#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <stdatomic.h>
#include <inttypes.h>

typedef struct {
    unsigned int thread_num;
    unsigned int values_cnt;
} arguments;

typedef struct Item {
  _Atomic (struct Item *)next;
  int64_t  value;
} item_t;

_Atomic (item_t*) head = NULL;

void add_value(int64_t  value) {
   item_t* new_item = (item_t*) malloc(sizeof(item_t));
   new_item->next = atomic_exchange(&head, new_item);
   new_item->value = value;
}

void* add_values(void* arg) {
    arguments args = * (arguments*)arg;
    for (int64_t i = (args.thread_num + 1)*args.values_cnt-1; i >= args.thread_num*args.values_cnt; i--) {
        add_value(i);
        sched_yield();
    }
}

int main(int argc, char* argv[]) {
    int thread_cnt = atoi(argv[1]);
    int values_cnt = atoi(argv[2]);

    item_t* last_item = (item_t*) malloc(sizeof(item_t));
    last_item->next = NULL;
    last_item->value = -1;
    atomic_store(&head, last_item);

    pthread_t thread[thread_cnt];
    arguments args_to_pass[thread_cnt];

    pthread_attr_t threads_attr;
    int page_size = 4096;
    pthread_attr_init(&threads_attr);
    pthread_attr_setstacksize(&threads_attr, page_size);

    for (int  i = 0; i < thread_cnt; i++) {
        args_to_pass[i].thread_num = i;
        args_to_pass[i].values_cnt = values_cnt;
        pthread_create(&thread[i], &threads_attr, add_values, (void*) &args_to_pass[i]);
    }

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

    pthread_attr_destroy(&threads_attr);

    for (item_t* i = atomic_load(&head); i->value != -1;) {
        printf("%" PRId64 "\n", i->value);
        item_t* cur = i;
        i = cur->next;
        free(cur);
    }
    free(last_item);
}