[fio.git] / rate-submit.c

/*
 * Rated submission helpers
 *
 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
 *
 */
#include "fio.h"
#include "ioengines.h"
#include "lib/getrusage.h"
#include "rate-submit.h"

static int io_workqueue_fn(struct submit_worker *sw,
			   struct workqueue_work *work)
{
	struct io_u *io_u = container_of(work, struct io_u, work);
	const enum fio_ddir ddir = io_u->ddir;
	struct thread_data *td = sw->priv;
	int ret;

	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());

	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);

	td->cur_depth++;

	do {
		ret = td_io_queue(td, io_u);
		if (ret != FIO_Q_BUSY)
			break;
		ret = io_u_queued_complete(td, 1);
		if (ret > 0)
			td->cur_depth -= ret;
		io_u_clear(td, io_u, IO_U_F_FLIGHT);
	} while (1);

	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());

	io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);

	if (ret == FIO_Q_COMPLETED)
		td->cur_depth--;
	else if (ret == FIO_Q_QUEUED) {
		unsigned int min_evts;

		if (td->o.iodepth == 1)
			min_evts = 1;
		else
			min_evts = 0;

		ret = io_u_queued_complete(td, min_evts);
		if (ret > 0)
			td->cur_depth -= ret;
	} else if (ret == FIO_Q_BUSY) {
		ret = io_u_queued_complete(td, td->cur_depth);
		if (ret > 0)
			td->cur_depth -= ret;
	}

	return 0;
}

static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
{
	struct thread_data *td = sw->priv;

	if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
		return true;

	return false;
}

static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
{
	struct thread_data *td = sw->priv;
	int ret;

	ret = io_u_quiesce(td);
	if (ret > 0)
		td->cur_depth -= ret;
}

static int io_workqueue_alloc_fn(struct submit_worker *sw)
{
	struct thread_data *td;

	td = calloc(1, sizeof(*td));
	sw->priv = td;
	return 0;
}

static void io_workqueue_free_fn(struct submit_worker *sw)
{
	free(sw->priv);
	sw->priv = NULL;
}

static int io_workqueue_init_worker_fn(struct submit_worker *sw)
{
	struct thread_data *parent = sw->wq->td;
	struct thread_data *td = sw->priv;

	memcpy(&td->o, &parent->o, sizeof(td->o));
	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	td->o.uid = td->o.gid = -1U;
	dup_files(td, parent);
	td->eo = parent->eo;
	fio_options_mem_dupe(td);

	if (ioengine_load(td))
		goto err;

	td->pid = gettid();

	INIT_FLIST_HEAD(&td->io_log_list);
	INIT_FLIST_HEAD(&td->io_hist_list);
	INIT_FLIST_HEAD(&td->verify_list);
	INIT_FLIST_HEAD(&td->trim_list);
	td->io_hist_tree = RB_ROOT;

	td->o.iodepth = 1;
	if (td_io_init(td))
		goto err_io_init;

	set_epoch_time(td, td->o.log_unix_epoch);
	fio_getrusage(&td->ru_start);
	clear_io_state(td, 1);

	td_set_runstate(td, TD_RUNNING);
	td->flags |= TD_F_CHILD;
	td->parent = parent;
	return 0;

err_io_init:
	close_ioengine(td);
err:
	return 1;

}

static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
					unsigned int *sum_cnt)
{
	struct thread_data *td = sw->priv;

	(*sum_cnt)++;
	sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);

	fio_options_free(td);
	close_and_free_files(td);
	if (td->io_ops)
		close_ioengine(td);
	td_set_runstate(td, TD_EXITED);
}

#ifdef CONFIG_SFAA
static void sum_val(uint64_t *dst, uint64_t *src)
{
	if (*src) {
		__sync_fetch_and_add(dst, *src);
		*src = 0;
	}
}
#else
static void sum_val(uint64_t *dst, uint64_t *src)
{
	if (*src) {
		*dst += *src;
		*src = 0;
	}
}
#endif

static void pthread_double_unlock(pthread_mutex_t *lock1,
				  pthread_mutex_t *lock2)
{
#ifndef CONFIG_SFAA
	pthread_mutex_unlock(lock1);
	pthread_mutex_unlock(lock2);
#endif
}

static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
{
#ifndef CONFIG_SFAA
	if (lock1 < lock2) {
		pthread_mutex_lock(lock1);
		pthread_mutex_lock(lock2);
	} else {
		pthread_mutex_lock(lock2);
		pthread_mutex_lock(lock1);
	}
#endif
}

static void sum_ddir(struct thread_data *dst, struct thread_data *src,
		     enum fio_ddir ddir)
{
	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);

	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);

	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
}

static void io_workqueue_update_acct_fn(struct submit_worker *sw)
{
	struct thread_data *src = sw->priv;
	struct thread_data *dst = sw->wq->td;

	if (td_read(src))
		sum_ddir(dst, src, DDIR_READ);
	if (td_write(src))
		sum_ddir(dst, src, DDIR_WRITE);
	if (td_trim(src))
		sum_ddir(dst, src, DDIR_TRIM);

}

static struct workqueue_ops rated_wq_ops = {
	.fn			= io_workqueue_fn,
	.pre_sleep_flush_fn	= io_workqueue_pre_sleep_flush_fn,
	.pre_sleep_fn		= io_workqueue_pre_sleep_fn,
	.update_acct_fn		= io_workqueue_update_acct_fn,
	.alloc_worker_fn	= io_workqueue_alloc_fn,
	.free_worker_fn		= io_workqueue_free_fn,
	.init_worker_fn		= io_workqueue_init_worker_fn,
	.exit_worker_fn		= io_workqueue_exit_worker_fn,
};

int rate_submit_init(struct thread_data *td, struct sk_out *sk_out)
{
	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
		return 0;

	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
}

void rate_submit_exit(struct thread_data *td)
{
	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
		return;

	workqueue_exit(&td->io_wq);
}
Commit	Line	Data
51575029 JA	1	/*
	2	* Rated submission helpers
	3	*
	4	* Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
	5	*
	6	*/
40511301	7	#include "fio.h"
618ee94c	8	#include "ioengines.h"
40511301	9	#include "lib/getrusage.h"
24660963	10	#include "rate-submit.h"
40511301	11
155f2f02 JA	12	static int io_workqueue_fn(struct submit_worker *sw,
155f2f02 JA	13	struct workqueue_work *work)
40511301 JA	14	{
	15	struct io_u *io_u = container_of(work, struct io_u, work);
	16	const enum fio_ddir ddir = io_u->ddir;
b86ad8f1	17	struct thread_data *td = sw->priv;
40511301 JA	18	int ret;
	19
	20	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
	21
1651e431	22	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
40511301 JA	23
	24	td->cur_depth++;
	25
	26	do {
	27	ret = td_io_queue(td, io_u);
	28	if (ret != FIO_Q_BUSY)
	29	break;
	30	ret = io_u_queued_complete(td, 1);
	31	if (ret > 0)
	32	td->cur_depth -= ret;
1651e431	33	io_u_clear(td, io_u, IO_U_F_FLIGHT);
40511301 JA	34	} while (1);
	35
	36	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
	37
	38	io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
	39
	40	if (ret == FIO_Q_COMPLETED)
	41	td->cur_depth--;
	42	else if (ret == FIO_Q_QUEUED) {
	43	unsigned int min_evts;
	44
	45	if (td->o.iodepth == 1)
	46	min_evts = 1;
	47	else
	48	min_evts = 0;
	49
	50	ret = io_u_queued_complete(td, min_evts);
	51	if (ret > 0)
	52	td->cur_depth -= ret;
	53	} else if (ret == FIO_Q_BUSY) {
	54	ret = io_u_queued_complete(td, td->cur_depth);
	55	if (ret > 0)
	56	td->cur_depth -= ret;
	57	}
155f2f02 JA	58
155f2f02 JA	59	return 0;
40511301 JA	60	}
	61
	62	static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
	63	{
b86ad8f1	64	struct thread_data *td = sw->priv;
40511301 JA	65
	66	if (td->io_u_queued \|\| td->cur_depth \|\| td->io_u_in_flight)
	67	return true;
	68
	69	return false;
	70	}
	71
	72	static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
	73	{
b86ad8f1	74	struct thread_data *td = sw->priv;
40511301 JA	75	int ret;
	76
	77	ret = io_u_quiesce(td);
	78	if (ret > 0)
	79	td->cur_depth -= ret;
	80	}
	81
	82	static int io_workqueue_alloc_fn(struct submit_worker *sw)
	83	{
	84	struct thread_data *td;
	85
	86	td = calloc(1, sizeof(*td));
b86ad8f1	87	sw->priv = td;
40511301 JA	88	return 0;
	89	}
	90
	91	static void io_workqueue_free_fn(struct submit_worker *sw)
	92	{
b86ad8f1 CB	93	free(sw->priv);
b86ad8f1 CB	94	sw->priv = NULL;
40511301 JA	95	}
	96
	97	static int io_workqueue_init_worker_fn(struct submit_worker *sw)
	98	{
	99	struct thread_data *parent = sw->wq->td;
b86ad8f1	100	struct thread_data *td = sw->priv;
40511301 JA	101
	102	memcpy(&td->o, &parent->o, sizeof(td->o));
	103	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	104	td->o.uid = td->o.gid = -1U;
	105	dup_files(td, parent);
	106	td->eo = parent->eo;
	107	fio_options_mem_dupe(td);
	108
	109	if (ioengine_load(td))
	110	goto err;
	111
40511301 JA	112	td->pid = gettid();
	113
	114	INIT_FLIST_HEAD(&td->io_log_list);
	115	INIT_FLIST_HEAD(&td->io_hist_list);
	116	INIT_FLIST_HEAD(&td->verify_list);
	117	INIT_FLIST_HEAD(&td->trim_list);
40511301 JA	118	td->io_hist_tree = RB_ROOT;
	119
	120	td->o.iodepth = 1;
	121	if (td_io_init(td))
	122	goto err_io_init;
	123
3aea75b1	124	set_epoch_time(td, td->o.log_unix_epoch);
40511301 JA	125	fio_getrusage(&td->ru_start);
	126	clear_io_state(td, 1);
	127
	128	td_set_runstate(td, TD_RUNNING);
	129	td->flags \|= TD_F_CHILD;
	130	td->parent = parent;
	131	return 0;
	132
	133	err_io_init:
	134	close_ioengine(td);
	135	err:
	136	return 1;
	137
	138	}
	139
	140	static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
	141	unsigned int *sum_cnt)
	142	{
b86ad8f1	143	struct thread_data *td = sw->priv;
40511301 JA	144
	145	(*sum_cnt)++;
	146	sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
	147
	148	fio_options_free(td);
	149	close_and_free_files(td);
	150	if (td->io_ops)
	151	close_ioengine(td);
	152	td_set_runstate(td, TD_EXITED);
	153	}
	154
	155	#ifdef CONFIG_SFAA
	156	static void sum_val(uint64_t dst, uint64_t src)
	157	{
	158	if (*src) {
	159	__sync_fetch_and_add(dst, *src);
	160	*src = 0;
	161	}
	162	}
	163	#else
	164	static void sum_val(uint64_t dst, uint64_t src)
	165	{
	166	if (*src) {
	167	dst += src;
	168	*src = 0;
	169	}
	170	}
	171	#endif
	172
	173	static void pthread_double_unlock(pthread_mutex_t *lock1,
	174	pthread_mutex_t *lock2)
	175	{
	176	#ifndef CONFIG_SFAA
	177	pthread_mutex_unlock(lock1);
	178	pthread_mutex_unlock(lock2);
	179	#endif
	180	}
	181
	182	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	183	{
	184	#ifndef CONFIG_SFAA
	185	if (lock1 < lock2) {
	186	pthread_mutex_lock(lock1);
	187	pthread_mutex_lock(lock2);
	188	} else {
	189	pthread_mutex_lock(lock2);
	190	pthread_mutex_lock(lock1);
	191	}
	192	#endif
	193	}
	194
	195	static void sum_ddir(struct thread_data dst, struct thread_data src,
	196	enum fio_ddir ddir)
	197	{
	198	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	199
	200	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	201	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	202	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	203	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	204	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
	205
	206	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	207	}
208
209	static void io_workqueue_update_acct_fn(struct submit_worker *sw)
210	{
b86ad8f1	211	struct thread_data *src = sw->priv;
40511301 JA	212	struct thread_data *dst = sw->wq->td;
	213
	214	if (td_read(src))
	215	sum_ddir(dst, src, DDIR_READ);
	216	if (td_write(src))
	217	sum_ddir(dst, src, DDIR_WRITE);
	218	if (td_trim(src))
	219	sum_ddir(dst, src, DDIR_TRIM);
	220
	221	}
	222
103b174e	223	static struct workqueue_ops rated_wq_ops = {
40511301 JA	224	.fn = io_workqueue_fn,
	225	.pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
	226	.pre_sleep_fn = io_workqueue_pre_sleep_fn,
	227	.update_acct_fn = io_workqueue_update_acct_fn,
	228	.alloc_worker_fn = io_workqueue_alloc_fn,
	229	.free_worker_fn = io_workqueue_free_fn,
	230	.init_worker_fn = io_workqueue_init_worker_fn,
	231	.exit_worker_fn = io_workqueue_exit_worker_fn,
	232	};
103b174e	233
24660963	234	int rate_submit_init(struct thread_data td, struct sk_out sk_out)
103b174e JA	235	{
	236	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	237	return 0;
	238
24660963	239	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
103b174e JA	240	}
	241
	242	void rate_submit_exit(struct thread_data *td)
	243	{
	244	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	245	return;
	246
	247	workqueue_exit(&td->io_wq);
	248	}