[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 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);

		assert(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++;
}

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->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;
	}

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

	fio_gettime(&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)
{
	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;

		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);

		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;

	fio_gettime(&icd->time, NULL);

	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
	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
	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 int min_bs = td->rw_min_bs;
	31	unsigned long long block;
	32	unsigned int blocks;
	33	unsigned int nr_blocks;
	34
	35	block = io_u->offset / (unsigned long long) min_bs;
	36	blocks = 0;
	37	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
	38
	39	while (blocks < nr_blocks) {
	40	unsigned int idx, bit;
	41
	42	if (!random_map_free(td, f, block))
	43	break;
	44
	45	idx = RAND_MAP_IDX(td, f, block);
	46	bit = RAND_MAP_BIT(td, f, block);
	47
	48	assert(idx < f->num_maps);
	49
	50	f->file_map[idx] \|= (1UL << bit);
	51	block++;
	52	blocks++;
	53	}
	54
	55	if ((blocks * min_bs) < io_u->buflen)
	56	io_u->buflen = blocks * min_bs;
	57	}
	58
	59	/*
	60	* Return the next free block in the map.
	61	*/
	62	static int get_next_free_block(struct thread_data td, struct fio_file f,
	63	unsigned long long *b)
	64	{
	65	int i;
	66
	67	i = f->last_free_lookup;
	68	b = (i BLOCKS_PER_MAP);
	69	while ((b) td->rw_min_bs < f->real_file_size) {
	70	if (f->file_map[i] != -1UL) {
	71	*b += ffz(f->file_map[i]);
	72	f->last_free_lookup = i;
	73	return 0;
	74	}
	75
	76	*b += BLOCKS_PER_MAP;
	77	i++;
	78	}
	79
	80	return 1;
	81	}
	82
	83	/*
	84	* For random io, generate a random new block and see if it's used. Repeat
	85	* until we find a free one. For sequential io, just return the end of
	86	* the last io issued.
	87	*/
	88	static int get_next_offset(struct thread_data td, struct fio_file f,
	89	struct io_u *io_u)
	90	{
	91	const int ddir = io_u->ddir;
	92	unsigned long long b, rb;
	93	long r;
	94
	95	if (!td->sequential) {
	96	unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
	97	int loops = 5;
	98
	99	do {
	100	r = os_random_long(&td->random_state);
	101	b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
	102	if (td->norandommap)
	103	break;
	104	rb = b + (f->file_offset / td->min_bs[ddir]);
	105	loops--;
	106	} while (!random_map_free(td, f, rb) && loops);
	107
	108	/*
	109	* if we failed to retrieve a truly random offset within
	110	* the loops assigned, see if there are free ones left at all
	111	*/
	112	if (!loops && get_next_free_block(td, f, &b))
	113	return 1;
	114	} else
	115	b = f->last_pos / td->min_bs[ddir];
	116
	117	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
	118	if (io_u->offset >= f->real_file_size)
	119	return 1;
	120
	121	return 0;
	122	}
	123
	124	static unsigned int get_next_buflen(struct thread_data td, struct fio_file f,
	125	struct io_u *io_u)
	126	{
	127	const int ddir = io_u->ddir;
	128	unsigned int buflen;
	129	long r;
	130
	131	if (td->min_bs[ddir] == td->max_bs[ddir])
	132	buflen = td->min_bs[ddir];
	133	else {
	134	r = os_random_long(&td->bsrange_state);
	135	buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
	136	if (!td->bs_unaligned)
	137	buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
	138	}
	139
	140	while (buflen + io_u->offset > f->real_file_size) {
	141	if (buflen == td->min_bs[ddir])
	142	return 0;
	143
	144	buflen = td->min_bs[ddir];
	145	}
	146
	147	return buflen;
	148	}
	149
	150	/*
	151	* Return the data direction for the next io_u. If the job is a
	152	* mixed read/write workload, check the rwmix cycle and switch if
	153	* necessary.
	154	*/
	155	static enum fio_ddir get_rw_ddir(struct thread_data *td)
	156	{
	157	if (td_rw(td)) {
	158	struct timeval now;
	159	unsigned long elapsed;
	160
	161	fio_gettime(&now, NULL);
	162	elapsed = mtime_since_now(&td->rwmix_switch);
	163
	164	/*
	165	* Check if it's time to seed a new data direction.
	166	*/
	167	if (elapsed >= td->rwmixcycle) {
	168	unsigned int v;
	169	long r;
	170
	171	r = os_random_long(&td->rwmix_state);
	172	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
	173	if (v < td->rwmixread)
	174	td->rwmix_ddir = DDIR_READ;
	175	else
	176	td->rwmix_ddir = DDIR_WRITE;
	177	memcpy(&td->rwmix_switch, &now, sizeof(now));
	178	}
	179	return td->rwmix_ddir;
	180	} else if (td_read(td))
	181	return DDIR_READ;
	182	else
	183	return DDIR_WRITE;
	184	}
	185
	186	void put_io_u(struct thread_data td, struct io_u io_u)
	187	{
	188	io_u->file = NULL;
	189	list_del(&io_u->list);
	190	list_add(&io_u->list, &td->io_u_freelist);
	191	td->cur_depth--;
	192	}
	193
	194	static int fill_io_u(struct thread_data td, struct fio_file f,
	195	struct io_u *io_u)
	196	{
	197	/*
	198	* If using an iolog, grab next piece if any available.
	199	*/
	200	if (td->read_iolog)
	201	return read_iolog_get(td, io_u);
	202
	203	/*
	204	* see if it's time to sync
	205	*/
	206	if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
	207	&& should_fsync(td)) {
	208	io_u->ddir = DDIR_SYNC;
	209	io_u->file = f;
	210	return 0;
	211	}
	212
	213	io_u->ddir = get_rw_ddir(td);
	214
	215	/*
	216	* No log, let the seq/rand engine retrieve the next buflen and
	217	* position.
	218	*/
	219	if (get_next_offset(td, f, io_u))
	220	return 1;
	221
	222	io_u->buflen = get_next_buflen(td, f, io_u);
	223	if (!io_u->buflen)
	224	return 1;
	225
	226	/*
	227	* mark entry before potentially trimming io_u
	228	*/
	229	if (!td->read_iolog && !td->sequential && !td->norandommap)
	230	mark_random_map(td, f, io_u);
	231
	232	/*
	233	* If using a write iolog, store this entry.
	234	*/
	235	if (td->write_iolog_file)
	236	write_iolog_put(td, io_u);
	237
	238	io_u->file = f;
	239	return 0;
	240	}
	241
	242	static void io_u_mark_depth(struct thread_data *td)
	243	{
	244	int index = 0;
	245
	246	switch (td->cur_depth) {
	247	default:
	248	index++;
	249	case 32 ... 63:
	250	index++;
	251	case 16 ... 31:
	252	index++;
	253	case 8 ... 15:
	254	index++;
	255	case 4 ... 7:
	256	index++;
	257	case 2 ... 3:
	258	index++;
	259	case 1:
	260	break;
	261	}
	262
	263	td->io_u_map[index]++;
	264	td->total_io_u++;
	265	}
	266
	267	struct io_u __get_io_u(struct thread_data td)
	268	{
	269	struct io_u *io_u = NULL;
	270
	271	if (!queue_full(td)) {
	272	io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
	273
	274	io_u->buflen = 0;
	275	io_u->error = 0;
	276	io_u->resid = 0;
	277	list_del(&io_u->list);
	278	list_add(&io_u->list, &td->io_u_busylist);
	279	td->cur_depth++;
	280	io_u_mark_depth(td);
	281	}
	282
	283	return io_u;
	284	}
	285
	286	/*
	287	* Return an io_u to be processed. Gets a buflen and offset, sets direction,
	288	* etc. The returned io_u is fully ready to be prepped and submitted.
	289	*/
	290	struct io_u get_io_u(struct thread_data td, struct fio_file *f)
	291	{
	292	struct io_u *io_u;
	293
	294	io_u = __get_io_u(td);
	295	if (!io_u)
	296	return NULL;
	297
	298	if (td->zone_bytes >= td->zone_size) {
	299	td->zone_bytes = 0;
	300	f->last_pos += td->zone_skip;
	301	}
	302
	303	if (fill_io_u(td, f, io_u)) {
	304	put_io_u(td, io_u);
	305	return NULL;
	306	}
	307
	308	if (io_u->buflen + io_u->offset > f->real_file_size) {
	309	if (td->io_ops->flags & FIO_RAWIO) {
	310	put_io_u(td, io_u);
	311	return NULL;
	312	}
	313
	314	io_u->buflen = f->real_file_size - io_u->offset;
	315	}
	316
	317	if (io_u->ddir != DDIR_SYNC) {
	318	if (!io_u->buflen) {
	319	put_io_u(td, io_u);
	320	return NULL;
	321	}
	322
	323	f->last_pos += io_u->buflen;
	324
	325	if (td->verify != VERIFY_NONE)
	326	populate_verify_io_u(td, io_u);
	327	}
	328
	329	if (td_io_prep(td, io_u)) {
	330	put_io_u(td, io_u);
	331	return NULL;
	332	}
	333
	334	/*
	335	* Set io data pointers.
	336	*/
	337	io_u->xfer_buf = io_u->buf;
	338	io_u->xfer_buflen = io_u->buflen;
	339
	340	fio_gettime(&io_u->start_time, NULL);
	341	return io_u;
	342	}
	343
	344	void io_completed(struct thread_data td, struct io_u io_u,
	345	struct io_completion_data *icd)
	346	{
	347	unsigned long msec;
	348
	349	if (io_u->ddir == DDIR_SYNC) {
	350	td->last_was_sync = 1;
	351	return;
	352	}
	353
	354	td->last_was_sync = 0;
	355
	356	if (!io_u->error) {
	357	unsigned int bytes = io_u->buflen - io_u->resid;
	358	const enum fio_ddir idx = io_u->ddir;
	359
	360	td->io_blocks[idx]++;
	361	td->io_bytes[idx] += bytes;
	362	td->zone_bytes += bytes;
	363	td->this_io_bytes[idx] += bytes;
	364
	365	io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
	366
	367	msec = mtime_since(&io_u->issue_time, &icd->time);
	368
	369	add_clat_sample(td, idx, msec);
	370	add_bw_sample(td, idx, &icd->time);
	371
	372	if ((td_rw(td) \|\| td_write(td)) && idx == DDIR_WRITE)
	373	log_io_piece(td, io_u);
	374
	375	icd->bytes_done[idx] += bytes;
	376	} else
	377	icd->error = io_u->error;
	378	}
	379
	380	void ios_completed(struct thread_data td, struct io_completion_data icd)
	381	{
	382	struct io_u *io_u;
	383	int i;
	384
	385	fio_gettime(&icd->time, NULL);
	386
	387	icd->error = 0;
	388	icd->bytes_done[0] = icd->bytes_done[1] = 0;
	389
	390	for (i = 0; i < icd->nr; i++) {
	391	io_u = td->io_ops->event(td, i);
	392
	393	io_completed(td, io_u, icd);
	394	put_io_u(td, io_u);
	395	}
	396	}