#if 0
/* done in Makefile */
#define _GNU_SOURCE        /* or _BSD_SOURCE or _SVID_SOURCE */
#endif

#include "asm/unistd.h"		/* from kernel source tree */
#include <unistd.h>		/* for syscall */

#include <sys/ioctl.h>
#include <stdio.h>
#include <stdlib.h>
#include <sched.h>
#include <stdint.h>	/* rt_param needs uint32 */

#include "../../litmus-rt/include/linux/perf_event.h"

#include <litmus/rt_param.h> /* page size macro */

/* from kernel tools/perf/perf.h */
static inline int sys_perf_event_open(struct perf_event_attr *attr, pid_t pid,
		int cpu, int group_fd, unsigned long flags)
{
        attr->size = sizeof(*attr);
	return syscall(__NR_perf_event_open, attr, pid, cpu, group_fd, flags);
}

#define C(x)	(PERF_COUNT_HW_CACHE_##x)
#define ATTR_CONFIG_CACHE(cache, op, result) \
	(((C(cache)  & 0xffULL) <<  0) | \
	 ((C(op)     & 0xffULL) <<  8) | \
	 ((C(result) & 0xffULL) << 16))

#define ATTR_CONFIG(event, umask)   \
	((((event) & 0xffULL) << 0) | \
	 (((umask) & 0xffULL) << 8))

struct perf_event_attr perf_event_attr = {
	.type            = 0,     /* set per initilized event */
	.size            = 0,     /* set later */
	.config          = 0,     /* set per initilized event */
	{ .sample_period = 0, }, /* is a counter, so no period */
	.disabled        = 0,     /* event is enabled */
	.inherit         = 0,     /* children don't inherit */
	.pinned          = 0,     /* set per initilized event */
	.exclusive       = 0,     /* set per initilized event */
	.exclude_user    = 0,     /* don't count user      */
	.exclude_kernel  = 0,     /* ditto kernel          */
	.exclude_hv      = 0,     /* ditto hypervisor      */
	.exclude_idle    = 0,     /* don't count when idle */
	.mmap            = 0,     /* include mmap data     */
	.comm            = 0,     /* include comm data     */
};

/* Pound */
#define NR_CPUS		4
#define CACHE_SIZE_MB	8
#define ASSOC		16
#define LINE_SIZE	64
#define CACHE_SIZE	(CACHE_SIZE_MB * 1024 * 1024)

/* arena size in bytes */
//#define ARENA_SIZE	(CACHE_SIZE * 14 / 16)
#define ARENA_SIZE	(CACHE_SIZE * 1)

/* number of pages in arena */
#define ARENA_PAGES	(ARENA_SIZE / PAGE_SIZE)

/* number of cache lines per page */
#define PAGE_LINES	(PAGE_SIZE / LINE_SIZE)

/* number of cache lines in arena */
#define ARENA_LINES	(ARENA_SIZE / LINE_SIZE)

/* number of integers in arena */
#define ARENA_INTS	(ARENA_SIZE / sizeof(int))

/* number of integers in a page */
#define PAGE_INTS	(PAGE_SIZE / sizeof(int))

/* number of integers in a cache line */
#define LINE_INTS	(LINE_SIZE / sizeof(int))

/* convert page number and cache line number to an integer index */
#define PAGE_AND_LINE_TO_IDX(page, line) \
	(((page) * PAGE_INTS) + ((line) * LINE_INTS))


/* not really a good way to do this */
inline int randrange(const int max)
{
	return (rand() / (RAND_MAX / max + 1));
}

void sequential(int *items, const int len)
{
	int i;
	for (i = 0; i < len; i++)
		items[i] = (i + 1) % len;
}

/* Sattolo's algorithm makes a cycle. */
void sattolo(int *items, const int len)
{
	int i;
	for (i = 0; i < len; i++)
		items[i] = i;
	while (1 < i--) {
		/* 0 <= j < i */
		int t, j = randrange(i);
		t = items[i];
		items[i] = items[j];
		items[j] = t;
	}
}

/*
 * Write the order to read the arena into the arena. Each page in the arena is
 * read back, but the page is read in a random order to prevent the prefetcher
 * from working.
 */
static void init_arena_page_line_order(int *arena, int *page_line_order)
{
	int cur_page;
	for (cur_page = 0; cur_page < ARENA_PAGES; cur_page++) {
		/* for each page in the arena */
		int cur_line;
		for (cur_line = 0; cur_line < PAGE_LINES; cur_line++) {
			/* for each line in the page */
			const int idx = PAGE_AND_LINE_TO_IDX(cur_page,
					cur_line);
			const int next_line = page_line_order[cur_line];
			int next_idx = PAGE_AND_LINE_TO_IDX(cur_page,
					next_line);

			if (!next_line) {
				/* special case: last line in the page */
				if (cur_page < ARENA_PAGES - 1) {
					/* arena has more pages: go to next */
					next_idx = PAGE_AND_LINE_TO_IDX(
							(cur_page + 1), 0);
				} else {
					/* the very last element */
					next_idx = 0;
				}
			}
			arena[idx] = next_idx;
		}
	}
}

static int loop_once(const int perf_fd, int *arena)
{
	int i = 0, j;
	do {
		i = arena[i];
		j = i;
	} while (i);
	return j;
}

static int set_affinity(int cpu)
{
	cpu_set_t cpu_set;
	CPU_ZERO(&cpu_set);
	CPU_SET(cpu, &cpu_set);
	return sched_setaffinity(0, sizeof(cpu_set_t), &cpu_set);
}

struct perf_fd {
	int			fd;
	char			*name;
	enum perf_type_id	type;
	__u64			config;
	__u64			exclusive    :  1,
				pinned       :  1,
				__reserved_1 : 62;
};

#define PERF_FD_EMPTY(p) \
	((p)->fd == 0 && (p)->name == NULL && \
	 (p)->type == 0 && (p)->config == 0)
#define PERF_FD_NON_EMPTY(p) (!PERF_FD_EMPTY(p))


#if 0
/* these events are always zero */
static struct perf_fd perf_fds[] = {
	{
		.fd	   = -1,
		.name	   = "MEM_UNCORE_RETIRED.REMOTE_CACHE_LOCAL_HOME_HIT",
		.type	   = PERF_TYPE_RAW,
		.config	   = ATTR_CONFIG(0x0f, 0x08),
		.exclusive = 0,
		.pinned    = 0,
	},
	{
		.fd	= -1,
		.name	= "MEM_UNCORE_RETIRED.REMOTE_DRAM",
		.type	= PERF_TYPE_RAW,
		.config	= ATTR_CONFIG(0x0f, 0x10),
		.exclusive = 0, /* child events cannot be exclusive */
		.pinned    = 0, /* child events cannot be pinned */
	},
	{ },
};
#endif

static struct perf_fd perf_fds[] = {
	/* first element is assumed to be group leader */
#if 0
	{
		.fd	= -1,
		.name	= "MEM_UNCORE_RETIRED.LOCAL_DRAM",
		.type	= PERF_TYPE_RAW,
		.config	= ATTR_CONFIG(0x0f, 0x20),
		.exclusive = 1, /* group leader is scheduled exclusively */
		.pinned    = 1, /* group leader is pinnned to CPU (always on) */
	},
#endif
	{
		.fd	= -1,
		.name	= "L2_RQSTS.PREFETCH_HIT",
		.type	= PERF_TYPE_RAW,
		.config	= ATTR_CONFIG(0x24, 0x40),
#if 0
		.exclusive = 0,
		.pinned    = 0,
#endif
		.exclusive = 1, /* group leader is scheduled exclusively */
		.pinned    = 1, /* group leader is pinnned to CPU (always on) */
	},
	{
		.fd	= -1,
		.name	= "L2_RQSTS.PREFETCH_MISS",
		.type	= PERF_TYPE_RAW,
		.config	= ATTR_CONFIG(0x24, 0x80),
		.exclusive = 0,
		.pinned    = 0,
	},
	{
		.fd	= -1,
		.name	= "MEM_LOAD_RETIRED.L3_MISS",
		.type	= PERF_TYPE_RAW,
		.config	= ATTR_CONFIG(0xcb, 0x10),
		.exclusive = 0,
		.pinned    = 0,
	},
	{
		.fd	   = -1,
		.name	   = "Off Core Response Counter",
		.type	   = PERF_TYPE_HW_CACHE,
		.config	   = ATTR_CONFIG_CACHE(LL, OP_READ, RESULT_MISS),
#if 0
		/* read misses */
		.config	= ATTR_CONFIG_CACHE(LL, OP_READ, RESULT_MISS),
		/* write misses */
		.config	= ATTR_CONFIG_CACHE(LL, OP_WRITE, RESULT_MISS),
		/* prefetch misses */
		.config	= ATTR_CONFIG_CACHE(LL, OP_PREFETCH, RESULT_MISS),
#endif
		.exclusive = 0,
		.pinned    = 0,
	},
	{ },
};


static inline void events_ioctl(const int request)
{
	ioctl(perf_fds[0].fd, request);
}

static void do_read(double divide)
{
	struct perf_fd *perf_fd;
	for (perf_fd = perf_fds; PERF_FD_NON_EMPTY(perf_fd); perf_fd++) {
		__u64 perf_val;
		ssize_t ret;
		ret = read(perf_fd->fd, &perf_val, sizeof(perf_val));
		if (0 >= ret)
			printf("%50s: ERROR\n", perf_fd->name);
		else
			printf("%50s: %10.3f\n",
					perf_fd->name, (perf_val / divide));
		ioctl(perf_fd->fd, PERF_EVENT_IOC_RESET);
	}
}
  
static void write_global_perf_attr(struct perf_fd *perf_fd)
{
		perf_event_attr.type      = perf_fd->type;
		perf_event_attr.config    = perf_fd->config;
		perf_event_attr.exclusive = perf_fd->exclusive;
		perf_event_attr.pinned    = perf_fd->pinned;
}

#define CPU 0
static int setup_perf(void)
{
	/* cannot have pid == -1 and cpu == -1 */
	const int perf_pid	= -1;	/* -1: all tasks, 0: this task */
	const int perf_cpu	= CPU;	/* -1: all CPUs (follow task) */
	struct perf_fd *perf_fd;
	int err = 0;

	for (perf_fd = perf_fds; PERF_FD_NON_EMPTY(perf_fd); perf_fd++) {
		/* make a group whose leader is the zeroth element */
		const int perf_group = perf_fds[0].fd;

		/* setup the attributes to pass in */
		write_global_perf_attr(perf_fd);

		perf_fd->fd = sys_perf_event_open(&perf_event_attr, perf_pid,
				perf_cpu, perf_group, 0);

		if (0 > perf_fd->fd) {
			fprintf(stderr, "could not setup %s\n", perf_fd->name);
			err = -1;
			goto out;
		}
	}
out:
	return err;
}

int main(int argc, char **argv)
{

	const int task_cpu = CPU;
	int ret = 0, i;
	int *arena, *page_line_order;

	if (set_affinity(task_cpu)) {
		fprintf(stderr, "could not set affinity\n");
		ret = -1;
		goto out;
	}

	arena = malloc(ARENA_SIZE);
	if (!arena) {
		fprintf(stderr, "could not allocate memory\n");
		ret = -1;
		goto out;
	}

	page_line_order = malloc(PAGE_LINES * sizeof(*page_line_order));
	if (!page_line_order) {
		fprintf(stderr, "could not allocate memory\n");
		ret = -1;
		goto out;
	}

	sattolo(page_line_order, PAGE_LINES);
	//sequential(page_line_order, PAGE_LINES);
	init_arena_page_line_order(arena, page_line_order);

	if (setup_perf()) {
		ret = -1;
		goto out;
	}

	printf("arena_size:  %d\n", ARENA_SIZE);
	printf("arena_lines: %d\n", ARENA_LINES);

	printf("initially\n");
	do_read(1.0);

	events_ioctl(PERF_EVENT_IOC_ENABLE);
	loop_once(perf_fds[0].fd, arena);
	events_ioctl(PERF_EVENT_IOC_DISABLE);
	printf("after a loop\n");
	do_read(1.0);

	events_ioctl(PERF_EVENT_IOC_ENABLE);
	loop_once(perf_fds[0].fd, arena);
	events_ioctl(PERF_EVENT_IOC_DISABLE);
	printf("after another loop\n");
	do_read(1.0);

	events_ioctl(PERF_EVENT_IOC_ENABLE);
	loop_once(perf_fds[0].fd, arena);
	events_ioctl(PERF_EVENT_IOC_DISABLE);
	printf("after another loop\n");
	do_read(1.0);

	events_ioctl(PERF_EVENT_IOC_ENABLE);
	loop_once(perf_fds[0].fd, arena);
	events_ioctl(PERF_EVENT_IOC_DISABLE);
	printf("after another loop\n");
	do_read(1.0);

	events_ioctl(PERF_EVENT_IOC_ENABLE);
	for (i = 0; i < 100; i++)
		loop_once(perf_fds[0].fd, arena);
	events_ioctl(PERF_EVENT_IOC_DISABLE);
	printf("after 100 loops\n");
	do_read(100.0);

out:
	return ret;
}