[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->min_bs[io_u->ddir];
	unsigned long long block;
	unsigned int blocks;

	block = io_u->offset / (unsigned long long) min_bs;
	blocks = 0;
	while (blocks < (io_u->buflen / 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 * 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;

	*b = 0;
	i = 0;
	while ((*b) * td->rw_min_bs < f->real_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, int 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 = 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[ddir]);
			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[ddir];

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

	return 0;
}

static unsigned int get_next_buflen(struct thread_data *td, int 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);
	}

	if (buflen > td->io_size - td->this_io_bytes[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[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 position.
	 */
	if (!get_next_offset(td, f, &io_u->offset, io_u->ddir)) {
		io_u->buflen = get_next_buflen(td, io_u->ddir);
		if (io_u->buflen) {
			/*
			 * 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;
}

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;
		}

		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;
	}

	/*
	 * 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
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	{
a00735e6 JA	30	unsigned int min_bs = td->min_bs[io_u->ddir];
	31	unsigned long long block;
	32	unsigned int blocks;
df415585	33
a00735e6 JA	34	block = io_u->offset / (unsigned long long) min_bs;
	35	blocks = 0;
	36	while (blocks < (io_u->buflen / min_bs)) {
df415585 JA	37	unsigned int idx, bit;
	38
	39	if (!random_map_free(td, f, block))
	40	break;
	41
	42	idx = RAND_MAP_IDX(td, f, block);
	43	bit = RAND_MAP_BIT(td, f, block);
	44
	45	assert(idx < f->num_maps);
	46
	47	f->file_map[idx] \|= (1UL << bit);
	48	block++;
	49	blocks++;
	50	}
	51
a00735e6 JA	52	if ((blocks * min_bs) < io_u->buflen)
a00735e6 JA	53	io_u->buflen = blocks * min_bs;
df415585 JA	54	}
df415585 JA	55
10ba535a JA	56	/*
	57	* Return the next free block in the map.
	58	*/
	59	static int get_next_free_block(struct thread_data td, struct fio_file f,
	60	unsigned long long *b)
	61	{
	62	int i;
	63
	64	*b = 0;
	65	i = 0;
ee88470c	66	while ((b) td->rw_min_bs < f->real_file_size) {
10ba535a JA	67	if (f->file_map[i] != -1UL) {
	68	*b += ffz(f->file_map[i]);
	69	return 0;
	70	}
	71
	72	*b += BLOCKS_PER_MAP;
	73	i++;
	74	}
	75
	76	return 1;
	77	}
	78
	79	/*
	80	* For random io, generate a random new block and see if it's used. Repeat
	81	* until we find a free one. For sequential io, just return the end of
	82	* the last io issued.
	83	*/
	84	static int get_next_offset(struct thread_data td, struct fio_file f,
a00735e6	85	unsigned long long *offset, int ddir)
10ba535a JA	86	{
	87	unsigned long long b, rb;
	88	long r;
	89
	90	if (!td->sequential) {
f697125a	91	unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
10ba535a JA	92	int loops = 50;
	93
	94	do {
	95	r = os_random_long(&td->random_state);
	96	b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
bb8895e0 JA	97	if (td->norandommap)
bb8895e0 JA	98	break;
a00735e6	99	rb = b + (f->file_offset / td->min_bs[ddir]);
10ba535a JA	100	loops--;
	101	} while (!random_map_free(td, f, rb) && loops);
	102
	103	if (!loops) {
	104	if (get_next_free_block(td, f, &b))
	105	return 1;
	106	}
	107	} else
a00735e6	108	b = f->last_pos / td->min_bs[ddir];
10ba535a	109
a00735e6	110	offset = (b td->min_bs[ddir]) + f->file_offset;
ee88470c	111	if (*offset > f->real_file_size)
10ba535a JA	112	return 1;
	113
	114	return 0;
	115	}
	116
a00735e6	117	static unsigned int get_next_buflen(struct thread_data *td, int ddir)
10ba535a JA	118	{
	119	unsigned int buflen;
	120	long r;
	121
a00735e6 JA	122	if (td->min_bs[ddir] == td->max_bs[ddir])
a00735e6 JA	123	buflen = td->min_bs[ddir];
10ba535a JA	124	else {
10ba535a JA	125	r = os_random_long(&td->bsrange_state);
1e97cce9	126	buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
690adba3	127	if (!td->bs_unaligned)
a00735e6	128	buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
10ba535a JA	129	}
10ba535a JA	130
a00735e6	131	if (buflen > td->io_size - td->this_io_bytes[ddir]) {
10ba535a JA	132	/*
	133	* if using direct/raw io, we may not be able to
	134	* shrink the size. so just fail it.
	135	*/
	136	if (td->io_ops->flags & FIO_RAWIO)
	137	return 0;
	138
a00735e6	139	buflen = td->io_size - td->this_io_bytes[ddir];
10ba535a JA	140	}
	141
	142	return buflen;
	143	}
	144
	145	/*
	146	* Return the data direction for the next io_u. If the job is a
	147	* mixed read/write workload, check the rwmix cycle and switch if
	148	* necessary.
	149	*/
1e97cce9	150	static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a JA	151	{
	152	if (td_rw(td)) {
	153	struct timeval now;
	154	unsigned long elapsed;
	155
02bcaa8c	156	fio_gettime(&now, NULL);
10ba535a JA	157	elapsed = mtime_since_now(&td->rwmix_switch);
	158
	159	/*
	160	* Check if it's time to seed a new data direction.
	161	*/
	162	if (elapsed >= td->rwmixcycle) {
	163	unsigned int v;
	164	long r;
	165
	166	r = os_random_long(&td->rwmix_state);
	167	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
	168	if (v < td->rwmixread)
	169	td->rwmix_ddir = DDIR_READ;
	170	else
	171	td->rwmix_ddir = DDIR_WRITE;
	172	memcpy(&td->rwmix_switch, &now, sizeof(now));
	173	}
	174	return td->rwmix_ddir;
	175	} else if (td_read(td))
	176	return DDIR_READ;
	177	else
	178	return DDIR_WRITE;
	179	}
	180
10ba535a JA	181	void put_io_u(struct thread_data td, struct io_u io_u)
	182	{
	183	io_u->file = NULL;
	184	list_del(&io_u->list);
	185	list_add(&io_u->list, &td->io_u_freelist);
	186	td->cur_depth--;
	187	}
	188
	189	static int fill_io_u(struct thread_data td, struct fio_file f,
	190	struct io_u *io_u)
	191	{
	192	/*
	193	* If using an iolog, grab next piece if any available.
	194	*/
	195	if (td->read_iolog)
	196	return read_iolog_get(td, io_u);
	197
87dc1ab1 JA	198	/*
	199	* see if it's time to sync
	200	*/
	201	if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
	202	&& should_fsync(td)) {
	203	io_u->ddir = DDIR_SYNC;
	204	io_u->file = f;
	205	return 0;
	206	}
	207
a00735e6 JA	208	io_u->ddir = get_rw_ddir(td);
a00735e6 JA	209
10ba535a JA	210	/*
	211	* No log, let the seq/rand engine retrieve the next position.
	212	*/
a00735e6 JA	213	if (!get_next_offset(td, f, &io_u->offset, io_u->ddir)) {
a00735e6 JA	214	io_u->buflen = get_next_buflen(td, io_u->ddir);
10ba535a	215	if (io_u->buflen) {
10ba535a JA	216	/*
	217	* If using a write iolog, store this entry.
	218	*/
076efc7c	219	if (td->write_iolog_file)
10ba535a JA	220	write_iolog_put(td, io_u);
	221
	222	io_u->file = f;
	223	return 0;
	224	}
	225	}
	226
	227	return 1;
	228	}
	229
71619dc2 JA	230	static void io_u_mark_depth(struct thread_data *td)
	231	{
	232	int index = 0;
	233
	234	switch (td->cur_depth) {
	235	default:
	236	index++;
	237	case 32 ... 63:
	238	index++;
	239	case 16 ... 31:
	240	index++;
	241	case 8 ... 15:
	242	index++;
	243	case 4 ... 7:
	244	index++;
	245	case 2 ... 3:
	246	index++;
	247	case 1:
	248	break;
	249	}
	250
	251	td->io_u_map[index]++;
	252	td->total_io_u++;
	253	}
	254
10ba535a JA	255	struct io_u __get_io_u(struct thread_data td)
	256	{
	257	struct io_u *io_u = NULL;
	258
	259	if (!queue_full(td)) {
	260	io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
	261
6040dabc	262	io_u->buflen = 0;
10ba535a JA	263	io_u->error = 0;
	264	io_u->resid = 0;
	265	list_del(&io_u->list);
	266	list_add(&io_u->list, &td->io_u_busylist);
	267	td->cur_depth++;
71619dc2	268	io_u_mark_depth(td);
10ba535a JA	269	}
	270
	271	return io_u;
	272	}
	273
	274	/*
	275	* Return an io_u to be processed. Gets a buflen and offset, sets direction,
	276	* etc. The returned io_u is fully ready to be prepped and submitted.
	277	*/
	278	struct io_u get_io_u(struct thread_data td, struct fio_file *f)
	279	{
	280	struct io_u *io_u;
	281
	282	io_u = __get_io_u(td);
	283	if (!io_u)
	284	return NULL;
	285
	286	if (td->zone_bytes >= td->zone_size) {
	287	td->zone_bytes = 0;
	288	f->last_pos += td->zone_skip;
	289	}
	290
	291	if (fill_io_u(td, f, io_u)) {
	292	put_io_u(td, io_u);
	293	return NULL;
	294	}
	295
ee88470c	296	if (io_u->buflen + io_u->offset > f->real_file_size) {
10ba535a JA	297	if (td->io_ops->flags & FIO_RAWIO) {
	298	put_io_u(td, io_u);
	299	return NULL;
	300	}
	301
ee88470c	302	io_u->buflen = f->real_file_size - io_u->offset;
10ba535a JA	303	}
10ba535a JA	304
87dc1ab1 JA	305	if (io_u->ddir != DDIR_SYNC) {
	306	if (!io_u->buflen) {
	307	put_io_u(td, io_u);
	308	return NULL;
	309	}
10ba535a	310
bb8895e0	311	if (!td->read_iolog && !td->sequential && !td->norandommap)
87dc1ab1	312	mark_random_map(td, f, io_u);
10ba535a	313
87dc1ab1	314	f->last_pos += io_u->buflen;
10ba535a	315
87dc1ab1 JA	316	if (td->verify != VERIFY_NONE)
	317	populate_verify_io_u(td, io_u);
	318	}
10ba535a JA	319
	320	if (td_io_prep(td, io_u)) {
	321	put_io_u(td, io_u);
	322	return NULL;
	323	}
	324
165faf16 JA	325	/*
	326	* Set io data pointers.
	327	*/
cec6b55d JA	328	io_u->xfer_buf = io_u->buf;
cec6b55d JA	329	io_u->xfer_buflen = io_u->buflen;
165faf16	330
02bcaa8c	331	fio_gettime(&io_u->start_time, NULL);
10ba535a JA	332	return io_u;
	333	}
	334
	335	void io_completed(struct thread_data td, struct io_u io_u,
	336	struct io_completion_data *icd)
	337	{
10ba535a JA	338	unsigned long msec;
10ba535a JA	339
87dc1ab1 JA	340	if (io_u->ddir == DDIR_SYNC) {
	341	td->last_was_sync = 1;
	342	return;
	343	}
	344
	345	td->last_was_sync = 0;
	346
10ba535a JA	347	if (!io_u->error) {
10ba535a JA	348	unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9	349	const enum fio_ddir idx = io_u->ddir;
10ba535a JA	350
	351	td->io_blocks[idx]++;
	352	td->io_bytes[idx] += bytes;
	353	td->zone_bytes += bytes;
	354	td->this_io_bytes[idx] += bytes;
	355
02bcaa8c JA	356	io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
	357
	358	msec = mtime_since(&io_u->issue_time, &icd->time);
10ba535a JA	359
10ba535a JA	360	add_clat_sample(td, idx, msec);
02bcaa8c	361	add_bw_sample(td, idx, &icd->time);
10ba535a JA	362
	363	if ((td_rw(td) \|\| td_write(td)) && idx == DDIR_WRITE)
	364	log_io_piece(td, io_u);
	365
	366	icd->bytes_done[idx] += bytes;
	367	} else
	368	icd->error = io_u->error;
	369	}
	370
	371	void ios_completed(struct thread_data td, struct io_completion_data icd)
	372	{
	373	struct io_u *io_u;
	374	int i;
	375
02bcaa8c JA	376	fio_gettime(&icd->time, NULL);
02bcaa8c JA	377
10ba535a JA	378	icd->error = 0;
	379	icd->bytes_done[0] = icd->bytes_done[1] = 0;
	380
	381	for (i = 0; i < icd->nr; i++) {
	382	io_u = td->io_ops->event(td, i);
	383
	384	io_completed(td, io_u, icd);
	385	put_io_u(td, io_u);
	386	}
	387	}