#include <numaif.h>
#include <sys/mman.h>
#include <sys/time.h>
#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>

#define BITS_PER_LONG 64
#define MAX_THREADS 256

static void *p;
static int maxnode;
static size_t per_node_len;

pthread_t threads[MAX_THREADS];

void *alloc_memory(void *arg)
{
	size_t i;
	int v, r, ret;
	unsigned long nid = (unsigned long)arg;
	unsigned long nodemask[4];
	void *q = (char *)p + per_node_len * nid;

	nodemask[0] = nodemask[1] = nodemask[2] = nodemask[3] = 0;
	v = nid / BITS_PER_LONG;
	r = nid % BITS_PER_LONG;
	nodemask[v] |= 1UL << r;
	printf("node=%d, nodemask=0x%llx 0x%llx 0x%llx 0x%llx\n", nid,
			nodemask[0], nodemask[1], nodemask[2], nodemask[3]);
	ret = mbind(q, per_node_len, MPOL_BIND, nodemask, maxnode, 0);
	if (ret == -1) {
		perror("mbind");
		return (void *)-1;
	}
	for (i = 0; i < per_node_len; i += 0x1000)
		*((char *)q + i) = nid;

	return NULL;
}

int main(int argc, char *argv[])
{
	int ret, node_nr;
	struct timeval t1, t2;
	unsigned long i, deltaus;

	if (argc != 3) {
		fprintf(stderr, "usage: ./node_alloc node_nr per_node_len_in_MB\n");
		return 0;
	}

	node_nr = atoi(argv[1]);
	per_node_len = atoi(argv[2]);
	printf("node_nr=%d, per_node_len=%dMB\n", node_nr, per_node_len);
	per_node_len <<= 20;
	printf("per_node_len=0x%lx\n", per_node_len);
	
	p = mmap(NULL, node_nr * per_node_len, PROT_READ | PROT_WRITE,
			MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
	if (p == MAP_FAILED) {
		perror("mmap");
		return -1;
	}

	maxnode = node_nr + 1;
	for (i = 0; i < node_nr; i++)
		pthread_create(&threads[i], NULL, alloc_memory, (void *)i);

	for (i = 0; i < node_nr; i++)
		pthread_join(threads[i], NULL);

	printf("allocation done, press enter to start free\n");
	getchar();

	gettimeofday(&t1, NULL);
	munmap(p, node_nr * per_node_len);
	gettimeofday(&t2, NULL);
	deltaus = (t2.tv_sec - t1.tv_sec) * 1000000 + t2.tv_usec - t1.tv_usec;
	printf("time spent: %fs\n", (float)deltaus / 1000000);

	return 0;
}