again

#include <stdio.h>
#include <pthread.h>

typedef struct {
    unsigned long state;
} apc_lock_t;

int apc_lock_init(apc_lock_t* lock)
{
    lock->state = 0;
}

int apc_lock_try(apc_lock_t* lock)
{
    int failed = 1;

    asm volatile
    (
        "xchgl %0, 0(%1)" :
        "=r" (failed) : "r" (&lock->state),
        "0" (failed)
    );

    return failed;
}

int apc_lock_get(apc_lock_t* lock)
{
    int failed = 1;

    do {
        failed = apc_lock_try(
            lock);
        usleep(0);
    } while (failed);

    return failed;
}

int apc_lock_release(apc_lock_t* lock)
{
    int released = 0;

    asm volatile (
        "xchg %0, 0(%1)" : "=r" (released) : "r" (&lock->state),
        "0" (released)
    );

    return !released;
}

void* pthread_test(void *input) {
    apc_lock_t* lock = (apc_lock_t*) input;

    if (lock) {
        int runs = 0;

        while (runs++ < 10) {
            printf("%lu (%d) get lock: %d\n", pthread_self(), runs, apc_lock_get(lock));
            printf("%lu has lock ...\n", pthread_self());
            sleep(1);
            printf("%lu (%d) release lock: %d\n", pthread_self(), runs, apc_lock_release(lock));
            fflush(stdout);
        }
    }

    pthread_exit(NULL);
}

int main(char **argv) {
  apc_lock_t lock;

  apc_lock_init(&lock);

  pthread_t threads[8];
  int next = 0;

  while (next < 8) {
    pthread_create(
        &threads[next], NULL,
        pthread_test, (void*) &lock
    );
    next++;
  }

  while (next--) {
    pthread_join(
        threads[next], NULL);
  }

  return 0;
}