[fio.git] / rate-submit.c

/*
 * Rated submission helpers
 *
 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
 *
 */
#include <assert.h>
#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, res;
	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.
	 */
	res = pthread_mutex_lock(&overlap_check);
	assert(res == 0);

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;

		res = pthread_mutex_unlock(&overlap_check);
		assert(res == 0);
		res = pthread_mutex_lock(&overlap_check);
		assert(res == 0);
		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	*/
a38c70a8	7	#include <assert.h>
40511301	8	#include "fio.h"
618ee94c	9	#include "ioengines.h"
40511301	10	#include "lib/getrusage.h"
24660963	11	#include "rate-submit.h"
40511301	12
c06379a6 VF	13	static void check_overlap(struct io_u *io_u)
c06379a6 VF	14	{
a38c70a8	15	int i, res;
c06379a6	16	struct thread_data *td;
c06379a6	17
c76b661c BVA	18	/*
	19	* Allow only one thread to check for overlap at a time to prevent two
	20	* threads from thinking the coast is clear and then submitting IOs
	21	* that overlap with each other.
	22	*
	23	* If an overlap is found, release the lock and re-acquire it before
	24	* checking again to give other threads a chance to make progress.
	25	*
	26	* If no overlap is found, release the lock when the io_u's
	27	* IO_U_F_FLIGHT flag is set so that this io_u can be checked by other
	28	* threads as they assess overlap.
	29	*/
a38c70a8 BVA	30	res = pthread_mutex_lock(&overlap_check);
a38c70a8 BVA	31	assert(res == 0);
c76b661c BVA	32
	33	retry:
	34	for_each_td(td, i) {
	35	if (td->runstate <= TD_SETTING_UP \|\|
	36	td->runstate >= TD_FINISHING \|\|
	37	!td->o.serialize_overlap \|\|
	38	td->o.io_submit_mode != IO_MODE_OFFLOAD)
	39	continue;
c06379a6	40
c76b661c BVA	41	if (!in_flight_overlap(&td->io_u_all, io_u))
	42	continue;
	43
a38c70a8 BVA	44	res = pthread_mutex_unlock(&overlap_check);
	45	assert(res == 0);
	46	res = pthread_mutex_lock(&overlap_check);
	47	assert(res == 0);
c76b661c BVA	48	goto retry;
c76b661c BVA	49	}
c06379a6 VF	50	}
c06379a6 VF	51
155f2f02 JA	52	static int io_workqueue_fn(struct submit_worker *sw,
155f2f02 JA	53	struct workqueue_work *work)
40511301 JA	54	{
	55	struct io_u *io_u = container_of(work, struct io_u, work);
	56	const enum fio_ddir ddir = io_u->ddir;
b86ad8f1	57	struct thread_data *td = sw->priv;
d28174f0	58	int ret, error;
40511301	59
c06379a6 VF	60	if (td->o.serialize_overlap)
	61	check_overlap(io_u);
	62
40511301 JA	63	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
40511301 JA	64
1651e431	65	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
40511301 JA	66
	67	td->cur_depth++;
	68
	69	do {
	70	ret = td_io_queue(td, io_u);
	71	if (ret != FIO_Q_BUSY)
	72	break;
	73	ret = io_u_queued_complete(td, 1);
	74	if (ret > 0)
	75	td->cur_depth -= ret;
d28174f0 JA	76	else if (ret < 0)
d28174f0 JA	77	break;
1651e431	78	io_u_clear(td, io_u, IO_U_F_FLIGHT);
40511301 JA	79	} while (1);
	80
	81	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
	82
d28174f0	83	error = io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
40511301 JA	84
	85	if (ret == FIO_Q_COMPLETED)
	86	td->cur_depth--;
	87	else if (ret == FIO_Q_QUEUED) {
	88	unsigned int min_evts;
	89
	90	if (td->o.iodepth == 1)
	91	min_evts = 1;
	92	else
	93	min_evts = 0;
	94
	95	ret = io_u_queued_complete(td, min_evts);
	96	if (ret > 0)
	97	td->cur_depth -= ret;
40511301	98	}
155f2f02	99
d28174f0 JA	100	if (error \|\| td->error)
	101	pthread_cond_signal(&td->parent->free_cond);
	102
155f2f02	103	return 0;
40511301 JA	104	}
	105
	106	static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
	107	{
b86ad8f1	108	struct thread_data *td = sw->priv;
40511301	109
d28174f0 JA	110	if (td->error)
d28174f0 JA	111	return false;
40511301 JA	112	if (td->io_u_queued \|\| td->cur_depth \|\| td->io_u_in_flight)
	113	return true;
	114
	115	return false;
	116	}
	117
	118	static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
	119	{
b86ad8f1	120	struct thread_data *td = sw->priv;
40511301 JA	121	int ret;
	122
	123	ret = io_u_quiesce(td);
	124	if (ret > 0)
	125	td->cur_depth -= ret;
	126	}
	127
	128	static int io_workqueue_alloc_fn(struct submit_worker *sw)
	129	{
	130	struct thread_data *td;
	131
	132	td = calloc(1, sizeof(*td));
b86ad8f1	133	sw->priv = td;
40511301 JA	134	return 0;
	135	}
	136
	137	static void io_workqueue_free_fn(struct submit_worker *sw)
	138	{
b86ad8f1 CB	139	free(sw->priv);
b86ad8f1 CB	140	sw->priv = NULL;
40511301 JA	141	}
	142
	143	static int io_workqueue_init_worker_fn(struct submit_worker *sw)
	144	{
	145	struct thread_data *parent = sw->wq->td;
b86ad8f1	146	struct thread_data *td = sw->priv;
40511301 JA	147
	148	memcpy(&td->o, &parent->o, sizeof(td->o));
	149	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	150	td->o.uid = td->o.gid = -1U;
	151	dup_files(td, parent);
	152	td->eo = parent->eo;
	153	fio_options_mem_dupe(td);
	154
	155	if (ioengine_load(td))
	156	goto err;
	157
40511301 JA	158	td->pid = gettid();
	159
	160	INIT_FLIST_HEAD(&td->io_log_list);
	161	INIT_FLIST_HEAD(&td->io_hist_list);
	162	INIT_FLIST_HEAD(&td->verify_list);
	163	INIT_FLIST_HEAD(&td->trim_list);
40511301 JA	164	td->io_hist_tree = RB_ROOT;
	165
	166	td->o.iodepth = 1;
	167	if (td_io_init(td))
	168	goto err_io_init;
	169
4c085cf2 VF	170	if (td->io_ops->post_init && td->io_ops->post_init(td))
	171	goto err_io_init;
	172
3aea75b1	173	set_epoch_time(td, td->o.log_unix_epoch);
40511301 JA	174	fio_getrusage(&td->ru_start);
	175	clear_io_state(td, 1);
	176
	177	td_set_runstate(td, TD_RUNNING);
b7aae4ba	178	td->flags \|= TD_F_CHILD \| TD_F_NEED_LOCK;
40511301 JA	179	td->parent = parent;
	180	return 0;
	181
	182	err_io_init:
	183	close_ioengine(td);
	184	err:
	185	return 1;
	186
	187	}
	188
	189	static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
	190	unsigned int *sum_cnt)
	191	{
b86ad8f1	192	struct thread_data *td = sw->priv;
40511301 JA	193
	194	(*sum_cnt)++;
	195	sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
	196
	197	fio_options_free(td);
	198	close_and_free_files(td);
	199	if (td->io_ops)
	200	close_ioengine(td);
	201	td_set_runstate(td, TD_EXITED);
	202	}
	203
	204	#ifdef CONFIG_SFAA
	205	static void sum_val(uint64_t dst, uint64_t src)
	206	{
	207	if (*src) {
	208	__sync_fetch_and_add(dst, *src);
	209	*src = 0;
	210	}
	211	}
	212	#else
	213	static void sum_val(uint64_t dst, uint64_t src)
	214	{
	215	if (*src) {
	216	dst += src;
	217	*src = 0;
	218	}
	219	}
	220	#endif
	221
	222	static void pthread_double_unlock(pthread_mutex_t *lock1,
	223	pthread_mutex_t *lock2)
	224	{
	225	#ifndef CONFIG_SFAA
	226	pthread_mutex_unlock(lock1);
	227	pthread_mutex_unlock(lock2);
	228	#endif
	229	}
	230
	231	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	232	{
	233	#ifndef CONFIG_SFAA
	234	if (lock1 < lock2) {
	235	pthread_mutex_lock(lock1);
	236	pthread_mutex_lock(lock2);
	237	} else {
	238	pthread_mutex_lock(lock2);
	239	pthread_mutex_lock(lock1);
	240	}
	241	#endif
	242	}
	243
	244	static void sum_ddir(struct thread_data dst, struct thread_data src,
	245	enum fio_ddir ddir)
	246	{
	247	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	248
	249	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	250	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	251	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	252	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	253	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
	254
	255	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	256	}
257
258	static void io_workqueue_update_acct_fn(struct submit_worker *sw)
259	{
b86ad8f1	260	struct thread_data *src = sw->priv;
40511301 JA	261	struct thread_data *dst = sw->wq->td;
	262
	263	if (td_read(src))
	264	sum_ddir(dst, src, DDIR_READ);
	265	if (td_write(src))
	266	sum_ddir(dst, src, DDIR_WRITE);
	267	if (td_trim(src))
	268	sum_ddir(dst, src, DDIR_TRIM);
	269
	270	}
	271
103b174e	272	static struct workqueue_ops rated_wq_ops = {
40511301 JA	273	.fn = io_workqueue_fn,
	274	.pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
	275	.pre_sleep_fn = io_workqueue_pre_sleep_fn,
	276	.update_acct_fn = io_workqueue_update_acct_fn,
	277	.alloc_worker_fn = io_workqueue_alloc_fn,
	278	.free_worker_fn = io_workqueue_free_fn,
	279	.init_worker_fn = io_workqueue_init_worker_fn,
	280	.exit_worker_fn = io_workqueue_exit_worker_fn,
	281	};
103b174e	282
24660963	283	int rate_submit_init(struct thread_data td, struct sk_out sk_out)
103b174e JA	284	{
	285	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	286	return 0;
	287
24660963	288	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
103b174e JA	289	}
	290
	291	void rate_submit_exit(struct thread_data *td)
	292	{
	293	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	294	return;
	295
	296	workqueue_exit(&td->io_wq);
	297	}