[fio.git] / io_u.c

#include <unistd.h>
#include <fcntl.h>
#include <string.h>
#include <signal.h>
#include <time.h>
#include <assert.h>

#include "fio.h"
#include "os.h"

/*
 * Change this define to play with the timeout handling
 */
#undef FIO_USE_TIMEOUT

struct io_completion_data {
	int nr;				/* input */

	int error;			/* output */
	unsigned long bytes_done[2];	/* output */
	struct timeval time;		/* output */
};

/*
 * The ->file_map[] contains a map of blocks we have or have not done io
 * to yet. Used to make sure we cover the entire range in a fair fashion.
 */
static int random_map_free(struct thread_data *td, struct fio_file *f,
			   unsigned long long block)
{
	unsigned int idx = RAND_MAP_IDX(td, f, block);
	unsigned int bit = RAND_MAP_BIT(td, f, block);

	return (f->file_map[idx] & (1UL << bit)) == 0;
}

/*
 * Mark a given offset as used in the map.
 */
static void mark_random_map(struct thread_data *td, struct fio_file *f,
			    struct io_u *io_u)
{
	unsigned int min_bs = td->rw_min_bs;
	unsigned long long block;
	unsigned int blocks;
	unsigned int nr_blocks;

	block = io_u->offset / (unsigned long long) min_bs;
	blocks = 0;
	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;

	while (blocks < nr_blocks) {
		unsigned int idx, bit;

		if (!random_map_free(td, f, block))
			break;

		idx = RAND_MAP_IDX(td, f, block);
		bit = RAND_MAP_BIT(td, f, block);

		fio_assert(td, idx < f->num_maps);

		f->file_map[idx] |= (1UL << bit);
		block++;
		blocks++;
	}

	if ((blocks * min_bs) < io_u->buflen)
		io_u->buflen = blocks * min_bs;
}

/*
 * Return the next free block in the map.
 */
static int get_next_free_block(struct thread_data *td, struct fio_file *f,
			       unsigned long long *b)
{
	int i;

	i = f->last_free_lookup;
	*b = (i * BLOCKS_PER_MAP);
	while ((*b) * td->rw_min_bs < f->real_file_size) {
		if (f->file_map[i] != -1UL) {
			*b += ffz(f->file_map[i]);
			f->last_free_lookup = i;
			return 0;
		}

		*b += BLOCKS_PER_MAP;
		i++;
	}

	return 1;
}

/*
 * For random io, generate a random new block and see if it's used. Repeat
 * until we find a free one. For sequential io, just return the end of
 * the last io issued.
 */
static int get_next_offset(struct thread_data *td, struct fio_file *f,
			   struct io_u *io_u)
{
	const int ddir = io_u->ddir;
	unsigned long long b, rb;
	long r;

	if (td_random(td)) {
		unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
		int loops = 5;

		do {
			r = os_random_long(&td->random_state);
			b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
			if (td->norandommap)
				break;
			rb = b + (f->file_offset / td->min_bs[ddir]);
			loops--;
		} while (!random_map_free(td, f, rb) && loops);

		/*
		 * if we failed to retrieve a truly random offset within
		 * the loops assigned, see if there are free ones left at all
		 */
		if (!loops && get_next_free_block(td, f, &b))
			return 1;
	} else
		b = f->last_pos / td->min_bs[ddir];

	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
	if (io_u->offset >= f->real_file_size)
		return 1;

	return 0;
}

static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f,
				    struct io_u *io_u)
{
	const int ddir = io_u->ddir;
	unsigned int buflen;
	long r;

	if (td->min_bs[ddir] == td->max_bs[ddir])
		buflen = td->min_bs[ddir];
	else {
		r = os_random_long(&td->bsrange_state);
		buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
		if (!td->bs_unaligned)
			buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
	}

	while (buflen + io_u->offset > f->real_file_size) {
		if (buflen == td->min_bs[ddir])
			return 0;

		buflen = td->min_bs[ddir];
	}

	return buflen;
}

/*
 * Return the data direction for the next io_u. If the job is a
 * mixed read/write workload, check the rwmix cycle and switch if
 * necessary.
 */
static enum fio_ddir get_rw_ddir(struct thread_data *td)
{
	if (td_rw(td)) {
		struct timeval now;
		unsigned long elapsed;

		fio_gettime(&now, NULL);
	 	elapsed = mtime_since_now(&td->rwmix_switch);

		/*
		 * Check if it's time to seed a new data direction.
		 */
		if (elapsed >= td->rwmixcycle) {
			unsigned int v;
			long r;

			r = os_random_long(&td->rwmix_state);
			v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
			if (v < td->rwmixread)
				td->rwmix_ddir = DDIR_READ;
			else
				td->rwmix_ddir = DDIR_WRITE;
			memcpy(&td->rwmix_switch, &now, sizeof(now));
		}
		return td->rwmix_ddir;
	} else if (td_read(td))
		return DDIR_READ;
	else
		return DDIR_WRITE;
}

void put_io_u(struct thread_data *td, struct io_u *io_u)
{
	assert((io_u->flags & IO_U_F_FREE) == 0);
	io_u->flags |= IO_U_F_FREE;

	io_u->file = NULL;
	list_del(&io_u->list);
	list_add(&io_u->list, &td->io_u_freelist);
	td->cur_depth--;
}

void requeue_io_u(struct thread_data *td, struct io_u **io_u)
{
	struct io_u *__io_u = *io_u;

	__io_u->flags |= IO_U_F_FREE;
	__io_u->flags &= ~IO_U_F_FLIGHT;

	list_del(&__io_u->list);
	list_add_tail(&__io_u->list, &td->io_u_requeues);
	td->cur_depth--;
	*io_u = NULL;
}

static int fill_io_u(struct thread_data *td, struct fio_file *f,
		     struct io_u *io_u)
{
	/*
	 * If using an iolog, grab next piece if any available.
	 */
	if (td->read_iolog)
		return read_iolog_get(td, io_u);

	/*
	 * see if it's time to sync
	 */
	if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
	    && td->io_issues[DDIR_WRITE] && should_fsync(td)) {
		io_u->ddir = DDIR_SYNC;
		io_u->file = f;
		return 0;
	}

	io_u->ddir = get_rw_ddir(td);

	/*
	 * No log, let the seq/rand engine retrieve the next buflen and
	 * position.
	 */
	if (get_next_offset(td, f, io_u))
		return 1;

	io_u->buflen = get_next_buflen(td, f, io_u);
	if (!io_u->buflen)
		return 1;

	/*
	 * mark entry before potentially trimming io_u
	 */
	if (!td->read_iolog && td_random(td) && !td->norandommap)
		mark_random_map(td, f, io_u);

	/*
	 * If using a write iolog, store this entry.
	 */
	if (td->write_iolog_file)
		write_iolog_put(td, io_u);

	io_u->file = f;
	return 0;
}

static void io_u_mark_depth(struct thread_data *td)
{
	int index = 0;

	switch (td->cur_depth) {
	default:
		index++;
	case 32 ... 63:
		index++;
	case 16 ... 31:
		index++;
	case 8 ... 15:
		index++;
	case 4 ... 7:
		index++;
	case 2 ... 3:
		index++;
	case 1:
		break;
	}

	td->io_u_map[index]++;
	td->total_io_u++;
}

static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
{
	int index = 0;

	switch (msec) {
	default:
		index++;
	case 1000 ... 1999:
		index++;
	case 750 ... 999:
		index++;
	case 500 ... 749:
		index++;
	case 250 ... 499:
		index++;
	case 100 ... 249:
		index++;
	case 50 ... 99:
		index++;
	case 20 ... 49:
		index++;
	case 10 ... 19:
		index++;
	case 4 ... 9:
		index++;
	case 2 ... 3:
		index++;
	case 0 ... 1:
		break;
	}

	td->io_u_lat[index]++;
}

/*
 * Get next file to service by choosing one at random
 */
static struct fio_file *get_next_file_rand(struct thread_data *td)
{
	long r = os_random_long(&td->next_file_state);
	unsigned int fileno;
	struct fio_file *f;

	do {
		fileno = (unsigned int) ((double) (td->nr_files - 1) * r / (RAND_MAX + 1.0));
		f = &td->files[fileno];
		if (f->fd != -1)
			return f;
	} while (1);
}

/*
 * Get next file to service by doing round robin between all available ones
 */
static struct fio_file *get_next_file_rr(struct thread_data *td)
{
	unsigned int old_next_file = td->next_file;
	struct fio_file *f;

	do {
		f = &td->files[td->next_file];

		td->next_file++;
		if (td->next_file >= td->nr_files)
			td->next_file = 0;

		if (f->fd != -1)
			break;

		f = NULL;
	} while (td->next_file != old_next_file);

	return f;
}

struct io_u *__get_io_u(struct thread_data *td)
{
	struct io_u *io_u = NULL;

	if (!list_empty(&td->io_u_requeues))
		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
	else if (!queue_full(td)) {
		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);

		io_u->buflen = 0;
		io_u->resid = 0;
		io_u->file = NULL;
		io_u->end_io = NULL;
	}

	if (io_u) {
		assert(io_u->flags & IO_U_F_FREE);
		io_u->flags &= ~IO_U_F_FREE;

		io_u->error = 0;
		list_del(&io_u->list);
		list_add(&io_u->list, &td->io_u_busylist);
		td->cur_depth++;
		io_u_mark_depth(td);
	}

	return io_u;
}

/*
 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
 * etc. The returned io_u is fully ready to be prepped and submitted.
 */
struct io_u *get_io_u(struct thread_data *td)
{
	struct fio_file *f;
	struct io_u *io_u;

	io_u = __get_io_u(td);
	if (!io_u)
		return NULL;

	/*
	 * from a requeue, io_u already setup
	 */
	if (io_u->file)
		goto out;

	if (td->file_service_type == FIO_FSERVICE_RR)
		f = get_next_file_rr(td);
	else
		f = get_next_file_rand(td);

	if (!f) {
		put_io_u(td, io_u);
		return NULL;
	}

	io_u->file = f;

	if (td->zone_bytes >= td->zone_size) {
		td->zone_bytes = 0;
		f->last_pos += td->zone_skip;
	}

	if (fill_io_u(td, f, io_u)) {
		put_io_u(td, io_u);
		return NULL;
	}

	if (io_u->buflen + io_u->offset > f->real_file_size) {
		if (td->io_ops->flags & FIO_RAWIO) {
			put_io_u(td, io_u);
			return NULL;
		}

		io_u->buflen = f->real_file_size - io_u->offset;
	}

	if (io_u->ddir != DDIR_SYNC) {
		if (!io_u->buflen) {
			put_io_u(td, io_u);
			return NULL;
		}

		f->last_pos = io_u->offset + io_u->buflen;

		if (td->verify != VERIFY_NONE)
			populate_verify_io_u(td, io_u);
	}

	/*
	 * Set io data pointers.
	 */
out:
	io_u->xfer_buf = io_u->buf;
	io_u->xfer_buflen = io_u->buflen;

	if (td_io_prep(td, io_u)) {
		put_io_u(td, io_u);
		return NULL;
	}

	fio_gettime(&io_u->start_time, NULL);
	return io_u;
}

static void io_completed(struct thread_data *td, struct io_u *io_u,
			 struct io_completion_data *icd)
{
	unsigned long msec;

	assert(io_u->flags & IO_U_F_FLIGHT);
	io_u->flags &= ~IO_U_F_FLIGHT;

	if (io_u->ddir == DDIR_SYNC) {
		td->last_was_sync = 1;
		return;
	}

	td->last_was_sync = 0;

	if (!io_u->error) {
		unsigned int bytes = io_u->buflen - io_u->resid;
		const enum fio_ddir idx = io_u->ddir;
		int ret;

		td->io_blocks[idx]++;
		td->io_bytes[idx] += bytes;
		td->zone_bytes += bytes;
		td->this_io_bytes[idx] += bytes;

		io_u->file->last_completed_pos = io_u->offset + io_u->buflen;

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

		add_clat_sample(td, idx, msec);
		add_bw_sample(td, idx, &icd->time);
		io_u_mark_latency(td, msec);

		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
			log_io_piece(td, io_u);

		icd->bytes_done[idx] += bytes;

		if (io_u->end_io) {
			ret = io_u->end_io(io_u);
			if (ret && !icd->error)
				icd->error = ret;
		}
	} else
		icd->error = io_u->error;
}

static void init_icd(struct io_completion_data *icd, int nr)
{
	fio_gettime(&icd->time, NULL);

	icd->nr = nr;

	icd->error = 0;
	icd->bytes_done[0] = icd->bytes_done[1] = 0;
}

static void ios_completed(struct thread_data *td,
			  struct io_completion_data *icd)
{
	struct io_u *io_u;
	int i;

	for (i = 0; i < icd->nr; i++) {
		io_u = td->io_ops->event(td, i);

		io_completed(td, io_u, icd);
		put_io_u(td, io_u);
	}
}

/*
 * Complete a single io_u for the sync engines.
 */
long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
{
	struct io_completion_data icd;

	init_icd(&icd, 1);
	io_completed(td, io_u, &icd);
	put_io_u(td, io_u);

	if (!icd.error)
		return icd.bytes_done[0] + icd.bytes_done[1];

	return -1;
}

/*
 * Called to complete min_events number of io for the async engines.
 */
long io_u_queued_complete(struct thread_data *td, int min_events)
{
	struct io_completion_data icd;
	struct timespec *tvp = NULL;
	int ret;

	if (!min_events) {
		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };

		tvp = &ts;
	}

	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
	if (ret < 0) {
		td_verror(td, -ret, "td_io_getevents");
		return ret;
	} else if (!ret)
		return ret;

	init_icd(&icd, ret);
	ios_completed(td, &icd);
	if (!icd.error)
		return icd.bytes_done[0] + icd.bytes_done[1];

	return -1;
}

/*
 * Call when io_u is really queued, to update the submission latency.
 */
void io_u_queued(struct thread_data *td, struct io_u *io_u)
{
	unsigned long slat_time;

	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
	add_slat_sample(td, io_u->ddir, slat_time);
}

#ifdef FIO_USE_TIMEOUT
void io_u_set_timeout(struct thread_data *td)
{
	assert(td->cur_depth);

	td->timer.it_interval.tv_sec = 0;
	td->timer.it_interval.tv_usec = 0;
	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
	td->timer.it_value.tv_usec = 0;
	setitimer(ITIMER_REAL, &td->timer, NULL);
	fio_gettime(&td->timeout_end, NULL);
}

static void io_u_dump(struct io_u *io_u)
{
	unsigned long t_start = mtime_since_now(&io_u->start_time);
	unsigned long t_issue = mtime_since_now(&io_u->issue_time);

	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
}
#else
void io_u_set_timeout(struct thread_data fio_unused *td)
{
}
#endif

#ifdef FIO_USE_TIMEOUT
static void io_u_timeout_handler(int fio_unused sig)
{
	struct thread_data *td, *__td;
	pid_t pid = getpid();
	struct list_head *entry;
	struct io_u *io_u;
	int i;

	log_err("fio: io_u timeout\n");

	/*
	 * TLS would be nice...
	 */
	td = NULL;
	for_each_td(__td, i) {
		if (__td->pid == pid) {
			td = __td;
			break;
		}
	}

	if (!td) {
		log_err("fio: io_u timeout, can't find job\n");
		exit(1);
	}

	if (!td->cur_depth) {
		log_err("fio: timeout without pending work?\n");
		return;
	}

	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);

	list_for_each(entry, &td->io_u_busylist) {
		io_u = list_entry(entry, struct io_u, list);

		io_u_dump(io_u);
	}

	td_verror(td, ETIMEDOUT, "io_u timeout");
	exit(1);
}
#endif

void io_u_init_timeout(void)
{
#ifdef FIO_USE_TIMEOUT
	signal(SIGALRM, io_u_timeout_handler);
#endif
}
Commit	Line	Data
10ba535a JA	1	#include <unistd.h>
	2	#include <fcntl.h>
	3	#include <string.h>
	4	#include <signal.h>
	5	#include <time.h>
0c6e7517	6	#include <assert.h>
10ba535a JA	7
	8	#include "fio.h"
	9	#include "os.h"
	10
5945b9b4 JA	11	/*
	12	* Change this define to play with the timeout handling
	13	*/
	14	#undef FIO_USE_TIMEOUT
	15
97601024 JA	16	struct io_completion_data {
97601024 JA	17	int nr; /* input */
97601024 JA	18
	19	int error; /* output */
	20	unsigned long bytes_done[2]; /* output */
	21	struct timeval time; /* output */
	22	};
	23
10ba535a JA	24	/*
	25	* The ->file_map[] contains a map of blocks we have or have not done io
	26	* to yet. Used to make sure we cover the entire range in a fair fashion.
	27	*/
	28	static int random_map_free(struct thread_data td, struct fio_file f,
	29	unsigned long long block)
	30	{
	31	unsigned int idx = RAND_MAP_IDX(td, f, block);
	32	unsigned int bit = RAND_MAP_BIT(td, f, block);
	33
	34	return (f->file_map[idx] & (1UL << bit)) == 0;
	35	}
	36
df415585 JA	37	/*
	38	* Mark a given offset as used in the map.
	39	*/
	40	static void mark_random_map(struct thread_data td, struct fio_file f,
	41	struct io_u *io_u)
	42	{
afdbe580	43	unsigned int min_bs = td->rw_min_bs;
a00735e6 JA	44	unsigned long long block;
a00735e6 JA	45	unsigned int blocks;
c685b5b2	46	unsigned int nr_blocks;
df415585	47
a00735e6 JA	48	block = io_u->offset / (unsigned long long) min_bs;
a00735e6 JA	49	blocks = 0;
c685b5b2 JA	50	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
	51
	52	while (blocks < nr_blocks) {
df415585 JA	53	unsigned int idx, bit;
	54
	55	if (!random_map_free(td, f, block))
	56	break;
	57
	58	idx = RAND_MAP_IDX(td, f, block);
	59	bit = RAND_MAP_BIT(td, f, block);
	60
0032bf9f	61	fio_assert(td, idx < f->num_maps);
df415585 JA	62
	63	f->file_map[idx] \|= (1UL << bit);
	64	block++;
	65	blocks++;
	66	}
	67
a00735e6 JA	68	if ((blocks * min_bs) < io_u->buflen)
a00735e6 JA	69	io_u->buflen = blocks * min_bs;
df415585 JA	70	}
df415585 JA	71
10ba535a JA	72	/*
	73	* Return the next free block in the map.
	74	*/
	75	static int get_next_free_block(struct thread_data td, struct fio_file f,
	76	unsigned long long *b)
	77	{
	78	int i;
	79
c685b5b2 JA	80	i = f->last_free_lookup;
c685b5b2 JA	81	b = (i BLOCKS_PER_MAP);
ee88470c	82	while ((b) td->rw_min_bs < f->real_file_size) {
10ba535a JA	83	if (f->file_map[i] != -1UL) {
10ba535a JA	84	*b += ffz(f->file_map[i]);
c685b5b2	85	f->last_free_lookup = i;
10ba535a JA	86	return 0;
	87	}
	88
	89	*b += BLOCKS_PER_MAP;
	90	i++;
	91	}
	92
	93	return 1;
	94	}
	95
	96	/*
	97	* For random io, generate a random new block and see if it's used. Repeat
	98	* until we find a free one. For sequential io, just return the end of
	99	* the last io issued.
	100	*/
	101	static int get_next_offset(struct thread_data td, struct fio_file f,
c685b5b2	102	struct io_u *io_u)
10ba535a	103	{
c685b5b2	104	const int ddir = io_u->ddir;
10ba535a JA	105	unsigned long long b, rb;
	106	long r;
	107
413dd459	108	if (td_random(td)) {
f697125a	109	unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
c685b5b2	110	int loops = 5;
10ba535a JA	111
	112	do {
	113	r = os_random_long(&td->random_state);
	114	b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
bb8895e0 JA	115	if (td->norandommap)
bb8895e0 JA	116	break;
a00735e6	117	rb = b + (f->file_offset / td->min_bs[ddir]);
10ba535a JA	118	loops--;
	119	} while (!random_map_free(td, f, rb) && loops);
	120
bca4ed4d JA	121	/*
	122	* if we failed to retrieve a truly random offset within
	123	* the loops assigned, see if there are free ones left at all
	124	*/
	125	if (!loops && get_next_free_block(td, f, &b))
	126	return 1;
10ba535a	127	} else
a00735e6	128	b = f->last_pos / td->min_bs[ddir];
10ba535a	129
c685b5b2	130	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
bca4ed4d	131	if (io_u->offset >= f->real_file_size)
10ba535a JA	132	return 1;
	133
	134	return 0;
	135	}
	136
bca4ed4d JA	137	static unsigned int get_next_buflen(struct thread_data td, struct fio_file f,
bca4ed4d JA	138	struct io_u *io_u)
10ba535a	139	{
bca4ed4d	140	const int ddir = io_u->ddir;
10ba535a JA	141	unsigned int buflen;
	142	long r;
	143
a00735e6 JA	144	if (td->min_bs[ddir] == td->max_bs[ddir])
a00735e6 JA	145	buflen = td->min_bs[ddir];
10ba535a JA	146	else {
10ba535a JA	147	r = os_random_long(&td->bsrange_state);
1e97cce9	148	buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
690adba3	149	if (!td->bs_unaligned)
a00735e6	150	buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
10ba535a JA	151	}
10ba535a JA	152
bca4ed4d	153	while (buflen + io_u->offset > f->real_file_size) {
bca4ed4d JA	154	if (buflen == td->min_bs[ddir])
	155	return 0;
	156
	157	buflen = td->min_bs[ddir];
10ba535a JA	158	}
	159
	160	return buflen;
	161	}
	162
	163	/*
	164	* Return the data direction for the next io_u. If the job is a
	165	* mixed read/write workload, check the rwmix cycle and switch if
	166	* necessary.
	167	*/
1e97cce9	168	static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a JA	169	{
	170	if (td_rw(td)) {
	171	struct timeval now;
	172	unsigned long elapsed;
	173
02bcaa8c	174	fio_gettime(&now, NULL);
10ba535a JA	175	elapsed = mtime_since_now(&td->rwmix_switch);
	176
	177	/*
	178	* Check if it's time to seed a new data direction.
	179	*/
	180	if (elapsed >= td->rwmixcycle) {
	181	unsigned int v;
	182	long r;
	183
	184	r = os_random_long(&td->rwmix_state);
	185	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
	186	if (v < td->rwmixread)
	187	td->rwmix_ddir = DDIR_READ;
	188	else
	189	td->rwmix_ddir = DDIR_WRITE;
	190	memcpy(&td->rwmix_switch, &now, sizeof(now));
	191	}
	192	return td->rwmix_ddir;
	193	} else if (td_read(td))
	194	return DDIR_READ;
	195	else
	196	return DDIR_WRITE;
	197	}
	198
10ba535a JA	199	void put_io_u(struct thread_data td, struct io_u io_u)
10ba535a JA	200	{
0c6e7517 JA	201	assert((io_u->flags & IO_U_F_FREE) == 0);
	202	io_u->flags \|= IO_U_F_FREE;
	203
10ba535a JA	204	io_u->file = NULL;
	205	list_del(&io_u->list);
	206	list_add(&io_u->list, &td->io_u_freelist);
	207	td->cur_depth--;
	208	}
	209
755200a3 JA	210	void requeue_io_u(struct thread_data td, struct io_u *io_u)
	211	{
	212	struct io_u __io_u = io_u;
	213
4d2e0f49 JA	214	__io_u->flags \|= IO_U_F_FREE;
	215	__io_u->flags &= ~IO_U_F_FLIGHT;
	216
755200a3 JA	217	list_del(&__io_u->list);
	218	list_add_tail(&__io_u->list, &td->io_u_requeues);
	219	td->cur_depth--;
	220	*io_u = NULL;
	221	}
	222
10ba535a JA	223	static int fill_io_u(struct thread_data td, struct fio_file f,
	224	struct io_u *io_u)
	225	{
	226	/*
	227	* If using an iolog, grab next piece if any available.
	228	*/
	229	if (td->read_iolog)
	230	return read_iolog_get(td, io_u);
	231
87dc1ab1 JA	232	/*
	233	* see if it's time to sync
	234	*/
755200a3 JA	235	if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
755200a3 JA	236	&& td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 JA	237	io_u->ddir = DDIR_SYNC;
	238	io_u->file = f;
	239	return 0;
	240	}
	241
a00735e6 JA	242	io_u->ddir = get_rw_ddir(td);
a00735e6 JA	243
10ba535a	244	/*
c685b5b2 JA	245	* No log, let the seq/rand engine retrieve the next buflen and
c685b5b2 JA	246	* position.
10ba535a	247	*/
bca4ed4d JA	248	if (get_next_offset(td, f, io_u))
bca4ed4d JA	249	return 1;
10ba535a	250
bca4ed4d JA	251	io_u->buflen = get_next_buflen(td, f, io_u);
	252	if (!io_u->buflen)
	253	return 1;
	254
	255	/*
	256	* mark entry before potentially trimming io_u
	257	*/
413dd459	258	if (!td->read_iolog && td_random(td) && !td->norandommap)
bca4ed4d JA	259	mark_random_map(td, f, io_u);
	260
	261	/*
	262	* If using a write iolog, store this entry.
	263	*/
	264	if (td->write_iolog_file)
	265	write_iolog_put(td, io_u);
	266
	267	io_u->file = f;
	268	return 0;
10ba535a JA	269	}
10ba535a JA	270
71619dc2 JA	271	static void io_u_mark_depth(struct thread_data *td)
	272	{
	273	int index = 0;
	274
	275	switch (td->cur_depth) {
	276	default:
	277	index++;
	278	case 32 ... 63:
	279	index++;
	280	case 16 ... 31:
	281	index++;
	282	case 8 ... 15:
	283	index++;
	284	case 4 ... 7:
	285	index++;
	286	case 2 ... 3:
	287	index++;
	288	case 1:
	289	break;
	290	}
	291
	292	td->io_u_map[index]++;
	293	td->total_io_u++;
	294	}
	295
ec118304 JA	296	static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
	297	{
	298	int index = 0;
	299
	300	switch (msec) {
	301	default:
	302	index++;
8abdce66	303	case 1000 ... 1999:
ec118304	304	index++;
8abdce66	305	case 750 ... 999:
ec118304	306	index++;
8abdce66	307	case 500 ... 749:
ec118304	308	index++;
8abdce66	309	case 250 ... 499:
ec118304	310	index++;
8abdce66	311	case 100 ... 249:
ec118304	312	index++;
8abdce66	313	case 50 ... 99:
ec118304	314	index++;
8abdce66	315	case 20 ... 49:
ec118304	316	index++;
8abdce66	317	case 10 ... 19:
ec118304	318	index++;
8abdce66	319	case 4 ... 9:
ec118304 JA	320	index++;
	321	case 2 ... 3:
	322	index++;
	323	case 0 ... 1:
	324	break;
	325	}
	326
	327	td->io_u_lat[index]++;
	328	}
	329
0aabe160 JA	330	/*
	331	* Get next file to service by choosing one at random
	332	*/
	333	static struct fio_file get_next_file_rand(struct thread_data td)
	334	{
	335	long r = os_random_long(&td->next_file_state);
	336	unsigned int fileno;
	337	struct fio_file *f;
	338
	339	do {
	340	fileno = (unsigned int) ((double) (td->nr_files - 1) * r / (RAND_MAX + 1.0));
	341	f = &td->files[fileno];
	342	if (f->fd != -1)
	343	return f;
	344	} while (1);
	345	}
	346
	347	/*
	348	* Get next file to service by doing round robin between all available ones
	349	*/
	350	static struct fio_file get_next_file_rr(struct thread_data td)
3d7c391d JA	351	{
	352	unsigned int old_next_file = td->next_file;
	353	struct fio_file *f;
	354
	355	do {
	356	f = &td->files[td->next_file];
	357
	358	td->next_file++;
	359	if (td->next_file >= td->nr_files)
	360	td->next_file = 0;
	361
	362	if (f->fd != -1)
	363	break;
	364
	365	f = NULL;
	366	} while (td->next_file != old_next_file);
	367
	368	return f;
	369	}
	370
10ba535a JA	371	struct io_u __get_io_u(struct thread_data td)
	372	{
	373	struct io_u *io_u = NULL;
	374
755200a3 JA	375	if (!list_empty(&td->io_u_requeues))
	376	io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
	377	else if (!queue_full(td)) {
10ba535a JA	378	io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a JA	379
6040dabc	380	io_u->buflen = 0;
10ba535a	381	io_u->resid = 0;
755200a3	382	io_u->file = NULL;
d7762cf8	383	io_u->end_io = NULL;
755200a3 JA	384	}
	385
	386	if (io_u) {
0c6e7517 JA	387	assert(io_u->flags & IO_U_F_FREE);
	388	io_u->flags &= ~IO_U_F_FREE;
	389
755200a3	390	io_u->error = 0;
10ba535a JA	391	list_del(&io_u->list);
	392	list_add(&io_u->list, &td->io_u_busylist);
	393	td->cur_depth++;
71619dc2	394	io_u_mark_depth(td);
10ba535a JA	395	}
	396
	397	return io_u;
	398	}
	399
	400	/*
	401	* Return an io_u to be processed. Gets a buflen and offset, sets direction,
	402	* etc. The returned io_u is fully ready to be prepped and submitted.
	403	*/
3d7c391d	404	struct io_u get_io_u(struct thread_data td)
10ba535a	405	{
3d7c391d	406	struct fio_file *f;
10ba535a JA	407	struct io_u *io_u;
	408
	409	io_u = __get_io_u(td);
	410	if (!io_u)
	411	return NULL;
	412
755200a3 JA	413	/*
	414	* from a requeue, io_u already setup
	415	*/
	416	if (io_u->file)
77f392bf	417	goto out;
755200a3	418
0aabe160 JA	419	if (td->file_service_type == FIO_FSERVICE_RR)
	420	f = get_next_file_rr(td);
	421	else
	422	f = get_next_file_rand(td);
	423
3d7c391d JA	424	if (!f) {
	425	put_io_u(td, io_u);
	426	return NULL;
	427	}
	428
	429	io_u->file = f;
	430
10ba535a JA	431	if (td->zone_bytes >= td->zone_size) {
	432	td->zone_bytes = 0;
	433	f->last_pos += td->zone_skip;
	434	}
	435
	436	if (fill_io_u(td, f, io_u)) {
	437	put_io_u(td, io_u);
	438	return NULL;
	439	}
	440
ee88470c	441	if (io_u->buflen + io_u->offset > f->real_file_size) {
10ba535a JA	442	if (td->io_ops->flags & FIO_RAWIO) {
	443	put_io_u(td, io_u);
	444	return NULL;
	445	}
	446
ee88470c	447	io_u->buflen = f->real_file_size - io_u->offset;
10ba535a JA	448	}
10ba535a JA	449
87dc1ab1 JA	450	if (io_u->ddir != DDIR_SYNC) {
	451	if (!io_u->buflen) {
	452	put_io_u(td, io_u);
	453	return NULL;
	454	}
10ba535a	455
36167d82	456	f->last_pos = io_u->offset + io_u->buflen;
10ba535a	457
87dc1ab1 JA	458	if (td->verify != VERIFY_NONE)
	459	populate_verify_io_u(td, io_u);
	460	}
10ba535a	461
165faf16 JA	462	/*
	463	* Set io data pointers.
	464	*/
77f392bf	465	out:
cec6b55d JA	466	io_u->xfer_buf = io_u->buf;
cec6b55d JA	467	io_u->xfer_buflen = io_u->buflen;
165faf16	468
36167d82 JA	469	if (td_io_prep(td, io_u)) {
	470	put_io_u(td, io_u);
	471	return NULL;
	472	}
	473
02bcaa8c	474	fio_gettime(&io_u->start_time, NULL);
10ba535a JA	475	return io_u;
	476	}
	477
97601024 JA	478	static void io_completed(struct thread_data td, struct io_u io_u,
97601024 JA	479	struct io_completion_data *icd)
10ba535a	480	{
10ba535a JA	481	unsigned long msec;
10ba535a JA	482
0c6e7517 JA	483	assert(io_u->flags & IO_U_F_FLIGHT);
	484	io_u->flags &= ~IO_U_F_FLIGHT;
	485
87dc1ab1 JA	486	if (io_u->ddir == DDIR_SYNC) {
	487	td->last_was_sync = 1;
	488	return;
	489	}
	490
	491	td->last_was_sync = 0;
	492
10ba535a JA	493	if (!io_u->error) {
10ba535a JA	494	unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9	495	const enum fio_ddir idx = io_u->ddir;
3af6ef39	496	int ret;
10ba535a JA	497
	498	td->io_blocks[idx]++;
	499	td->io_bytes[idx] += bytes;
	500	td->zone_bytes += bytes;
	501	td->this_io_bytes[idx] += bytes;
	502
02bcaa8c JA	503	io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
	504
	505	msec = mtime_since(&io_u->issue_time, &icd->time);
10ba535a JA	506
10ba535a JA	507	add_clat_sample(td, idx, msec);
02bcaa8c	508	add_bw_sample(td, idx, &icd->time);
ec118304	509	io_u_mark_latency(td, msec);
10ba535a JA	510
	511	if ((td_rw(td) \|\| td_write(td)) && idx == DDIR_WRITE)
	512	log_io_piece(td, io_u);
	513
	514	icd->bytes_done[idx] += bytes;
3af6ef39	515
d7762cf8 JA	516	if (io_u->end_io) {
d7762cf8 JA	517	ret = io_u->end_io(io_u);
3af6ef39 JA	518	if (ret && !icd->error)
	519	icd->error = ret;
	520	}
10ba535a JA	521	} else
	522	icd->error = io_u->error;
	523	}
	524
d7762cf8	525	static void init_icd(struct io_completion_data *icd, int nr)
10ba535a	526	{
02bcaa8c JA	527	fio_gettime(&icd->time, NULL);
02bcaa8c JA	528
3af6ef39 JA	529	icd->nr = nr;
3af6ef39 JA	530
10ba535a JA	531	icd->error = 0;
10ba535a JA	532	icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82 JA	533	}
36167d82 JA	534
97601024 JA	535	static void ios_completed(struct thread_data *td,
97601024 JA	536	struct io_completion_data *icd)
36167d82 JA	537	{
	538	struct io_u *io_u;
	539	int i;
	540
10ba535a JA	541	for (i = 0; i < icd->nr; i++) {
	542	io_u = td->io_ops->event(td, i);
	543
	544	io_completed(td, io_u, icd);
	545	put_io_u(td, io_u);
	546	}
	547	}
97601024	548
e7e6cfb4 JA	549	/*
	550	* Complete a single io_u for the sync engines.
	551	*/
d7762cf8	552	long io_u_sync_complete(struct thread_data td, struct io_u io_u)
97601024 JA	553	{
	554	struct io_completion_data icd;
	555
d7762cf8	556	init_icd(&icd, 1);
97601024 JA	557	io_completed(td, io_u, &icd);
	558	put_io_u(td, io_u);
	559
	560	if (!icd.error)
	561	return icd.bytes_done[0] + icd.bytes_done[1];
	562
97601024 JA	563	return -1;
	564	}
	565
e7e6cfb4 JA	566	/*
	567	* Called to complete min_events number of io for the async engines.
	568	*/
d7762cf8	569	long io_u_queued_complete(struct thread_data *td, int min_events)
97601024	570	{
97601024	571	struct io_completion_data icd;
00de55ef	572	struct timespec *tvp = NULL;
97601024 JA	573	int ret;
97601024 JA	574
face81b2	575	if (!min_events) {
00de55ef JA	576	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
	577
	578	tvp = &ts;
755200a3	579	}
97601024	580
00de55ef	581	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
97601024	582	if (ret < 0) {
e1161c32	583	td_verror(td, -ret, "td_io_getevents");
97601024 JA	584	return ret;
	585	} else if (!ret)
	586	return ret;
	587
d7762cf8	588	init_icd(&icd, ret);
97601024 JA	589	ios_completed(td, &icd);
	590	if (!icd.error)
	591	return icd.bytes_done[0] + icd.bytes_done[1];
	592
97601024 JA	593	return -1;
97601024 JA	594	}
7e77dd02 JA	595
	596	/*
	597	* Call when io_u is really queued, to update the submission latency.
	598	*/
	599	void io_u_queued(struct thread_data td, struct io_u io_u)
	600	{
	601	unsigned long slat_time;
	602
	603	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
	604	add_slat_sample(td, io_u->ddir, slat_time);
	605	}
433afcb4	606
55bc9728	607	#ifdef FIO_USE_TIMEOUT
433afcb4 JA	608	void io_u_set_timeout(struct thread_data *td)
	609	{
	610	assert(td->cur_depth);
	611
	612	td->timer.it_interval.tv_sec = 0;
	613	td->timer.it_interval.tv_usec = 0;
	614	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
	615	td->timer.it_value.tv_usec = 0;
	616	setitimer(ITIMER_REAL, &td->timer, NULL);
	617	fio_gettime(&td->timeout_end, NULL);
	618	}
5945b9b4 JA	619
	620	static void io_u_dump(struct io_u *io_u)
	621	{
	622	unsigned long t_start = mtime_since_now(&io_u->start_time);
	623	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
	624
	625	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
	626	log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
	627	log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
	628	}
55bc9728 JA	629	#else
	630	void io_u_set_timeout(struct thread_data fio_unused *td)
	631	{
	632	}
	633	#endif
433afcb4	634
55bc9728	635	#ifdef FIO_USE_TIMEOUT
433afcb4 JA	636	static void io_u_timeout_handler(int fio_unused sig)
	637	{
	638	struct thread_data td, __td;
	639	pid_t pid = getpid();
5945b9b4 JA	640	struct list_head *entry;
5945b9b4 JA	641	struct io_u *io_u;
433afcb4 JA	642	int i;
	643
	644	log_err("fio: io_u timeout\n");
	645
	646	/*
	647	* TLS would be nice...
	648	*/
	649	td = NULL;
	650	for_each_td(__td, i) {
	651	if (__td->pid == pid) {
	652	td = __td;
	653	break;
	654	}
	655	}
	656
	657	if (!td) {
	658	log_err("fio: io_u timeout, can't find job\n");
	659	exit(1);
	660	}
	661
	662	if (!td->cur_depth) {
	663	log_err("fio: timeout without pending work?\n");
	664	return;
	665	}
	666
	667	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
5945b9b4 JA	668
	669	list_for_each(entry, &td->io_u_busylist) {
	670	io_u = list_entry(entry, struct io_u, list);
	671
	672	io_u_dump(io_u);
	673	}
	674
	675	td_verror(td, ETIMEDOUT, "io_u timeout");
433afcb4 JA	676	exit(1);
433afcb4 JA	677	}
55bc9728	678	#endif
433afcb4 JA	679
	680	void io_u_init_timeout(void)
	681	{
55bc9728	682	#ifdef FIO_USE_TIMEOUT
433afcb4	683	signal(SIGALRM, io_u_timeout_handler);
55bc9728	684	#endif
433afcb4	685	}