[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"

/*
 * 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 long long block = io_u->offset / (unsigned long long) td->min_bs;
	unsigned int blocks = 0;

	while (blocks < (io_u->buflen / td->min_bs)) {
		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);

		assert(idx < f->num_maps);

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

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

	*b = 0;
	i = 0;
	while ((*b) * td->min_bs < f->file_size) {
		if (f->file_map[i] != -1UL) {
			*b += ffz(f->file_map[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,
			   unsigned long long *offset)
{
	unsigned long long b, rb;
	long r;

	if (!td->sequential) {
		unsigned long long max_blocks = td->io_size / td->min_bs;
		int loops = 50;

		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);
			loops--;
		} while (!random_map_free(td, f, rb) && loops);

		if (!loops) {
			if (get_next_free_block(td, f, &b))
				return 1;
		}
	} else
		b = f->last_pos / td->min_bs;

	*offset = (b * td->min_bs) + f->file_offset;
	if (*offset > f->file_size)
		return 1;

	return 0;
}

static unsigned int get_next_buflen(struct thread_data *td)
{
	unsigned int buflen;
	long r;

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

	if (buflen > td->io_size - td->this_io_bytes[td->ddir]) {
		/*
		 * if using direct/raw io, we may not be able to
		 * shrink the size. so just fail it.
		 */
		if (td->io_ops->flags & FIO_RAWIO)
			return 0;

		buflen = td->io_size - td->this_io_bytes[td->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 int get_rw_ddir(struct thread_data *td)
{
	if (td_rw(td)) {
		struct timeval now;
		unsigned long elapsed;

		gettimeofday(&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;
	}

	/*
	 * No log, let the seq/rand engine retrieve the next position.
	 */
	if (!get_next_offset(td, f, &io_u->offset)) {
		io_u->buflen = get_next_buflen(td);

		if (io_u->buflen) {
			io_u->ddir = get_rw_ddir(td);

			/*
			 * 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;
		}
	}

	return 1;
}

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++;
	}

	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->file_size) {
		if (td->io_ops->flags & FIO_RAWIO) {
			put_io_u(td, io_u);
			return NULL;
		}

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

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

		if (!td->read_iolog && !td->sequential && !td->norandommap)
			mark_random_map(td, f, io_u);

		f->last_pos += io_u->buflen;

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

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

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

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

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

	td->last_was_sync = 0;

	gettimeofday(&e, NULL);

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

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

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

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

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

		icd->bytes_done[idx] += bytes;
	} else
		icd->error = io_u->error;
}

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

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

	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);
	}
}
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>
	6	#include <assert.h>
	7
	8	#include "fio.h"
	9	#include "os.h"
	10
	11	/*
	12	* The ->file_map[] contains a map of blocks we have or have not done io
	13	* to yet. Used to make sure we cover the entire range in a fair fashion.
	14	*/
	15	static int random_map_free(struct thread_data td, struct fio_file f,
	16	unsigned long long block)
	17	{
	18	unsigned int idx = RAND_MAP_IDX(td, f, block);
	19	unsigned int bit = RAND_MAP_BIT(td, f, block);
	20
	21	return (f->file_map[idx] & (1UL << bit)) == 0;
	22	}
	23
df415585 JA	24	/*
	25	* Mark a given offset as used in the map.
	26	*/
	27	static void mark_random_map(struct thread_data td, struct fio_file f,
	28	struct io_u *io_u)
	29	{
	30	unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
	31	unsigned int blocks = 0;
	32
	33	while (blocks < (io_u->buflen / td->min_bs)) {
	34	unsigned int idx, bit;
	35
	36	if (!random_map_free(td, f, block))
	37	break;
	38
	39	idx = RAND_MAP_IDX(td, f, block);
	40	bit = RAND_MAP_BIT(td, f, block);
	41
	42	assert(idx < f->num_maps);
	43
	44	f->file_map[idx] \|= (1UL << bit);
	45	block++;
	46	blocks++;
	47	}
	48
	49	if ((blocks * td->min_bs) < io_u->buflen)
	50	io_u->buflen = blocks * td->min_bs;
	51	}
	52
10ba535a JA	53	/*
	54	* Return the next free block in the map.
	55	*/
	56	static int get_next_free_block(struct thread_data td, struct fio_file f,
	57	unsigned long long *b)
	58	{
	59	int i;
	60
	61	*b = 0;
	62	i = 0;
	63	while ((b) td->min_bs < f->file_size) {
	64	if (f->file_map[i] != -1UL) {
	65	*b += ffz(f->file_map[i]);
	66	return 0;
	67	}
	68
	69	*b += BLOCKS_PER_MAP;
	70	i++;
	71	}
	72
	73	return 1;
	74	}
	75
	76	/*
	77	* For random io, generate a random new block and see if it's used. Repeat
	78	* until we find a free one. For sequential io, just return the end of
	79	* the last io issued.
	80	*/
	81	static int get_next_offset(struct thread_data td, struct fio_file f,
	82	unsigned long long *offset)
	83	{
	84	unsigned long long b, rb;
	85	long r;
	86
	87	if (!td->sequential) {
	88	unsigned long long max_blocks = td->io_size / td->min_bs;
	89	int loops = 50;
	90
	91	do {
	92	r = os_random_long(&td->random_state);
	93	b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
bb8895e0 JA	94	if (td->norandommap)
bb8895e0 JA	95	break;
10ba535a JA	96	rb = b + (f->file_offset / td->min_bs);
	97	loops--;
	98	} while (!random_map_free(td, f, rb) && loops);
	99
	100	if (!loops) {
	101	if (get_next_free_block(td, f, &b))
	102	return 1;
	103	}
	104	} else
	105	b = f->last_pos / td->min_bs;
	106
	107	offset = (b td->min_bs) + f->file_offset;
	108	if (*offset > f->file_size)
	109	return 1;
	110
	111	return 0;
	112	}
	113
	114	static unsigned int get_next_buflen(struct thread_data *td)
	115	{
	116	unsigned int buflen;
	117	long r;
	118
	119	if (td->min_bs == td->max_bs)
	120	buflen = td->min_bs;
	121	else {
	122	r = os_random_long(&td->bsrange_state);
	123	buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
690adba3 JA	124	if (!td->bs_unaligned)
690adba3 JA	125	buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
10ba535a JA	126	}
	127
	128	if (buflen > td->io_size - td->this_io_bytes[td->ddir]) {
	129	/*
	130	* if using direct/raw io, we may not be able to
	131	* shrink the size. so just fail it.
	132	*/
	133	if (td->io_ops->flags & FIO_RAWIO)
	134	return 0;
	135
	136	buflen = td->io_size - td->this_io_bytes[td->ddir];
	137	}
	138
	139	return buflen;
	140	}
	141
	142	/*
	143	* Return the data direction for the next io_u. If the job is a
	144	* mixed read/write workload, check the rwmix cycle and switch if
	145	* necessary.
	146	*/
	147	static int get_rw_ddir(struct thread_data *td)
	148	{
	149	if (td_rw(td)) {
	150	struct timeval now;
	151	unsigned long elapsed;
	152
	153	gettimeofday(&now, NULL);
	154	elapsed = mtime_since_now(&td->rwmix_switch);
	155
	156	/*
	157	* Check if it's time to seed a new data direction.
	158	*/
	159	if (elapsed >= td->rwmixcycle) {
	160	unsigned int v;
	161	long r;
	162
	163	r = os_random_long(&td->rwmix_state);
	164	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
	165	if (v < td->rwmixread)
	166	td->rwmix_ddir = DDIR_READ;
	167	else
	168	td->rwmix_ddir = DDIR_WRITE;
	169	memcpy(&td->rwmix_switch, &now, sizeof(now));
	170	}
	171	return td->rwmix_ddir;
	172	} else if (td_read(td))
	173	return DDIR_READ;
	174	else
	175	return DDIR_WRITE;
	176	}
	177
10ba535a JA	178	void put_io_u(struct thread_data td, struct io_u io_u)
	179	{
	180	io_u->file = NULL;
	181	list_del(&io_u->list);
	182	list_add(&io_u->list, &td->io_u_freelist);
	183	td->cur_depth--;
	184	}
	185
	186	static int fill_io_u(struct thread_data td, struct fio_file f,
	187	struct io_u *io_u)
	188	{
	189	/*
	190	* If using an iolog, grab next piece if any available.
	191	*/
	192	if (td->read_iolog)
	193	return read_iolog_get(td, io_u);
	194
87dc1ab1 JA	195	/*
	196	* see if it's time to sync
	197	*/
	198	if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
	199	&& should_fsync(td)) {
	200	io_u->ddir = DDIR_SYNC;
	201	io_u->file = f;
	202	return 0;
	203	}
	204
10ba535a JA	205	/*
	206	* No log, let the seq/rand engine retrieve the next position.
	207	*/
	208	if (!get_next_offset(td, f, &io_u->offset)) {
	209	io_u->buflen = get_next_buflen(td);
	210
	211	if (io_u->buflen) {
	212	io_u->ddir = get_rw_ddir(td);
	213
	214	/*
	215	* If using a write iolog, store this entry.
	216	*/
076efc7c	217	if (td->write_iolog_file)
10ba535a JA	218	write_iolog_put(td, io_u);
	219
	220	io_u->file = f;
	221	return 0;
	222	}
	223	}
	224
	225	return 1;
	226	}
	227
	228	struct io_u __get_io_u(struct thread_data td)
	229	{
	230	struct io_u *io_u = NULL;
	231
	232	if (!queue_full(td)) {
	233	io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
	234
6040dabc	235	io_u->buflen = 0;
10ba535a JA	236	io_u->error = 0;
	237	io_u->resid = 0;
	238	list_del(&io_u->list);
	239	list_add(&io_u->list, &td->io_u_busylist);
	240	td->cur_depth++;
	241	}
	242
	243	return io_u;
	244	}
	245
	246	/*
	247	* Return an io_u to be processed. Gets a buflen and offset, sets direction,
	248	* etc. The returned io_u is fully ready to be prepped and submitted.
	249	*/
	250	struct io_u get_io_u(struct thread_data td, struct fio_file *f)
	251	{
	252	struct io_u *io_u;
	253
	254	io_u = __get_io_u(td);
	255	if (!io_u)
	256	return NULL;
	257
	258	if (td->zone_bytes >= td->zone_size) {
	259	td->zone_bytes = 0;
	260	f->last_pos += td->zone_skip;
	261	}
	262
	263	if (fill_io_u(td, f, io_u)) {
	264	put_io_u(td, io_u);
	265	return NULL;
	266	}
	267
	268	if (io_u->buflen + io_u->offset > f->file_size) {
	269	if (td->io_ops->flags & FIO_RAWIO) {
	270	put_io_u(td, io_u);
	271	return NULL;
	272	}
	273
	274	io_u->buflen = f->file_size - io_u->offset;
	275	}
	276
87dc1ab1 JA	277	if (io_u->ddir != DDIR_SYNC) {
	278	if (!io_u->buflen) {
	279	put_io_u(td, io_u);
	280	return NULL;
	281	}
10ba535a	282
bb8895e0	283	if (!td->read_iolog && !td->sequential && !td->norandommap)
87dc1ab1	284	mark_random_map(td, f, io_u);
10ba535a	285
87dc1ab1	286	f->last_pos += io_u->buflen;
10ba535a	287
87dc1ab1 JA	288	if (td->verify != VERIFY_NONE)
	289	populate_verify_io_u(td, io_u);
	290	}
10ba535a JA	291
	292	if (td_io_prep(td, io_u)) {
	293	put_io_u(td, io_u);
	294	return NULL;
	295	}
	296
	297	gettimeofday(&io_u->start_time, NULL);
	298	return io_u;
	299	}
	300
	301	void io_completed(struct thread_data td, struct io_u io_u,
	302	struct io_completion_data *icd)
	303	{
	304	struct timeval e;
	305	unsigned long msec;
	306
87dc1ab1 JA	307	if (io_u->ddir == DDIR_SYNC) {
	308	td->last_was_sync = 1;
	309	return;
	310	}
	311
	312	td->last_was_sync = 0;
	313
10ba535a JA	314	gettimeofday(&e, NULL);
	315
	316	if (!io_u->error) {
	317	unsigned int bytes = io_u->buflen - io_u->resid;
	318	const int idx = io_u->ddir;
	319
	320	td->io_blocks[idx]++;
	321	td->io_bytes[idx] += bytes;
	322	td->zone_bytes += bytes;
	323	td->this_io_bytes[idx] += bytes;
	324
	325	msec = mtime_since(&io_u->issue_time, &e);
	326
	327	add_clat_sample(td, idx, msec);
	328	add_bw_sample(td, idx);
	329
	330	if ((td_rw(td) \|\| td_write(td)) && idx == DDIR_WRITE)
	331	log_io_piece(td, io_u);
	332
	333	icd->bytes_done[idx] += bytes;
	334	} else
	335	icd->error = io_u->error;
	336	}
	337
	338	void ios_completed(struct thread_data td, struct io_completion_data icd)
	339	{
	340	struct io_u *io_u;
	341	int i;
	342
	343	icd->error = 0;
	344	icd->bytes_done[0] = icd->bytes_done[1] = 0;
	345
	346	for (i = 0; i < icd->nr; i++) {
	347	io_u = td->io_ops->event(td, i);
	348
	349	io_completed(td, io_u, icd);
	350	put_io_u(td, io_u);
	351	}
	352	}
	353
	354