#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <fcntl.h>
#include <sys/mman.h>
#include <sys/wait.h>
#include <string.h>
#include <unistd.h>
#include <stdbool.h>
#include <pthread.h>

#define LEN	(1UL*512*1024*1024)

typedef long double t;

static t gett(void)
{
	struct timespec ts;
	if (clock_gettime(CLOCK_MONOTONIC, &ts))
		perror("clock_gettime"), exit(1);
	return (t)ts.tv_sec + (t)ts.tv_nsec * (t)1./(t)1000000000.;
}

static int h;
static char *m;

#define m_l	((unsigned long *)m)
#define LEN_L	(LEN / sizeof(unsigned long))

#define m_p	((void **)m)
#define LEN_P	(LEN / sizeof(void *))

static int n_threads;
static int tid = -1;
static bool mt = false;

static pthread_barrier_t *bar;
static pthread_barrierattr_t bar_attr;

static long double *global;

static void map(int argc, char **argv)
{
	if (argc < 2)
		fprintf(stderr, "args\n"), exit(1);
	h = open(argv[1], O_RDWR);
	if (h < 0)
		perror(argv[1]), exit(1);

	if (argc >= 3) {
		int i;
		n_threads = atoi(argv[2]);
		if (n_threads <= 0)
			fprintf(stderr, "threads\n"), exit(1);

		mt = true;

		global = mmap(NULL, n_threads * sizeof(long double), PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
		if (global == MAP_FAILED)
			perror("mmap"), exit(1);

		bar = mmap(NULL, sizeof(pthread_barrier_t), PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
		if (bar == MAP_FAILED)
			perror("mmap"), exit(1);
		if (pthread_barrierattr_init(&bar_attr))
			fprintf(stderr, "pthread_barrierattr_init\n"), exit(1);
		if (pthread_barrierattr_setpshared(&bar_attr, PTHREAD_PROCESS_SHARED))
			fprintf(stderr, "pthread_barrierattr_setpshared\n"), exit(1);
		if (pthread_barrier_init(bar, &bar_attr, n_threads))
			fprintf(stderr, "pthread_barrier_init\n"), exit(1);

		for (i = 1; i < n_threads; i++) {
			if (!fork()) {
				tid = i;
				goto child;
			}
		}
		tid = 0;
	}

child:
	m = mmap(NULL, LEN, PROT_READ | PROT_WRITE, MAP_SHARED_VALIDATE, h, LEN * (getpid() & 63));
	if (m == MAP_FAILED)
		perror("mmap"), exit(1);
	memset(m, 0, LEN);

	if (mt) {
		int r = pthread_barrier_wait(bar);
		if (r > 0)
			fprintf(stderr, "pthread_barrier_wait: %s\n", strerror(r)), exit(1);
	}
}

#define start_loop			\
{					\
	long double lat, thrpt;		\
	__attribute__((unused)) unsigned long i;		\
	unsigned long loops = 0;	\
	t t1 = gett(), t2;		\
	do {

#define end_loop(q, q2)							\
		loops++;						\
		t2 = gett();						\
	} while (t2 - t1 < 5);						\
	lat = ((t2 - t1) / loops / (q)) * 1000000000.;			\
	thrpt = (t)(q2) * loops / (t2 - t1) / 1000000000.;		\
	if (mt) global[tid] = thrpt;					\
	printf("thr: %Lf GB/s, lat: %Lf nsec\n", thrpt, lat);		\
	/*printf("loops: %lu, time: %Lf, q2: %lu\n", loops, t2 - t1, q2);*/\
	while (wait(NULL) > 0);						\
	if (!tid && mt) {						\
		long double thrpt_sum = 0;				\
		int i;							\
		for (i = 0; i < n_threads; i++)				\
			thrpt_sum += global[i];				\
		printf("sum: %Lf GB/s\n", thrpt_sum);			\
	}								\
}