[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_mutex_lock(&td->io_u_lock);
		pthread_cond_signal(&td->parent->free_cond);
		pthread_mutex_unlock(&td->io_u_lock);
	}

	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);
	td->iolog_f = parent->iolog_f;

	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 | td->o.log_alternate_epoch, td->o.log_alternate_epoch_clock_id);
	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)++;

	/*
	 * io_workqueue_update_acct_fn() doesn't support per prio stats, and
	 * even if it did, offload can't be used with all async IO engines.
	 * If group reporting is set in the parent td, the group result
	 * generated by __show_run_stats() can still contain multiple prios
	 * from different offloaded jobs.
	 */
	sw->wq->td->ts.disable_prio_stat = 1;
	sum_thread_stats(&sw->wq->td->ts, &td->ts);

	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]);
	if (ddir == DDIR_READ)
		sum_val(&dst->bytes_verified, &src->bytes_verified);

	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
d89ee9f4 BVA	100	if (error \|\| td->error) {
d89ee9f4 BVA	101	pthread_mutex_lock(&td->io_u_lock);
d28174f0	102	pthread_cond_signal(&td->parent->free_cond);
d89ee9f4 BVA	103	pthread_mutex_unlock(&td->io_u_lock);
d89ee9f4 BVA	104	}
d28174f0	105
155f2f02	106	return 0;
40511301 JA	107	}
	108
	109	static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
	110	{
b86ad8f1	111	struct thread_data *td = sw->priv;
40511301	112
d28174f0 JA	113	if (td->error)
d28174f0 JA	114	return false;
40511301 JA	115	if (td->io_u_queued \|\| td->cur_depth \|\| td->io_u_in_flight)
	116	return true;
	117
	118	return false;
	119	}
	120
	121	static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
	122	{
b86ad8f1	123	struct thread_data *td = sw->priv;
40511301 JA	124	int ret;
	125
	126	ret = io_u_quiesce(td);
	127	if (ret > 0)
	128	td->cur_depth -= ret;
	129	}
	130
	131	static int io_workqueue_alloc_fn(struct submit_worker *sw)
	132	{
	133	struct thread_data *td;
	134
	135	td = calloc(1, sizeof(*td));
b86ad8f1	136	sw->priv = td;
40511301 JA	137	return 0;
	138	}
	139
	140	static void io_workqueue_free_fn(struct submit_worker *sw)
	141	{
b86ad8f1 CB	142	free(sw->priv);
b86ad8f1 CB	143	sw->priv = NULL;
40511301 JA	144	}
	145
	146	static int io_workqueue_init_worker_fn(struct submit_worker *sw)
	147	{
	148	struct thread_data *parent = sw->wq->td;
b86ad8f1	149	struct thread_data *td = sw->priv;
40511301 JA	150
	151	memcpy(&td->o, &parent->o, sizeof(td->o));
	152	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	153	td->o.uid = td->o.gid = -1U;
	154	dup_files(td, parent);
	155	td->eo = parent->eo;
	156	fio_options_mem_dupe(td);
c83579f0	157	td->iolog_f = parent->iolog_f;
40511301 JA	158
	159	if (ioengine_load(td))
	160	goto err;
	161
40511301 JA	162	td->pid = gettid();
	163
	164	INIT_FLIST_HEAD(&td->io_log_list);
	165	INIT_FLIST_HEAD(&td->io_hist_list);
	166	INIT_FLIST_HEAD(&td->verify_list);
	167	INIT_FLIST_HEAD(&td->trim_list);
40511301 JA	168	td->io_hist_tree = RB_ROOT;
	169
	170	td->o.iodepth = 1;
	171	if (td_io_init(td))
	172	goto err_io_init;
	173
4c085cf2 VF	174	if (td->io_ops->post_init && td->io_ops->post_init(td))
	175	goto err_io_init;
	176
d5b3cfd4	177	set_epoch_time(td, td->o.log_unix_epoch \| td->o.log_alternate_epoch, td->o.log_alternate_epoch_clock_id);
40511301 JA	178	fio_getrusage(&td->ru_start);
	179	clear_io_state(td, 1);
	180
	181	td_set_runstate(td, TD_RUNNING);
b7aae4ba	182	td->flags \|= TD_F_CHILD \| TD_F_NEED_LOCK;
40511301 JA	183	td->parent = parent;
	184	return 0;
	185
	186	err_io_init:
	187	close_ioengine(td);
	188	err:
	189	return 1;
	190
	191	}
	192
	193	static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
	194	unsigned int *sum_cnt)
	195	{
b86ad8f1	196	struct thread_data *td = sw->priv;
40511301 JA	197
40511301 JA	198	(*sum_cnt)++;
691310e2 NC	199
	200	/*
	201	* io_workqueue_update_acct_fn() doesn't support per prio stats, and
	202	* even if it did, offload can't be used with all async IO engines.
	203	* If group reporting is set in the parent td, the group result
	204	* generated by __show_run_stats() can still contain multiple prios
	205	* from different offloaded jobs.
	206	*/
	207	sw->wq->td->ts.disable_prio_stat = 1;
016869be	208	sum_thread_stats(&sw->wq->td->ts, &td->ts);
40511301 JA	209
	210	fio_options_free(td);
	211	close_and_free_files(td);
	212	if (td->io_ops)
	213	close_ioengine(td);
	214	td_set_runstate(td, TD_EXITED);
	215	}
	216
	217	#ifdef CONFIG_SFAA
	218	static void sum_val(uint64_t dst, uint64_t src)
	219	{
	220	if (*src) {
	221	__sync_fetch_and_add(dst, *src);
	222	*src = 0;
	223	}
	224	}
	225	#else
	226	static void sum_val(uint64_t dst, uint64_t src)
	227	{
	228	if (*src) {
	229	dst += src;
	230	*src = 0;
	231	}
	232	}
	233	#endif
	234
	235	static void pthread_double_unlock(pthread_mutex_t *lock1,
	236	pthread_mutex_t *lock2)
	237	{
	238	#ifndef CONFIG_SFAA
	239	pthread_mutex_unlock(lock1);
	240	pthread_mutex_unlock(lock2);
	241	#endif
	242	}
	243
	244	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	245	{
	246	#ifndef CONFIG_SFAA
	247	if (lock1 < lock2) {
	248	pthread_mutex_lock(lock1);
	249	pthread_mutex_lock(lock2);
	250	} else {
	251	pthread_mutex_lock(lock2);
	252	pthread_mutex_lock(lock1);
	253	}
	254	#endif
	255	}
	256
	257	static void sum_ddir(struct thread_data dst, struct thread_data src,
	258	enum fio_ddir ddir)
	259	{
	260	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	261
	262	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	263	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	264	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	265	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	266	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
191d6634 SK	267	if (ddir == DDIR_READ)
191d6634 SK	268	sum_val(&dst->bytes_verified, &src->bytes_verified);
40511301 JA	269
	270	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	271	}
	272
	273	static void io_workqueue_update_acct_fn(struct submit_worker *sw)
	274	{
b86ad8f1	275	struct thread_data *src = sw->priv;
40511301 JA	276	struct thread_data *dst = sw->wq->td;
	277
	278	if (td_read(src))
	279	sum_ddir(dst, src, DDIR_READ);
	280	if (td_write(src))
	281	sum_ddir(dst, src, DDIR_WRITE);
	282	if (td_trim(src))
	283	sum_ddir(dst, src, DDIR_TRIM);
	284
	285	}
	286
103b174e	287	static struct workqueue_ops rated_wq_ops = {
40511301 JA	288	.fn = io_workqueue_fn,
	289	.pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
	290	.pre_sleep_fn = io_workqueue_pre_sleep_fn,
	291	.update_acct_fn = io_workqueue_update_acct_fn,
	292	.alloc_worker_fn = io_workqueue_alloc_fn,
	293	.free_worker_fn = io_workqueue_free_fn,
	294	.init_worker_fn = io_workqueue_init_worker_fn,
	295	.exit_worker_fn = io_workqueue_exit_worker_fn,
	296	};
103b174e	297
24660963	298	int rate_submit_init(struct thread_data td, struct sk_out sk_out)
103b174e JA	299	{
	300	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	301	return 0;
	302
24660963	303	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
103b174e JA	304	}
	305
	306	void rate_submit_exit(struct thread_data *td)
	307	{
	308	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	309	return;
	310
	311	workqueue_exit(&td->io_wq);
	312	}