[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 res;

	/*
	 * 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) {
		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;
	} end_for_each();
}

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	{
da8f124f	15	int res;
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	*/
a38c70a8 BVA	29	res = pthread_mutex_lock(&overlap_check);
a38c70a8 BVA	30	assert(res == 0);
c76b661c BVA	31
c76b661c BVA	32	retry:
da8f124f	33	for_each_td(td) {
c76b661c BVA	34	if (td->runstate <= TD_SETTING_UP \|\|
	35	td->runstate >= TD_FINISHING \|\|
	36	!td->o.serialize_overlap \|\|
	37	td->o.io_submit_mode != IO_MODE_OFFLOAD)
	38	continue;
c06379a6	39
c76b661c BVA	40	if (!in_flight_overlap(&td->io_u_all, io_u))
	41	continue;
	42
a38c70a8 BVA	43	res = pthread_mutex_unlock(&overlap_check);
	44	assert(res == 0);
	45	res = pthread_mutex_lock(&overlap_check);
	46	assert(res == 0);
c76b661c	47	goto retry;
da8f124f	48	} end_for_each();
c06379a6 VF	49	}
c06379a6 VF	50
155f2f02 JA	51	static int io_workqueue_fn(struct submit_worker *sw,
155f2f02 JA	52	struct workqueue_work *work)
40511301 JA	53	{
	54	struct io_u *io_u = container_of(work, struct io_u, work);
	55	const enum fio_ddir ddir = io_u->ddir;
b86ad8f1	56	struct thread_data *td = sw->priv;
d28174f0	57	int ret, error;
40511301	58
c06379a6 VF	59	if (td->o.serialize_overlap)
	60	check_overlap(io_u);
	61
40511301 JA	62	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
40511301 JA	63
1651e431	64	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
40511301 JA	65
	66	td->cur_depth++;
	67
	68	do {
	69	ret = td_io_queue(td, io_u);
	70	if (ret != FIO_Q_BUSY)
	71	break;
	72	ret = io_u_queued_complete(td, 1);
	73	if (ret > 0)
	74	td->cur_depth -= ret;
d28174f0 JA	75	else if (ret < 0)
d28174f0 JA	76	break;
1651e431	77	io_u_clear(td, io_u, IO_U_F_FLIGHT);
40511301 JA	78	} while (1);
	79
	80	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
	81
d28174f0	82	error = io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
40511301 JA	83
	84	if (ret == FIO_Q_COMPLETED)
	85	td->cur_depth--;
	86	else if (ret == FIO_Q_QUEUED) {
	87	unsigned int min_evts;
	88
	89	if (td->o.iodepth == 1)
	90	min_evts = 1;
	91	else
	92	min_evts = 0;
	93
	94	ret = io_u_queued_complete(td, min_evts);
	95	if (ret > 0)
	96	td->cur_depth -= ret;
40511301	97	}
155f2f02	98
d89ee9f4 BVA	99	if (error \|\| td->error) {
d89ee9f4 BVA	100	pthread_mutex_lock(&td->io_u_lock);
d28174f0	101	pthread_cond_signal(&td->parent->free_cond);
d89ee9f4 BVA	102	pthread_mutex_unlock(&td->io_u_lock);
d89ee9f4 BVA	103	}
d28174f0	104
155f2f02	105	return 0;
40511301 JA	106	}
	107
	108	static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
	109	{
b86ad8f1	110	struct thread_data *td = sw->priv;
40511301	111
d28174f0 JA	112	if (td->error)
d28174f0 JA	113	return false;
40511301 JA	114	if (td->io_u_queued \|\| td->cur_depth \|\| td->io_u_in_flight)
	115	return true;
	116
	117	return false;
	118	}
	119
	120	static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
	121	{
b86ad8f1	122	struct thread_data *td = sw->priv;
40511301 JA	123	int ret;
	124
	125	ret = io_u_quiesce(td);
	126	if (ret > 0)
	127	td->cur_depth -= ret;
	128	}
	129
	130	static int io_workqueue_alloc_fn(struct submit_worker *sw)
	131	{
	132	struct thread_data *td;
	133
	134	td = calloc(1, sizeof(*td));
b86ad8f1	135	sw->priv = td;
40511301 JA	136	return 0;
	137	}
	138
	139	static void io_workqueue_free_fn(struct submit_worker *sw)
	140	{
b86ad8f1 CB	141	free(sw->priv);
b86ad8f1 CB	142	sw->priv = NULL;
40511301 JA	143	}
	144
	145	static int io_workqueue_init_worker_fn(struct submit_worker *sw)
	146	{
	147	struct thread_data *parent = sw->wq->td;
b86ad8f1	148	struct thread_data *td = sw->priv;
40511301 JA	149
	150	memcpy(&td->o, &parent->o, sizeof(td->o));
	151	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	152	td->o.uid = td->o.gid = -1U;
	153	dup_files(td, parent);
	154	td->eo = parent->eo;
	155	fio_options_mem_dupe(td);
c83579f0	156	td->iolog_f = parent->iolog_f;
40511301 JA	157
	158	if (ioengine_load(td))
	159	goto err;
	160
40511301 JA	161	td->pid = gettid();
	162
	163	INIT_FLIST_HEAD(&td->io_log_list);
	164	INIT_FLIST_HEAD(&td->io_hist_list);
	165	INIT_FLIST_HEAD(&td->verify_list);
	166	INIT_FLIST_HEAD(&td->trim_list);
40511301 JA	167	td->io_hist_tree = RB_ROOT;
	168
	169	td->o.iodepth = 1;
	170	if (td_io_init(td))
	171	goto err_io_init;
	172
4c085cf2 VF	173	if (td->io_ops->post_init && td->io_ops->post_init(td))
	174	goto err_io_init;
	175
d5b3cfd4	176	set_epoch_time(td, td->o.log_unix_epoch \| td->o.log_alternate_epoch, td->o.log_alternate_epoch_clock_id);
40511301 JA	177	fio_getrusage(&td->ru_start);
	178	clear_io_state(td, 1);
	179
	180	td_set_runstate(td, TD_RUNNING);
b7aae4ba	181	td->flags \|= TD_F_CHILD \| TD_F_NEED_LOCK;
40511301 JA	182	td->parent = parent;
	183	return 0;
	184
	185	err_io_init:
	186	close_ioengine(td);
	187	err:
	188	return 1;
	189
	190	}
	191
	192	static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
	193	unsigned int *sum_cnt)
	194	{
b86ad8f1	195	struct thread_data *td = sw->priv;
40511301 JA	196
40511301 JA	197	(*sum_cnt)++;
691310e2 NC	198
	199	/*
	200	* io_workqueue_update_acct_fn() doesn't support per prio stats, and
	201	* even if it did, offload can't be used with all async IO engines.
	202	* If group reporting is set in the parent td, the group result
	203	* generated by __show_run_stats() can still contain multiple prios
	204	* from different offloaded jobs.
	205	*/
	206	sw->wq->td->ts.disable_prio_stat = 1;
016869be	207	sum_thread_stats(&sw->wq->td->ts, &td->ts);
40511301 JA	208
	209	fio_options_free(td);
	210	close_and_free_files(td);
	211	if (td->io_ops)
	212	close_ioengine(td);
	213	td_set_runstate(td, TD_EXITED);
	214	}
	215
	216	#ifdef CONFIG_SFAA
	217	static void sum_val(uint64_t dst, uint64_t src)
	218	{
	219	if (*src) {
	220	__sync_fetch_and_add(dst, *src);
	221	*src = 0;
	222	}
	223	}
	224	#else
	225	static void sum_val(uint64_t dst, uint64_t src)
	226	{
	227	if (*src) {
	228	dst += src;
	229	*src = 0;
	230	}
	231	}
	232	#endif
	233
	234	static void pthread_double_unlock(pthread_mutex_t *lock1,
	235	pthread_mutex_t *lock2)
	236	{
	237	#ifndef CONFIG_SFAA
	238	pthread_mutex_unlock(lock1);
	239	pthread_mutex_unlock(lock2);
	240	#endif
	241	}
	242
	243	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	244	{
	245	#ifndef CONFIG_SFAA
	246	if (lock1 < lock2) {
	247	pthread_mutex_lock(lock1);
	248	pthread_mutex_lock(lock2);
	249	} else {
	250	pthread_mutex_lock(lock2);
	251	pthread_mutex_lock(lock1);
	252	}
	253	#endif
	254	}
	255
	256	static void sum_ddir(struct thread_data dst, struct thread_data src,
	257	enum fio_ddir ddir)
	258	{
	259	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	260
	261	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	262	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	263	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	264	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	265	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
191d6634 SK	266	if (ddir == DDIR_READ)
191d6634 SK	267	sum_val(&dst->bytes_verified, &src->bytes_verified);
40511301 JA	268
	269	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	270	}
	271
	272	static void io_workqueue_update_acct_fn(struct submit_worker *sw)
	273	{
b86ad8f1	274	struct thread_data *src = sw->priv;
40511301 JA	275	struct thread_data *dst = sw->wq->td;
	276
	277	if (td_read(src))
	278	sum_ddir(dst, src, DDIR_READ);
	279	if (td_write(src))
	280	sum_ddir(dst, src, DDIR_WRITE);
	281	if (td_trim(src))
	282	sum_ddir(dst, src, DDIR_TRIM);
	283
	284	}
	285
103b174e	286	static struct workqueue_ops rated_wq_ops = {
40511301 JA	287	.fn = io_workqueue_fn,
	288	.pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
	289	.pre_sleep_fn = io_workqueue_pre_sleep_fn,
	290	.update_acct_fn = io_workqueue_update_acct_fn,
	291	.alloc_worker_fn = io_workqueue_alloc_fn,
	292	.free_worker_fn = io_workqueue_free_fn,
	293	.init_worker_fn = io_workqueue_init_worker_fn,
	294	.exit_worker_fn = io_workqueue_exit_worker_fn,
	295	};
103b174e	296
24660963	297	int rate_submit_init(struct thread_data td, struct sk_out sk_out)
103b174e JA	298	{
	299	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	300	return 0;
	301
24660963	302	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
103b174e JA	303	}
	304
	305	void rate_submit_exit(struct thread_data *td)
	306	{
	307	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	308	return;
	309
	310	workqueue_exit(&td->io_wq);
	311	}