[PATCH] fio: more async fixes

[disktools.git] / fio.c

/*
 * fio - the flexible io tester
 *
 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
 *
 *  This program is free software; you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License as published by
 *  the Free Software Foundation; either version 2 of the License, or
 *  (at your option) any later version.
 *
 *  This program is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with this program; if not, write to the Free Software
 *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 *
 */
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <time.h>
#include <ctype.h>
#include <sched.h>
#include <libaio.h>
#include <sys/time.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/wait.h>
#include <semaphore.h>
#include <sys/ipc.h>
#include <sys/shm.h>
#include <asm/unistd.h>

#define MAX_JOBS	(1024)

/*
 * assume we don't have _get either, if _set isn't defined
 */
#ifndef __NR_ioprio_set

#if defined(__i386__)
#define __NR_ioprio_set		289
#define __NR_ioprio_get		290
#elif defined(__powerpc__) || defined(__powerpc64__)
#define __NR_ioprio_set		273
#define __NR_ioprio_get		274
#elif defined(__x86_64__)
#define __NR_ioprio_set		251
#define __NR_ioprio_get		252
#elif defined(__ia64__)
#define __NR_ioprio_set		1274
#define __NR_ioprio_get		1275
#elif defined(__alpha__)
#define __NR_ioprio_set		442
#define __NR_ioprio_get		443
#elif defined(__s390x__) || defined(__s390__)
#define __NR_ioprio_set		282
#define __NR_ioprio_get		283
#else
#error "Unsupported arch"
#endif

#endif

static int ioprio_set(int which, int who, int ioprio)
{
	return syscall(__NR_ioprio_set, which, who, ioprio);
}

enum {
	IOPRIO_WHO_PROCESS = 1,
	IOPRIO_WHO_PGRP,
	IOPRIO_WHO_USER,
};

#define IOPRIO_CLASS_SHIFT	13

#define MASK	(4095)

#define DEF_BS		(4096)
#define DEF_TIMEOUT	(30)
#define DEF_RATE_CYCLE	(1000)
#define DEF_ODIRECT	(1)
#define DEF_SEQUENTIAL	(1)
#define DEF_WRITESTAT	(0)
#define DEF_RAND_REPEAT	(1)

#define ALIGN(buf)	(char *) (((unsigned long) (buf) + MASK) & ~(MASK))

static int sequential = DEF_SEQUENTIAL;
static int write_stat = DEF_WRITESTAT;
static int repeatable = DEF_RAND_REPEAT;
static int timeout = DEF_TIMEOUT;
static int odirect = DEF_ODIRECT;
static int global_bs = DEF_BS;
static int rate_quit = 1;

static int thread_number;
static char *ini_file;

static int shm_id;

static cpu_set_t def_cpumask;

enum {
	DDIR_READ = 0,
	DDIR_WRITE,
};

/*
 * thread life cycle
 */
enum {
	TD_NOT_CREATED = 0,
	TD_CREATED,
	TD_STARTED,
	TD_EXITED,
	TD_REAPED,
};

struct thread_data {
	char file_name[256];
	int thread_number;
	int error;
	int fd;
	int stat_fd;
	pid_t pid;
	volatile int terminate;
	volatile int runstate;
	unsigned int ddir;
	unsigned int ioprio;
	unsigned int sequential;
	unsigned int bs;
	unsigned int odirect;
	unsigned int delay_sleep;
	unsigned int fsync_blocks;
	unsigned int start_delay;
	unsigned int use_aio;
	cpu_set_t cpumask;

	io_context_t *aio_ctx;
	struct iocb *aio_iocbs;
	unsigned int aio_depth;
	unsigned int aio_cur_depth;
	struct io_event *aio_events;
	char *aio_iocbs_status;

	unsigned int rate;
	unsigned int ratemin;
	unsigned int ratecycle;
	unsigned long rate_usec_cycle;
	long rate_pending_usleep;
	unsigned long rate_blocks;
	struct timeval lastrate;

	unsigned long max_latency;	/* msec */
	unsigned long min_latency;	/* msec */
	unsigned long runtime;		/* sec */
	unsigned long blocks;
	unsigned long io_blocks;
	unsigned long last_block;
	sem_t mutex;
	struct drand48_data random_state;

	/*
	 * bandwidth stat
	 */
	unsigned long stat_time;
	unsigned long stat_time_last;
	unsigned long stat_blocks_last;

	struct timeval start;
};

static struct thread_data *threads;

static sem_t startup_sem;

static void sig_handler(int sig)
{
	int i;

	for (i = 0; i < thread_number; i++) {
		struct thread_data *td = &threads[i];

		td->terminate = 1;
		td->start_delay = 0;
	}
}

static int init_random_state(struct thread_data *td)
{
	unsigned long seed = 123;

	if (td->sequential)
		return 0;

	if (!repeatable) {
		int fd = open("/dev/random", O_RDONLY);

		if (fd == -1) {
			td->error = errno;
			return 1;
		}

		if (read(fd, &seed, sizeof(seed)) < (int) sizeof(seed)) {
			td->error = EIO;
			close(fd);
			return 1;
		}

		close(fd);
	}

	srand48_r(seed, &td->random_state);
	return 0;
}

static void shutdown_stat_file(struct thread_data *td)
{
	if (td->stat_fd != -1) {
		fsync(td->stat_fd);
		close(td->stat_fd);
	}
}

static int init_stat_file(struct thread_data *td)
{
	char n[256];

	if (!write_stat)
		return 0;

	sprintf(n, "%s.stat", td->file_name);
	td->stat_fd = open(n, O_WRONLY | O_CREAT | O_TRUNC, 0644);
	if (td->stat_fd == -1) {
		td->error = errno;
		return 1;
	}

	return 0;
}

static unsigned long utime_since(struct timeval *s, struct timeval *e)
{
	double sec, usec;

	sec = e->tv_sec - s->tv_sec;
	usec = e->tv_usec - s->tv_usec;
	if (sec > 0 && usec < 0) {
		sec--;
		usec += 1000000;
	}

	sec *= (double) 1000000;

	return sec + usec;
}

static unsigned long mtime_since(struct timeval *s, struct timeval *e)
{
	double sec, usec;

	sec = e->tv_sec - s->tv_sec;
	usec = e->tv_usec - s->tv_usec;
	if (sec > 0 && usec < 0) {
		sec--;
		usec += 1000000;
	}

	sec *= (double) 1000;
	usec /= (double) 1000;

	return sec + usec;
}

static inline unsigned long msec_now(struct timeval *s)
{
	return s->tv_sec * 1000 + s->tv_usec / 1000;
}

static unsigned long get_next_offset(struct thread_data *td)
{
	unsigned long b;
	long r;

	if (!td->sequential) {
		lrand48_r(&td->random_state, &r);
		b = (1+(double) (td->blocks-1) * r / (RAND_MAX+1.0));
	} else {
		b = td->last_block;
		td->last_block++;
	}

	return b * td->bs;
}

static void add_stat_sample(struct thread_data *td, unsigned long msec)
{
	char sample[256];

	if (!td->stat_fd)
		return;

#if 0
	sprintf(sample, "%lu, %lu\n", td->io_blocks, msec);
	write(td->stat_fd, sample, strlen(sample));
#else
	td->stat_time += msec;
	td->stat_time_last += msec;
	td->stat_blocks_last++;

	if (td->stat_time_last >= 500) {
		unsigned long rate = td->stat_blocks_last * td->bs / (td->stat_time_last);

		td->stat_time_last = 0;
		td->stat_blocks_last = 0;
		sprintf(sample, "%lu, %lu\n", td->stat_time, rate);
		//sprintf(sample, "%lu, %lu\n", td->io_blocks, msec);
		write(td->stat_fd, sample, strlen(sample));
	}
#endif
}

static void usec_sleep(int usec)
{
	struct timespec req = { .tv_sec = 0, .tv_nsec = usec * 1000 };
	struct timespec rem;

	do {
		rem.tv_sec = rem.tv_nsec = 0;
		nanosleep(&req, &rem);
		if (!rem.tv_nsec)
			break;

		req.tv_nsec = rem.tv_nsec;
	} while (1);
}

static void rate_throttle(struct thread_data *td, unsigned long time_spent)
{
	if (!td->rate)
		return;

	if (time_spent < td->rate_usec_cycle) {
		unsigned long s = td->rate_usec_cycle - time_spent;

		td->rate_pending_usleep += s;
		if (td->rate_pending_usleep >= 100000) {
			usec_sleep(td->rate_pending_usleep);
			td->rate_pending_usleep = 0;
		}
	} else {
		long overtime = time_spent - td->rate_usec_cycle;

		td->rate_pending_usleep -= overtime;
	}
}

static int check_min_rate(struct thread_data *td, struct timeval *now)
{
	unsigned long spent = mtime_since(&td->start, now);
	unsigned long rate;

	/*
	 * allow a 2 second settle period in the beginning
	 */
	if (spent < 2000)
		return 0;

	/*
	 * if rate blocks is set, sample is running
	 */
	if (td->rate_blocks) {
		spent = mtime_since(&td->lastrate, now);
		if (spent < td->ratecycle)
			return 0;

		rate = ((td->io_blocks - td->rate_blocks) * td->bs) / spent;
		if (rate < td->ratemin) {
			printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
			if (rate_quit)
				sig_handler(0);
			return 1;
		}
	}

	td->rate_blocks = td->io_blocks;
	memcpy(&td->lastrate, now, sizeof(*now));
	return 0;
}

#define should_fsync(td)	((td)->ddir == DDIR_WRITE && !(td)->odirect)

static void do_sync_io(struct thread_data *td)
{
	struct timeval s, e;
	char *buffer, *ptr;
	unsigned long blocks, msec, usec;

	ptr = malloc(td->bs + MASK);
	buffer = ALIGN(ptr);

	gettimeofday(&td->start, NULL);

	if (td->ratemin)
		memcpy(&td->lastrate, &td->start, sizeof(td->start));

	for (blocks = 0; blocks < td->blocks; blocks++) {
		off_t offset = get_next_offset(td);
		int ret;

		if (td->terminate)
			break;

		if (lseek(td->fd, offset, SEEK_SET) == -1) {
			td->error = errno;
			break;
		}

		if (td->delay_sleep)
			usec_sleep(td->delay_sleep);

		gettimeofday(&s, NULL);

		if (td->ddir == DDIR_READ)
			ret = read(td->fd, buffer, td->bs);
		else
			ret = write(td->fd, buffer, td->bs);

		if (ret < (int) td->bs) {
			if (ret == -1)
				td->error = errno;
			break;
		}

		td->io_blocks++;

		if (should_fsync(td) && td->fsync_blocks &&
		    (td->io_blocks % td->fsync_blocks) == 0)
			fsync(td->fd);

		gettimeofday(&e, NULL);

		usec = utime_since(&s, &e);

		rate_throttle(td, usec);

		if (check_min_rate(td, &e)) {
			td->error = ENODATA;
			break;
		}

		msec = usec / 1000;
		add_stat_sample(td, msec);

		if (msec < td->min_latency)
			td->min_latency = msec;
		if (msec > td->max_latency)
			td->max_latency = msec;
	}

	if (should_fsync(td))
		fsync(td->fd);

	gettimeofday(&e, NULL);
	td->runtime = mtime_since(&td->start, &e);

	free(ptr);
}

static void aio_put_iocb(struct thread_data *td, struct iocb *iocb)
{
	long offset = ((long) iocb - (long) td->aio_iocbs)/ sizeof(struct iocb);

	td->aio_iocbs_status[offset] = 0;
	td->aio_cur_depth--;
}

static struct iocb *aio_get_iocb(struct thread_data *td, char *buffer,
				 struct timeval *t)
{
	struct iocb *iocb = NULL;
	unsigned int i;

	for (i = 0; i < td->aio_depth; i++) {
		if (td->aio_iocbs_status[i] == 0) {
			td->aio_iocbs_status[i] = 1;
			iocb = &td->aio_iocbs[i];
			break;
		}
	}

	if (iocb) {
		off_t off = get_next_offset(td);
		char *p = buffer + i * td->bs;

		if (td->ddir == DDIR_READ)
			io_prep_pread(iocb, td->fd, p, td->bs, off);
		else
			io_prep_pwrite(iocb, td->fd, p, td->bs, off);

		io_set_callback(iocb, (io_callback_t) msec_now(t));
	}

	return iocb;
}

static int aio_submit(struct thread_data *td, struct iocb *iocb)
{
	int ret;

	do {
		ret = io_submit(*td->aio_ctx, 1, &iocb);
		if (ret == 1)
			return 0;

		if (errno == EINTR)
			continue;
		else if (errno == EAGAIN)
			usleep(100);
		else
			break;
	} while (1);

	return 1;
}

#define iocb_time(iocb)	((unsigned long) (iocb)->data)

static void do_async_io(struct thread_data *td)
{
	struct timeval s, e;
	char *buf, *ptr;
	unsigned long blocks, msec, usec;

	ptr = malloc(td->bs * td->aio_depth + MASK);
	buf = ALIGN(ptr);

	gettimeofday(&td->start, NULL);

	if (td->ratemin)
		memcpy(&td->lastrate, &td->start, sizeof(td->start));

	for (blocks = 0; blocks < td->blocks; blocks++) {
		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
		struct timespec *timeout;
		int ret, i, min_evts = 0;
		struct iocb *iocb;

		if (td->terminate)
			break;

		if (td->delay_sleep)
			usec_sleep(td->delay_sleep);

		gettimeofday(&s, NULL);

		iocb = aio_get_iocb(td, buf, &s);

		ret = aio_submit(td, iocb);
		if (ret) {
			td->error = errno;
			break;
		}

		td->aio_cur_depth++;

		if (td->aio_cur_depth < td->aio_depth) {
			timeout = &ts;
			min_evts = 0;
		} else {
			timeout = NULL;
			min_evts = 1;
		}

		ret = io_getevents(*td->aio_ctx, min_evts, td->aio_cur_depth, td->aio_events, timeout);
		if (ret < 0) {
			td->error = errno;
			break;
		} else if (!ret)
			continue;

		gettimeofday(&e, NULL);

		for (i = 0; i < ret; i++) {
			struct io_event *ev = td->aio_events + i;

			td->io_blocks++;

			iocb = ev->obj;

			msec = msec_now(&e) - iocb_time(iocb);
			add_stat_sample(td, msec);

			if (msec < td->min_latency)
				td->min_latency = msec;
			if (msec > td->max_latency)
				td->max_latency = msec;

			aio_put_iocb(td, iocb);
		}

		/*
		 * the rate is batched for now, it should work for batches
		 * of completions except the very first one which may look
		 * a little bursty
		 */
		usec = utime_since(&s, &e);

		rate_throttle(td, usec);

		if (check_min_rate(td, &e)) {
			td->error = ENODATA;
			break;
		}
	}

	gettimeofday(&e, NULL);
	td->runtime = mtime_since(&td->start, &e);

	free(ptr);
}

static void cleanup_pending_aio(struct thread_data *td)
{
	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
	unsigned int i;
	int r;

	/*
	 * get immediately available events, if any
	 */
	r = io_getevents(*td->aio_ctx, 0, td->aio_cur_depth, td->aio_events, &ts);
	if (r > 0) {
		for (i = 0; i < r; i++)
			aio_put_iocb(td, &td->aio_iocbs[i]);
	}

	/*
	 * now cancel remaining active events
	 */
	for (i = 0; i < td->aio_depth; i++) {
		if (td->aio_iocbs_status[i] == 0)
			continue;

		r = io_cancel(*td->aio_ctx, &td->aio_iocbs[i], td->aio_events);
		if (!r)
			aio_put_iocb(td, &td->aio_iocbs[i]);
	}

	if (td->aio_cur_depth)
		io_getevents(*td->aio_ctx, td->aio_cur_depth, td->aio_cur_depth, td->aio_events, NULL);
}

static void cleanup_aio(struct thread_data *td)
{
	if (td->aio_cur_depth)
		cleanup_pending_aio(td);

	if (td->aio_ctx) {
		io_destroy(*td->aio_ctx);
		free(td->aio_ctx);
	}
	if (td->aio_iocbs)
		free(td->aio_iocbs);
	if (td->aio_events)
		free(td->aio_events);
	if (td->aio_iocbs_status)
		free(td->aio_iocbs_status);
}

static int init_aio(struct thread_data *td)
{
	td->aio_ctx = malloc(sizeof(*td->aio_ctx));

	if (io_queue_init(td->aio_depth, td->aio_ctx)) {
		td->error = errno;
		return 1;
	}

	td->aio_iocbs = malloc(td->aio_depth * sizeof(struct iocb));
	td->aio_events = malloc(td->aio_depth * sizeof(struct io_event));
	td->aio_iocbs_status = malloc(td->aio_depth * sizeof(char));
	return 0;
}

static void *thread_main(int shm_id, int offset, char *argv[])
{
	struct thread_data *td;
	void *data;
	struct stat st;
	int ret = 1, flags;

	data = shmat(shm_id, NULL, 0);
	td = data + offset * sizeof(struct thread_data);
	td->pid = getpid();

	td->fd = -1;

	if (sched_setaffinity(td->pid, sizeof(td->cpumask), &td->cpumask) == -1) {
		td->error = errno;
		goto err;
	}

	printf("Thread (%s) (pid=%u) (f=%s) (aio=%d) started\n", td->ddir == DDIR_READ ? "read" : "write", td->pid, td->file_name, td->use_aio);
	fflush(stdout);

	sprintf(argv[0], "fio%d", offset);

	flags = 0;
	if (td->odirect)
		flags |= O_DIRECT;

	if (td->ddir == DDIR_READ)
		td->fd = open(td->file_name, flags | O_RDONLY);
	else
		td->fd = open(td->file_name, flags | O_WRONLY | O_CREAT | O_TRUNC, 0644);

	if (td->fd == -1) {
		td->error = errno;
		goto err;
	}

	if (td->use_aio && init_aio(td))
		goto err;

	if (init_random_state(td))
		goto out;
	if (init_stat_file(td))
		goto out;

	if (td->ddir == DDIR_READ) {
		if (fstat(td->fd, &st) == -1) {
			td->error = errno;
			goto out;
		}

		td->blocks = st.st_size / td->bs;
		if (!td->blocks) {
			td->error = EINVAL;
			goto out;
		}
	} else
		td->blocks = 1024 * 1024 * 1024 / td->bs;

	if (td->ioprio) {
		if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
			td->error = errno;
			goto out;
		}
	}

	sem_post(&startup_sem);
	sem_wait(&td->mutex);

	if (!td->use_aio)
		do_sync_io(td);
	else
		do_async_io(td);

	ret = 0;

out:
	shutdown_stat_file(td);
err:
	if (td->fd != -1)
		close(td->fd);
	if (td->use_aio)
		cleanup_aio(td);
	if (ret)
		sem_post(&startup_sem);

	td->runstate = TD_EXITED;
	shmdt(data);
	return NULL;
}

static void free_shm(void)
{
	shmdt(threads);
}

static void show_thread_status(struct thread_data *td)
{
	int prio, prio_class;
	unsigned long bw = 0;

	if (!td->io_blocks && !td->error)
		return;

	if (td->runtime)
		bw = (td->io_blocks * td->bs) / td->runtime;

	prio = td->ioprio & 0xff;
	prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;

	printf("thread%d (%s): err=%2d, prio=%1d/%1d maxl=%5lumsec, io=%6luMiB, bw=%6luKiB/sec\n", td->thread_number, td->ddir == DDIR_READ ? " read": "write", td->error, prio_class, prio, td->max_latency, td->io_blocks * td->bs >> 20, bw);
}

static int setup_rate(struct thread_data *td)
{
	int nr_reads_per_sec;

	if (!td->rate)
		return 0;

	if (td->rate < td->ratemin) {
		fprintf(stderr, "min rate larger than nominal rate\n");
		return -1;
	}

	nr_reads_per_sec = td->rate * 1024 / td->bs;
	td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
	td->rate_pending_usleep = 0;
	return 0;
}

static struct thread_data *get_new_job(void)
{
	struct thread_data *td;

	if (thread_number >= MAX_JOBS)
		return NULL;

	td = &threads[thread_number++];
	memset(td, 0, sizeof(*td));

	td->thread_number = thread_number;
	td->ddir = DDIR_READ;
	td->bs = global_bs;
	td->odirect = 1;
	td->delay_sleep = 0;
	td->rate = 0;
	td->ratecycle = DEF_RATE_CYCLE;
	td->sequential = sequential;
	td->ioprio = 0;
	td->use_aio = 0;
	td->aio_depth = 0;
	td->aio_cur_depth = 0;
	memcpy(&td->cpumask, &def_cpumask, sizeof(td->cpumask));

	return td;
}

static void put_job(struct thread_data *td)
{
	memset(&threads[td->thread_number - 1], 0, sizeof(*td));
	thread_number--;
}

static int add_job(struct thread_data *td, const char *filename, int prioclass,
		   int prio)
{
	strcpy(td->file_name, filename);
	td->stat_fd = -1;
	sem_init(&td->mutex, 1, 0);
	td->min_latency = 10000000;
	td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;

	if (td->use_aio && !td->aio_depth)
		td->aio_depth = 1;

	if (setup_rate(td))
		return -1;

	printf("Client%d: file=%s, rw=%d, prio=%d, seq=%d, odir=%d, bs=%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, filename, td->ddir, td->ioprio, td->sequential, td->odirect, td->bs, td->rate, td->use_aio, td->aio_depth);
	return 0;
}

static void fill_cpu_mask(cpu_set_t cpumask, int cpu)
{
	unsigned int i;

	CPU_ZERO(&cpumask);

	for (i = 0; i < sizeof(int) * 8; i++) {
		if ((1 << i) & cpu)
			CPU_SET(i, &cpumask);
	}
}

static void fill_option(const char *input, char *output)
{
	int i;

	i = 0;
	while (input[i] != ',' && input[i] != '}' && input[i] != '\0') {
		output[i] = input[i];
		i++;
	}

	output[i] = '\0';
}

/*
 * job key words:
 *
 * file=
 * bs=
 * rw=
 * direct=
 */
static void parse_jobs_cmd(int argc, char *argv[], int index)
{
	struct thread_data *td;
	unsigned int prio, prioclass, cpu;
	char *string, *filename, *p, *c;
	int i;

	string = malloc(256);
	filename = malloc(256);

	for (i = index; i < argc; i++) {
		p = argv[i];

		c = strpbrk(p, "{");
		if (!c)
			break;

		filename[0] = 0;

		td = get_new_job();
		if (!td)
			break;

		prioclass = 2;
		prio = 4;

		c = strstr(p, "rw=");
		if (c) {
			c += 3;
			if (*c == '0')
				td->ddir = DDIR_READ;
			else
				td->ddir = DDIR_WRITE;
		}

		c = strstr(p, "prio=");
		if (c) {
			c += 5;
			prio = *c - '0';
		}

		c = strstr(p, "prioclass=");
		if (c) {
			c += 10;
			prioclass = *c - '0';
		}

		c = strstr(p, "file=");
		if (c) {
			c += 5;
			fill_option(c, filename);
		}

		c = strstr(p, "bs=");
		if (c) {
			c += 3;
			fill_option(c, string);
			td->bs = strtoul(string, NULL, 10);
			td->bs <<= 10;
		}

		c = strstr(p, "direct=");
		if (c) {
			c += 7;
			if (*c != '0')
				td->odirect = 1;
			else
				td->odirect = 0;
		}

		c = strstr(p, "delay=");
		if (c) {
			c += 6;
			fill_option(c, string);
			td->delay_sleep = strtoul(string, NULL, 10);
		}

		c = strstr(p, "rate=");
		if (c) {
			c += 5;
			fill_option(c, string);
			td->rate = strtoul(string, NULL, 10);
		}

		c = strstr(p, "ratemin=");
		if (c) {
			c += 8;
			fill_option(c, string);
			td->ratemin = strtoul(string, NULL, 10);
		}

		c = strstr(p, "ratecycle=");
		if (c) {
			c += 10;
			fill_option(c, string);
			td->ratecycle = strtoul(string, NULL, 10);
		}

		c = strstr(p, "cpumask=");
		if (c) {
			c += 8;
			fill_option(c, string);
			cpu = strtoul(string, NULL, 10);
			fill_cpu_mask(td->cpumask, cpu);
		}

		c = strstr(p, "fsync=");
		if (c) {
			c += 6;
			fill_option(c, string);
			td->fsync_blocks = strtoul(string, NULL, 10);
		}

		c = strstr(p, "startdelay=");
		if (c) {
			c += 11;
			fill_option(c, string);
			td->start_delay = strtoul(string, NULL, 10);
		}

		c = strstr(p, "aio_depth=");
		if (c) {
			c += 10;
			fill_option(c, string);
			td->aio_depth = strtoul(string, NULL, 10);
		}

		c = strstr(p, "aio");
		if (c)
			td->use_aio = 1;

		c = strstr(p, "random");
		if (c)
			td->sequential = 0;
		c = strstr(p, "sequential");
		if (c)
			td->sequential = 1;

		if (add_job(td, filename, prioclass, prio))
			put_job(td);
	}

	free(string);
	free(filename);
}

static int check_int(char *p, char *name, unsigned int *val)
{
	char str[128];

	sprintf(str, "%s=%%d", name);
	if (sscanf(p, str, val) == 1)
		return 0;

	sprintf(str, "%s = %%d", name);
	if (sscanf(p, str, val) == 1)
		return 0;

	return 1;
}

static int is_empty(char *line)
{
	unsigned int i;

	for (i = 0; i < strlen(line); i++)
		if (!isspace(line[i]) && !iscntrl(line[i]))
			return 0;

	return 1;
}

static int parse_jobs_ini(char *file)
{
	unsigned int prioclass, prio, cpu;
	struct thread_data *td;
	char *string, *name;
	fpos_t off;
	FILE *f;
	char *p;

	f = fopen(file, "r");
	if (!f) {
		perror("fopen");
		return 1;
	}

	string = malloc(4096);
	name = malloc(256);

	while ((p = fgets(string, 4096, f)) != NULL) {
		if (sscanf(p, "[%s]", name) != 1)
			continue;

		name[strlen(name) - 1] = '\0';

		td = get_new_job();
		if (!td)
			break;

		prioclass = 2;
		prio = 4;

		fgetpos(f, &off);
		while ((p = fgets(string, 4096, f)) != NULL) {
			if (is_empty(p))
				continue;
			if (strstr(p, "["))
				break;
			if (!check_int(p, "bs", &td->bs)) {
				td->bs <<= 10;
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "rw", &td->ddir)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "prio", &prio)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "prioclass", &prioclass)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "direct", &td->odirect)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "rate", &td->rate)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "ratemin", &td->ratemin)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "ratecycle", &td->ratecycle)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "delay", &td->delay_sleep)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "cpumask", &cpu)) {
				fill_cpu_mask(td->cpumask, cpu);
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "fsync", &td->fsync_blocks)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "startdelay", &td->start_delay)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "aio_depth", &td->aio_depth)) {
				fgetpos(f, &off);
				continue;
			}
			if (!strncmp(p, "sequential", 10)) {
				td->sequential = 1;
				fgetpos(f, &off);
				continue;
			}
			if (!strncmp(p, "random", 6)) {
				td->sequential = 0;
				fgetpos(f, &off);
				continue;
			}
			if (!strncmp(p, "aio", 3)) {
				td->use_aio = 1;
				fgetpos(f, &off);
				continue;
			}

			printf("Client%d: bad option %s\n",td->thread_number,p);
		}
		fsetpos(f, &off);

		if (add_job(td, name, prioclass, prio))
			put_job(td);
	}

	free(string);
	free(name);
	fclose(f);
	return 0;
}

static int parse_options(int argc, char *argv[])
{
	int i;

	for (i = 1; i < argc; i++) {
		char *parm = argv[i];

		if (parm[0] != '-')
			break;

		parm++;
		switch (*parm) {
			case 's':
				parm++;
				sequential = !!atoi(parm);
				break;
			case 'b':
				parm++;
				global_bs = atoi(parm);
				global_bs <<= 10;
				if (!global_bs) {
					printf("bad block size\n");
					global_bs = DEF_BS;
				}
				break;
			case 't':
				parm++;
				timeout = atoi(parm);
				break;
			case 'w':
				parm++;
				write_stat = !!atoi(parm);
				break;
			case 'r':
				parm++;
				repeatable = !!atoi(parm);
				break;
			case 'R':
				parm++;
				rate_quit = !!atoi(parm);
				break;
			case 'o':
				parm++;
				odirect = !!atoi(parm);
				break;
			case 'f':
				if (i + 1 >= argc) {
					printf("-f needs file as arg\n");
					break;
				}
				ini_file = strdup(argv[i+1]);
				break;
			default:
				printf("bad option %s\n", argv[i]);
				break;
		}
	}

	return i;
}

static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
{
	int i;

	for (i = 0; i < thread_number; i++) {
		struct thread_data *td = &threads[i];

		if (td->runstate != TD_EXITED)
			continue;

		td->runstate = TD_REAPED;
		waitpid(td->pid, NULL, 0);
		(*nr_running)--;
		(*m_rate) -= td->ratemin;
		(*t_rate) -= td->rate;

		if (td->terminate)
			continue;

		printf("Threads now running: %d", *nr_running);
		if (*m_rate || *t_rate)
			printf(", rate %d/%dKiB/sec", *t_rate, *m_rate);
		printf("\n");
	}
}

static void run_threads(char *argv[])
{
	struct timeval genesis, now;
	struct thread_data *td;
	unsigned long spent;
	int i, todo, nr_running, m_rate, t_rate;

	gettimeofday(&genesis, NULL);

	printf("Starting %d threads\n", thread_number);
	fflush(stdout);

	if (timeout) {
		signal(SIGALRM, sig_handler);
		signal(SIGINT, sig_handler);
		alarm(timeout);
	}

	todo = thread_number;
	nr_running = 0;
	m_rate = t_rate = 0;

	while (todo) {
		for (i = 0; i < thread_number; i++) {
			td = &threads[i];

			if (td->runstate != TD_NOT_CREATED)
				continue;

			/*
			 * never got a chance to start, killed by other
			 * thread for some reason
			 */
			if (td->terminate) {
				todo--;
				continue;
			}

			if (td->start_delay) {
				gettimeofday(&now, NULL);
				spent = mtime_since(&genesis, &now);

				if (td->start_delay * 1000 > spent)
					continue;
			}

			td->runstate = TD_CREATED;
			sem_init(&startup_sem, 1, 1);
			todo--;

			if (fork())
				sem_wait(&startup_sem);
			else {
				thread_main(shm_id, i, argv);
				exit(0);
			}
		}

		for (i = 0; i < thread_number; i++) {
			struct thread_data *td = &threads[i];

			if (td->runstate == TD_CREATED) {
				td->runstate = TD_STARTED;
				nr_running++;
				m_rate += td->ratemin;
				t_rate += td->rate;
				sem_post(&td->mutex);

				printf("Threads now running: %d", nr_running);
				if (m_rate || t_rate)
					printf(", rate %d/%dKiB/sec", t_rate, m_rate);
				printf("\n");
			}
		}

		reap_threads(&nr_running, &t_rate, &m_rate);

		if (todo)
			usleep(100000);
	}

	while (nr_running) {
		reap_threads(&nr_running, &t_rate, &m_rate);
		usleep(10000);
	}
}

int main(int argc, char *argv[])
{
	static unsigned long max_run[2], min_run[2], total_blocks[2];
	static unsigned long max_bw[2], min_bw[2], maxl[2], minl[2];
	static unsigned long read_mb, write_mb, read_agg, write_agg;
	int i;

	shm_id = shmget(0, MAX_JOBS * sizeof(struct thread_data), IPC_CREAT | 0600);
	if (shm_id == -1) {
		perror("shmget");
		return 1;
	}

	threads = shmat(shm_id, NULL, 0);
	if (threads == (void *) -1 ) {
		perror("shmat");
		return 1;
	}

	atexit(free_shm);

	if (sched_getaffinity(getpid(), sizeof(def_cpumask), &def_cpumask) == -1) {
		perror("sched_getaffinity");
		return 1;
	}

	i = parse_options(argc, argv);

	if (ini_file) {
		if (parse_jobs_ini(ini_file))
			return 1;
	} else
		parse_jobs_cmd(argc, argv, i);

	if (!thread_number) {
		printf("Nothing to do\n");
		return 1;
	}

	printf("%s: %s, bs=%uKiB, timeo=%u, write_stat=%u, odirect=%d\n", argv[0], sequential ? "sequential" : "random", global_bs >> 10, timeout, write_stat, odirect);

	run_threads(argv);

	min_bw[0] = min_run[0] = ~0UL;
	min_bw[1] = min_run[1] = ~0UL;
	minl[0] = minl[1] = ~0UL;
	for (i = 0; i < thread_number; i++) {
		struct thread_data *td = &threads[i];
		unsigned long bw = 0;

		if (td->error)
			goto show_stat;

		if (td->runtime < min_run[td->ddir])
			min_run[td->ddir] = td->runtime;
		if (td->runtime > max_run[td->ddir])
			max_run[td->ddir] = td->runtime;

		if (td->runtime)
			bw = (td->io_blocks * td->bs) / td->runtime;
		if (bw < min_bw[td->ddir])
			min_bw[td->ddir] = bw;
		if (bw > max_bw[td->ddir])
			max_bw[td->ddir] = bw;
		if (td->max_latency < minl[td->ddir])
			minl[td->ddir] = td->max_latency;
		if (td->max_latency > maxl[td->ddir])
			maxl[td->ddir] = td->max_latency;

		total_blocks[td->ddir] += td->io_blocks;

		if (td->ddir == DDIR_READ) {
			read_mb += (td->bs * td->io_blocks) >> 20;
			if (td->runtime)
				read_agg += (td->io_blocks * td->bs) / td->runtime;
		}
		if (td->ddir == DDIR_WRITE) {
			write_mb += (td->bs * td->io_blocks) >> 20;
			if (td->runtime)
				write_agg += (td->io_blocks * td->bs) / td->runtime;
		}

show_stat:
		show_thread_status(td);
	}

	printf("Run status:\n");
	if (max_run[DDIR_READ])
		printf("   READ: io=%luMiB, aggrb=%lu, minl=%lu, maxl=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", read_mb, read_agg, minl[0], maxl[0], min_bw[0], max_bw[0], min_run[0], max_run[0]);
	if (max_run[DDIR_WRITE])
		printf("  WRITE: io=%luMiB, aggrb=%lu, minl=%lu, maxl=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", write_mb, write_agg, minl[1], maxl[1], min_bw[1], max_bw[1], min_run[1], max_run[1]);

	return 0;
}