Segmentation fault when allocating second array

Question

I quit. This is the most frustrating thing I've ever had to do. Even a non dynamic int array causes segfault. But if I declare it as a float/char whatever array, it works alright.

Update: If i remove the line MPI_Scatter(A[0], N, MPI_INT, A_row, N, MPI_INT, 0, MPI_COMM_WORLD); it works fine. Problem is I need it...

I'm working on a program but I have a bizarre problem.

The following code works fine (if we suppose that N is a multiple of p):

#include <stdio.h>
#include <stdlib.h>
#include "mpi.h"


void main(int argc, char** argv)  
{
   int my_rank, p, N, **A, *diagonals, *A_row;
   MPI_Status status;

   MPI_Init(&argc, &argv);
   MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
   MPI_Comm_size(MPI_COMM_WORLD, &p);

    if (my_rank == 0)  {

        N = 4;
        int *mem = malloc(N * N * sizeof(int));
        A = malloc(N * sizeof(int*));
        for(int i = 0; i < N; i++) 
            A[i] = mem + N*i;    

    }
    MPI_Bcast(&N, 1, MPI_INT, 0, MPI_COMM_WORLD);

    A_row = malloc (N * sizeof(int));

    MPI_Scatter(A[0], N, MPI_INT, A_row, N, MPI_INT, 0, MPI_COMM_WORLD);

    MPI_Finalize();
}

However, I need to allocate another array (diagonals), like this:

#include <stdio.h>
#include <stdlib.h>
#include "mpi.h"


void main(int argc, char** argv)  
{
   int my_rank, p, N, **A, *diagonals, *A_row;
   MPI_Status status;

   MPI_Init(&argc, &argv);
   MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
   MPI_Comm_size(MPI_COMM_WORLD, &p);

    if (my_rank == 0)  {

        N = 4;
        int *mem = malloc(N * N * sizeof(int));
        A = malloc(N * sizeof(int*));
        for(int i = 0; i < N; i++) 
            A[i] = mem + N*i;

        diagonals = malloc (N * sizeof(int));    
    }
    MPI_Bcast(&N, 1, MPI_INT, 0, MPI_COMM_WORLD);

    A_row = malloc (N * sizeof(int));

    MPI_Scatter(A[0], N, MPI_INT, A_row, N, MPI_INT, 0, MPI_COMM_WORLD);

    MPI_Finalize();
}

I get this segmentation fault (if it helps at all):

[teo-VirtualBox:02582] *** Process received signal ***
[teo-VirtualBox:02582] Signal: Segmentation fault (11)
[teo-VirtualBox:02582] Signal code: Address not mapped (1)
[teo-VirtualBox:02582] Failing at address: 0x1
[teo-VirtualBox:02582] [ 0] /lib/x86_64-linux-gnu/libpthread.so.0(+0x113d0)[0x7faecc8d23d0]
[teo-VirtualBox:02582] [ 1] a[0x400c85]
[teo-VirtualBox:02582] [ 2] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf0)[0x7faecc511830]
[teo-VirtualBox:02582] [ 3] a[0x4009a9]
[teo-VirtualBox:02582] *** End of error message ***

Am I missing something obvious?

By the way, I'm not using free(), or doing anything specific because this is not the complete code. It's just a side file that I created for testing.

For the future, load the program with GDB, and it will tell you what line seg faulted, and you can test the values of certain variables. — Dellowar, Dec 09 '16 at 19:05
@SanchkeDellowar I'm not familiar with gdb. I tried `mpiexec -np 4 a.out -gdb` but it doesn't make any difference in the error message. — Teo Alivanoglou, Dec 09 '16 at 19:11
@TeoAlivanoglou Could you try and whittle your code down to a minimum verifiable example i.e. take out everything (mainly the mpi stuff) that you don't need so that we can try running it ourselves? — gowrath, Dec 09 '16 at 19:16
you should check what `malloc` returns to make sure it's not returning NULL. That being said you're not asking for a whole lot of memory here, so I wouldn't expect any NULLs returned. — yano, Dec 09 '16 at 19:42
are you sure `my_rank` always equals 0? If not, `A` never gets `malloc`ed, and the called to `MPI_Scatter` will be broken. Same for `N`. The call to `malloc` `A_row` will be called with junk for `N` unless `my_rank==0` (although I don't know what `MPI_Bcast` does with `N`) — yano, Dec 09 '16 at 20:02
@yano `MPI_Bcast` broadcasts some value to all processes (in my case, `N`). And `MPI_Scatter` sends stuff from `**A` to the other processes, and saves it in `*A_row` — Teo Alivanoglou, Dec 09 '16 at 20:10
If it's broadcasting the value of `N` I find it odd that it wants the address of `N`. But that just re-enforces my thought more. If `MPI_Bcast` isn't setting the value of `N`, then you have undefined behavior starting at `A_row = malloc(..);` if `my_rank != 0`. If `my_rank` always equals 0, then why even check? Otherwise, you should have an `else` condition that handles that, or everything involving `N` and `A` should be moved into the `if (my_rank==0)` block. — yano, Dec 09 '16 at 20:18
@yano First of all, [MPI_Bcast](http://www.mpich.org/static/docs/v3.2/www3/MPI_Bcast.html). It seems to me that you don't really understand how MPI works. Everything that is outside of the `my_rank == 0` block gets executed by all process. — Teo Alivanoglou, Dec 09 '16 at 20:23
no I don't know a thing about MPI. That's irrelevant for what I've suggested. Unless `my_rank==0`, you're going to use uninitialized memory, which is UB. — yano, Dec 09 '16 at 20:29
@yano incorrect. only the process with `rank == 0` accesses `**A` and everything else (except for `diagonals`) is initialized in all processes — Teo Alivanoglou, Dec 09 '16 at 20:45
You say the *Segmentation fault seems to occur before `malloc()`* but at the same time it works if you remove the `MPI_Scatter`? So did you extend the code to check the return values of all `malloc` calls? In any case, guessing is not productive, please [use a debugger (correctly)](https://www.open-mpi.org/faq/?category=debugging). — Zulan, Dec 09 '16 at 22:00
fair enough .. I'll yield to MPI mystery black box ... hope you figure it out — yano, Dec 09 '16 at 22:02

score 1 · Answer 1 · edited May 23 '17 at 11:53

To be honest, I cannot reproduce:

linux21:/home/users/grad1459/Desktop/parallel>mpiexec -np 4 a.out
linux21:/home/users/grad1459/Desktop/parallel>mpicc -Wall -std=c99 main.c
main.c: In function ‘main’:
main.c:9:15: warning: unused variable ‘status’ [-Wunused-variable]
main.c:8:29: warning: variable ‘diagonals’ set but not used [-Wunused-but-set-variable]
linux21:/home/users/grad1459/Desktop/parallel>mpiexec -np 4 a.out
ALL OK
ALL OK
ALL OK
ALL OK
linux21:/home/users/grad1459/Desktop/parallel>

with very similar code to yours:

#include <stdio.h>
#include <stdlib.h>
#include "mpi.h"


int main(int argc, char** argv)  
{
   int my_rank, p, N, **A, *diagonals, *A_row;
   MPI_Status status;

   MPI_Init(&argc, &argv);
   MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
   MPI_Comm_size(MPI_COMM_WORLD, &p);

    if (my_rank == 0)  {

        N = 4;
        int *mem = malloc(N * N * sizeof(int));
        A = malloc(N * sizeof(int*));
        for(int i = 0; i < N; i++) 
            A[i] = mem + N*i;

        diagonals = malloc (N * sizeof(int));
    }
    MPI_Bcast(&N, 1, MPI_INT, 0, MPI_COMM_WORLD);

    A_row = malloc (N * sizeof(int));

    MPI_Scatter(A[0], N, MPI_INT, A_row, N, MPI_INT, 0, MPI_COMM_WORLD);

    MPI_Finalize();
    printf("ALL OK\n");
    return 0;
}

As a result, I think that your virtualbox has some memory limitations and your malloc() fails, check its return value to make sure it's not NULL, like this: How detect malloc failure?

Here is my version:

linux21:/home/users/grad1459/Desktop/parallel>mpiexec --version
HYDRA build details:
    Version:                                 3.1.3
    Release Date:                            Wed Oct  8 09:37:19 CDT 2014
    CC:                              gcc    
    CXX:                             g++    
    F77:                             gfortran   
    F90:                             gfortran   
    Configure options:                       '--disable-option-checking' '--prefix=/usr/local/mpich3' '--cache-file=/dev/null' '--srcdir=.' 'CC=gcc' 'CFLAGS= -O2' 'LDFLAGS= ' 'LIBS=-lpthread ' 'CPPFLAGS= -I/usr/local/USB/mpich-3.1.3/src/mpl/include -I/usr/local/USB/mpich-3.1.3/src/mpl/include -I/usr/local/USB/mpich-3.1.3/src/openpa/src -I/usr/local/USB/mpich-3.1.3/src/openpa/src -D_REENTRANT -I/usr/local/USB/mpich-3.1.3/src/mpi/romio/include'
    Process Manager:                         pmi
    Launchers available:                     ssh rsh fork slurm ll lsf sge manual persist
    Topology libraries available:            hwloc
    Resource management kernels available:   user slurm ll lsf sge pbs cobalt
    Checkpointing libraries available:       
    Demux engines available:                 poll select

Maybe the problem is that you don't free() your memory? Did you try that?

In general, when using mpi, try allocating the 2D dynamic array in contiguous memory cells (so that can MPI can freely use its stride, etc.). In general you can do this with these functions:

int** allocate2D(int** A, const int N, const int M) {
    int i;
    int *t0;

    A = malloc(M * sizeof (int*)); /* Allocating pointers */
    t0 = malloc(N * M * sizeof (int)); /* Allocating data */
    for (i = 0; i < M; i++)
        A[i] = t0 + i * (N);

    return A;
}

void free2Darray(int** p, const int N) {
    free(p[0]);
    free(p);
}

as I explain in 2D dynamic array in continuous memory locations (C) .

Unrelated to your runtime error: Why do we need to use `int main` and not `void main` in C++?

Isn't it the same as what I've done? Just `mem` instead of `t0`. Also, thanks for the int vs void page. Also, what could I free without messing up my data? This program obviously does more things than that. — Teo Alivanoglou, Dec 09 '16 at 19:33
Oops @TeoAlivanoglou you do pretty much the same thing, but with the code lines swapped, sorry. Well, you generally `free()` your memory when you no longer needed it, in other words when you don't care about the data. You could that now to see if the complete minimal example runs. I tried in the machines of my university and it worked fine, even without free'ing the memory! — gsamaras, Dec 09 '16 at 19:40
Wow. I even copy-pasted your code, just to make sure it's 100% the same. I still get the error. Is there any chance that my mpi or gcc are broken? Or maybe VirtualBox limitations? — Teo Alivanoglou, Dec 09 '16 at 19:43
I think VirtualBox limitation @TeoAlivanoglou, see my update. — gsamaras, Dec 09 '16 at 19:45
I can't check if malloc succeded. Segmentation fault seems to occur before `malloc()` returns. — Teo Alivanoglou, Dec 09 '16 at 19:51
@TeoAlivanoglou Do you even have to `malloc`? In the code you've shown, you don't. You know `N`, so you know the size of everything at compile time. — yano, Dec 09 '16 at 20:03
@yano I normally don't know `N`. It has to be dynamic. But I used `N=4;` just for testing purposes. — Teo Alivanoglou, Dec 09 '16 at 20:08
@gsamaras I tried it in a normal linux machine. Still no luck. Any chance you can tell me which implementation of mpi you're using? — Teo Alivanoglou, Dec 09 '16 at 20:39
@TeoAlivanoglou I updated my answer, good luck, I will upvote your question so that more people come into play, if you find the answer helpful do the same. :) — gsamaras, Dec 09 '16 at 20:50
@gsamaras Thank you. I'm installing `mpich` as well. I have a feeling it's all `openmpi`'s fault — Teo Alivanoglou, Dec 09 '16 at 21:01

Segmentation fault when allocating second array

1 Answers1