shm_open.c

/* # shm_open
 *
 * # shm_unlink
 *
 * Like open, but creates a fake file that can be passed to mmap.
 *
 * The file is never written to disk.
 *
 * Application: let two processes share memory.
 *
 * Requires `-lrt` for Linux.
 *
 * ## shm_open vs mmap
 *
 * There is also the Linux specific `MAP_ANONYMOUS` `mmap` flag which seems
 * to reach a similar effect, but with an unamed `shm_mem`. TODO check.
 *
 * - https://stackoverflow.com/questions/21311080/linux-shared-memory-shmget-vs-mmap
 * - https://stackoverflow.com/questions/5656530/how-to-use-shared-memory-with-linux-in-c
 * - https://stackoverflow.com/questions/16560401/anonymous-shared-memory
 */

#include "common.h"

int main(void) {
    int fd;
    int *map;
    char *name = "/" TMPFILE();
    enum Constexpr { SIZE = sizeof(*map) };

    /* The name creates a virtual file under `/dev/shm` in Linux.
     *
     * There seems to be no POSIX way to list the objects. E.g. in FreeBDS there isn't:
     * http://stackoverflow.com/questions/12430351/how-to-get-list-of-open-posix-shared-memory-segments-in-freebsd
     */
    fd = shm_open(name, O_RDWR | O_CREAT | O_TRUNC, S_IRUSR | S_IWUSR);
    /* Without ftruncate Linux gives SIGBUS. */
    ftruncate(fd, SIZE);
    map = mmap(NULL, SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    *map = 0;
    if (fork() == 0) {
        *map = 1;
        /* `man msync` says that only other thing that guarantees write is munmap.
         *
         * So I think `wait` is not enough for memory synchronization.
         */
        msync(map, SIZE, MS_SYNC);
        exit(EXIT_SUCCESS);
    }
    wait(NULL);
    /* Memory changed across processes! */
    assert(*map == 1);
    /* You *must* do this! If you do not, the object survive (under /dev/shm in Linux). */
    munmap(map, SIZE);
    close(fd);
    shm_unlink(name);
    return EXIT_SUCCESS;
}