#include <stdlib.h>
#include <stdio.h>
#include <string.h>

extern "C" {
	#include <locale.h>
	#include <getopt.h>
	#include <fcntl.h>
	#include <sys/uio.h>
}

#include <unistd.h>

#include <clock.hfa>
#include <kernel.hfa>
#include <thread.hfa>
#include <time.hfa>

extern bool traceHeapOn();
extern ssize_t cfa_preadv2(int fd, const struct iovec *iov, int iovcnt, off_t offset, int flags);

int fd;
volatile bool run = false;
volatile size_t count = 0;

unsigned long int buflen = 50;

cluster * the_cluster;

thread Reader {};
void ?{}( Reader & this ) {
	((thread&)this){ "Reader Thread", *the_cluster };
}

struct my_processor {
	processor p;
};

void ?{}( my_processor & this ) {
	(this.p){ "I/O Processor", *the_cluster };
}

void main( Reader & ) {
	while(!__atomic_load_n(&run, __ATOMIC_RELAXED)) yield();

	char data[buflen];
	struct iovec iov = { data, buflen };

	while(__atomic_load_n(&run, __ATOMIC_RELAXED)) {
		int r = cfa_preadv2(fd, &iov, 1, 0, 0);
		if(r < 0) abort("%s\n", strerror(-r));

		__atomic_fetch_add( &count, 1, __ATOMIC_SEQ_CST );
	}
}

int main(int argc, char * argv[]) {
	double duration   = 5.0;
	unsigned long int nthreads = 2;
	unsigned long int nprocs   = 1;
	unsigned flags = 0;
	unsigned sublen = 16;

	arg_loop:
	for(;;) {
		static struct option options[] = {
			{"duration",     required_argument, 0, 'd'},
			{"nthreads",     required_argument, 0, 't'},
			{"nprocs",       required_argument, 0, 'p'},
			{"bufsize",      required_argument, 0, 'b'},
			{"userthread",   no_argument      , 0, 'u'},
			{"submitthread", no_argument      , 0, 's'},
			{"submitlength", required_argument, 0, 'l'},
			{0, 0, 0, 0}
		};

		int idx = 0;
		int opt = getopt_long(argc, argv, "d:t:p:b:usl:", options, &idx);

		const char * arg = optarg ? optarg : "";
		char * end;
		switch(opt) {
			// Exit Case
			case -1:
				break arg_loop;
			// Numeric Arguments
			case 'd':
				duration = strtod(arg, &end);
				if(*end != '\0') {
					fprintf(stderr, "Duration must be a valid double, was %s\n", arg);
					goto usage;
				}
				break;
			case 't':
				nthreads = strtoul(arg, &end, 10);
				if(*end != '\0' || nthreads < 1) {
					fprintf(stderr, "Number of threads must be a positive integer, was %s\n", arg);
					goto usage;
				}
				break;
			case 'p':
				nprocs = strtoul(arg, &end, 10);
				if(*end != '\0' || nprocs < 1) {
					fprintf(stderr, "Number of processors must be a positive integer, was %s\n", arg);
					goto usage;
				}
				break;
			case 'b':
				buflen = strtoul(arg, &end, 10);
				if(*end != '\0' && buflen < 10) {
					fprintf(stderr, "Buffer size must be at least 10, was %s\n", arg);
					goto usage;
				}
				break;
			case 'u':
				flags |= CFA_CLUSTER_IO_POLLER_USER_THREAD;
				break;
			case 's':
				flags |= CFA_CLUSTER_IO_POLLER_THREAD_SUBMITS;
				break;
			case 'l':
				sublen = strtoul(arg, &end, 10);
				if(*end != '\0' && sublen < 16) {
					fprintf(stderr, "Submit length must be at least 16, was %s\n", arg);
					goto usage;
				}
				flags |= (sublen << CFA_CLUSTER_IO_BUFFLEN_OFFSET);
				break;
			// Other cases
			default: /* ? */
				fprintf(stderr, "%d\n", opt);
			usage:
				fprintf(stderr, "Usage: %s : [options]\n", argv[0]);
				fprintf(stderr, "\n");
				fprintf(stderr, "  -d, --duration=DURATION  Duration of the experiment, in seconds\n");
				fprintf(stderr, "  -t, --nthreads=NTHREADS  Number of user threads\n");
				fprintf(stderr, "  -p, --nprocs=NPROCS      Number of kernel threads\n");
				fprintf(stderr, "  -b, --buflen=SIZE        Number of bytes to read per request\n");
				fprintf(stderr, "  -u, --userthread         If set, cluster uses user-thread to poll I/O\n");
				fprintf(stderr, "  -s, --submitthread       If set, cluster uses polling thread to submit I/O\n");
				exit(EXIT_FAILURE);
		}
	}

	fd = open(__FILE__, 0);
	if(fd < 0) {
		fprintf(stderr, "Could not open source file\n");
		exit(EXIT_FAILURE);
	}

	printf("Running %lu threads, reading %lu bytes each, over %lu processors for %lf seconds\n", nthreads, buflen, nprocs, duration);

	{
		Time start, end;
		cluster cl = { "IO Cluster", flags };
		the_cluster = &cl;
		#if !defined(__CFA_NO_STATISTICS__)
			print_stats_at_exit( cl );
		#endif
		{
			my_processor procs[nprocs];
			{
				Reader threads[nthreads];

				printf("Starting\n");
				start = getTime();
				run = true;
				do {
					sleep(500`ms);
					end = getTime();
				} while( (end - start) < duration`s );
				run = false;
				end = getTime();
				printf("Done\n");
			}
		}
		printf("Took %'ld ms\n", (end - start)`ms);
		printf("Total reads      : %'15zu\n", count);
		printf("Reads per second : %'18.2lf\n", ((double)count) / (end - start)`s);
		printf("Total read size  : %'15zu\n", buflen * count);
		printf("Bytes per second : %'18.2lf\n", ((double)count * buflen) / (end - start)`s);
	}

	close(fd);
}