                                                                                                                                                                                                                                                                                                         
        printf("*** ERROR *** PAPI failed to start the event for thread %d: %s.\n", thread_id, error_string);
    }                                                                                                                                                                                                                                                                                                                                                                             
}  
void end_papi_thread(int* eventSet)                                                                                                                                                       
{                                                                                                                                                                                         
    int thread_id = pthread_self();                                                                                                                                                       
    int i;                                                                                                                                                                                

    long long * count_values = (long long*)malloc(sizeof(long long) * event_num);                                                                                                         
    if (PAPI_read(*eventSet, count_values) != PAPI_OK)                                                                                                                                    
        printf("*** ERROR *** Failed to load count values.\n");                                                                                                                           

    if (PAPI_stop(*eventSet, &dummy_values) != PAPI_OK) {
        PAPI_perror(return_value, error_string, PAPI_MAX_STR_LEN);
        printf("*** ERROR *** PAPI failed to stop the event for thread %d: %s.\n", thread_id, error_string);
        return;
    }
    if(PAPI_cleanup_eventset(*eventSet) != PAPI_OK)
        printf("*** ERROR *** Clean up failed for the thread %d.\n", thread_id);                                                                                                          
} 

Answer 1

I don't think you've posted enough code to really understand your problem, but it looks suspicious that you've declared is_center global. I assume you're using it in more than one place, possibly by multiple threads (localSearchSub mentions it, which is your worker thread function).

If is_center is being read or written by multiple threads, you probably want to protect it with a pthread mutex. You say it is "freed when the thread finishes", but you should be aware that there are nprocs threads, and it looks like they're all working on an array of is_center[points] bools. If points != nproc, this could b e a bad thing[1]. Each thread should probably work on its own array, and localSearch should aggregate the results.

The xxx_papi_thread functions don't get any hits on Google, so I can only imagine it's your own... unlikely we'll be able to help you, if the problem is in there :)

[1]: Even if points == nproc, it is not necessarily OK to write to different elements of an array from multiple threads (it's compiler and processor dependent). Be safe, use a mutex.

Also, this is tagged C++. Can you replace the calloc and dynamic arrays (using new) with vectors? It might end up easier to debug, and it certainly ends up easier to maintain. Why do you hate and want to punish the readers of your code? ;)