[fio.git] / io_u.c

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

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

struct io_completion_data {
	int nr;				/* input */
	endio_handler *handler;		/* 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->sequential) {
		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)
{
	io_u->file = NULL;
	list_del(&io_u->list);
	list_add(&io_u->list, &td->io_u_freelist);
	td->cur_depth--;
}

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_blocks[DDIR_WRITE] % td->fsync_blocks)
	    && 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->sequential && !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 1024 ... 2047:
		index++;
	case 512 ... 1023:
		index++;
	case 256 ... 511:
		index++;
	case 128 ... 255:
		index++;
	case 64 ... 127:
		index++;
	case 32 ... 63:
		index++;
	case 16 ... 31:
		index++;
	case 8 ... 15:
		index++;
	case 4 ... 7:
		index++;
	case 2 ... 3:
		index++;
	case 0 ... 1:
		break;
	}

	td->io_u_lat[index]++;
}

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

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

		io_u->buflen = 0;
		io_u->error = 0;
		io_u->resid = 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;

	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.
	 */
	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;

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

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

	icd->handler = handler;
	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);
	}
}

long io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
			endio_handler *handler)
{
	struct io_completion_data icd;

	init_icd(&icd, handler, 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;
}

long io_u_queued_complete(struct thread_data *td, int min_events,
			  endio_handler *handler)

{
	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
	struct timespec *tsp = NULL;
	struct io_completion_data icd;
	int ret;

	if (min_events > 0)
		tsp = &ts;

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

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

	return -1;
}
Commit	Line	Data
	1	#include <unistd.h>
	2	#include <fcntl.h>
	3	#include <string.h>
	4	#include <signal.h>
	5	#include <time.h>
	6
	7	#include "fio.h"
	8	#include "os.h"
	9
	10	struct io_completion_data {
	11	int nr; /* input */
	12	endio_handler handler; / input */
	13
	14	int error; /* output */
	15	unsigned long bytes_done[2]; /* output */
	16	struct timeval time; /* output */
	17	};
	18
	19	/*
	20	* The ->file_map[] contains a map of blocks we have or have not done io
	21	* to yet. Used to make sure we cover the entire range in a fair fashion.
	22	*/
	23	static int random_map_free(struct thread_data td, struct fio_file f,
	24	unsigned long long block)
	25	{
	26	unsigned int idx = RAND_MAP_IDX(td, f, block);
	27	unsigned int bit = RAND_MAP_BIT(td, f, block);
	28
	29	return (f->file_map[idx] & (1UL << bit)) == 0;
	30	}
	31
	32	/*
	33	* Mark a given offset as used in the map.
	34	*/
	35	static void mark_random_map(struct thread_data td, struct fio_file f,
	36	struct io_u *io_u)
	37	{
	38	unsigned int min_bs = td->rw_min_bs;
	39	unsigned long long block;
	40	unsigned int blocks;
	41	unsigned int nr_blocks;
	42
	43	block = io_u->offset / (unsigned long long) min_bs;
	44	blocks = 0;
	45	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
	46
	47	while (blocks < nr_blocks) {
	48	unsigned int idx, bit;
	49
	50	if (!random_map_free(td, f, block))
	51	break;
	52
	53	idx = RAND_MAP_IDX(td, f, block);
	54	bit = RAND_MAP_BIT(td, f, block);
	55
	56	fio_assert(td, idx < f->num_maps);
	57
	58	f->file_map[idx] \|= (1UL << bit);
	59	block++;
	60	blocks++;
	61	}
	62
	63	if ((blocks * min_bs) < io_u->buflen)
	64	io_u->buflen = blocks * min_bs;
	65	}
	66
	67	/*
	68	* Return the next free block in the map.
	69	*/
	70	static int get_next_free_block(struct thread_data td, struct fio_file f,
	71	unsigned long long *b)
	72	{
	73	int i;
	74
	75	i = f->last_free_lookup;
	76	b = (i BLOCKS_PER_MAP);
	77	while ((b) td->rw_min_bs < f->real_file_size) {
	78	if (f->file_map[i] != -1UL) {
	79	*b += ffz(f->file_map[i]);
	80	f->last_free_lookup = i;
	81	return 0;
	82	}
	83
	84	*b += BLOCKS_PER_MAP;
	85	i++;
	86	}
	87
	88	return 1;
	89	}
	90
	91	/*
	92	* For random io, generate a random new block and see if it's used. Repeat
	93	* until we find a free one. For sequential io, just return the end of
	94	* the last io issued.
	95	*/
	96	static int get_next_offset(struct thread_data td, struct fio_file f,
	97	struct io_u *io_u)
	98	{
	99	const int ddir = io_u->ddir;
	100	unsigned long long b, rb;
	101	long r;
	102
	103	if (!td->sequential) {
	104	unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
	105	int loops = 5;
	106
	107	do {
	108	r = os_random_long(&td->random_state);
	109	b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
	110	if (td->norandommap)
	111	break;
	112	rb = b + (f->file_offset / td->min_bs[ddir]);
	113	loops--;
	114	} while (!random_map_free(td, f, rb) && loops);
	115
	116	/*
	117	* if we failed to retrieve a truly random offset within
	118	* the loops assigned, see if there are free ones left at all
	119	*/
	120	if (!loops && get_next_free_block(td, f, &b))
	121	return 1;
	122	} else
	123	b = f->last_pos / td->min_bs[ddir];
	124
	125	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
	126	if (io_u->offset >= f->real_file_size)
	127	return 1;
	128
	129	return 0;
	130	}
	131
	132	static unsigned int get_next_buflen(struct thread_data td, struct fio_file f,
	133	struct io_u *io_u)
	134	{
	135	const int ddir = io_u->ddir;
	136	unsigned int buflen;
	137	long r;
	138
	139	if (td->min_bs[ddir] == td->max_bs[ddir])
	140	buflen = td->min_bs[ddir];
	141	else {
	142	r = os_random_long(&td->bsrange_state);
	143	buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
	144	if (!td->bs_unaligned)
	145	buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
	146	}
	147
	148	while (buflen + io_u->offset > f->real_file_size) {
	149	if (buflen == td->min_bs[ddir])
	150	return 0;
	151
	152	buflen = td->min_bs[ddir];
	153	}
	154
	155	return buflen;
	156	}
	157
	158	/*
	159	* Return the data direction for the next io_u. If the job is a
	160	* mixed read/write workload, check the rwmix cycle and switch if
	161	* necessary.
	162	*/
	163	static enum fio_ddir get_rw_ddir(struct thread_data *td)
	164	{
	165	if (td_rw(td)) {
	166	struct timeval now;
	167	unsigned long elapsed;
	168
	169	fio_gettime(&now, NULL);
	170	elapsed = mtime_since_now(&td->rwmix_switch);
	171
	172	/*
	173	* Check if it's time to seed a new data direction.
	174	*/
	175	if (elapsed >= td->rwmixcycle) {
	176	unsigned int v;
	177	long r;
	178
	179	r = os_random_long(&td->rwmix_state);
	180	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
	181	if (v < td->rwmixread)
	182	td->rwmix_ddir = DDIR_READ;
	183	else
	184	td->rwmix_ddir = DDIR_WRITE;
	185	memcpy(&td->rwmix_switch, &now, sizeof(now));
	186	}
	187	return td->rwmix_ddir;
	188	} else if (td_read(td))
	189	return DDIR_READ;
	190	else
	191	return DDIR_WRITE;
	192	}
	193
	194	void put_io_u(struct thread_data td, struct io_u io_u)
	195	{
	196	io_u->file = NULL;
	197	list_del(&io_u->list);
	198	list_add(&io_u->list, &td->io_u_freelist);
	199	td->cur_depth--;
	200	}
	201
	202	static int fill_io_u(struct thread_data td, struct fio_file f,
	203	struct io_u *io_u)
	204	{
	205	/*
	206	* If using an iolog, grab next piece if any available.
	207	*/
	208	if (td->read_iolog)
	209	return read_iolog_get(td, io_u);
	210
	211	/*
	212	* see if it's time to sync
	213	*/
	214	if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
	215	&& should_fsync(td)) {
	216	io_u->ddir = DDIR_SYNC;
	217	io_u->file = f;
	218	return 0;
	219	}
	220
	221	io_u->ddir = get_rw_ddir(td);
	222
	223	/*
	224	* No log, let the seq/rand engine retrieve the next buflen and
	225	* position.
	226	*/
	227	if (get_next_offset(td, f, io_u))
	228	return 1;
	229
	230	io_u->buflen = get_next_buflen(td, f, io_u);
	231	if (!io_u->buflen)
	232	return 1;
	233
	234	/*
	235	* mark entry before potentially trimming io_u
	236	*/
	237	if (!td->read_iolog && !td->sequential && !td->norandommap)
	238	mark_random_map(td, f, io_u);
	239
	240	/*
	241	* If using a write iolog, store this entry.
	242	*/
	243	if (td->write_iolog_file)
	244	write_iolog_put(td, io_u);
	245
	246	io_u->file = f;
	247	return 0;
	248	}
	249
	250	static void io_u_mark_depth(struct thread_data *td)
	251	{
	252	int index = 0;
	253
	254	switch (td->cur_depth) {
	255	default:
	256	index++;
	257	case 32 ... 63:
	258	index++;
	259	case 16 ... 31:
	260	index++;
	261	case 8 ... 15:
	262	index++;
	263	case 4 ... 7:
	264	index++;
	265	case 2 ... 3:
	266	index++;
	267	case 1:
	268	break;
	269	}
	270
	271	td->io_u_map[index]++;
	272	td->total_io_u++;
	273	}
	274
	275	static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
	276	{
	277	int index = 0;
	278
	279	switch (msec) {
	280	default:
	281	index++;
	282	case 1024 ... 2047:
	283	index++;
	284	case 512 ... 1023:
	285	index++;
	286	case 256 ... 511:
	287	index++;
	288	case 128 ... 255:
	289	index++;
	290	case 64 ... 127:
	291	index++;
	292	case 32 ... 63:
	293	index++;
	294	case 16 ... 31:
	295	index++;
	296	case 8 ... 15:
	297	index++;
	298	case 4 ... 7:
	299	index++;
	300	case 2 ... 3:
	301	index++;
	302	case 0 ... 1:
	303	break;
	304	}
	305
	306	td->io_u_lat[index]++;
	307	}
	308
	309	struct io_u __get_io_u(struct thread_data td)
	310	{
	311	struct io_u *io_u = NULL;
	312
	313	if (!queue_full(td)) {
	314	io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
	315
	316	io_u->buflen = 0;
	317	io_u->error = 0;
	318	io_u->resid = 0;
	319	list_del(&io_u->list);
	320	list_add(&io_u->list, &td->io_u_busylist);
	321	td->cur_depth++;
	322	io_u_mark_depth(td);
	323	}
	324
	325	return io_u;
	326	}
	327
	328	/*
	329	* Return an io_u to be processed. Gets a buflen and offset, sets direction,
	330	* etc. The returned io_u is fully ready to be prepped and submitted.
	331	*/
	332	struct io_u get_io_u(struct thread_data td, struct fio_file *f)
	333	{
	334	struct io_u *io_u;
	335
	336	io_u = __get_io_u(td);
	337	if (!io_u)
	338	return NULL;
	339
	340	if (td->zone_bytes >= td->zone_size) {
	341	td->zone_bytes = 0;
	342	f->last_pos += td->zone_skip;
	343	}
	344
	345	if (fill_io_u(td, f, io_u)) {
	346	put_io_u(td, io_u);
	347	return NULL;
	348	}
	349
	350	if (io_u->buflen + io_u->offset > f->real_file_size) {
	351	if (td->io_ops->flags & FIO_RAWIO) {
	352	put_io_u(td, io_u);
	353	return NULL;
	354	}
	355
	356	io_u->buflen = f->real_file_size - io_u->offset;
	357	}
	358
	359	if (io_u->ddir != DDIR_SYNC) {
	360	if (!io_u->buflen) {
	361	put_io_u(td, io_u);
	362	return NULL;
	363	}
	364
	365	f->last_pos = io_u->offset + io_u->buflen;
	366
	367	if (td->verify != VERIFY_NONE)
	368	populate_verify_io_u(td, io_u);
	369	}
	370
	371	/*
	372	* Set io data pointers.
	373	*/
	374	io_u->xfer_buf = io_u->buf;
	375	io_u->xfer_buflen = io_u->buflen;
	376
	377	if (td_io_prep(td, io_u)) {
	378	put_io_u(td, io_u);
	379	return NULL;
	380	}
	381
	382	fio_gettime(&io_u->start_time, NULL);
	383	return io_u;
	384	}
	385
	386	static void io_completed(struct thread_data td, struct io_u io_u,
	387	struct io_completion_data *icd)
	388	{
	389	unsigned long msec;
	390
	391	if (io_u->ddir == DDIR_SYNC) {
	392	td->last_was_sync = 1;
	393	return;
	394	}
	395
	396	td->last_was_sync = 0;
	397
	398	if (!io_u->error) {
	399	unsigned int bytes = io_u->buflen - io_u->resid;
	400	const enum fio_ddir idx = io_u->ddir;
	401	int ret;
	402
	403	td->io_blocks[idx]++;
	404	td->io_bytes[idx] += bytes;
	405	td->zone_bytes += bytes;
	406	td->this_io_bytes[idx] += bytes;
	407
	408	io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
	409
	410	msec = mtime_since(&io_u->issue_time, &icd->time);
	411
	412	add_clat_sample(td, idx, msec);
	413	add_bw_sample(td, idx, &icd->time);
	414	io_u_mark_latency(td, msec);
	415
	416	if ((td_rw(td) \|\| td_write(td)) && idx == DDIR_WRITE)
	417	log_io_piece(td, io_u);
	418
	419	icd->bytes_done[idx] += bytes;
	420
	421	if (icd->handler) {
	422	ret = icd->handler(io_u);
	423	if (ret && !icd->error)
	424	icd->error = ret;
	425	}
	426	} else
	427	icd->error = io_u->error;
	428	}
	429
	430	static void init_icd(struct io_completion_data icd, endio_handler handler,
	431	int nr)
	432	{
	433	fio_gettime(&icd->time, NULL);
	434
	435	icd->handler = handler;
	436	icd->nr = nr;
	437
	438	icd->error = 0;
	439	icd->bytes_done[0] = icd->bytes_done[1] = 0;
	440	}
	441
	442	static void ios_completed(struct thread_data *td,
	443	struct io_completion_data *icd)
	444	{
	445	struct io_u *io_u;
	446	int i;
	447
	448	for (i = 0; i < icd->nr; i++) {
	449	io_u = td->io_ops->event(td, i);
	450
	451	io_completed(td, io_u, icd);
	452	put_io_u(td, io_u);
	453	}
	454	}
	455
	456	long io_u_sync_complete(struct thread_data td, struct io_u io_u,
	457	endio_handler *handler)
	458	{
	459	struct io_completion_data icd;
	460
	461	init_icd(&icd, handler, 1);
	462	io_completed(td, io_u, &icd);
	463	put_io_u(td, io_u);
	464
	465	if (!icd.error)
	466	return icd.bytes_done[0] + icd.bytes_done[1];
	467
	468	return -1;
	469	}
	470
	471	long io_u_queued_complete(struct thread_data *td, int min_events,
	472	endio_handler *handler)
	473
	474	{
	475	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
	476	struct timespec *tsp = NULL;
	477	struct io_completion_data icd;
	478	int ret;
	479
	480	if (min_events > 0)
	481	tsp = &ts;
	482
	483	ret = td_io_getevents(td, min_events, td->cur_depth, tsp);
	484	if (ret < 0) {
	485	td_verror(td, -ret);
	486	return ret;
	487	} else if (!ret)
	488	return ret;
	489
	490	init_icd(&icd, handler, ret);
	491	ios_completed(td, &icd);
	492	if (!icd.error)
	493	return icd.bytes_done[0] + icd.bytes_done[1];
	494
	495	return -1;
	496	}