[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 */
	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_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;

	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;
	}

	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 (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);
	}
}

/*
 * Complete a single io_u for the sync engines.
 */
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;
}

/*
 * Called to complete min_events number of io for the async engines.
 */
long io_u_queued_complete(struct thread_data *td, int min_events,
			  endio_handler *handler)

{
	struct io_completion_data icd;
	struct timespec *tvp = NULL;
	int ret;

	if (min_events > 0) {
		ret = td_io_commit(td);
		if (ret < 0) {
			td_verror(td, -ret, "td_io_commit");
			return ret;
		}
	} else {
		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, handler, 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 {
	17	int nr; /* input */
	18	endio_handler handler; / input */
	19
	20	int error; /* output */
	21	unsigned long bytes_done[2]; /* output */
	22	struct timeval time; /* output */
	23	};
	24
10ba535a JA	25	/*
	26	* The ->file_map[] contains a map of blocks we have or have not done io
	27	* to yet. Used to make sure we cover the entire range in a fair fashion.
	28	*/
	29	static int random_map_free(struct thread_data td, struct fio_file f,
	30	unsigned long long block)
	31	{
	32	unsigned int idx = RAND_MAP_IDX(td, f, block);
	33	unsigned int bit = RAND_MAP_BIT(td, f, block);
	34
	35	return (f->file_map[idx] & (1UL << bit)) == 0;
	36	}
	37
df415585 JA	38	/*
	39	* Mark a given offset as used in the map.
	40	*/
	41	static void mark_random_map(struct thread_data td, struct fio_file f,
	42	struct io_u *io_u)
	43	{
afdbe580	44	unsigned int min_bs = td->rw_min_bs;
a00735e6 JA	45	unsigned long long block;
a00735e6 JA	46	unsigned int blocks;
c685b5b2	47	unsigned int nr_blocks;
df415585	48
a00735e6 JA	49	block = io_u->offset / (unsigned long long) min_bs;
a00735e6 JA	50	blocks = 0;
c685b5b2 JA	51	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
	52
	53	while (blocks < nr_blocks) {
df415585 JA	54	unsigned int idx, bit;
	55
	56	if (!random_map_free(td, f, block))
	57	break;
	58
	59	idx = RAND_MAP_IDX(td, f, block);
	60	bit = RAND_MAP_BIT(td, f, block);
	61
0032bf9f	62	fio_assert(td, idx < f->num_maps);
df415585 JA	63
	64	f->file_map[idx] \|= (1UL << bit);
	65	block++;
	66	blocks++;
	67	}
	68
a00735e6 JA	69	if ((blocks * min_bs) < io_u->buflen)
a00735e6 JA	70	io_u->buflen = blocks * min_bs;
df415585 JA	71	}
df415585 JA	72
10ba535a JA	73	/*
	74	* Return the next free block in the map.
	75	*/
	76	static int get_next_free_block(struct thread_data td, struct fio_file f,
	77	unsigned long long *b)
	78	{
	79	int i;
	80
c685b5b2 JA	81	i = f->last_free_lookup;
c685b5b2 JA	82	b = (i BLOCKS_PER_MAP);
ee88470c	83	while ((b) td->rw_min_bs < f->real_file_size) {
10ba535a JA	84	if (f->file_map[i] != -1UL) {
10ba535a JA	85	*b += ffz(f->file_map[i]);
c685b5b2	86	f->last_free_lookup = i;
10ba535a JA	87	return 0;
	88	}
	89
	90	*b += BLOCKS_PER_MAP;
	91	i++;
	92	}
	93
	94	return 1;
	95	}
	96
	97	/*
	98	* For random io, generate a random new block and see if it's used. Repeat
	99	* until we find a free one. For sequential io, just return the end of
	100	* the last io issued.
	101	*/
	102	static int get_next_offset(struct thread_data td, struct fio_file f,
c685b5b2	103	struct io_u *io_u)
10ba535a	104	{
c685b5b2	105	const int ddir = io_u->ddir;
10ba535a JA	106	unsigned long long b, rb;
	107	long r;
	108
413dd459	109	if (td_random(td)) {
f697125a	110	unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
c685b5b2	111	int loops = 5;
10ba535a JA	112
	113	do {
	114	r = os_random_long(&td->random_state);
	115	b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
bb8895e0 JA	116	if (td->norandommap)
bb8895e0 JA	117	break;
a00735e6	118	rb = b + (f->file_offset / td->min_bs[ddir]);
10ba535a JA	119	loops--;
	120	} while (!random_map_free(td, f, rb) && loops);
	121
bca4ed4d JA	122	/*
	123	* if we failed to retrieve a truly random offset within
	124	* the loops assigned, see if there are free ones left at all
	125	*/
	126	if (!loops && get_next_free_block(td, f, &b))
	127	return 1;
10ba535a	128	} else
a00735e6	129	b = f->last_pos / td->min_bs[ddir];
10ba535a	130
c685b5b2	131	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
bca4ed4d	132	if (io_u->offset >= f->real_file_size)
10ba535a JA	133	return 1;
	134
	135	return 0;
	136	}
	137
bca4ed4d JA	138	static unsigned int get_next_buflen(struct thread_data td, struct fio_file f,
bca4ed4d JA	139	struct io_u *io_u)
10ba535a	140	{
bca4ed4d	141	const int ddir = io_u->ddir;
10ba535a JA	142	unsigned int buflen;
	143	long r;
	144
a00735e6 JA	145	if (td->min_bs[ddir] == td->max_bs[ddir])
a00735e6 JA	146	buflen = td->min_bs[ddir];
10ba535a JA	147	else {
10ba535a JA	148	r = os_random_long(&td->bsrange_state);
1e97cce9	149	buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
690adba3	150	if (!td->bs_unaligned)
a00735e6	151	buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
10ba535a JA	152	}
10ba535a JA	153
bca4ed4d	154	while (buflen + io_u->offset > f->real_file_size) {
bca4ed4d JA	155	if (buflen == td->min_bs[ddir])
	156	return 0;
	157
	158	buflen = td->min_bs[ddir];
10ba535a JA	159	}
	160
	161	return buflen;
	162	}
	163
	164	/*
	165	* Return the data direction for the next io_u. If the job is a
	166	* mixed read/write workload, check the rwmix cycle and switch if
	167	* necessary.
	168	*/
1e97cce9	169	static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a JA	170	{
	171	if (td_rw(td)) {
	172	struct timeval now;
	173	unsigned long elapsed;
	174
02bcaa8c	175	fio_gettime(&now, NULL);
10ba535a JA	176	elapsed = mtime_since_now(&td->rwmix_switch);
	177
	178	/*
	179	* Check if it's time to seed a new data direction.
	180	*/
	181	if (elapsed >= td->rwmixcycle) {
	182	unsigned int v;
	183	long r;
	184
	185	r = os_random_long(&td->rwmix_state);
	186	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
	187	if (v < td->rwmixread)
	188	td->rwmix_ddir = DDIR_READ;
	189	else
	190	td->rwmix_ddir = DDIR_WRITE;
	191	memcpy(&td->rwmix_switch, &now, sizeof(now));
	192	}
	193	return td->rwmix_ddir;
	194	} else if (td_read(td))
	195	return DDIR_READ;
	196	else
	197	return DDIR_WRITE;
	198	}
	199
10ba535a JA	200	void put_io_u(struct thread_data td, struct io_u io_u)
10ba535a JA	201	{
0c6e7517 JA	202	assert((io_u->flags & IO_U_F_FREE) == 0);
	203	io_u->flags \|= IO_U_F_FREE;
	204
10ba535a JA	205	io_u->file = NULL;
	206	list_del(&io_u->list);
	207	list_add(&io_u->list, &td->io_u_freelist);
	208	td->cur_depth--;
	209	}
	210
755200a3 JA	211	void requeue_io_u(struct thread_data td, struct io_u *io_u)
	212	{
	213	struct io_u __io_u = io_u;
	214
	215	list_del(&__io_u->list);
	216	list_add_tail(&__io_u->list, &td->io_u_requeues);
	217	td->cur_depth--;
	218	*io_u = NULL;
	219	}
	220
10ba535a JA	221	static int fill_io_u(struct thread_data td, struct fio_file f,
	222	struct io_u *io_u)
	223	{
	224	/*
	225	* If using an iolog, grab next piece if any available.
	226	*/
	227	if (td->read_iolog)
	228	return read_iolog_get(td, io_u);
	229
87dc1ab1 JA	230	/*
	231	* see if it's time to sync
	232	*/
755200a3 JA	233	if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
755200a3 JA	234	&& td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 JA	235	io_u->ddir = DDIR_SYNC;
	236	io_u->file = f;
	237	return 0;
	238	}
	239
a00735e6 JA	240	io_u->ddir = get_rw_ddir(td);
a00735e6 JA	241
10ba535a	242	/*
c685b5b2 JA	243	* No log, let the seq/rand engine retrieve the next buflen and
c685b5b2 JA	244	* position.
10ba535a	245	*/
bca4ed4d JA	246	if (get_next_offset(td, f, io_u))
bca4ed4d JA	247	return 1;
10ba535a	248
bca4ed4d JA	249	io_u->buflen = get_next_buflen(td, f, io_u);
	250	if (!io_u->buflen)
	251	return 1;
	252
	253	/*
	254	* mark entry before potentially trimming io_u
	255	*/
413dd459	256	if (!td->read_iolog && td_random(td) && !td->norandommap)
bca4ed4d JA	257	mark_random_map(td, f, io_u);
	258
	259	/*
	260	* If using a write iolog, store this entry.
	261	*/
	262	if (td->write_iolog_file)
	263	write_iolog_put(td, io_u);
	264
	265	io_u->file = f;
	266	return 0;
10ba535a JA	267	}
10ba535a JA	268
71619dc2 JA	269	static void io_u_mark_depth(struct thread_data *td)
	270	{
	271	int index = 0;
	272
	273	switch (td->cur_depth) {
	274	default:
	275	index++;
	276	case 32 ... 63:
	277	index++;
	278	case 16 ... 31:
	279	index++;
	280	case 8 ... 15:
	281	index++;
	282	case 4 ... 7:
	283	index++;
	284	case 2 ... 3:
	285	index++;
	286	case 1:
	287	break;
	288	}
	289
	290	td->io_u_map[index]++;
	291	td->total_io_u++;
	292	}
	293
ec118304 JA	294	static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
	295	{
	296	int index = 0;
	297
	298	switch (msec) {
	299	default:
	300	index++;
8abdce66	301	case 1000 ... 1999:
ec118304	302	index++;
8abdce66	303	case 750 ... 999:
ec118304	304	index++;
8abdce66	305	case 500 ... 749:
ec118304	306	index++;
8abdce66	307	case 250 ... 499:
ec118304	308	index++;
8abdce66	309	case 100 ... 249:
ec118304	310	index++;
8abdce66	311	case 50 ... 99:
ec118304	312	index++;
8abdce66	313	case 20 ... 49:
ec118304	314	index++;
8abdce66	315	case 10 ... 19:
ec118304	316	index++;
8abdce66	317	case 4 ... 9:
ec118304 JA	318	index++;
	319	case 2 ... 3:
	320	index++;
	321	case 0 ... 1:
	322	break;
	323	}
	324
	325	td->io_u_lat[index]++;
	326	}
	327
0aabe160 JA	328	/*
	329	* Get next file to service by choosing one at random
	330	*/
	331	static struct fio_file get_next_file_rand(struct thread_data td)
	332	{
	333	long r = os_random_long(&td->next_file_state);
	334	unsigned int fileno;
	335	struct fio_file *f;
	336
	337	do {
	338	fileno = (unsigned int) ((double) (td->nr_files - 1) * r / (RAND_MAX + 1.0));
	339	f = &td->files[fileno];
	340	if (f->fd != -1)
	341	return f;
	342	} while (1);
	343	}
	344
	345	/*
	346	* Get next file to service by doing round robin between all available ones
	347	*/
	348	static struct fio_file get_next_file_rr(struct thread_data td)
3d7c391d JA	349	{
	350	unsigned int old_next_file = td->next_file;
	351	struct fio_file *f;
	352
	353	do {
	354	f = &td->files[td->next_file];
	355
	356	td->next_file++;
	357	if (td->next_file >= td->nr_files)
	358	td->next_file = 0;
	359
	360	if (f->fd != -1)
	361	break;
	362
	363	f = NULL;
	364	} while (td->next_file != old_next_file);
	365
	366	return f;
	367	}
	368
10ba535a JA	369	struct io_u __get_io_u(struct thread_data td)
	370	{
	371	struct io_u *io_u = NULL;
	372
755200a3 JA	373	if (!list_empty(&td->io_u_requeues))
	374	io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
	375	else if (!queue_full(td)) {
10ba535a JA	376	io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a JA	377
6040dabc	378	io_u->buflen = 0;
10ba535a	379	io_u->resid = 0;
755200a3 JA	380	io_u->file = NULL;
	381	}
	382
	383	if (io_u) {
0c6e7517 JA	384	assert(io_u->flags & IO_U_F_FREE);
	385	io_u->flags &= ~IO_U_F_FREE;
	386
755200a3	387	io_u->error = 0;
10ba535a JA	388	list_del(&io_u->list);
	389	list_add(&io_u->list, &td->io_u_busylist);
	390	td->cur_depth++;
71619dc2	391	io_u_mark_depth(td);
10ba535a JA	392	}
	393
	394	return io_u;
	395	}
	396
	397	/*
	398	* Return an io_u to be processed. Gets a buflen and offset, sets direction,
	399	* etc. The returned io_u is fully ready to be prepped and submitted.
	400	*/
3d7c391d	401	struct io_u get_io_u(struct thread_data td)
10ba535a	402	{
3d7c391d	403	struct fio_file *f;
10ba535a JA	404	struct io_u *io_u;
	405
	406	io_u = __get_io_u(td);
	407	if (!io_u)
	408	return NULL;
	409
755200a3 JA	410	/*
	411	* from a requeue, io_u already setup
	412	*/
	413	if (io_u->file)
77f392bf	414	goto out;
755200a3	415
0aabe160 JA	416	if (td->file_service_type == FIO_FSERVICE_RR)
	417	f = get_next_file_rr(td);
	418	else
	419	f = get_next_file_rand(td);
	420
3d7c391d JA	421	if (!f) {
	422	put_io_u(td, io_u);
	423	return NULL;
	424	}
	425
	426	io_u->file = f;
	427
10ba535a JA	428	if (td->zone_bytes >= td->zone_size) {
	429	td->zone_bytes = 0;
	430	f->last_pos += td->zone_skip;
	431	}
	432
	433	if (fill_io_u(td, f, io_u)) {
	434	put_io_u(td, io_u);
	435	return NULL;
	436	}
	437
ee88470c	438	if (io_u->buflen + io_u->offset > f->real_file_size) {
10ba535a JA	439	if (td->io_ops->flags & FIO_RAWIO) {
	440	put_io_u(td, io_u);
	441	return NULL;
	442	}
	443
ee88470c	444	io_u->buflen = f->real_file_size - io_u->offset;
10ba535a JA	445	}
10ba535a JA	446
87dc1ab1 JA	447	if (io_u->ddir != DDIR_SYNC) {
	448	if (!io_u->buflen) {
	449	put_io_u(td, io_u);
	450	return NULL;
	451	}
10ba535a	452
36167d82	453	f->last_pos = io_u->offset + io_u->buflen;
10ba535a	454
87dc1ab1 JA	455	if (td->verify != VERIFY_NONE)
	456	populate_verify_io_u(td, io_u);
	457	}
10ba535a	458
165faf16 JA	459	/*
	460	* Set io data pointers.
	461	*/
77f392bf	462	out:
cec6b55d JA	463	io_u->xfer_buf = io_u->buf;
cec6b55d JA	464	io_u->xfer_buflen = io_u->buflen;
165faf16	465
36167d82 JA	466	if (td_io_prep(td, io_u)) {
	467	put_io_u(td, io_u);
	468	return NULL;
	469	}
	470
02bcaa8c	471	fio_gettime(&io_u->start_time, NULL);
10ba535a JA	472	return io_u;
	473	}
	474
97601024 JA	475	static void io_completed(struct thread_data td, struct io_u io_u,
97601024 JA	476	struct io_completion_data *icd)
10ba535a	477	{
10ba535a JA	478	unsigned long msec;
10ba535a JA	479
0c6e7517 JA	480	assert(io_u->flags & IO_U_F_FLIGHT);
	481	io_u->flags &= ~IO_U_F_FLIGHT;
	482
87dc1ab1 JA	483	if (io_u->ddir == DDIR_SYNC) {
	484	td->last_was_sync = 1;
	485	return;
	486	}
	487
	488	td->last_was_sync = 0;
	489
10ba535a JA	490	if (!io_u->error) {
10ba535a JA	491	unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9	492	const enum fio_ddir idx = io_u->ddir;
3af6ef39	493	int ret;
10ba535a JA	494
	495	td->io_blocks[idx]++;
	496	td->io_bytes[idx] += bytes;
	497	td->zone_bytes += bytes;
	498	td->this_io_bytes[idx] += bytes;
	499
02bcaa8c JA	500	io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
	501
	502	msec = mtime_since(&io_u->issue_time, &icd->time);
10ba535a JA	503
10ba535a JA	504	add_clat_sample(td, idx, msec);
02bcaa8c	505	add_bw_sample(td, idx, &icd->time);
ec118304	506	io_u_mark_latency(td, msec);
10ba535a JA	507
	508	if ((td_rw(td) \|\| td_write(td)) && idx == DDIR_WRITE)
	509	log_io_piece(td, io_u);
	510
	511	icd->bytes_done[idx] += bytes;
3af6ef39 JA	512
	513	if (icd->handler) {
	514	ret = icd->handler(io_u);
	515	if (ret && !icd->error)
	516	icd->error = ret;
	517	}
10ba535a JA	518	} else
	519	icd->error = io_u->error;
	520	}
	521
97601024 JA	522	static void init_icd(struct io_completion_data icd, endio_handler handler,
97601024 JA	523	int nr)
10ba535a	524	{
02bcaa8c JA	525	fio_gettime(&icd->time, NULL);
02bcaa8c JA	526
3af6ef39 JA	527	icd->handler = handler;
	528	icd->nr = nr;
	529
10ba535a JA	530	icd->error = 0;
10ba535a JA	531	icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82 JA	532	}
36167d82 JA	533
97601024 JA	534	static void ios_completed(struct thread_data *td,
97601024 JA	535	struct io_completion_data *icd)
36167d82 JA	536	{
	537	struct io_u *io_u;
	538	int i;
	539
10ba535a JA	540	for (i = 0; i < icd->nr; i++) {
	541	io_u = td->io_ops->event(td, i);
	542
	543	io_completed(td, io_u, icd);
	544	put_io_u(td, io_u);
	545	}
	546	}
97601024	547
e7e6cfb4 JA	548	/*
	549	* Complete a single io_u for the sync engines.
	550	*/
97601024 JA	551	long io_u_sync_complete(struct thread_data td, struct io_u io_u,
	552	endio_handler *handler)
	553	{
	554	struct io_completion_data icd;
	555
	556	init_icd(&icd, handler, 1);
	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	*/
97601024 JA	569	long io_u_queued_complete(struct thread_data *td, int min_events,
	570	endio_handler *handler)
	571
	572	{
97601024	573	struct io_completion_data icd;
00de55ef	574	struct timespec *tvp = NULL;
97601024 JA	575	int ret;
97601024 JA	576
755200a3	577	if (min_events > 0) {
755200a3 JA	578	ret = td_io_commit(td);
755200a3 JA	579	if (ret < 0) {
e1161c32	580	td_verror(td, -ret, "td_io_commit");
755200a3 JA	581	return ret;
755200a3 JA	582	}
00de55ef JA	583	} else {
	584	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
	585
	586	tvp = &ts;
755200a3	587	}
97601024	588
00de55ef	589	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
97601024	590	if (ret < 0) {
e1161c32	591	td_verror(td, -ret, "td_io_getevents");
97601024 JA	592	return ret;
	593	} else if (!ret)
	594	return ret;
	595
	596	init_icd(&icd, handler, ret);
	597	ios_completed(td, &icd);
	598	if (!icd.error)
	599	return icd.bytes_done[0] + icd.bytes_done[1];
	600
97601024 JA	601	return -1;
97601024 JA	602	}
7e77dd02 JA	603
	604	/*
	605	* Call when io_u is really queued, to update the submission latency.
	606	*/
	607	void io_u_queued(struct thread_data td, struct io_u io_u)
	608	{
	609	unsigned long slat_time;
	610
	611	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
	612	add_slat_sample(td, io_u->ddir, slat_time);
	613	}
433afcb4	614
55bc9728	615	#ifdef FIO_USE_TIMEOUT
433afcb4 JA	616	void io_u_set_timeout(struct thread_data *td)
	617	{
	618	assert(td->cur_depth);
	619
	620	td->timer.it_interval.tv_sec = 0;
	621	td->timer.it_interval.tv_usec = 0;
	622	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
	623	td->timer.it_value.tv_usec = 0;
	624	setitimer(ITIMER_REAL, &td->timer, NULL);
	625	fio_gettime(&td->timeout_end, NULL);
	626	}
5945b9b4 JA	627
	628	static void io_u_dump(struct io_u *io_u)
	629	{
	630	unsigned long t_start = mtime_since_now(&io_u->start_time);
	631	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
	632
	633	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
	634	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);
	635	log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
	636	}
55bc9728 JA	637	#else
	638	void io_u_set_timeout(struct thread_data fio_unused *td)
	639	{
	640	}
	641	#endif
433afcb4	642
55bc9728	643	#ifdef FIO_USE_TIMEOUT
433afcb4 JA	644	static void io_u_timeout_handler(int fio_unused sig)
	645	{
	646	struct thread_data td, __td;
	647	pid_t pid = getpid();
5945b9b4 JA	648	struct list_head *entry;
5945b9b4 JA	649	struct io_u *io_u;
433afcb4 JA	650	int i;
	651
	652	log_err("fio: io_u timeout\n");
	653
	654	/*
	655	* TLS would be nice...
	656	*/
	657	td = NULL;
	658	for_each_td(__td, i) {
	659	if (__td->pid == pid) {
	660	td = __td;
	661	break;
	662	}
	663	}
	664
	665	if (!td) {
	666	log_err("fio: io_u timeout, can't find job\n");
	667	exit(1);
	668	}
	669
	670	if (!td->cur_depth) {
	671	log_err("fio: timeout without pending work?\n");
	672	return;
	673	}
	674
	675	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
5945b9b4 JA	676
	677	list_for_each(entry, &td->io_u_busylist) {
	678	io_u = list_entry(entry, struct io_u, list);
	679
	680	io_u_dump(io_u);
	681	}
	682
	683	td_verror(td, ETIMEDOUT, "io_u timeout");
433afcb4 JA	684	exit(1);
433afcb4 JA	685	}
55bc9728	686	#endif
433afcb4 JA	687
	688	void io_u_init_timeout(void)
	689	{
55bc9728	690	#ifdef FIO_USE_TIMEOUT
433afcb4	691	signal(SIGALRM, io_u_timeout_handler);
55bc9728	692	#endif
433afcb4	693	}