[PATCH] fio: allow buffed aio. check shmat return value

[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 <math.h>
#include <limits.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>

#include "list.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

#ifndef __NR_fadvise64
#if defined(__i386__)
#define __NR_fadvise64		250
#elif defined(__powerpc__) || defined(__powerpc64__)
#define __NR_fadvise64		233
#elif defined(__x86_64__)
#define __NR_fadvise64		221
#elif defined(__ia64__)
#define __NR_fadvise64		1234
#elif defined(__alpha__)
#define __NR_fadvise64		413
#elif defined(__s390x__) || defined(__s390__)
#define __NR_fadvise64		253
#else
#error "Unsupported arch"
#endif
#endif

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

/*
 * we want fadvise64 really, but it's so tangled... later
 */
static int fadvise(int fd, loff_t offset, size_t len, int advice)
{
#if 0
	return syscall(__NR_fadvise64, fd, offset, offset >> 32, len, advice);
#else
	return posix_fadvise(fd, (off_t) offset, len, advice);
#endif
}

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_RAND_REPEAT	(1)
#define DEF_OVERWRITE	(0)
#define DEF_CREATE	(1)
#define DEF_INVALIDATE	(1)

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

static int repeatable = DEF_RAND_REPEAT;
static int rate_quit = 1;
static int write_lat_log;
static int write_bw_log;

static int thread_number;
static char *ini_file;

static int max_jobs = MAX_JOBS;

static char run_str[MAX_JOBS + 1];

static int shm_id;

enum {
	DDIR_READ = 0,
	DDIR_WRITE,
};

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

/*
 * The io unit
 */
struct io_u {
	struct iocb iocb;
	struct timeval start_time;
	struct timeval issue_time;

	void *mem;
	char *buf;
	unsigned int buflen;
	unsigned long long offset;

	struct list_head list;
};

struct io_stat {
	unsigned long val;
	unsigned long val_sq;
	unsigned long max_val;
	unsigned long min_val;
	unsigned long samples;
};

struct io_sample {
	unsigned long time;
	unsigned long val;
};

struct io_log {
	unsigned long nr_samples;
	unsigned long max_samples;
	struct io_sample *log;
};

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

struct thread_data {
	char file_name[256];
	int thread_number;
	int error;
	int 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 timeout;
	unsigned int use_aio;
	unsigned int create_file;
	unsigned int overwrite;
	unsigned int invalidate_cache;
	unsigned long long file_size;
	unsigned long long file_offset;
	cpu_set_t cpumask;

	off_t cur_off;

	io_context_t aio_ctx;
	unsigned int aio_depth;
	struct io_event *aio_events;

	unsigned int cur_depth;
	struct list_head io_u_freelist;
	struct list_head io_u_busylist;

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

	/*
	 * bandwidth and latency stats
	 */
	struct io_stat clat_stat;		/* completion latency */
	struct io_stat slat_stat;		/* submission latency */

	struct io_stat bw_stat;			/* bandwidth stats */
	unsigned long stat_io_blocks;
	struct timeval stat_sample_time;

	struct io_log *lat_log;
	struct io_log *bw_log;

	struct timeval start;
};

static struct thread_data *threads;
static struct thread_data def_thread;

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 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 unsigned long mtime_since_now(struct timeval *s)
{
	struct timeval t;

	gettimeofday(&t, NULL);
	return mtime_since(s, &t);
}

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

static unsigned long long get_next_offset(struct thread_data *td)
{
	unsigned long 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 + td->file_offset;
}

static inline void add_stat_sample(struct thread_data *td, struct io_stat *is,
				   unsigned long val)
{
	if (val > is->max_val)
		is->max_val = val;
	if (val < is->min_val)
		is->min_val = val;

	is->val += val;
	is->val_sq += val * val;
	is->samples++;
}

static void add_log_sample(struct thread_data *td, struct io_log *log,
			   unsigned long val)
{
	if (log->nr_samples == log->max_samples) {
		int new_size = sizeof(struct io_sample) * log->max_samples * 2;

		log->log = realloc(log->log, new_size);
		log->max_samples <<= 1;
	}

	log->log[log->nr_samples].val = val;
	log->log[log->nr_samples].time = mtime_since_now(&td->start);
	log->nr_samples++;
}

static void add_clat_sample(struct thread_data *td, unsigned long msec)
{
	add_stat_sample(td, &td->clat_stat, msec);

	if (td->lat_log)
		add_log_sample(td, td->lat_log, msec);
}

static void add_slat_sample(struct thread_data *td, unsigned long msec)
{
	add_stat_sample(td, &td->slat_stat, msec);
}

static void add_bw_sample(struct thread_data *td, unsigned long msec)
{
	unsigned long spent = mtime_since_now(&td->stat_sample_time);
	unsigned long rate;

	if (spent < 500)
		return;

	rate = ((td->io_blocks - td->stat_io_blocks) * td->bs) / spent;
	add_stat_sample(td, &td->bw_stat, rate);

	if (td->bw_log)
		add_log_sample(td, td->bw_log, rate);

	gettimeofday(&td->stat_sample_time, NULL);
	td->stat_io_blocks = td->io_blocks;
}

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;
	unsigned long rate;

	/*
	 * allow a 2 second settle period in the beginning
	 */
	if (mtime_since(&td->start, now) < 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;
}

static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
{
	if (mtime_since(&td->start, t) >= td->timeout * 1000)
		return 1;

	return 0;
}

static void put_io_u(struct thread_data *td, struct io_u *io_u)
{
	list_del(&io_u->list);
	list_add(&io_u->list, &td->io_u_freelist);
	td->cur_depth--;
}

static struct io_u *get_io_u(struct thread_data *td)
{
	struct io_u *io_u;

	if (list_empty(&td->io_u_freelist))
		return NULL;

	io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
	list_del(&io_u->list);
	list_add(&io_u->list, &td->io_u_busylist);

	io_u->offset = get_next_offset(td);

	if (td->use_aio) {
		if (td_read(td))
			io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
		else
			io_prep_pwrite(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
	}

	gettimeofday(&io_u->start_time, NULL);
	td->cur_depth++;
	return io_u;
}

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

	td->cur_off = 0;

	for (blocks = 0; blocks < td->blocks; blocks++) {
		struct io_u *io_u;
		int ret;

		if (td->terminate)
			break;

		io_u = get_io_u(td);

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

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

		if (td_read(td))
			ret = read(td->fd, io_u->buf, io_u->buflen);
		else
			ret = write(td->fd, io_u->buf, io_u->buflen);

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

		td->io_blocks++;
		td->cur_off = io_u->offset + io_u->buflen;

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

		gettimeofday(&e, NULL);

		usec = utime_since(&io_u->start_time, &e);

		rate_throttle(td, usec);

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

		msec = usec / 1000;
		add_clat_sample(td, msec);
		add_bw_sample(td, msec);

		if (runtime_exceeded(td, &e))
			break;

		put_io_u(td, io_u);
	}

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

static int io_u_queue(struct thread_data *td, struct io_u *io_u)
{
	struct iocb *iocb = &io_u->iocb;
	int ret;

	do {
		ret = io_submit(td->aio_ctx, 1, &iocb);
		if (ret == 1)
			return 0;
		else if (ret == EAGAIN)
			usleep(100);
		else if (ret == EINTR)
			continue;
		else
			break;
	} while (1);

	return ret;
}

#define iocb_time(iocb)	((unsigned long) (iocb)->data)
#define ev_to_iou(ev)	(struct io_u *) ((unsigned long) (ev)->obj)

static void ios_completed(struct thread_data *td, int nr)
{
	unsigned long msec;
	struct io_u *io_u;
	struct timeval e;
	int i;

	gettimeofday(&e, NULL);

	for (i = 0; i < nr; i++) {
		td->io_blocks++;

		io_u = ev_to_iou(td->aio_events + i);

		msec = mtime_since(&io_u->issue_time, &e);

		add_clat_sample(td, msec);
		add_bw_sample(td, msec);

		put_io_u(td, io_u);
	}
}

static void cleanup_pending_aio(struct thread_data *td)
{
	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
	struct list_head *entry, *n;
	struct io_u *io_u;
	int r;

	/*
	 * get immediately available events, if any
	 */
	r = io_getevents(td->aio_ctx, 0, td->cur_depth, td->aio_events, &ts);
	if (r > 0)
		ios_completed(td, r);

	/*
	 * now cancel remaining active events
	 */
	list_for_each_safe(entry, n, &td->io_u_busylist) {
		io_u = list_entry(entry, struct io_u, list);

		r = io_cancel(td->aio_ctx, &io_u->iocb, td->aio_events);
		if (!r)
			put_io_u(td, io_u);
	}

	if (td->cur_depth) {
		r = io_getevents(td->aio_ctx, td->cur_depth, td->cur_depth, td->aio_events, NULL);
		if (r > 0)
			ios_completed(td, r);
	}
}

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

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

		if (td->terminate)
			break;

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

		io_u = get_io_u(td);

		memcpy(&s, &io_u->start_time, sizeof(s));

		ret = io_u_queue(td, io_u);
		if (ret) {
			put_io_u(td, io_u);
			td->error = ret;
			break;
		}

		gettimeofday(&io_u->issue_time, NULL);
		add_slat_sample(td, mtime_since(&io_u->start_time, &io_u->issue_time));

		if (td->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->cur_depth, td->aio_events, timeout);
		if (ret < 0) {
			td->error = errno;
			break;
		} else if (!ret)
			continue;

		ios_completed(td, ret);

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

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

		rate_throttle(td, usec);

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

		if (runtime_exceeded(td, &e))
			break;
	}

	if (td->cur_depth)
		cleanup_pending_aio(td);

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

static void cleanup_aio(struct thread_data *td)
{
	io_destroy(td->aio_ctx);

	if (td->aio_events)
		free(td->aio_events);
}

static int init_aio(struct thread_data *td)
{
	if (io_queue_init(td->aio_depth, &td->aio_ctx)) {
		td->error = errno;
		return 1;
	}

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

static void cleanup_io_u(struct thread_data *td)
{
	struct list_head *entry, *n;
	struct io_u *io_u;

	list_for_each_safe(entry, n, &td->io_u_freelist) {
		io_u = list_entry(entry, struct io_u, list);

		list_del(&io_u->list);
		free(io_u->mem);
		free(io_u);
	}
}

static void init_io_u(struct thread_data *td)
{
	struct io_u *io_u;
	int i, max_units;

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

	INIT_LIST_HEAD(&td->io_u_freelist);
	INIT_LIST_HEAD(&td->io_u_busylist);

	for (i = 0; i < max_units; i++) {
		io_u = malloc(sizeof(*io_u));
		memset(io_u, 0, sizeof(*io_u));
		INIT_LIST_HEAD(&io_u->list);

		io_u->mem = malloc(td->bs + MASK);
		io_u->buf = ALIGN(io_u->mem);
		io_u->buflen = td->bs;

		list_add(&io_u->list, &td->io_u_freelist);
	}
}

static void setup_log(struct io_log **log)
{
	struct io_log *l = malloc(sizeof(*l));

	l->nr_samples = 0;
	l->max_samples = 1024;
	l->log = malloc(l->max_samples * sizeof(struct io_sample));
	*log = l;
}

static void finish_log(struct thread_data *td, struct io_log *log, char *name)
{
	char file_name[128];
	FILE *f;
	int i;

	sprintf(file_name, "client%d_%s.log", td->thread_number, name);
	f = fopen(file_name, "w");
	if (!f) {
		perror("fopen log");
		return;
	}

	for (i = 0; i < log->nr_samples; i++)
		fprintf(f, "%lu, %lu\n", log->log[i].time, log->log[i].val);

	fclose(f);
	free(log->log);
	free(log);
}

static int create_file(struct thread_data *td)
{
	unsigned int i;
	char *b;

	/*
	 * unless specifically asked for overwrite, let normal io extend it
	 */
	if (!td_read(td) && !td->overwrite)
		return 0;

	if (!td->file_size) {
		fprintf(stderr, "Need size for create\n");
		td->error = EINVAL;
		return 1;
	}

	td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
	if (td->fd < 0) {
		td->error = errno;
		return 1;
	}

	td->blocks = td->file_size / td->bs;
	b = malloc(td->bs);
	memset(b, 0, td->bs);

	for (i = 0; i < td->blocks; i++) {
		int r = write(td->fd, b, td->bs);

		if (r == td->bs)
			continue;
		else {
			if (r < 0)
				td->error = errno;
			else
				td->error = EIO;

			break;
		}
	}

	fsync(td->fd);
	close(td->fd);
	td->fd = -1;
	free(b);
	return 0;
}

static int file_exists(struct thread_data *td)
{
	struct stat st;

	if (stat(td->file_name, &st) != -1)
		return 1;

	return errno != ENOENT;
}

static int setup_file(struct thread_data *td)
{
	struct stat st;
	int flags = 0;

	if (!file_exists(td)) {
		if (!td->create_file) {
			td->error = ENOENT;
			return 1;
		}
		if (create_file(td))
			return 1;
	}

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

	if (td_read(td))
		td->fd = open(td->file_name, flags | O_RDONLY);
	else {
		if (!td->overwrite)
			flags |= O_TRUNC;

		td->fd = open(td->file_name, flags | O_WRONLY | O_CREAT, 0600);
	}

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

	if (td_read(td)) {
		if (fstat(td->fd, &st) == -1) {
			td->error = errno;
			return 1;
		}

		if (td->file_size > st.st_size)
			st.st_size = td->file_size;
	} else {
		if (!td->file_size)
			td->file_size = 1024 * 1024 * 1024;

		st.st_size = td->file_size;
	}

	td->blocks = (st.st_size - td->file_offset) / td->bs;
	if (!td->blocks) {
		fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
		td->error = EINVAL;
		return 1;
	}

	if (td->invalidate_cache) {
		if (fadvise(td->fd, 0, st.st_size, POSIX_FADV_DONTNEED) < 0) {
			td->error = errno;
			return 1;
		}
	}

	return 0;
}

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

	setsid();

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

	td = data + offset * sizeof(struct thread_data);
	td->pid = getpid();

	init_io_u(td);

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

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

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

	if (init_random_state(td))
		goto err;

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

	if (setup_file(td))
		goto err;

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

	gettimeofday(&td->start, NULL);

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

	memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));

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

	td->runtime = mtime_since_now(&td->start);
	ret = 0;

	if (td->bw_log)
		finish_log(td, td->bw_log, "bw");
	if (td->lat_log)
		finish_log(td, td->lat_log, "lat");

err:
	if (td->fd != -1) {
		close(td->fd);
		td->fd = -1;
	}
	if (td->use_aio)
		cleanup_aio(td);
	cleanup_io_u(td);
	if (ret) {
		sem_post(&startup_sem);
		sem_wait(&td->mutex);
	}
	td->runstate = TD_EXITED;
	shmdt(data);
	return NULL;
}

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

static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
		    double *mean, double *dev)
{
	double n;

	if (is->samples == 0)
		return 0;

	*min = is->min_val;
	*max = is->max_val;

	n = (double) is->samples;
	*mean = (double) is->val / n;
	*dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
	return 1;
}

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

	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("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s\n", td->thread_number, td->error, td->io_blocks * td->bs >> 20, bw);

	if (calc_lat(&td->slat_stat, &min, &max, &mean, &dev))
		printf("  slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
	if (calc_lat(&td->clat_stat, &min, &max, &mean, &dev))
		printf("  clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
	if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev))
		printf("  bw (KiB/s) : min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
}

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(int global)
{
	struct thread_data *td;

	if (global)
		return &def_thread;
	if (thread_number >= max_jobs)
		return NULL;

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

	td->fd = -1;
	td->thread_number = thread_number;

	td->ddir = def_thread.ddir;
	td->ioprio = def_thread.ioprio;
	td->sequential = def_thread.sequential;
	td->bs = def_thread.bs;
	td->odirect = def_thread.odirect;
	td->delay_sleep = def_thread.delay_sleep;
	td->fsync_blocks = def_thread.fsync_blocks;
	td->start_delay = def_thread.start_delay;
	td->timeout = def_thread.timeout;
	td->use_aio = def_thread.use_aio;
	td->create_file = def_thread.create_file;
	td->overwrite = def_thread.overwrite;
	td->invalidate_cache = def_thread.invalidate_cache;
	td->file_size = def_thread.file_size;
	td->file_offset = def_thread.file_offset;
	td->rate = def_thread.rate;
	td->ratemin = def_thread.ratemin;
	td->ratecycle = def_thread.ratecycle;
	td->aio_depth = def_thread.aio_depth;
	memcpy(&td->cpumask, &def_thread.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)
{
	if (td == &def_thread)
		return 0;

	strcpy(td->file_name, filename);
	sem_init(&td->mutex, 1, 0);
	td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;

	td->clat_stat.min_val = ULONG_MAX;
	td->slat_stat.min_val = ULONG_MAX;
	td->bw_stat.min_val = ULONG_MAX;

	run_str[td->thread_number - 1] = 'P';

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

	if (setup_rate(td))
		return -1;

	if (write_lat_log)
		setup_log(&td->lat_log);
	if (write_bw_log)
		setup_log(&td->bw_log);

	printf("Client%d: file=%s, rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, filename, td->ddir, prioclass, prio, 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';
}

/*
 * convert string after '=' into decimal value, noting any size suffix
 */
static int str_cnv(char *p, unsigned long long *val)
{
	unsigned long mult;
	char *str;
	int len;

	str = strstr(p, "=");
	if (!str)
		return 1;

	str++;
	len = strlen(str);
	mult = 1;

	switch (str[len - 2]) {
		case 'k':
		case 'K':
			mult = 1024;
			break;
		case 'm':
		case 'M':
			mult = 1024 * 1024;
			break;
		case 'g':
		case 'G':
			mult = 1024 * 1024 * 1024;
			break;
	}

	*val = strtoul(str, NULL, 10);
	if (*val == ULONG_MAX && errno == ERANGE)
		return 1;

	*val *= mult;
	return 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(0);
		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, "timeout=");
		if (c) {
			c += 8;
			fill_option(c, string);
			td->timeout = strtoul(string, NULL, 10);
		}

		c = strstr(p, "invalidate=");
		if (c) {
			c += 11;
			if (*c != '0')
				td->invalidate_cache = 1;
			else
				td->invalidate_cache = 0;
		}

		c = strstr(p, "size=");
		if (c) {
			c += 5;
			str_cnv(c, &td->file_size);
		}

		c = strstr(p, "offset=");
		if (c) {
			c += 7;
			str_cnv(c, &td->file_offset);
		}

		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, "create");
		if (c)
			td->create_file = 1;

		c = strstr(p, "overwrite");
		if (c)
			td->overwrite = 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_strcnv(char *p, char *name, unsigned long long *val)
{
	if (!strstr(p, name))
		return 1;

	return str_cnv(p, val);
}

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_or_comment(char *line)
{
	unsigned int i;

	for (i = 0; i < strlen(line); i++) {
		if (line[i] == ';')
			return 1;
		if (!isspace(line[i]) && !iscntrl(line[i]))
			return 0;
	}

	return 1;
}

static int parse_jobs_ini(char *file)
{
	unsigned int prioclass, prio, cpu, global;
	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 (is_empty_or_comment(p))
			continue;
		if (sscanf(p, "[%s]", name) != 1)
			continue;

		global = !strncmp(name, "global", 6);

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

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

		prioclass = 2;
		prio = 4;

		fgetpos(f, &off);
		while ((p = fgets(string, 4096, f)) != NULL) {
			if (is_empty_or_comment(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, "timeout", &td->timeout)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "invalidate",&td->invalidate_cache)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_int(p, "aio_depth", &td->aio_depth)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_strcnv(p, "size", &td->file_size)) {
				fgetpos(f, &off);
				continue;
			}
			if (!check_strcnv(p, "offset", &td->file_offset)) {
				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;
			}
			if (!strncmp(p, "create", 6)) {
				td->create_file = 1;
				fgetpos(f, &off);
				continue;
			}
			if (!strncmp(p, "overwrite", 9)) {
				td->overwrite = 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++;
				def_thread.sequential = !!atoi(parm);
				break;
			case 'b':
				parm++;
				def_thread.bs = atoi(parm);
				def_thread.bs <<= 10;
				if (!def_thread.bs) {
					printf("bad block size\n");
					def_thread.bs = DEF_BS;
				}
				break;
			case 't':
				parm++;
				def_thread.timeout = atoi(parm);
				break;
			case 'r':
				parm++;
				repeatable = !!atoi(parm);
				break;
			case 'R':
				parm++;
				rate_quit = !!atoi(parm);
				break;
			case 'o':
				parm++;
				def_thread.odirect = !!atoi(parm);
				break;
			case 'f':
				if (i + 1 >= argc) {
					printf("-f needs file as arg\n");
					break;
				}
				ini_file = strdup(argv[i+1]);
				i++;
				break;
			case 'l':
				write_lat_log = 1;
				break;
			case 'w':
				write_bw_log = 1;
				break;
			default:
				printf("bad option %s\n", argv[i]);
				break;
		}
	}

	return i;
}

static void print_thread_status(struct thread_data *td, int nr_running,
				int t_rate, int m_rate)
{
	printf("Threads now running: %d", nr_running);
	if (m_rate || t_rate)
		printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
	printf(" : [%s]\r", run_str);
	fflush(stdout);
}

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

	/*
	 * reap exited threads (TD_EXITED -> TD_REAPED)
	 */
	for (i = 0; i < thread_number; i++) {
		struct thread_data *td = &threads[i];

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

		td->runstate = TD_REAPED;
		run_str[td->thread_number - 1] = '_';
		waitpid(td->pid, NULL, 0);
		(*nr_running)--;
		(*m_rate) -= td->ratemin;
		(*t_rate) -= td->rate;

		if (td->terminate)
			continue;

		print_thread_status(td, *nr_running, *t_rate, *m_rate);
	}
}

static void run_threads(char *argv[])
{
	struct timeval genesis;
	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);

	signal(SIGINT, sig_handler);

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

	while (todo) {
		/*
		 * create threads (TD_NOT_CREATED -> TD_CREATED)
		 */
		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) {
				spent = mtime_since_now(&genesis);

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

			td->runstate = TD_CREATED;
			run_str[td->thread_number - 1] = 'C';
			sem_init(&startup_sem, 1, 1);
			todo--;

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

		/*
		 * start created threads (TD_CREATED -> TD_STARTED)
		 */
		for (i = 0; i < thread_number; i++) {
			struct thread_data *td = &threads[i];

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

			td->runstate = TD_STARTED;
			run_str[td->thread_number - 1] = '+';
			nr_running++;
			m_rate += td->ratemin;
			t_rate += td->rate;
			sem_post(&td->mutex);

			print_thread_status(td, nr_running, t_rate, m_rate);
		}

		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 setup_thread_area(void)
{
	/*
	 * 1024 is too much on some machines, scale max_jobs if
	 * we get a failure that looks like too large a shm segment
	 */
	do {
		int s = max_jobs * sizeof(struct thread_data);

		shm_id = shmget(0, s, IPC_CREAT | 0600);
		if (shm_id != -1)
			break;
		if (errno != EINVAL) {
			perror("shmget");
			break;
		}

		max_jobs >>= 1;
	} while (max_jobs);

	if (shm_id == -1)
		return 1;

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

	atexit(free_shm);
	return 0;
}

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];
	static unsigned long read_mb, write_mb, read_agg, write_agg;
	int i;

	if (setup_thread_area())
		return 1;

	if (sched_getaffinity(getpid(), sizeof(cpu_set_t), &def_thread.cpumask) == -1) {
		perror("sched_getaffinity");
		return 1;
	}

	/*
	 * fill globals
	 */
	def_thread.ddir = DDIR_READ;
	def_thread.bs = DEF_BS;
	def_thread.odirect = DEF_ODIRECT;
	def_thread.ratecycle = DEF_RATE_CYCLE;
	def_thread.sequential = DEF_SEQUENTIAL;
	def_thread.timeout = DEF_TIMEOUT;
	def_thread.create_file = DEF_CREATE;
	def_thread.overwrite = DEF_OVERWRITE;
	def_thread.invalidate_cache = DEF_INVALIDATE;

	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;
	}

	run_threads(argv);

	min_bw[0] = min_run[0] = ~0UL;
	min_bw[1] = min_run[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;

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

		if (td_read(td)) {
			read_mb += (td->bs * td->io_blocks) >> 20;
			if (td->runtime)
				read_agg += (td->io_blocks * td->bs) / td->runtime;
		} else {
			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("\nRun status:\n");
	if (max_run[DDIR_READ])
		printf("   READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", read_mb, read_agg, min_bw[0], max_bw[0], min_run[0], max_run[0]);
	if (max_run[DDIR_WRITE])
		printf("  WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", write_mb, write_agg, min_bw[1], max_bw[1], min_run[1], max_run[1]);

	return 0;
}