[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 void check_overlap(struct io_u *io_u)
{
	int i;
	struct thread_data *td;

	/*
	 * Allow only one thread to check for overlap at a time to prevent two
	 * threads from thinking the coast is clear and then submitting IOs
	 * that overlap with each other.
	 *
	 * If an overlap is found, release the lock and re-acquire it before
	 * checking again to give other threads a chance to make progress.
	 *
	 * If no overlap is found, release the lock when the io_u's
	 * IO_U_F_FLIGHT flag is set so that this io_u can be checked by other
	 * threads as they assess overlap.
	 */
	pthread_mutex_lock(&overlap_check);

retry:
	for_each_td(td, i) {
		if (td->runstate <= TD_SETTING_UP ||
		    td->runstate >= TD_FINISHING ||
		    !td->o.serialize_overlap ||
		    td->o.io_submit_mode != IO_MODE_OFFLOAD)
			continue;

		if (!in_flight_overlap(&td->io_u_all, io_u))
			continue;

		pthread_mutex_unlock(&overlap_check);
		pthread_mutex_lock(&overlap_check);
		goto retry;
	}
}

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, error;

	if (td->o.serialize_overlap)
		check_overlap(io_u);

	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;
		else if (ret < 0)
			break;
		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());

	error = 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;
	}

	if (error || td->error)
		pthread_cond_signal(&td->parent->free_cond);

	return 0;
}

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

	if (td->error)
		return false;
	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;

	if (td->io_ops->post_init && td->io_ops->post_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_F_NEED_LOCK;
	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
c06379a6 VF	12	static void check_overlap(struct io_u *io_u)
	13	{
	14	int i;
	15	struct thread_data *td;
c06379a6	16
c76b661c BVA	17	/*
	18	* Allow only one thread to check for overlap at a time to prevent two
	19	* threads from thinking the coast is clear and then submitting IOs
	20	* that overlap with each other.
	21	*
	22	* If an overlap is found, release the lock and re-acquire it before
	23	* checking again to give other threads a chance to make progress.
	24	*
	25	* If no overlap is found, release the lock when the io_u's
	26	* IO_U_F_FLIGHT flag is set so that this io_u can be checked by other
	27	* threads as they assess overlap.
	28	*/
	29	pthread_mutex_lock(&overlap_check);
	30
	31	retry:
	32	for_each_td(td, i) {
	33	if (td->runstate <= TD_SETTING_UP \|\|
	34	td->runstate >= TD_FINISHING \|\|
	35	!td->o.serialize_overlap \|\|
	36	td->o.io_submit_mode != IO_MODE_OFFLOAD)
	37	continue;
c06379a6	38
c76b661c BVA	39	if (!in_flight_overlap(&td->io_u_all, io_u))
	40	continue;
	41
	42	pthread_mutex_unlock(&overlap_check);
	43	pthread_mutex_lock(&overlap_check);
	44	goto retry;
	45	}
c06379a6 VF	46	}
c06379a6 VF	47
155f2f02 JA	48	static int io_workqueue_fn(struct submit_worker *sw,
155f2f02 JA	49	struct workqueue_work *work)
40511301 JA	50	{
	51	struct io_u *io_u = container_of(work, struct io_u, work);
	52	const enum fio_ddir ddir = io_u->ddir;
b86ad8f1	53	struct thread_data *td = sw->priv;
d28174f0	54	int ret, error;
40511301	55
c06379a6 VF	56	if (td->o.serialize_overlap)
	57	check_overlap(io_u);
	58
40511301 JA	59	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
40511301 JA	60
1651e431	61	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
40511301 JA	62
	63	td->cur_depth++;
	64
	65	do {
	66	ret = td_io_queue(td, io_u);
	67	if (ret != FIO_Q_BUSY)
	68	break;
	69	ret = io_u_queued_complete(td, 1);
	70	if (ret > 0)
	71	td->cur_depth -= ret;
d28174f0 JA	72	else if (ret < 0)
d28174f0 JA	73	break;
1651e431	74	io_u_clear(td, io_u, IO_U_F_FLIGHT);
40511301 JA	75	} while (1);
	76
	77	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
	78
d28174f0	79	error = io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
40511301 JA	80
	81	if (ret == FIO_Q_COMPLETED)
	82	td->cur_depth--;
	83	else if (ret == FIO_Q_QUEUED) {
	84	unsigned int min_evts;
	85
	86	if (td->o.iodepth == 1)
	87	min_evts = 1;
	88	else
	89	min_evts = 0;
	90
	91	ret = io_u_queued_complete(td, min_evts);
	92	if (ret > 0)
	93	td->cur_depth -= ret;
40511301	94	}
155f2f02	95
d28174f0 JA	96	if (error \|\| td->error)
	97	pthread_cond_signal(&td->parent->free_cond);
	98
155f2f02	99	return 0;
40511301 JA	100	}
	101
	102	static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
	103	{
b86ad8f1	104	struct thread_data *td = sw->priv;
40511301	105
d28174f0 JA	106	if (td->error)
d28174f0 JA	107	return false;
40511301 JA	108	if (td->io_u_queued \|\| td->cur_depth \|\| td->io_u_in_flight)
	109	return true;
	110
	111	return false;
	112	}
	113
	114	static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
	115	{
b86ad8f1	116	struct thread_data *td = sw->priv;
40511301 JA	117	int ret;
	118
	119	ret = io_u_quiesce(td);
	120	if (ret > 0)
	121	td->cur_depth -= ret;
	122	}
	123
	124	static int io_workqueue_alloc_fn(struct submit_worker *sw)
	125	{
	126	struct thread_data *td;
	127
	128	td = calloc(1, sizeof(*td));
b86ad8f1	129	sw->priv = td;
40511301 JA	130	return 0;
	131	}
	132
	133	static void io_workqueue_free_fn(struct submit_worker *sw)
	134	{
b86ad8f1 CB	135	free(sw->priv);
b86ad8f1 CB	136	sw->priv = NULL;
40511301 JA	137	}
	138
	139	static int io_workqueue_init_worker_fn(struct submit_worker *sw)
	140	{
	141	struct thread_data *parent = sw->wq->td;
b86ad8f1	142	struct thread_data *td = sw->priv;
40511301 JA	143
	144	memcpy(&td->o, &parent->o, sizeof(td->o));
	145	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	146	td->o.uid = td->o.gid = -1U;
	147	dup_files(td, parent);
	148	td->eo = parent->eo;
	149	fio_options_mem_dupe(td);
	150
	151	if (ioengine_load(td))
	152	goto err;
	153
40511301 JA	154	td->pid = gettid();
	155
	156	INIT_FLIST_HEAD(&td->io_log_list);
	157	INIT_FLIST_HEAD(&td->io_hist_list);
	158	INIT_FLIST_HEAD(&td->verify_list);
	159	INIT_FLIST_HEAD(&td->trim_list);
40511301 JA	160	td->io_hist_tree = RB_ROOT;
	161
	162	td->o.iodepth = 1;
	163	if (td_io_init(td))
	164	goto err_io_init;
	165
4c085cf2 VF	166	if (td->io_ops->post_init && td->io_ops->post_init(td))
	167	goto err_io_init;
	168
3aea75b1	169	set_epoch_time(td, td->o.log_unix_epoch);
40511301 JA	170	fio_getrusage(&td->ru_start);
	171	clear_io_state(td, 1);
	172
	173	td_set_runstate(td, TD_RUNNING);
b7aae4ba	174	td->flags \|= TD_F_CHILD \| TD_F_NEED_LOCK;
40511301 JA	175	td->parent = parent;
	176	return 0;
	177
	178	err_io_init:
	179	close_ioengine(td);
	180	err:
	181	return 1;
	182
	183	}
	184
	185	static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
	186	unsigned int *sum_cnt)
	187	{
b86ad8f1	188	struct thread_data *td = sw->priv;
40511301 JA	189
	190	(*sum_cnt)++;
	191	sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
	192
	193	fio_options_free(td);
	194	close_and_free_files(td);
	195	if (td->io_ops)
	196	close_ioengine(td);
	197	td_set_runstate(td, TD_EXITED);
	198	}
	199
	200	#ifdef CONFIG_SFAA
	201	static void sum_val(uint64_t dst, uint64_t src)
	202	{
	203	if (*src) {
	204	__sync_fetch_and_add(dst, *src);
	205	*src = 0;
	206	}
	207	}
	208	#else
	209	static void sum_val(uint64_t dst, uint64_t src)
	210	{
	211	if (*src) {
	212	dst += src;
	213	*src = 0;
	214	}
	215	}
	216	#endif
	217
	218	static void pthread_double_unlock(pthread_mutex_t *lock1,
	219	pthread_mutex_t *lock2)
	220	{
	221	#ifndef CONFIG_SFAA
	222	pthread_mutex_unlock(lock1);
	223	pthread_mutex_unlock(lock2);
	224	#endif
	225	}
	226
	227	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	228	{
	229	#ifndef CONFIG_SFAA
	230	if (lock1 < lock2) {
	231	pthread_mutex_lock(lock1);
	232	pthread_mutex_lock(lock2);
	233	} else {
	234	pthread_mutex_lock(lock2);
	235	pthread_mutex_lock(lock1);
	236	}
	237	#endif
	238	}
	239
	240	static void sum_ddir(struct thread_data dst, struct thread_data src,
	241	enum fio_ddir ddir)
	242	{
	243	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	244
	245	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	246	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	247	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	248	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	249	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
	250
	251	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	252	}
253
254	static void io_workqueue_update_acct_fn(struct submit_worker *sw)
255	{
b86ad8f1	256	struct thread_data *src = sw->priv;
40511301 JA	257	struct thread_data *dst = sw->wq->td;
	258
	259	if (td_read(src))
	260	sum_ddir(dst, src, DDIR_READ);
	261	if (td_write(src))
	262	sum_ddir(dst, src, DDIR_WRITE);
	263	if (td_trim(src))
	264	sum_ddir(dst, src, DDIR_TRIM);
	265
	266	}
	267
103b174e	268	static struct workqueue_ops rated_wq_ops = {
40511301 JA	269	.fn = io_workqueue_fn,
	270	.pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
	271	.pre_sleep_fn = io_workqueue_pre_sleep_fn,
	272	.update_acct_fn = io_workqueue_update_acct_fn,
	273	.alloc_worker_fn = io_workqueue_alloc_fn,
	274	.free_worker_fn = io_workqueue_free_fn,
	275	.init_worker_fn = io_workqueue_init_worker_fn,
	276	.exit_worker_fn = io_workqueue_exit_worker_fn,
	277	};
103b174e	278
24660963	279	int rate_submit_init(struct thread_data td, struct sk_out sk_out)
103b174e JA	280	{
	281	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	282	return 0;
	283
24660963	284	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
103b174e JA	285	}
	286
	287	void rate_submit_exit(struct thread_data *td)
	288	{
	289	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	290	return;
	291
	292	workqueue_exit(&td->io_wq);
	293	}