fork exec and mmap issues

Question

For the application I'm developing (under Linux, but I'm trying to maintain portability) I need to switch to shared memory for sharing data across different processes (and threads inside processes). There is a father process generating different children

I need for example to get every process able to increment a shared counter using a named semaphore.

In this case everything is ok:

#include <sys/mman.h>
#include <sys/wait.h>
#include <semaphore.h>
#include <fcntl.h>
#include <iostream>
#include <stdlib.h>
#include <string.h>
using namespace std;

#define SEM_NAME "/mysem"
#define SM_NAME "tmp_sm.txt"

int main(){
    int fd, nloop, counter_reset;
    int *smo;
    sem_t *mutex;

    nloop = 100;
    counter_reset = 1000;

    if (fork() == 0) {
        /* child */
        /* create, initialize, and unlink semaphore */
        mutex = sem_open(SEM_NAME, O_CREAT, 0777, 1);
        //sem_unlink(SEM_NAME);
        /* open file, initialize to 0, map into memory */
        fd = open(SM_NAME, O_RDWR | O_CREAT);
        smo = (int *) mmap(NULL, sizeof(int), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
        close(fd);
        /* INCREMENT */
        for (int i = 0; i < nloop; i++) {
            sem_wait(mutex);
            cout << "child: " << (*smo)++ << endl;
            if(*smo>=counter_reset){
                (*smo)=0;
            }
            sem_post(mutex);
        }
        exit(0);
    }
    /* parent */
    /* create, initialize, and unlink semaphore */
    mutex = sem_open(SEM_NAME, O_CREAT, 0777, 1);
    sem_unlink(SEM_NAME);
    /* open file, initialize to 0, map into memory */
    fd = open(SM_NAME, O_RDWR | O_CREAT);
    smo = (int *) mmap(NULL, sizeof(int), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    close(fd);
    /* INCREMENT */
    for (int i = 0; i < nloop; i++) {
        sem_wait(mutex);
        cout << "parent: " << (*smo)++ << endl;
        if(*smo>=counter_reset){
            (*smo)=0;
        }
        sem_post(mutex);
    }
    exit(0);
}

So far so good: both semaphore and shared counter are ok (same address in memory) and increment and reset work fine.

The program fails simply by moving child source code into a new source file invoked by exec. Shared memory and named semaphore addresses are different therefore increment fails.

Any suggestion? I used named semaphores and named shared memory (using a file) to try to get the same pointer values.

---

UPDATE:

as requested by Joachim Pileborg, this is the "server side" improvements respect above original code:

...
if (fork() == 0) {
    /* child */
    /*spawn child by execl*/
    char cmd[] = "/path_to_bin/client";
    execl(cmd, cmd, (char *)0);
    cerr << "error while istantiating new process" << endl;
    exit(EXIT_FAILURE);
}
...

And this is the "client" source code:

#include <sys/mman.h>
#include <sys/wait.h>
#include <semaphore.h>
#include <fcntl.h>
#include <iostream>
#include <stdlib.h>
using namespace std;

#define SEM_NAME "/mysem"
#define SM_NAME "tmp_ssm.txt"

int main(){
    int nloop, counter_reset;
    int *smo;
    sem_t *mutex;
    /* create, initialize, and unlink semaphore */
    mutex = sem_open(SEM_NAME, O_CREAT, 0777, 1);
    //sem_unlink(SEM_NAME);
    /* open file, initialize to 0, map into memory */
    int fd = open(SM_NAME, O_RDWR | O_CREAT);
    smo = (int *) mmap(NULL, sizeof(int), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    close(fd);
    nloop=100;
    counter_reset=1000;
    /* INCREMENT */
    for (int i = 0; i < nloop; i++) {
        sem_wait(mutex);
        cout << "child: " << (*smo)++ << endl;
        if(*smo>=counter_reset){
            (*smo)=0;
        }
        sem_post(mutex);
    }
    exit(0);
}

executing this code cause the process to block (deadlock) and waiting for an infinite time. looking at addresses they are tipically found to be:

father semaphore: 0x7f2fe1813000
child semahpore: 0x7f0c4c793000
father shared memory: 0x7f2fe1811000
child shared memory: 0x7ffd175cb000

removing 'sem_post' and 'sem_wait' everything is fine but I need mutual exlusion while incrementing...

Answer 1

Don't unlink the semaphore. it actually removes the semaphore.

From the sem_unlink manual page:

sem_unlink() removes the named semaphore referred to by name. The semaphore name is removed immediately. The semaphore is destroyed once all other processes that have the semaphore open close it.

This means that once you've created the semaphore in the parent process, you immediately remove it. The child process then will not be able to find the semaphore, and instead creates a new one.